Datasets:

bzantium

/

LongPerplexity

like 0

astrophysical studies since the 1960s have provided a solid foundation for the physics of disk evolution and planetesimal formation . the core accretion model proposed by @xcite and @xcite , for example , has become part of the standard theory for planetary formation ( e.g. , @xcite ) . however , the problem of meter - size barrier , i.e. , infall of dust caused by headwind @xcite , still forms a considerable impediment to planetary formation . it is crucial to understand how to create planetesimals out of newly condensed dust particles , because the failure to form planetesimals would leave little chance for subsequent planetary formation . a number of models have been proposed to overcome this difficulty , including efficient sticking @xcite , mid - plane gravitational instability @xcite , and streaming instability @xcite , but no consensus has been reached yet . different mechanisms could result in disparate chemical structures and thus different implications for cosmochemistry @xcite . the very first prerequisite for planetesimal formation is the settling of particles towards the disk mid - plane . most of the models proposed to overcome the meter - size barrier require a higher dust - to - gas ratio than expected from the solar abundance , and dust concentration through vertical settling is one of the source mechanisms @xcite . evaluating the timescale for settling is thus essential when considering the subsequent evolution . calculations assuming a laminar disk suggest the time scale of @xmath210@xmath1 years , which is shorter than that of radial disk evolution @xcite . disk turbulence could potentially inhibit the dust growth , resulting in slow settling @xcite . the fragmentation of aggregates would also contribute to maintain small dust size and slow settling @xcite . the effect of chemistry on settling , however , has not been considered in previous studies , although it could play an important role in an early evolutionary stage . when dust grains settle through a vertical temperature gradient , the stable phases of dust - composing elements are likely to change , and even the existence of dust itself is not always guaranteed . using thermodynamic database , condensing species can be calculated at a given pressure and temperature @xcite . this allows us to predict the evolution of dust - composing species as the grains settle , but previous studies on disk evolution have rarely incorporated this thermodynamic constraint on dust settling . by building a simple thermodynamically - consistent model , we will quantify the effects of chemistry on vertical dust settling in protoplanetary disks , which is one of the most fundamental processes in planetary formation . our approach is notable at least for the following two points . first , dust properties are calculated through gibbs free energy minimization rather than using condensation temperature . second , the disk thermal structure is calculated using the opacity information consistent with results from gibbs free energy minimization . the existence of dust will affect the opacity by orders of magnitude , thus affecting the overall temperature structure . such temperature change will feed directly back to the chemistry , possibly altering the stable phases , and further modifying the opacity . thus , opacity is a key parameter connecting physics and chemistry , being crucial to make our calculations fully self - consistent . the purpose of this study is to demonstrate quantitatively this feedback between physics and chemistry during the dust settling process . the paper is organized as follows . first , a theoretical formulation for the interaction between mechanics and condensation is described in detail . modeling results then follow , exhibiting profound differences from classical calculations . implications for astronomical and cosmochemical studies are discussed , including the time scale for dust settling and the trends of element abundance recorded in various meteorites . previous radial evolution models involving both astrophysics and cosmochemistry are discussed as well . our model calculates one - dimensional ( 1-d ) dust settling , tracking the temporal and spatial evolution of dust amount and composition at an early stage of a protoplanetary disk in the terrestrial region . viscous dissipation is likely to cause an increase in temperature towards the mid - plane in the vertical direction . the chemical compositions of dust and coexisting gas at different heights will reflect such a temperature variation , and the amount of dust and its composition will also evolve as the grains settle . our calculations are centered on the following radiative heat transfer equation @xcite : @xmath3 where @xmath4 is the temperature , @xmath5 is the height from the mid - plane , @xmath6 is the rosseland mean opacity , @xmath7 is the background gas density , @xmath8 is the stefan - boltzmann constant , and @xmath9 is the vertical radiation flux . the rosseland mean opacity is sensitive to dust amount , and the physics and chemistry of dust settling are connected through this opacity term . our model considers a system spanning from the mid - plane to five times the initial pressure scale height . at the outer edge of our model , therefore , the gas density is only 10@xmath10 of that at the mid - plane , and the temperature is expected to become roughly constant above the outer edge . the evolution of gas density is solved using a radial evolution model , independently from dust settling . the model is spatially discretized uniformly ; density , temperature , and gas composition are calculated for each grid cell , whereas a tracer - based method is used to solve the settling of dust grains and their composition . tracer particles , representing the dust size and the composition of a nearby region , are allocated in each grid cell . the evolution of dust density distribution is tracked by tracer particles , whose evolution is solved through the combination of classical physics and condensation theory . in our model , the equation of motion , the radiative heat transfer equation , dust opacity , and gibbs free energy minimization are solved sequentially . at each time step , we ( 1 ) populate tracer particles carrying the composition of nearby grains , ( 2 ) settle dust grains according to the equation of motion , ( 3 ) solve radiative heat transfer and gibbs free energy minimization simultaneously to obtain a new temperature profile and chemical composition . the detailed description of each step is given below . dust motion is assumed to be controlled by the solar gravity and the epstein drag @xcite . all the dust grains are assumed to be spherical and have no porous space inside , to focus on the effect of condensation . for an early - stage , high - temperature regime under consideration , this assumption is likely to be valid . the settling speed of dust particles reaches a terminal velocity in a time scale of a few seconds for grain size of @xmath21 @xmath11 m . therefore , all the grains are assumed to settle in the terminal velocity proportional to grain size given by @xmath12 where @xmath13 is the material dust density , @xmath14 is the grain size , and @xmath15 is the local keplerian angular velocity @xcite . the mean thermal speed of molecules is given by @xmath16 , where @xmath17 is the mean gas molecular mass and @xmath18 is the boltzmann constant . the temperature distribution is solved using the radiative transfer equation : equation ( [ dt / dz ] ) supplemented with @xmath19 where @xmath20 denotes the kinematic viscosity of the disk . the viscosity is scaled using the @xmath21-prescription of @xcite as @xmath22 where @xmath23 is the isothermal sound speed , @xmath24 is the pressure scale height , and @xmath21 is the constant to account for the undetermined mechanism for viscosity . observations of t - tauri stars suggest @xmath25 @xcite , whereas the recent studies of purely hydrodynamic turbulence imply @xmath26 @xcite . we employ the values between @xmath27 and @xmath28 to account for this uncertainty . whereas the value of @xmath21 may vary vertically depending on the source mechanism of viscosity , including the existence of dead zone @xcite , a constant value of @xmath21 along all height is adopter here . during the early stage of disk evolution , the terrestrial region is hot enough so that the gas is partially ionized , and a vertically uniform turbulence is likely to be generated . the rosseland mean opacity of dust particles is calculated through the mie - scattering code of @xcite using the refractive index data of silicates @xcite . the opacity varies with the amount of dust and its grain size @xcite . the temperature dependence of opacity is ignored here , because it is weak compared to the effect of dust size and amount . the rosseland mean opacity is plotted as a function of grain size in figure [ fig_op ] . dust particles of 1 @xmath11 m size have the largest opacity for near - infrared radiation . the grain size affects both opacity and settling speed , playing a crucial role in our model . when the dust first condenses , dust grains are assumed to nucleate homogeneously . for the initial grain size , we test a range of values spanning from 0.1 @xmath11 m to 10 @xmath11 m . the lower bound is taken from a typical interstellar grain size , and the upper bound is estimated from the mean free path , @xmath29 , and the thermal velocity of molecules , @xmath30 , assuming perfect aggregation at each collision . here we considered that two molecules or particles collide at an every interval of @xmath31 , and they always form an aggregate without bouncing . as molecules collide and stick , the number density becomes smaller and the interval of collision takes longer . one - hundredth of the time step ( see section [ imple ] ) is considered as the nucleation time scale , which is @xmath25 to 100 years in our calculations , and this gives the upper bound of 10 @xmath11 m as an initial nucleation size . given the assumption of perfect aggregation , it should be considered as an unlikely end - member case . this range covers the typical grain size of matrix materials ( 50 nm - 5 @xmath11 m ) in chondrites @xcite , which could be considered as the typical size of unmelted grains . grain growth occurs through additional condensation on existing grains ; i.e. , the number of dust particle is conserved by assuming heterogeneous nucleation . grain fragmentation , sticking , and dust size distribution are not considered in our model , and this simplification turns out to be justified a posteriori for most cases , as discussed in section 4 . equations ( [ dt / dz ] ) and ( [ dfz / dz ] ) are integrated from the disk surface towards the mid - plane . the surface temperature , @xmath32 , is used as a boundary condition and is converted from the effective temperature of the disk , @xmath33 , assuming that the disk is opaque to blackbody radiation due to the existence of @xmath11m - size grains . these temperatures are given by @xmath34 and @xmath35 where @xmath36 is the disk surface density @xcite . here the effective temperature is calculated by balancing viscous dissipation and blackbody radiation . the amounts of dust and its average chemical compositions at different heights in the system are calculated through gibbs free energy minimization . the disk temperature and pressure are necessary to determine stable phases , and the pressure is calculated assuming the ideal gas law ( @xmath37 ) . previous settling models have fixed the total dust amount in the system ( e.g. , @xcite ) , but our model determines the amount based on condensation calculation , making the model consistent with evolving chemistry . we have developed an original optimization method to minimize the free energy using a non - linear conjugate gradient method ( see appendix a for detail ) . only major elements ( h , o , na , mg , al , si , ca , and fe ) are considered in equilibrium calculations in this study . mineral species considered include corundum , melilite , spinel , olivine , and pyroxene . the complete list of species considered is given in table [ spec ] . thermodynamical data for silicates phases are taken from @xcite , and those for the rest of phases are from the janaf thermochemical tables . tracers are distributed uniformly in each grid cell at each time step , and they carry the information of dust amount and component in the nearby region . unlike conventional tracer methods ( e.g. , @xcite ) , the number density of tracers in our model does not correspond to the concentration of any kind . suppose that we populate @xmath38 tracers uniformly in one grid cell at @xmath5 and denote each tracer as @xmath39 @xmath40 . the tracer @xmath39 contains the total mass , average size , and average composition of all dust grains existing between @xmath41 and @xmath42 , where @xmath43 is the grid size . the tracers advect according to the settling velocity at each height . unlike grid - based methods , tracer method would not cause any numerical diffusion , thereby being able to calculate the settling accurately . although heterogeneous nucleation could cause the grain size to increase , a change in size is relatively small in our calculations , and no grains would overrun preceding grains . the distributions of temperature and dust composition are calculated based on this advected composition profile , using radiative heat transfer and gibbs free energy minimization simultaneously . after the settling is completed , we compute the average dust composition of each grid cell , by taking into account all of tracer particles in each cell . at this point , tracers still carry the dust composition calculated from the temperature profile at the previous time step . gibbs free energy minimization is then performed using the temperature and pressure at the upper node of the cell in order to obtain the new dust composition consistent with the new temperature profile . as we integrate the heat transfer equation from the disk surface to the mid - plane , the upper node is always updated first , and the upper node temperature corresponds to the new temperature for the cell . the temperature of the lower node of that cell is acquired by integrating equations ( [ dt / dz ] ) and ( [ dfz / dz ] ) , using the opacity based on the dust amount obtained from free energy minimization . when a grid cell has high dust density , however , a large opacity value of the cell could result in a substantial temperature difference between the upper and lower nodes of the grid . dust grains might not be able to remain in a solid phase anymore and evaporate at the lower end of the cell . however , using a single large opacity value will not account for this compositional variation and will result in an unrealistically high temperature at the lower node . in this case , we subdivide the grid cell and recalculate temperature and composition with higher resolution . we continue this grid refinement until additional subdivision changes the temperature of the lower node by less than 1 k. in this manner , the dust profile will be consistent with both radiative heat transfer and chemical thermodynamics . we used 100 grid cells ( size of 0.05 scale height ) and 20 tracer particles per grid cell . this particular choice of discretization is sufficient for our purpose , as shown later . the time step , @xmath44 , is chosen so that @xmath45 for the bottom 95% of dust grains by mass . the surface density @xmath36 is solved using a 1-d radial disk evolution model , derived from mass and angular momentum conservation laws , @xmath46,\ ] ] where @xmath47 is time , and @xmath48 is the radial distance from the sun @xcite . the viscosity @xmath20 is again described by the @xmath21-prescription . the disk temperature , necessary to compute the sound speed and thus the viscosity , is calculated using equation ( [ teff ] ) ( e.g. , @xcite ) . equation ( [ sigma ] ) is solved independently from the settling calculation , because dust settling in our model only causes a small difference on the mid - disk temperature and thus the viscosity . the vertical gas density profile is assumed to satisfy a hydrostatic equilibrium , and the surface density is converted to the vertical profile through @xmath49 the background gas density is assumed to vary temporally in proportion to the change of the surface density . in our model calculations , the gas composition is modified only through in situ and vertical processes such as condensation and dust settling , and we do not explicitly model the effect of radial transport on the gas composition . this is equivalent to assuming a similar chemical composition in the nearby region , and thus the radial mass flux into the system would not modify the gas composition . the evolution of gas surface density is calculated first using the radial disk evolution model of equation ( [ sigma ] ) , before solving the settling . the surface temperature is converted from the surface density using equation ( [ teff ] ) , and its evolution at 1 au and 4 au is plotted in figure [ fig_te1 ] . the timescale for disk evolution decreases inversely with the value of @xmath21 , which is proportional to viscosity . the evolution depends strongly also on an initial density profile , which is poorly constrained as well . in this study , we test two two initial disk masses , 0.35 and 0.5 solar mass , with a surface density profile inversely proportional to the distance from the sun @xcite , spanning up to 25 au ( @xmath50 , where @xmath51 is the surface density at 1 au ) . we used three values of @xmath21 ( @xmath52 , @xmath53 , and @xmath28 ) combined with two values of surface density @xmath51 ( 2.0 @xmath54 10@xmath0 and 3.0 @xmath54 10@xmath0 kg m@xmath55 , corresponding to 0.35@xmath56 and 0.5@xmath56 , respectively ) . we calculated the behavior of dust settlings at 1 au to 5 au with various parameter combinations . initially , our model has the solar composition of @xcite at all heights . when the disk interior is hot enough , si , mg , and fe evaporate , and when it is even hotter , al and ca also start to exist as gas phases . as we are considering an early stage of disk evolution , viscous heating is a major heat source of the disk , and the radiation from the sun has a negligible effect . depending on the value of @xmath21 , viscous heating is one to three orders of magnitude larger than the heat flux from the solar radiation , which is 3 w m@xmath55 at 1 au assuming a flat disk @xcite . this leads to an increase in temperature towards the mid - plane . dust first precipitates in the cooler upper region of the disk , and it sinks toward the mid - plane due to the gravity from the sun . the settling of micron - size particles is slow , but it gradually increases the dust - to - gas ratio in the lower region . in the inner region of the disk , if the value of @xmath21 is high enough , dust particles re - evaporate as they sink into the hotter interior and are prevented from further settling . figure [ fig_set1 ] shows some snapshots for the profiles of temperature and dust composition . the evaporation of dust particles creates a concentration of dust - composing elements . this leads to the formation of a dust - rich layer with high opacity , creating a large temperature gradient . any dust grain trying to cross this temperature jump will be subjected to evaporation , which prevents the grain from settling further ( figure [ fig_set1](a ) ) . this forms a `` condensation front '' , where dust particles concentrate due to a temperature jump created by themselves . figure [ fig_set1](c ) shows that mg , si , and fe are concentrated above the condensation front at @xmath22.5 times the scale height at @xmath57 year , because these elements all condense in a short range of temperature . figure [ fig_set1](b ) shows a pair of condensation and evaporation occurring above and below the front . when the total evaporated amount is larger than 10% of the total condensed amount , a condensation front is defined to have formed . the tracer method allows us to resolve condensation and evaporation within a single grid cell . all of our calculations were made with 100 grid cells and 20 tracer particles per grid , but a finer grid resolution or a denser tracer population does not change our results significantly . table [ conv ] lists the values of the mid - disk temperature at @xmath58 years obtained by varying the grid size and the number of tracers per cell . the front descends slowly with the same settling speeds as dust particles . repeating evaporation and condensation will keep the grain size in its initial nucleation size . some grains might go through collisional sticking and grow in size , but such larger grains would settle faster into a hotter region , ending up with evaporation . the condensation front and the associated temperature jump will be sustained by the remaining smaller particles . this justifies our strategy of neglecting grain growth in our calculations . some representative temporal model evolutions at 1 au are shown in figure [ fig_tev ] . as time proceeds , the average temperature decreases , and dust mass increases . by comparing the cases with @xmath59 and @xmath60 kg m@xmath55 , it can be seen that a higher surface density results in a hotter interior , keeping more dust - composing elements in the gas phase . for both cases , however , the ratio of evaporated and condensed amounts converges to almost unity , indicating that the amounts evaporating below and condensing above the front are nearly equal . the amount of dust is self - regulated below a certain value so that it would not create too large a temperature increase to prohibit the existence of the dust altogether , resulting in the ratio of evaporated and condensed close to unity . therefore , the settling of the condensation front could be even slower than the settling of individual dust particles . a larger initial nucleation size of 10 @xmath11 m or a smaller viscosity of @xmath27 would result in a lower mid - disk temperature , and thus it would take longer to form a condensation front with the ratio close to unity , but the front still forms after a certain period of time . in the outer terrestrial region , a different kind of model evolution is observed ( figure [ fig_tev3 ] ) . the ratio of evaporated and condensed would never be larger than unity , but some fraction of dust is still evaporating due to the formation of the front . even for the case of @xmath61 and @xmath62 kg m@xmath55 , the ratio rapidly decreases , due to its short radial evolution timescale . this is the case with the highest viscosity and the highest initial disk density among the parameter combinations considered in this study , so it has the highest potential to sustain a condensation front . the model with @xmath63 has a lower mid - disk temperature from the beginning , so the ratio also quickly decreases close to zero . another rise in the ratio can be seen after 5 @xmath54 10@xmath0 years , and this is caused by the condensation front formed by na and si , which condense at lower temperatures ( figure [ fig_setn ] ) . the temperature decreases slowly compared to the case with @xmath61 , and this leaves a room for a lower - temperature condensation front . because na has a small mass fraction compared to other major elements , it only has a minor effect on the evolution of the mid - disk temperature . the settling slows down as it approaches the mid - plane , because the settling speed is inversely proportional to the background gas density ( equation ( [ vset ] ) ) . both the surface and mid - disk temperatures gradually cool down as viscous heating decreases along with the dissipation of the nebular gas . the disk interior , where most of the mass exists , remains hotter than the condensation temperatures of mg , si , and fe due to the condensation front , whereas the region closer to the surface cools down within a few thousands of years . in all the cases , highly refractory species ( corundum and melilite ) do not form a condensation front . this is because ca- and al - bearing grains do not have a sufficient amount to form a front , and also because they only condense in the hot inner region , where it has too little radiation flux to form a temperature jump . the temperature gradient is proportional to the radiation flux , which is smaller in the interior . the absence of a ca or al condensation front suggests that highly refractory species settle without being affected by evolving chemistry . it is likely that ca and al dust settle simply following classical mechanics , and it is necessary to consider the grain - size evolution in a more careful manner . the formation of the condensation front inhibits dust settling , resulting in an extended settling timescale , which is comparable to that of radial evolution . at 1 au , the condensation front would last as long as @xmath64 years for @xmath65 , and @xmath66 years for @xmath61 with the initial nucleation size of 1 @xmath11 m , whereas the radial evolution occur in @xmath67 years . whether dust accumulates near the mid - plane or not has a large impact on subsequent evolution processes ( e.g. , @xcite ) , and the effect of the condensation front could be potentially crucial . figure [ fig_sum ] summarizes the condition for the formation of a condensation front as a function of the value of @xmath21 and the distance from the sun . when the value of @xmath21 is larger than 10@xmath68 , it is highly likely that a condensation front forms in the terrestrial region . if the initial nucleation size is limited to around 1 @xmath11 m , the condensation front could form as far as in the asteroid belt region ( figure [ fig_sum](b ) ) . for @xmath61 , the region spans even further close to the jovian orbit . even if the initial nucleation size is as small as interstellar grains or as large as 10 @xmath11 m , the existence of front is likely at @xmath69 . the front forms with a higher possibility when grains with 1 @xmath11 m exist , because the opacity and thus the mid - disk temperature are the highest at 1 @xmath11 m ( figure [ fig_op ] ) . given that the initial nucleation size would exhibit some finite distribution in reality , a small amount of dust with 1 @xmath11 m is expected to always exist , which would contribute to the formation of the front . the formation of condensation front could well be ubiquitous in the inner terrestrial region , causing the slow settling of dust grains . the existence of dust grains , which is necessary for the subsequent evolution to take place , is controlled by thermodynamics . in early protoplanetary disks , it is likely that silicates evaporated in the terrestrial planet region @xcite , and it would be essential to consider the complex evolution caused by condensation . our results indicate a substantially longer timescale for dust settling compared with the standard theory of planetary formation @xcite . the standard theory suggests that dust grains settle in the order of a few thousand years and grow into mm - size by the time they reach the mid - plane . when condensation is considered , however , our study shows that the fate of dust settling can be more complex , affected by the formation of a condensation front and its slow descent . dust concentrates above the front forming a temperature jump , and dust would not be able to penetrate this temperature jump created by itself . evaporation and condensation would repeat at the front , keeping the dust size in its initial nucleation size . a condensation front survives until viscous dissipation becomes weak . eventually , metallic iron , forsterite , and enstatite become stable near the cold mid - plane , and no evaporation will happen after that . even with a large initial grain size of 10 @xmath11 m , the settling takes @xmath210@xmath0 years at 1 au , which is two orders of magnitude longer than classical estimates , and with an initial size of 1 @xmath11 m , it takes longer than half a million years . in the meantime , some fraction of dust composing elements would dissipate as gas , without becoming planetesimals . this could potentially change the estimate of the minimum solar nebular mass . the prolonged time scale for dust settling will also change the interval between the formation times of cai and planets , because planet formation is not likely to start until dust settles to the mid - plane . this may affect our understanding of the solar system chronology . repeating evaporation and condensation limit the dust size to around the initial nucleation size , which again differs from the grain size predicted from pure mechanics . this could possibly explain the dominance of @xmath11m - size particles in the disks inferred from some astronomical observations @xcite . the observation of a region where a condensation front forms should show the dominance of @xmath11m - dust . what we would observe from the earth should correspond to the surface of micron - size dust - rich layer since dust is depleted in the overlying region of the front . condensation front will last for an order of radial evolution time scale , which makes @xmath11m - size dust observable for a prolonged time . this explanation , however , holds only for disks with a hot interior , while many observations of @xmath11m - size particles in cold disks also exist . the fragmentation of large - size dust aggregates could also replenish small grains , resulting in the existence of @xmath11m - size dust and slow settling @xcite . although it is neglected in our calculation , turbulent mixing could relax the concentration of dust , along with the steep temperature jump , at the condensation front . a gradual increase of temperature will still cause dust grains to evaporate at the end of the temperature jump , and repeating condensation and evaporation are likely to happen , meaning that the suppression of grain size will still occur . if turbulence is weak enough in such a upper layer , the concentration of dust above the front may even cause the layer to be gravitationally unstable . the degree of viscous turbulence could be complex depending on the source mechanism , although it is simply characterized by a constant parameter @xmath21 in this study . this assumption is reasonable for most of our cases , since the regions considered here are the hot inner part of the disks , which remain hotter than 1000 k , leading the gas to be ionized through collision . this assumption , however , could be invalid in the outer cold region . future studies are warranted to address these important complications . various element abundances recorded in chondrites have shown a depletion trend correlating with their volatility , and its origin has been debated for the past several decades @xcite . the elements with condensation temperatures between that of silicon and sulfur in particular have shown a clear trend @xcite , and this trend has been called the depletion of moderately volatile elements ( move ) . quite a few cosmochemical models have tried to explain the move depletion based on the idea of incomplete condensation , i.e. , fractional condensation in a gradually cooling and dissipating protoplanetary disk ( e.g. , @xcite ) . @xcite , for example , used an analytical disk model to calculate the move depletion trend , and his model has succeeded in reproducing the observed trend . however , the result changes substantially just by incorporating the infall of dust caused by headwind @xcite , and other mechanisms suggested by astrophysical theory , such as turbulence and grain size evolution , have also been neglected in these analytical disk models . the abundance of refractory elements , such as the al / si and mg / si ratios , have not been fully explained as well . most of chondrites show similar values of al / si and mg / si to those of the solar photosphere @xcite . however , if , for some period of time , the nebula has been hotter than the si condensation temperature , gas containing si would have dissipated , while al would have remained in the nebula as solid . under the framework of incomplete condensation , therefore , it should leave a depletion trend not only in volatile elements , but also in refractory elements . when the al / si ratio is calculated using the approach of @xcite , al / si is predicted to be at least factor of 2 higher than that of ci chondrites . furthermore , ordinary and enstatite chondrites show depletion in al , which is one of the most refractory elements . this could not be explained by incomplete condensation . figure [ fig_me ] plots the molar ratios of al / si and mg / si for various types of chondrites . the lack of a depletion trend between al and si indicates either that the highest temperature achieved in the disk must have been below the condensation temperature of si ( e.g. , @xcite ) , or that the incomplete condensation hypothesis is neglecting some important mechanisms . astrophysical models , however , show that all the elements including refractory elements might have evaporated around 1 au when the mass accretion rate is larger than @xmath70 yr@xmath71 @xcite , which is a typical accretion rate for the disk at an early phase . our dust settling model may bring a new insight into the depletion trend observed in chondrites . in the framework of incomplete condensation , it seems inevitable to produce a depletion of si , if the terrestrial region exceeds the condensation temperature of si . as discussed in the previous section , corundum and gehlenite do not form a condensation front , which indicates that al and ca bearing refractory grains would not evaporate even if they settle below the condensation front . because their grain growth is not likely to be limited by the front , it will stick to each other and grow in size . particles larger than cm - size tend to decouple from the gas and drift inward in a short time scale @xcite . whereas dust bearing mg , si , and fe is suppressed to @xmath11m - size above the condensation front , refractory grains are likely to drift inward , and they would either concentrate in the inner region or fall into the sun . therefore , as long as the infall of materials does not end up in the region under consideration , dust grains remaining above the condensation front are likely to be the planet - forming materials in the region . for the sake of discussion , we assume that planets would have the same composition with what is left above the front and compare it with the al / si and mg / si ratios of chondrites . as seen in figure [ fig_me ] , the composition above the condensation front can exhibit a significant depletion in both al and mg , and its compositional variation can cover from ordinary to enstatite chondrites . most of al - rich dust particles would penetrate through the front and settle towards the mid - plane , though those associated with albite ( naalsi@xmath72o@xmath73 ) could remain above the front . the amount of al is thus determined by albite , but when condensation front forms further up from the mid - plane in the hot inner disk , and the formation and settling of corundum ( al@xmath74o@xmath72 ) would only leave a small fraction of al left for albite to form . therefore , al above the front will be smaller in those cases , resulting in a stronger depletion . the composition of earth has long been considered to be chondritic except for volatile elements @xcite , but quite a few geochemical data are showing the possibility of a non - chondritic origin @xcite . these studies suggested the necessity of a non - chondritic reservoir to explain the bulk silicate earth ( bse ) , but how to create such a reservoir has been controversial . however , our idea of condensation front might naturally explain a possible difference between chondrites and the earth . when refractory elements are transported inward , the enrichment of refractory elements in the inner region will occur together with the depletion in the outer region , the result of which might lead to the enrichment observed in bse @xcite and the depletion in ordinary chondrites . chondrites that formed further out where the condensation front did not form would not go through such enrichment nor depletion , and therefore , abundances close to ci chondrites are likely to be recorded . how far the condensation front formed can not sharply be constrained because of the uncertainty of disk parameters , but the boundary likely exists between 2 and 5 au , which corresponds to the source region of chondrites . in some cases , na and si also formed a condensation front ( figure [ fig_setn ] ) , potentially providing another mechanism to modify the abundance of move . with our 1-d dust settling model , the al / si and mg / si ratios can not be discussed in a quantitative manner , because the effect of radial drift has yet to be considered . our model , however , has shown the importance of the physical effect of condensation in the theory of planet formation , and we plan to consider the effects of grain growth and turbulence and to expand the model to 2-d in the future . although both turbulence and grain growth have small effects on vertical dust settling , they are likely to be important in radial models . turbulence would only smear out the temperature and dust peaks at the condensation front in our current vertical model , but it could cause a substantial transport of mass in the radial model . grain growth during vertical dust settling was suppressed by a condensation front , but it would not be the case in a 2-d model . grain growth could cause the variations of grain structure and surface composition , both of which could affect the opacity to some extent @xcite . some of the previous astrophysical models have already studied the effect of condensation on radial evolution . they showed that the effects of evaporation as dust grains drift inward could lead to an order - magnitude enhancement of evaporated element concentration in the vicinity of condensation temperature @xcite . however , most of the existing astrophysical models still treat condensation temperatures as if they were constant , although they may change by 100 - 200 k as pressure changes spatially and temporarily throughout the disk . the enhancement of dust / gas ratio or the concentration of certain elements could also affect condensation temperatures significantly ( @xcite ; see also appendix a ) . moreover , most of the astrophysical models classify dust into only two classes , i.e. , ice and silicate , although this treatment will lose the important information of condensing minerals , hindering the quantitative comparison with meteoritic data . some other radial models that tried to relate to cosmochemistry ( e.g. , @xcite ) have used an analytical disk model ( e.g. , @xcite ) . however , this ignores the effect of evolving dust composition on opacity . opacity is sensitive to grain size , shape , and its composition , and it should be calculated by honoring the disk environment @xcite . in our approach , the effect of chemistry on dust migration is taken into account accurately by thermodynamic calculations , and the evolution of compositional variation within the disk is fed back to the dynamics of disk evolution through opacity calculation , making the calculation fully self - consistent . by comparing results from our future 2-d models with the meteoritic data , we may be able to estimate some disk parameters such as initial density profile , viscosity , and grain size evolution , all of which are still highly unknown . to extract such astrophysical constraints from cosmochemical data , it is of utmost importance to develop a new theory of protoplanetary disks , which is consistent with both physics and chemistry . we constructed a vertical dust settling model by incorporating the effects of chemistry to the classical physical model of dust settling in a self - consistent manner . we showed that dust grains evaporate as they settle towards the hotter interior and cause the concentration of dust - composing elements . this leads to the formation of a condensation front . concentrated dust grains act as a high opacity layer , generating a large temperature increase , which hampers further dust settling by evaporation . the formation of a condensation front is ubiquitous at around 1 au and could exist even at as far as 4 au depending on the choice of disk parameters . the existence of the front would significantly change the timescale of dust settling and also cause highly refractory grains ( with al or ca ) and moderately refractory grains ( with mg , si , or fe ) to settle in different modes . moderately refractory grains with mg , si , and fe would form a condensation front and settle in a long timescale , but grains with al and ca would not form the front and are likely to settle quickly towards the mid - plane . al and ca could end up in a large grain size and experience radial drifting . this could potentially explain the chemical diversity observed among the bulk silicate earth , ordinary chondrites , and carbonatious chondrites . * acknowledgements . * this work was supported in part by the facilities and staff of the yale university faculty of arts and sciences high performance computing center . .chemical species considered in gibbs free energy minimization [ cols="^,<",options="header " , ] the total gibbs free energy of the system that contains @xmath75 different species may be expressed as @xmath76 where @xmath77 and @xmath78 denote the quantity and chemical potential of @xmath79-th species , respectively . for solid , the chemical potential is assumed to be independent of pressure , because the pressure effect is small in the low pressure condition . the chemical potential of gas is given by @xmath80 for gas , where @xmath81 is the chemical potential at a pressure of 1 atm and temperature @xmath4 , and @xmath38 is the total number of moles of gaseous species . equilibrium state is acquired by minimizing the gibbs free energy under the constraint of mass balance equation : @xmath82 where @xmath83 is the number of elements in the system , @xmath84 is the number of atoms of element @xmath85 in @xmath79-th species , and @xmath86 is the number of moles of element @xmath85 in the system . equation ( [ mass_b1 ] ) could be written in a matrix form : @xmath87 . in the system consisted of h , o , mg , and al , for example , the corresponding row of al@xmath74o@xmath72 in @xmath88 will be @xmath89 . we use a non - linear conjugate gradient method to minimize equation ( [ n - gibbs ] ) . we start with an initial composition of @xmath90 and incrementally update the composition along the direction calculated from the gradient at each iteration . the gradient of the free energy function , @xmath91 , is given by @xmath81 for a solid species , and by @xmath92 for a gas species . this gradient , however , does not satisfy the mass balance relationship , so we project the gradient to the null - space of @xmath93 using the projection matrix , @xmath94 , so that numbers of moles of the elements are conserved . by applying @xmath93 , mass conservation can be shown to hold : @xmath95 . the updated composition @xmath96 is searched as a minimum along the direction of @xmath97 : @xmath98 , where @xmath99 denotes an adjustable step length . the line search direction @xmath97 is calculated using the polak - ribire method @xcite : @xmath100 where @xmath101 is the projected gradient , @xmath102 . for the initial search , @xmath103 is set to @xmath104 . in order to calculate the step length @xmath99 that minimizes the free energy along the search direction @xmath97 , the bisection method is adopted , by using the fact that the dot product of @xmath97 and @xmath105 is zero at the minimum . other search methods that directly compare the free energy are less reliable , because subtle changes in free energy caused by metal elements can be so small compared with the free energy of hydrogen that their effect can not be represented by limited numerical precision . another constraint on minimizing the free energy is that the number of moles of every species should be non - negative . if the result of the line search returns a negative amount for a certain species , the step length @xmath99 is adjusted to satisfy the non - negative condition for all species . at a certain iteration , a number of moles @xmath77 could be zero for some @xmath79 , but the @xmath79-th component of the search direction @xmath97 could be negative , prohibiting further optimization . in this case , we remove the @xmath79-th species from our calculation for this step and re - optimize the number of moles . we shall , however , return @xmath79-th species into our calculation in the next iteration , because there is no guarantee that @xmath77 is zero at the global minimum . similarly , when the amounts of several species are zero and their search directions are pointing towards negative , we should not remove all the species from the calculation at once . even if the @xmath79-th component of the original projected gradient is positive , the sign of the component may change by excluding some species . we should thus check whether each component of the projected gradient is positive or not every time we remove any of the species from our calculations . it is important to randomize the order of removing the species from the system , because whether the projected gradient is positive or negative depends on the set of species we choose to compose @xmath93 . it is possible that when we remove the species from our calculations in a certain order , the projected gradient could end up with a zero vector , but if we change the order of removing the species , we may obtain a non - zero gradient that points to a lower energy state , satisfying the non - negative molar amount constraint . gas species may show a extremely small amount at global minimum , but its derivative might show a large value due to the steep nature of the derivative of log function even close to the minmum . when the gradient is large despite of the composition being close to the global minimum , it causes extra iterations to converge , so removing gas species with infinitesimal amount from the calculation at certain iterations is a way to make the calculation converge faster . this serious of procedures is continued until the norm of the projected gradient becomes sufficiently small . a flowchart for this procedure is shown in figure [ fig_flow ] . the validity of our code is checked against hsc chemistry ( version 8.1 ) using the solar abundance data from @xcite . for all of the temperature range considered , our code matches the result from hsc chemistry within the order of numerical error ( figure [ fig_seq](c ) ) . the condensation sequences starting with solar abundance gas and dust - enriched gas are calculated using our method , and they are plotted in figure [ fig_seq ] . the dust - enriched gas contains 20 times more na , mg , al , si , ca and fe relative to solar composition gas . it can be seen that the condensation temperatures of these elements are up to 100 - 200 k higher in the dust - enriched system . alexander , c. m. o. , hutchison , r. , barber , d.j . 1989 , epsl , 95 , 187 alexander , c. m. o. 2005 , m&ps , 7 , 943 anders , e. 1964 , ssr , 3 , 583 armitage , p. j. 2010 , astrophysics of planet formation ( cambridge university press ) bell , k. r. , cassen , p. m. , klahr , h. h. , and henning , th . 1997 , , 486 , 372 bland , p. a. , alard , o. , benedix , g. k. , kearsley , a. t. , menzies , o. n. , watt , l. e. , and rogers , n. w. 2005 , pnas , 102 , 39 calvet , n. , hartmann , l. , and strom , s. e. 2000 , in protostars and planets iv , ed . mannings , v. , boss , a. p. , and russell , s. s. ( tucson : univ . arizona press ) campbell , i. h. and oneill , h. st . c. 2012 , nature , 483 , 553 cassen , p. , 1996 , mps , 31 , 793 chambers , j. e. 2009 , , 705 , 1206 chambers , j. e. 2014 , , in treatise on geochemistry , 2nd edition , vol . 2 , ed . holland , h. and turekian , k. ( new york : elsevier ) chiang , e. and goldreich , p. 1997 , , 490 , 368 chiang , e. 2008 , , 675 , 1549 chick , k. m. and cassen , p. 1997 , , 477 , 398 cuzzi , j. n. and zanhle , k. j. 2004 , , 614 , 490 ciesla , f. j. and cuzzi , j. n. 2006 , icarus , 181 , 178 ciesla , f. j. 2007 , , 654 , l159 ciesla , f. j. 2008 , mps , 43 , 639 ciesla , f. j. 2011 , , 740 , 9 draine , b. t. 1985 , , 57 , 587 drake , m. j. and righter , k. 2002 , nature , 416 , 39 dullemond , c. p. and dominik c. 2005 , a&a , 434 , 971 ebel , d. s. and grossman , l. 2000 , gca , 64 , 339 estrada , p. r. , cuzzi , j. n. , and morgan d. a. 2016 , , 818 , 200 fleming , t. , and stone , j. m. 2003 , apj , 585 , 908 goldreich , p. and ward , w. r. 1973 , , 183 , 1051 grossman , l. 1972 , gca , 36 , 597 hartmann , k. , calvet , n. , gullbring , e. , and dalessio , p. 1998 , , 495 , 385 hayashi , c. , nakazawa k. , and nakagawa y. 1985 , in protostars and planets ii , ed . black , d. c. and matthews , m. s. ( tucson : univ . arizona press ) hubbard , a. and ebel , d. s. 2014 , icarus , 237 , 84 hueso , r. and guillot , r. 2005 , a&a , 442 , 703 larimer , j. w. 1967 , gca , 31 , 1215 johansen , a. , henning , t. , klahr , h. 2006 , apj , 643 , 1219 johansen , a. , oishi , j. s. , mac low , m. m. , klahr , h. , henning , t. , youdin , a. 2007 , nature , 448,1022 lewis , j. s. 1974 , science , 186 , 440 lodders , k. 2003 , , 591 , 1220 lyubetskaya , t. and korenaga , j. 2007 , , 112 , b03211 mtzler , c. 2002 , tech . rep . 2002 - 11 , institut fr angewandte physik , universitt bern monaghan , j. j. 1985 , comp . , 3 , 71 nakagawa , y. , sekiya , m. , and hayashi , c. 1986 , icarus , 67 , 375 nelson , r. p. , gressel , o. , umurhan , o. m. , 2013 , , 435 , 2610 palme , h. , larimer , j. w. , and lipschutz , m. e. 1988 , meteorites and the early solar syatem palme , h. and boynton , w. v. 1993 , in protostars and planets iii , ed . levy , e. h. and lunine , j. i. ( tucson : univ . arizona press ) palme , h. and oneill , h. s .c . 2014 , in treatise on geochemistry , 2nd edition , vol . 3 , ed . holland , h. and turekian , k. ( new york : elsevier ) palme , z. and zipfel , j. 2016 , 47th lpsc , 2252 , 2016 polak , b. and ribire , g. 1969 , rev . oper . , 16 , 35 pollack , j. b. , mckay , c. p. , and christofferson b. m. 1985 , icarus , 64 , 471 pollack , j. b. , hollenback , d. , beckwith , s. , simonelli , d. p. , and fong . w. 1994 , , 421 , 615 pringle , j. e. 1981 , araa , 19 , 137 robie , r. a. and hemingway , b. s. 1995 , u.s . geological survey bulletin 2131 ruden , s. p. , pollack , j. b. , 1991 , , 375 , 740 . safronov , v. s. 1972 , evolution of the protoplanetary cloud and formation of the earth and planets ( jerusalem : israel program for scientific translations , keter publishing house ) scott , e. r. d. and krot , a. n. 2014 , in treatise on geochemistry , 2nd edition , vol . 1 , ed . holland , h. and turekian , k. ( new york : elsevier ) semenov , d. , henning , th . , helling , ch . , ilgner , m. , and sedlmayr , e. 2003 , a&a , 410 , 611 shakura , n. i. and sunyaev , r. a. 1973 , a&a , 24 , 337 stoll , m. h. r. and kley , w. 2014 , a&a , 572 , a77 tackley , p. j. and king , s. d. 2003 , ggg , 4 , 4 van boekel , r. , dullemond , c. p. , and dominik , c. 2005 , a&a , 441 , 563 wasson , j. t. and kallemeyn , g. w. 1988 , rspta , 325 , 535 weidenschilling , s. j. 1977 , , 180 , 57 weidenschilling , s. j. 1984 , icarus , 60 , 553 weidenschilling , s. j. 2011 , icarus , 214 , 671 youdin , a. n. and shu , f. h. 2002 , , 580 , 494 youdin , a. n. and goodman , j. 2005 , 620 , 459 yoneda , s. and grossman , l. 1995 , gca , 59 , 3413

a very common statistical procedure is obtaining a confidence interval for a physics parameter of interest , when there are uncertainties in quantities such as the acceptance of the detector and/or the analysis procedure , the beam intensity , and the estimated background . these are known in statistics as nuisance parameters , or in particle physics as sources of systematic uncertainty . we assume that estimates of these quantities are available from subsidiary measurements . ; that is not enough information for a bayesian approach . alternatively the data relevant for the physics and nuisance parameters could be bound up in the main measurement , and not require a subsidiary one . ] a variant of this procedure which is particularly relevant for particle physics is the extraction of an upper limit on the rate of some hypothesized process or on a physical parameter , again with systematic uncertainties . to specify the problem in more detail , we assume that we are performing a counting experiment in which we observe @xmath0 counts , and that the acceptance has been estimated as @xmath1 and the background as @xmath2 . for a signal rate @xmath3 , @xmath0 is poisson distributed with mean @xmath4 . here @xmath5 contains factors like the intensity of the accelerator beam(s ) , the running time , and various efficiencies . it is constrained to be non - negative , but can be larger than unity . we aim to study and compare different approaches for determining confidence intervals for this problem . in general we are interested in pathologies in these areas : * * coverage . * this is a measure of how often the limits that we deduce would in fact include the true value of the parameter . this requires consideration of an ensemble of experiments like the one we actually performed , and hence is an essentially frequentist concept . nevertheless , it can be applied to a bayesian technique . + coverage is a property of the technique , and not of the particular limit deduced from a given measurement . it can , however , be a function of the true value of the parameter , which is in general unknown in a real measurement . + undercoverage ( i.e. the probability of containing the true value is less than the stated confidence level ) is regarded by frequentists as a serious defect . usually coverage is required for all possible values of the physical parameter . occur at values which are not fixed in @xmath3 , but which depend on the details of our experiment ( such as the values of @xmath5 and @xmath6 ) . these details vary from experiment to experiment . thus we could achieve ` no undercoverage for the ensemble of experiments measuring the parameter @xmath3 ' , even if the individual coverage plots did fall below the nominal coverage occasionally . thus in some sense ` average coverage ' would be sufficient ( see for example reference @xcite ) , although it is hard to quantify the exact meaning of ` average ' . it should be stated that this is not the accepted position of most high energy physics frequentists . ] in contrast , overcoverage is permissible , but the larger intervals result in less stringent tests of models that predict the value of the parameter . for measurements involving quantised data ( e.g. poisson counting ) , most methods have coverage which varies with the true value of the parameter of interest , and hence if undercoverage is to be avoided , overcoverage is inevitable . + frequentist methods by construction will not undercover for any values of the parameters . this is not guaranteed for other approaches . for example , even though the bayesian intervals shown here do not undercover , in other problems bayesian 95% credible intervals could even have zero coverage for some values of the parameter of interest.@xcite it should also be remarked that , although coverage is a very important property for frequentists , on its own exact coverage does not guarantee that intervals have desirable properties ( for many examples , see refs . @xcite and @xcite ) . * * interval length . * this is sometimes used as a criterion of accuracy of intervals , in the sense that shorter intervals have less probability of covering false values of the parameter of interest . however , one should keep in mind that short intervals are only desirable if they contain the true value of the parameter . thus empty intervals , which do occur in some frequentist constructions , are generally undesirable , even when their construction formally enjoys frequentist coverage . + intervals that fill the entire physically allowed range of the parameter of interest may also occur in some situations . examples of this behavior are given in @xcite and @xcite . an experimenter who requests a 68% confidence interval , but receives what appears to be a 100% confidence interval instead , may not be satisfied with the explanation that he is performing a useful service in helping to keep the coverage probability averaged over his measurement and his competitor s measurements from dropping below 68% . * * bayesian credibility . * in some situations it may be relevant to calculate the bayesian credibility of an interval , even when the latter was constructed by frequentist methods . this would of course require one to choose a prior for all the unknown parameters . the question is one of plausibility : given the type of measurement we are making , the resolution of the apparatus , etc . , how likely is it that the true value of the parameter we are interested in lies in the calculated interval ? does this differ dramatically from the nominal coverage probability of the interval ? in fact for different values of the observable(s ) , frequentist ranges are very likely to have different credibilities . some examples of this behavior are noted in ref . @xcite . + when calculating the bayesian credibility of frequentist intervals , `` uninformative '' priors appear advisable . note that severe interval length pathologies will automatically produce a large inconsistency between the nominal coverage and the bayesian credibility of an interval . except in a handful of very special cases , it is not possible to construct an interval scheme that has simultaneously constant bayesian credibility and constant frequentist coverage , even if one has total freedom in choosing the prior(s ) . although it is not at all clear exactly how large a level of disagreement is pathological , nevertheless it may be instructive to know how severely an interval scheme deviates from constant bayesian credibility ( and how sensitive this is to the choice of prior ) . * * bias . * in the context of interval selection , this means having a larger coverage @xmath7 for an incorrect value @xmath8 of the parameter than for the true value @xmath9 . this requires plots of coverage versus @xmath8 for different values of @xmath9 . for upper limits , @xmath10 if @xmath11 is less than @xmath12 , so methods are necessarily biassed for low @xmath8 . bias thus is not very interesting for upper limits . it will be discussed in later notes dealing with two - sided intervals . * * transformation under reparametrisation . * intervals that are not transformation - respecting can be problematic . for example , it is possible for the predicted value of the lifetime of a particle to be contained within the 90% interval determined from the data , but for the corresponding predicted value of the decay rate ( equal to the reciprocal of the predicted lifetime ) to be outside the 90% interval when the data is analysed by the same procedure , but in terms of decay rate . this would result in unwanted ambiguities about the compatibility of the data with the prediction . * * unphysical ranges . * the question here is whether the interval construction procedure can be made to respect the physical boundaries of the problem . for example , branching fractions should be in the range zero to one , masses should not be negative , etc . statements about the _ true _ value of a parameter should respect any physical bounds . in contrast , some methods give _ estimates _ of parameters which can themselves be unphysical , or which include unphysical values when the errors are taken into account . we do not recommend truncating ranges of _ estimates _ of parameters to obey such bounds . thus the fact that a branching fraction is estimated as @xmath13 conveys more information about the experimental result than does the statement that it lies in the range 0.9 to 1 . * * behavior with respect to nuisance parameter . * we would normally expect that the limits on a physical parameter would tighten as the uncertainty on a nuisance parameter decreases ; and that as this uncertainty tends to zero , the limits should agree with those obtained under the assumption that the `` nuisance parameter '' was exactly known . ( otherwise we could sometimes obtain a tighter limit simply by pretending that we knew less about the nuisance parameter than in fact is the case . ) these desiderata are not always satisfied by non - bayesian methods ( see @xcite and @xcite ) . although we are ultimately interested in comparing different approaches to this problem , in this note we investigate a bayesian technique for determining upper limits . our purpose is to spell out in some detail how this approach is used , and to discuss some of the properties of the resulting limits in this specific example . we believe that , for variants of this problem ( e.g. different choice of prior for @xmath3 ; alternative assumptions about the information on the nuisance parameters ; etc . ) , the reader could readily adapt the techniques described here ( and the associated software ) to their particular situation . we will report on two - sided intervals and also compare with other methods ( e.g. cousins highland , pure frequentist , profiled frequentist ) in later notes . before dealing with the problem of extracting and studying the limits on @xmath3 as deduced from observing @xmath0 events from a poisson distribution with mean @xmath4 in the presence of an uncertainty on @xmath5 , we recall the way the bayesian approach works for the simpler problem of a counting experiment with no background and with @xmath5 exactly known . then @xmath0 is poisson distributed with mean @xmath14 , and bayes theorem s denote prior p.d.f.s , lower case @xmath15 s denote other p.d.f.s , upper case @xmath16 s denote prior probabilities , and upper case @xmath17 s denote other probabilities . equation ( [ eqn : bayesth ] ) is true for probabilities , p.d.f.s , or mixtures depending on whether @xmath18 and/or @xmath19 are discrete or continuous variables . ] @xmath20 gives @xmath21 where @xmath22 is the prior probability density for @xmath3 ; @xmath23 is the posterior probability density function ( p.d.f . ) for @xmath3 , given the observed @xmath0 ; and @xmath24 is the probability of observing @xmath0 , given @xmath3 . we assume a constant prior for @xmath3 , . ) ] and that @xmath24 is given by the poisson @xmath25 then of the discrete distribution ( [ eqn : poisson ] ) and the integral over @xmath3 of the continuous distribution ( [ eqn : poisson2 ] ) are both equal to unity . this means that the probability @xmath24 and the probability density @xmath23 are correctly normalised . ] @xmath26 the limit is now obtained by integrating this posterior p.d.f . for @xmath3 until we achieve the required fraction @xmath27 of the total integral from zero to infinity . if @xmath27 is @xmath28 , the upper limit @xmath29 is given by @xmath30 @xmath27 is termed the credible or bayesian confidence level for the limit . for different observed @xmath0 , the upper limits are shown in the last two columns of table [ ltablex ] , for @xmath31 and for @xmath32 respectively . the gaussian approximation for the case @xmath31 , @xmath33 , would yield @xmath34 , which is roughly comparable to the corresponding @xmath35 of the table . for @xmath31 , it coincidentally turns out that , for this particular example , the bayesian upper limits are identical with those obtained in a frequentist calculation with the neyman construction and a simple ordering rule ( see later note on the frequentist approach to this problem ) . in general this is not so . other priors sometimes used for @xmath3 are @xmath36 @xcite or @xmath37 @xcite . having a prior peaked at smaller values of @xmath3 in general results in tighter limits for a given observed @xmath0 . if the whole procedure is now repeated with a background @xmath6 and a flat prior , the upper limits not surprisingly decrease for increasing @xmath6 at fixed @xmath0 ( except for the case @xmath38 where the limits can trivially be seen to be independent of @xmath6 ) . this is not inconsistent with the fact that the mean limit for a series of measurements increases with @xmath6 , i.e. experiments with larger expected backgrounds have poorer sensitivity . .upper 90% limits for @xmath0 observed events with @xmath6 background and @xmath39 ( @xmath40 and @xmath41 , as defined in section [ submeas ] ) . also shown are limits for @xmath42 and @xmath43 with fixed @xmath44 . [ ltablex ] [ cols= " > , > , > , > , > , > , > , > , > , > , > , > " , ] next we can investigate the frequentist coverage @xmath45 when the poisson variable is generated with @xmath46 . this differs from @xmath7 where the coverage is checked at @xmath47 when the generation value is @xmath9 . ] of this bayesian approach . that is , we can ask what the probability is , for a given value of @xmath9 , of our upper limit being larger than @xmath9 , and hence being consistent with it . this is equivalent to adding up the poisson probabilities of eqn . ( [ eqn : poisson ] ) for those values of @xmath0 for which @xmath48 i.e. @xmath49 as @xmath9 increases through any of the values of @xmath29 of the last two columns of table [ ltablex ] , the coverage drops sharply . for example , for the case of zero background and efficiency known to be unity , the 90% bayesian upper limits will include @xmath50 for @xmath51 or larger . but @xmath52 is no longer below the upper limit for @xmath51 . thus one term drops out of the summation of eqn . ( [ csum ] ) for the calculation of the coverage at @xmath52 , while the remaining terms change but little for the small change in @xmath9 ; this produces the abrupt fall in coverage . the coverage is plotted in fig . [ fig : bayes ] , where the drop at @xmath53 can be seen . _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ the calculation of @xmath45 can be done as follows : the identity @xmath54= -e^{-x}{x^n\over n ! } \qquad\mathrm{for}\qquad f(x)=e^{-x}\sum_{k=0}^n{x^k\over k!}\ ] ] allows us to write ( integrating @xmath55 ) @xmath56 from this , it follows that `` relevant @xmath0 '' is equivalent to any one of these inequalities : @xmath57 and our expression for the coverage becomes @xmath58 where @xmath59 means `` sum until the next term would cause the sum to exceed @xmath60 '' . this result proves that @xmath61 for all values of @xmath9 in this simple example . _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ it is seen that the coverage starts at @xmath62 for small @xmath9 . this is because even for @xmath38 the bayesian upper limit will include @xmath9 , and this is even more so for larger @xmath0 . bayesian methods can be shown to achieve average coverage . by this we mean that when the coverage is averaged over the parameter @xmath3 , weighted by the prior in @xmath3 , the result will agree with the nominal value @xmath27 , i.e. @xmath63 a proof of this theorem is given in the second appendix , section [ ac ] of this note . for a constant prior , the region at large @xmath3 tends to dominate the average , while in general we will be interested in the coverage at small @xmath3 . thus the `` average coverage '' result is of academic rather than practical interest , especially for the case of a flat prior . indeed it is possible to have a situation where the average coverage is , say , @xmath28 , while the coverage as a function of @xmath3 is always larger than or equal to @xmath28 . our actual problem differs from the simple case of section [ bayes ] in that + ( a ) we have a background @xmath6 , assumed for the time being to be accurately known ; and ( b ) we have an acceptance @xmath5 estimated in a subsidiary experiment as @xmath1 . what we are going to do is to use a multidimensional version of bayes theorem to express @xmath64 in terms of @xmath65 and the priors for @xmath3 and @xmath5 . the relationship is and + + @xmath66 . + + so with @xmath67 , @xmath68 and @xmath69 identified with @xmath3 , @xmath5 and @xmath0 respectively , and with the prior for @xmath3 and @xmath5 factorising into two separate priors for @xmath3 and for @xmath5 , we obtain @xmath70 . ] @xmath71 to obtain the posterior p.d.f . for @xmath3 , we now integrate this over @xmath5 : @xmath72 and finally we use this to set a limit on @xmath3 as in eqn . ( [ eqn:0.9 ] ) . the coverage for this procedure needs to be calculated as a function of @xmath9 and @xmath73 . the average coverage theorem of the previous section must be generalized to @xmath74 to implement the above procedure we need priors for @xmath3 and @xmath5 . as in the simple example of section [ bayes ] , for simplicity we assume that the prior for @xmath3 is constant . it will be interesting to look at the way the properties of this method change as other priors for @xmath3 are used . we assume that the prior for @xmath5 is extracted from some subsidiary measurement @xmath75 . we do * not * assume that this implies that our belief about @xmath73 is represented by a * gaussian * distribution centred on @xmath76 , as this would give trouble with the lower end of the gaussian extending to negative @xmath5 . instead , we specify some particular form of the subsidiary experiment that provides information about @xmath5 , and then assume that a bayesian analysis of this yields a posterior p.d.f . for @xmath5 . slightly confusingly , this posterior from the subsidiary experiment is used as the prior for the application of bayes theorem to extract the limit on @xmath3 ( see eqns.([eqn : extended ] ) and ( [ eqn : posterior ] ) ) . somewhat arbitrarily , we assume that , for a true acceptance @xmath73 , the probability for the measured value @xmath76 in the subsidiary experiment is given by a poisson distribution @xmath77 where @xmath78 , @xmath79 and @xmath80 is a scaling constant and @xmath81 as the mean and variance of the posterior p.d.f . of eqn . ( [ eqn : posterior_e ] ) . ] . we interpret this as the probability for @xmath76 . this is discrete because the observable @xmath82 is discrete , but the allowed values become closely spaced for large @xmath80 . for small @xmath83 ( i.e. for large @xmath82 ) , these probabilities approximate to a narrow gaussian ( see fig . [ fig : comparison ] ) . given our choice of probability in eqn . ( [ eqn : pdf ] ) , the likelihood for the parameter @xmath5 , given measured @xmath76 , is @xmath84 this is the same function of @xmath5 and @xmath76 as eqn . ( [ eqn : pdf ] ) , but now @xmath82 is regarded as fixed , and @xmath5 is the variable . the likelihood is a continuous function of @xmath5 . it is compared with a gaussian in fig . [ fig : comparison2 ] . finally in the bayes approach , with the choice of a constant prior for @xmath5 , the posterior probability density for @xmath5 after our subsidiary measurement is @xmath85 which is obtained by multiplying the right - hand side of eqn . ( [ eqn : likelihood ] ) by unity . this * posterior * probability density for @xmath5 will be used as our * prior * for @xmath5 in the next step of deducing the limit for @xmath3 . the details of the necessary analytical calculations are presented in the appendix of this note . in this section we investigate the behavior of the bayesian limits in this example , especially the shape of the frequentist coverage probability as a function of @xmath9 . the posterior p.d.f . for @xmath3 has the form @xmath861\,ds\ ] ] where the likelihood , the prior for @xmath5 , the ( constant ) prior for @xmath3 , and the marginalization integral over @xmath5 are all prominently displayed . the posterior probability density for @xmath3 gives the complete summary of the outcome of the measurement in the bayesian approach . it is therefore important to understand its shape before proceeding to use it to compute a limit ( or extract a central value and error - bars ) . figure [ pdfs ] illustrates the shape of the posterior for @xmath3 ( i.e. marginalized over @xmath5 ) in the case of a nominal 10% uncertainty on @xmath5 , and an expectation of 3 background events . plots are shown for 1 , 3 , 5 , and 10 observed events . the posterior evolves gracefully from being strongly peaked at @xmath87 to a roughly gaussian shape that excludes the neighborhood near @xmath87 with high probability . technically , the posterior would be described as a mixture of @xmath88 beta distributions of the 2nd kind '' ( i.e. `` beta prime '' ) , `` inverted beta '' , `` pearson type vi '' , `` variance - ratio '' , `` gamma - gamma '' , `` f '' , `` snedecor '' , `` fisher - snedecor'' . ] , giving it a tail at high @xmath3 that is heavier than that of a gaussian . in this note , our main goal is to obtain a bayesian upper limit @xmath29 from our observation of @xmath0 events . it is by integrating the posterior p.d.f . out to @xmath89 that an upper limit is calculated : a @xmath90 upper limit is defined so that the integral of the posterior from @xmath87 to @xmath89 is 0.9 . the probability ( in the bayesian sense ) of @xmath91 is then exactly @xmath27 . table [ ltablex ] shows the upper limits ( @xmath92 ) for @xmath9320 observed events with @xmath428 and @xmath39 . ( integer values of @xmath6 are chosen for illustration purposes only ; @xmath6 can , of course , take any real value @xmath94 . ) one notices that when @xmath93 , the limit is independent of the expected background @xmath6 . this is required in the bayesian approach : we know that exactly zero background events were produced ( when no events at all were produced ) , and this knowledge of what _ did _ happen makes what might have happened superfluous . an interesting corollary is , in the case of no events observed , uncertainties in estimating the background rate are of no consequence in the bayesian approach , and must not contribute any systematic uncertainty to the limit . this reasoning does not hold in the frequentist framework , where what might have happened definitely does influence the limit . for comparison , limits for fixed @xmath44 with @xmath42 or @xmath43 are also shown in table [ ltablex ] . it is interesting that these two columns start out equal at @xmath93 and differ by almost exactly 3 for @xmath95 . in contrast , the difference between the @xmath42 and @xmath43 columns for @xmath39 is already greater than 3 at @xmath96 , and continues to grow as @xmath0 increases ; it is not clear whether the difference approaches a finite value as @xmath97 . in any case , the limits for @xmath44 exactly are all smaller than the corresponding limits for @xmath39 , as expected . the main quantity of interest in this subsection is the frequentist coverage probability @xmath19 as a function of @xmath9 ( for fixed @xmath73 and @xmath6 ) . because both the main and the subsidiary measurements involve observing a discrete number of events , the function @xmath45 will have many discontinuities . on the other hand , @xmath98 will be continuous ( for fixed @xmath9 ) . the explanation of this effect is as follows : _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ the measured data are @xmath0 events in the main measurement and @xmath82 events in the subsidiary measurement . for each observed outcome @xmath99 there is a limit @xmath100 . this limit includes the effect of marginalization over @xmath5 . all @xmath99 with @xmath101 and @xmath102 are possible , and the probability @xmath17 of observing @xmath99 can be calculated as the product of two poissons . ( it will depend on @xmath9 , @xmath73 , ) if we look at all the possible limits we can obtain , @xmath103 and sort them in increasing @xmath29 , the @xmath29 are countably infinite in number and dense in the same way that rational numbers are dense in the reals . to compute the coverage as a function of @xmath9 , we simply add up all the probabilities of obtaining @xmath99 with @xmath104 : @xmath105 this sum is over a countably infinite number of terms . if we increase @xmath106 slightly to @xmath107 and recalculate the coverage , we have to drop all the terms @xmath108 from the previous sum ( the @xmath109 for each term also changes continuously with @xmath106 , but this is no problem ) . if there are @xmath110 such terms , there are @xmath111 discontinuities in the coverage in the interval @xmath112 $ ] , since @xmath109 for each of these is finite , and we lose them one by one as we sweep across the interval @xmath112 $ ] . but it seems that , in general , we can always find a solution to @xmath113 for finite @xmath114 by going out to larger and larger @xmath0 and @xmath82 . so , although the discontinuity may be tiny , we can always find a finite discontinuity in any finite interval of @xmath106 . on the other hand , if we keep @xmath106 fixed and vary @xmath115 , we always sum over the same set of @xmath99 , since the definition of @xmath116 does not involve @xmath73 , and @xmath109 is continuous in @xmath73 . so the coverage is continuous as a function of @xmath73 for @xmath9fixed . _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ plotting a curve that is discontinuous at every point is somewhat problematical . the solution adopted here is to plot the coverage as straight line segments between the discontinuities , ignoring any discontinuities with @xmath117 . figure [ l100 ] shows @xmath45 for the case @xmath90 , @xmath118 , nominal 10% uncertainty of the subsidiary measurement of @xmath5 , and @xmath43 . we observe that @xmath119 in this range , and it is not clear numerically whether @xmath120 as @xmath121 . the same conclusions hold for fig . [ l25 ] , which illustrates the same situation with a 20% nominal uncertainty for the @xmath5-measurement . figure [ ecov ] shows @xmath98 for @xmath90 , @xmath122 , @xmath40 , and @xmath43continuous as advertised . the shape of the curve is quite similar to that of figs . [ l100 ] and [ l25 ] , so it seems that the coverage probability ( with @xmath6 fixed ) is approximately a function of just the product of @xmath73 and @xmath9 . this approximate rule is likely to fail in the limit as @xmath123 and @xmath121 , for example , but it seems to hold when @xmath73 and @xmath9 are at least of the same order of magnitude . when @xmath124 is small , of order 1 or less , the coverage is @xmath125100% , as in the simple case of fig . [ fig : bayes ] . otherwise , the behavior of coverage in figs . [ l100][ecov ] is superior to that of fig . [ fig : bayes ] , which has a much larger amplitude of oscillation . another frequentist quantity that characterizes the performance of a limit scheme is the sensitivity , defined as the mean of @xmath29 . figure [ sens ] shows the sensitivity as a function of @xmath9 for the case of fig . [ l100 ] ; @xmath126 is observed to be nearly linearly dependent on @xmath9 . there is one complication here : when the subsidiary measurement observes @xmath127 events , and the prior for @xmath3 is flat , @xmath128 . since the poisson probability of obtaining @xmath127 is always finite , @xmath129 is consequently infinite . so we must exclude the @xmath127 case from the definition of @xmath129 . ( in fig . [ sens ] the probability of obtaining @xmath127 is @xmath130 . ) a weakness of the bayesian approach is that there is no universally accepted method to obtain a unique `` non - informative '' or `` objective '' prior p.d.f . reference @xcite , for example , states : _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ put bluntly : data can not ever speak entirely for themselves ; every prior specification has _ some _ informative posterior or predictive implications ; and `` vague '' is itself much too vague an idea to be useful . there is no `` objective '' prior that represents ignorance . _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ nevertheless , ref . @xcite does derive a @xmath131 `` reference prior '' for the simple poisson case , which is claimed to have `` a minimal effect , relative to the data , on the final inference '' . this is to be considered a `` _ default _ option when there are insufficient resources for detailed elicitation of actual prior knowledge '' . reference @xcite attempts to discover the optimal form for prior ignorance by considering the behavior of the prior under reparameterizations . for the case in question , the form @xmath37 clearly has the best properties in this respect . we are using an flat ( @xmath132 ) prior for this study , which seems to be the most popular choice , but the appendix works out the form of the posterior using an @xmath133 prior , so we can briefly here summarize the results for the @xmath37 and @xmath131 cases : the @xmath37 prior leads to an unnormalizable posterior for all observed @xmath0 when @xmath134 . the posterior becomes a @xmath135-function at @xmath87 , @xmath136 for any @xmath27 , and the coverage is consequently zero for all @xmath137 . this clearly is a disaster . the @xmath131 prior results in a posterior p.d.f . qualitatively similar in shape to those of fig [ pdfs ] , except that the p.d.f.is always infinite at @xmath87 . for @xmath138 , this produces an extremely thin `` spike '' at @xmath87 , which has a negligible contribution to the integral of the posterior p.d.f . a more significant difference ( for frequentists ) between the @xmath131 and the @xmath132 case is that the coverage probability is significantly reduced : for the case of fig . [ l100 ] the @xmath131 prior pushes the minimum coverage down to @xmath1250.87 . so the @xmath131 prior leads to violation of the frequentist coverage requirement ; it undercovers for some values of @xmath9 . one might also seek to further improve the coverage properties by adopting an intermediate prior . for example , an @xmath139 prior would reduce the level of overcoverage obtained with the @xmath132 prior . how acceptable this approach would be within the bayesian statistical community is an interesting question . it should be noted that all the prior p.d.f.s considered in this note are `` improper priors''they can not be correctly normalized : in the case of the @xmath132 and @xmath131 priors , the integral from 0 to any value @xmath140 is finite , while the integral from @xmath140 to infinity is infinite . the corresponding integrals of the @xmath37 prior are infinite on both sides for all @xmath141 . improper priors are dangerous but often useful ; `` improper posteriors '' are generally pathological . extra care must be taken when employing improper priors to verify the normalizability of the resulting posterior when using a numerical method to obtain the posterior , it is very easy to miss the fact that its integral is infinite . we summarize here the restrictions forced on the priors for @xmath3 and @xmath5see the appendix for the analytical causes . the discussion below assumes @xmath134 . the prior for @xmath3 being of the form @xmath133 , we must require @xmath142 , as discussed above . as specified in this note , the prior for @xmath5 , being taken from the posterior from the subsidiary measurement with a flat prior , has been given no freedom . should the subsidiary measurement observe @xmath127 events , the posterior for @xmath3 is not normalizable when @xmath143 : @xmath128 when @xmath127 and @xmath143 . this behavior is due to a well known effect : the @xmath5 prior becomes @xmath144 when @xmath127 , which remains finite as @xmath145 . all such cases is the standard example . ] yield @xmath128 when @xmath143 ; any positive @xmath146 cuts off the posterior at large @xmath3 sufficiently rapidly to render it normalizable . from this point of view , a @xmath131 prior may seem preferable , but on the other hand , having @xmath128 when @xmath127 seems intuitively reasonable . ( in general , we have @xmath128 for @xmath147 , but @xmath148 are not popular choices . ) there is another approach possible to the gamma prior for @xmath5 : we may simply specify by fiat the form of the prior as @xmath149 where @xmath150 is no longer required to be an integer . in practice , one then might obtain @xmath150 and @xmath80 from a subsidiary measurement whose result is approximated by the gamma distribution . in such cases , one must require @xmath151 to keep the posterior normalizable . note that in this form , @xmath152 is the mean of the @xmath5 prior , @xmath153 is the mode , and @xmath154 is the variance . the subsidiary measurement is often analysed by other experimenters , who chose statistics to quote for their central value and uncertainty ( omitting additional likelihood information ) . it is then important to obtain @xmath150 and @xmath80 in a consistent way from the information supplied by the subsidiary measurement . if @xmath5 , for example , were estimated by a maximum likelihood method , one would identify the estimate with @xmath153 rather than @xmath152 . results have been presented on the performance of a purely bayesian approach to the issue of setting upper limits on the rate of a process , when @xmath0 events have been observed in a situation where the expected background is @xmath6 and where the efficiency / acceptance factor @xmath155 has been determined in a subsidiary experiment . we find that this approach , when using a flat prior for the rate , results in modest overcoverage . plots of the expected sensitivity of such a measurement and of the coverage of the upper limits are given . it will be interesting to compare these with the corresponding plots for other methods of extracting upper limits , to be given in future notes . reference @xcite provides the limit calculating software associated with this study in the form of c functions . here we present the details of the analytical calculation of the posterior p.d.f . for @xmath3 . for generality , we work through the calculation with a @xmath133 prior ; a flat prior is then the special case @xmath156 . we measure @xmath0 events from a process with poisson rate @xmath157 , and we want the bayesian posterior for @xmath3 , given improper prior @xmath133 . we compute the posterior for fixed @xmath5 and @xmath6 in this subsection ; the calculation with our prior for @xmath5 follows in the next subsection . we have @xmath158 where all factors not depending on @xmath3 have already been absorbed into the normalization constant @xmath159 , which is defined by @xmath160 where we have performed the indicated change of variable . expanding @xmath161 in powers of @xmath162 using the binomial theorem , we get @xmath163 recognizing this as of the general hypergeometric form , we write it as @xmath164\ ] ] to make the hypergeometric nature more explicit . using the modern notation@xcite for the falling factorial @xmath165 this is expressed as @xmath166 where @xmath111 is the notation of @xcite . ( @xmath111 , a confluent hypergeometric function , is often written @xmath167 , and the relation given here is only valid for integer @xmath101 . ) note that @xmath168 is a polynomial of order @xmath0 in @xmath6 ( for @xmath0 a non - negative integer ) , and is related to the laguerre polynomials . when @xmath156 , we get @xmath169 , which is related to the incomplete gamma function . when @xmath170 , we get @xmath171 , which is infinite , so we require that @xmath142 . our posterior probability density for fixed @xmath5 is then given by @xmath172 the subsidiary measurement observes an integer number of events @xmath82 , poisson distributed as : @xmath173 where @xmath80 is a real number ( connecting the subsidiary measurement to the main measurement ) whose uncertainty is negligible , so @xmath80 can safely be treated as a fixed constant . @xmath80 might be thought of , for example , as based on a cross section that is exactly calculable by theory . there is negligible ( i.e. zero ) background in the subsidiary measurement . the prior for @xmath5 is specified to be flat . the bayesian posterior p.d.f . for @xmath5 is then @xmath174 ( or @xmath175 instead of @xmath176 in the denominator if you prefer ) . this is known as a gamma distribution . the mean and rms of this posterior p.d.f . summarize the result of the subsidiary measurement as : @xmath177 note that the observed data quantity in the subsidiary measurement is an integer @xmath82 , while the quantity being measured by the subsidiary measurement is a positive real number @xmath5 . next we compute the joint posterior @xmath178 using the @xmath133 prior for @xmath3 and our gamma distribution prior ( i.e. the posterior derived from the subsidiary measurement ) for @xmath179 @xmath180 where it is convenient to write @xmath150 for @xmath181 . we have for the joint posterior p.d.f . @xmath182 where @xmath183 we calculated @xmath159 above , so we have @xmath184 @xmath185 @xmath186 the marginalized posterior for @xmath3 can then be expressed as @xmath187ds= { s^{\alpha-1}\kappa^{\mu-\alpha}\mathcal{i}_\epsilon\over \gamma(\mu-\alpha)\gamma(\alpha+n)m(-n,1-n-\alpha , b)}ds\ ] ] where the integral @xmath188 is given by @xmath189 the same procedure that was used for the normalization integral can be applied here , producing @xmath190 @xmath191^k\ ] ] @xmath192 @xmath193 which has a particularly simple form when the background term is zero : @xmath194 a beta distribution of the 2nd kind . note that we must require @xmath195 to obtain a normalizable posterior . our posterior p.d.f . for @xmath3 with @xmath5 ( and @xmath6 ) fixed is recovered exactly by taking the limit of @xmath196 as @xmath197 . this means that the limit of @xmath198 as @xmath197 is identical to the value of @xmath198 when @xmath5 is known exactly . this property may seem obvious , but it is violated by some frequentist methods of setting limits , so it is worth mentioning . we need to integrate @xmath196 up to some limit @xmath29 , which can be done analytically as follows . @xmath199 where the substitution @xmath200 has been performed . re - expanding the polynomial @xmath111 and integrating term by term yields @xmath201 where @xmath202 is the standard notation for the incomplete beta function @xmath203 which also satisfies the following recursion : @xmath204 using the same technique as above , we can calculate the @xmath205th moment of the posterior p.d.f . as @xmath206 where we utilize the rising factorial notation@xcite @xmath207 the expression for the mean of the posterior when @xmath156 can be simplified using the identity @xmath208 obtaining @xmath209 note that the 2nd term is very small when @xmath138 . the recurrence relation@xcite @xmath210 leads to a recurrence relation between moments @xmath211 the special case @xmath156 then yields @xmath212\ ] ] which leads to this approximation for the variance of the posterior @xmath213 here we very briefly consider the case where the background parameter @xmath6 also acquires an uncertainty . this case is more general than the fixed @xmath6 case that is the main subject of this note : the fixed @xmath6 case will be the subject of additional studies employing various popular frequentist techniques , with the goal of comparing their performance . we judge the more general case considered in this subsection to be more complicated than necessary for the purpose of comparing the various methods , but it is instructive to document the fact that the bayesian method can easily handle the more general case . we assume a 2nd subsidiary measurement observing @xmath214 events ( poisson , as was the case for @xmath5 ) , which , when combined with a flat prior for @xmath6 , results in a gamma posterior for @xmath6 of the form @xmath215 where @xmath216 is a calibration constant ( analogous to @xmath80 in the subsidiary measurement for @xmath5 ) . the posterior for @xmath6 becomes the prior for @xmath6 in the measurement of @xmath3 . after determining the joint posterior @xmath217 by using our priors for @xmath3 , @xmath5 and @xmath6 , we marginalize with respect to @xmath5 and @xmath6 , resulting in @xmath218 where we write @xmath219 for convenience , and @xmath220 is the hypergeometric function@xcite . as long as @xmath0 is a non - negative integer and @xmath142 , @xmath221 is a polynomial of order @xmath0 in @xmath222 ( closely related to jacobi polynomials ) . this marginalized posterior for @xmath3 can then be integrated , with the result @xmath223 these two equations closely resemble the main results of sections [ postpdf ] and [ intpostpdf ] : to recover the fixed @xmath6 results , simply substitute @xmath224 for @xmath225 above , and take the limit @xmath226 . in this appendix we prove that bayesian credible intervals have average frequentist coverage , where the average is calculated with respect to the prior density . we start from the bayesian posterior density : @xmath227 for a given observed value of @xmath0 , a credibility-@xmath27 bayesian interval for @xmath3 is any interval @xmath228 $ ] that encloses a fraction @xmath27 of the total area under the posterior density . such an interval must therefore satisfy : @xmath229 or , using the definition of the posterior density : @xmath230 now for coverage . given a true value @xmath231 of @xmath3 , the coverage @xmath232 of @xmath233 $ ] is the frequentist probability that @xmath231 is included in that interval . we can write this as : @xmath234 s_\mathrm{l}(n)\le s\le s_\mathrm{u}(n ) } } p(n\,|\,s_\mathrm{t } ) . \label{eq : acbci2}\ ] ] next we calculate the average coverage @xmath235 , weighted by the prior @xmath22 : @xmath236\\[6 mm ] & \;=\ ; \int_{0}^{\infty}\sum_{\substack{\text{all $ n$ such that:}\\[1 mm ] s_\mathrm{l}(n)\le s\le s_\mathrm{u}(n ) } } p(n\,|\,s)\,\pi(s)\ , ds , & & \text{using equation ( \protect\ref{eq : acbci2 } ) , } \displaybreak[0]\\[6 mm ] & \;=\ ; \sum_{n=0}^{\infty}\;\int_{s_\mathrm{l}(n)}^{s_\mathrm{u}(n)}\ ! p(n\,|\,s)\,\pi(s)\,ds , & & \text{interchanging integral and sum,\protect\footnotemark}\displaybreak[0]\\[6 mm ] & \;=\ ; \beta\ ; \sum_{n=0}^{\infty}\;\int_{0}^{\infty}\ ! p(n\,|\,s)\,\pi(s)\,ds , & & \text{using equation ( \protect\ref{eq : acbci1 } ) , } \displaybreak[0]\\[6 mm ] & \;=\ ; \beta\ ; \int_{0}^{\infty}\sum_{n=0}^{\infty } p(n\,|\,s)\,\pi(s)\,ds , & & \text{interchanging sum and integral,}\displaybreak[0]\\[6 mm ] & \;=\ ; \beta\ ; \int_{0}^{\infty}\!\pi(s)\ , ds , & & \text{by the normalization of } p(n\,|\,s ) , \displaybreak[0]\\[6 mm ] & \;=\ ; \beta , & & \text{by the normalization of } \pi(s ) .\end{aligned}\ ] ] this completes the proof . we have assumed here that the prior @xmath22 is proper and normalized to 1 , but the proof can be generalized to improper priors such as those we considered in this note . a constant prior for example , can be regarded as the limit for @xmath237 of the proper prior : @xmath238 where @xmath239 is @xmath240 if @xmath241 and @xmath242 otherwise . we then _ define _ the average coverage for the constant prior as the limit : @xmath243 the previous proof can now be applied to the argument of the limit and leads to the same result . the average coverage theorem remains valid when @xmath3 is multidimensional , for example when it consists of a parameter of interest and one or more nuisance parameters . in that case one needs to average the coverage over _ all _ the parameters . m. j. bayarri and j. o. berger , `` the interplay of bayesian and frequentist analysis '' , statistical science 19 , p 58 ( 2004 ) , + projecteuclid.org/dienst/ui/1.0/summarize/euclid.ss/1089808273 , + www.isds.duke.edu/~berger/papers/interplay.html . peter clifford , `` interval estimation as viewed from the world of mathematical statistics '' , cern yellow report cern 2000 - 005 , p 157 ( 2000 ) , _ proceedings of the workshop on confidence limits at cern , 1718 january 2000 _ , edited by l. lyons , y. perrin , and f. james , + doc.cern.ch/yellowrep/2000/2000-005/p157.pdf . giovanni punzi , `` a stronger classical definition of confidence limits '' , hep - ex/9912048 , www.arxiv.org/abs/hep-ex/9912048 . gnter zech , `` confronting classical and bayesian confidence limits to examples '' , cern yellow report cern 2000 - 005 , p 141 ( 2000 ) , _ proceedings of the workshop on confidence limits at cern , 1718 january 2000 _ , edited by l. lyons , y. perrin , and f. james , doc.cern.ch/yellowrep/2000/2000-005/p141.pdf . d. karlen , `` credibility of confidence intervals '' , in _ proceedings of the conference on advanced techniques in particle physics , durham , 1822 march 2002 _ , edited by m. whalley and l. lyons , p 53 , ( 2002 ) , + www.ippp.dur.ac.uk/workshops/02/statistics/proceedings/karlen.pdf . gary feldman , `` multiple measurements and parameters in the unified approach '' , _ fermilab workshop on confidence limits 2728 march , 2000 _ , p 11 , + conferences.fnal.gov/cl2k/copies/feldman2.pdf , + huhepl.harvard.edu/~feldman/cl2k.pdf . joel heinrich , `` user guide to bayesian - limit software package '' , cdf internal note 7232 , ( 2004 ) , + www-cdf.fnal.gov/publications/cdf7232_blimitguide.pdf ; + cdf statistics committee software page , + www-cdf.fnal.gov/physics/statistics/statistics_software.html . r. l. graham , d. e. knuth , and o. patashnik , `` concrete mathematics : a foundation for computer science '' , 2nd ed . , ( addison - wesley , reading , ma , 1994 ) ; planetmath mathematics encyclopedia , + planetmath.org/encyclopedia/fallingfactorial.html . m. abramowitz and i.a . stegun , editors , `` handbook of mathematical functions '' , ( united states department of commerce , national bureau of standards , washington , d.c . 1964 ; and dover publications , new york , 1968 ) , chapter 13 . m. abramowitz and i.a . stegun , ibid . , chapter 15 ; william h. press , et al . , `` numerical recipes '' , 2nd edition , ( cambridge university press , cambridge , 1992 ) , 5.14 and 6.12 , lib-www.lanl.gov/numerical/bookcpdf/c5-14.pdf , + lib-www.lanl.gov/numerical/bookcpdf/c6-12.pdf . for bayes 90% limits , for the simple case of no background and no uncertainty on @xmath244 . the dotted line at @xmath245 is given to show that the coverage never falls below 90% ( in this simple case ) . ] vs @xmath3 for @xmath246 , 3 , 5 , 10 . in each case , @xmath43 and @xmath39 ( i.e. @xmath40 and @xmath82=99).,title="fig:",width=302]-0.2 in vs @xmath3 for @xmath246 , 3 , 5 , 10 . in each case , @xmath43 and @xmath39 ( i.e. @xmath40 and @xmath82=99).,title="fig:",width=302 ] + vs @xmath3 for @xmath246 , 3 , 5 , 10 . in each case , @xmath43 and @xmath39 ( i.e. @xmath40 and @xmath82=99).,title="fig:",width=302]-0.2 in vs @xmath3 for @xmath246 , 3 , 5 , 10 . in each case , @xmath43 and @xmath39 ( i.e. @xmath40 and @xmath82=99).,title="fig:",width=302 ] +

supernova remnants ( i.e. remnants of supernova explosions ) are commonly considered to be cosmic particle accelerators . in this review paper i will summarise experimental evidence , gathered through @xmath0-ray observations mainly in the vhe regime supporting this notion ( please note , that in the following @xmath0-ray will be used to stand for vhe @xmath0-rays ) . typically , snrs were detected through radio observations @xcite . recent advances in the understanding of these objects has been made through x - ray observations with instruments such as asca , bepposax , xmm - newton and chandra and through @xmath0-ray observations with instruments such as hegra ( high energy gamma ray astronomy ) , and h.e.s.s . ( high energy stereoscopic system ) . based on morphological properties , snrs can be classified into three broad categories : _ shell - like _ , _ plerionic _ ( also called pulsar wind nebulae or crab - like ) connected to a pulsar and _ composite _ ( in which both , a shell and a plerion are present ) , the later often showing markedly different radio and x - ray morphologies . for an excellent detailed review on plerionic snrs , see e.g. @xcite . snrs are thought to be responsible for the acceleration of cosmic rays up to energies around the `` knee '' ( @xmath1 ev ) at which the spectrum of cosmic rays significantly hardens from @xmath2 to @xmath3 . this statement is backed by experimental facts , as well as by theoretical considerations . _ experimental evidence _ is lent mainly by ( * 1 * ) x - ray observations of young shell - type snrs such as sn1006 @xcite , and cas a @xcite , in which sites dominated by hard non - thermal _ x - ray synchrotron _ emission were found , indicating an electron population extending up to @xmath4 tev , far beyond thermal energies . ( * 2 * ) _ very high energy ( vhe ) @xmath0-ray _ observations ( @xmath5 gev ) revealed sites of non - thermal particle populations @xcite . through theoretical considerations is has been known for a long time that supernova explosions release just about the right amount of energy into their surrounding to account for the energy budget of the cosmic rays ( assuming that they convert @xmath6 of their energy into kinetic energy of the cosmic rays ) @xcite . furthermore well - established theoretical models exists explaining how particles can be accelerated in supernova shock waves to energies approaching the _ knee_. in shell - type snrs particles are accelerated in the expanding shock waves through diffusive shock ( also called first order fermi ) acceleration @xcite . in plerionic snrs particles are accelerated to non - thermal energies in the termination shock between the relativistic outflow of electrons from the pulsar surface and the outer nebula . predictions on the @xmath0-ray visibility of snrs ( later confirmed by h.e.s.s . @xmath0-ray observations although there is still an experimental ambiguity in the underlying particle population resonsible for the @xmath0-ray emission ) were given by @xcite . since charged particles below the knee at @xmath7 ev are deflected in ubiquitous magnetic fields on their way from the origin to us , we have to turn to neutral messengers to reveal the acceleration sites ( the gyroradius of 1 tev cosmic rays in a magnetic field of @xmath8g - scale is of the order of 0.1 pc , much smaller than the thickness of the galaxy of 200300 pc ) . since neutrino detectors have not yet proved to be sensitive enough to detect neutrinos from astrophysical sources ( apart from the sun and the direct supernova explosion sn 1987a ) , observations in the radio , x - rays and @xmath0-ray wavebands are so far our best access to non - thermal acceleration processes in snrs . @xmath0-rays ( and neutrinos ) are produced in hadronic interactions with subsequent pionic decay and can reveal the acceleration sites since they travel un - deflected from their origin . however , @xmath0-rays not only reveal the sites of hadronic acceleration ; they also act as a tracer for energetic electrons that produce @xmath0-rays via ic scattering off background photon fields ( such as star - light or the cosmic microwave background ( cmbr ) ) . an ambiguity or duality therefore exists in the responsible radiating particle population in most cases when detecting @xmath0-rays from astrophysical objects . in spite of this ambiguity the detection of @xmath0-rays above @xmath9 1 gev from snrs gives us direct access to particle acceleration processes and the advantage of @xmath0-rays in comparison to other wavebands is that these are not affected by dust obscuration , which is particularly important for the population of snrs located within the galactic plane . a large volumne of the galaxy can thus be probed for @xmath0-ray emission from snrs by observations through the galactic disk . if snrs are indeed sites of particle acceleration , @xmath0-ray emission is expected and one of the puzzling aspects of previous @xmath0-ray observations of snrs was the rather low level of emission compared to model predictions @xcite . the history of soft @xmath0-ray ( or hard x - ray ) detection of snrs started with the detection of the crab nebula in 1964 with a scintillation counter detector flown on a balloon launched from palestine , texas @xcite . today hard x - rays up to 100 kev have been detected from various snrs both young shell - types such as cas a @xcite and sn 1006 @xcite and plerionic - types such as the vela - x pwn @xcite and msh 15 - 5_2 _ thin x - ray filaments in young shell - type snrs detected with high - angular resolution instruments such as xmm - newton and chandra point to regions with high - magnetic fields ( up to 0.5 mg ) in which electrons rapidly lose energy through synchrotron emission @xcite . in higher energies @xmath0-rays comptel detected the radioactive @xmath10ti - line at 1.157 mev from the two shell - type snrs cas a @xcite and rxj0852.04622 ( vela junior ) @xcite . however , it should be noted , that higher sensitivity integral observations provided a confirmation for this detection for cas a , but could not detect this @xmath10ti - line in rxj0852.04622 . therefore , these claims are still somewhat controversial . egret at energies above 100 mev did not detect prominent young shell - type snrs such as tycho , kepler , cas a or sn 1006 , noticed however several intriguing spatial coincidences of unidentified sources in the galactic plane with individual prominent radio snrs , such as w28 , and @xmath0-cygni @xcite . the combination of source confusion especially in the galactic plane , caused by the rather poor angular resolution of the egret instrument and the ambiguity in existing counterparts prevented individual identifications . however , a statistical assessment shows a 4 - 5@xmath11 effect when trying to correlate the population of egret unidentified sources with the population of radio snrs @xcite . also plerionic snrs have not been unambiguously identified with egret sources , although again intriguing associations of egret unidentified sources with prominent plerions such as psrb170644 , and the kookaburra complex exist @xcite . from a population point - of - view pwn are one of the best candidates to account for low - latitude slowly varying unidentified egret sources as proposed by @xcite . lately , using vhe @xmath0-ray source positions in the kookaburra region @xcite the re - analysis of egret data provided strong evidence of correlation of the pwn detected in this region with the confused unidentified egret source 3egj14206038 @xcite . this new approach might prove a useful template for connection future glast and vhe @xmath0-ray detections . all these possible associations of source classes with unidentified egret sources will hopefully be tested following the launch of the upcoming glast satellite in late 2007 . the history of vhe @xmath0-ray ( e@xmath12gev ) detections of snrs started again with the detection of the crab nebula , the first object to be reported in this waveband by the whipple collaboration @xcite . various claims of detections of shell - type snrs have been made before the advent of the h.e.s.s.telescope system . cas a was detected by hegra in a very deep ( @xmath13 hours ) exposure @xcite . detections of sn1006 @xcite and rxj1713.73946 @xcite have been reported by the cangaroo collaboration . with the advent of the h.e.s.s . telescope system , for the first time a number of galactic snrs , both shell - type and plerionic in nature could be established . in the following i will describe these populations of cosmic accelerators along with prospects for supernova remnant observations with the upcoming glast satellite . the outline of this paper is as follows : section [ sec::hess_shell ] provides a description of advances made through the vhe @xmath0-ray detections of shell - type supernova remnants , section [ sec::hess_pwn ] provides the corresponding description for pulsar wind nebulae . section [ sec::hess_new ] describes supernova remnants found in @xmath0-rays and later identified as supernova remnants , while section [ sec::glast ] summarises prospects for supernova remnant observations with the upcoming glast satellite . by means of data taken with the h.e.s.s . telescope system during the first few years of operation for the first time in vhe @xmath0-ray astronomy resolved images of shell - type snrs above 100 gev could be taken . in particular the snrs rxj1713.73946 and rxj0852.04622 , with diameters of @xmath14 and @xmath15 respectively could be resolved with unprecedented detail in this energy band . on the other hand , sn1006 , one of the snrs most expected to emit @xmath0-rays in the energy band ( due to the strong non - thermal x - ray emission from the rims ) was not detected in deep h.e.s.s.observations @xcite . the upper limit derived by this observations turned out to be an order of magnitude below the previously reported cangaroo detection . a reanalysis of the cangaroo data along with newer data from the cangaroo - iii detector is consistent with the h.e.s.s . upper limits @xcite . therefore in the following the h.e.s.s . upper limits will be used in the discussion of @xmath0-ray emission from sn1006 . since several papers on both rxj1713.73946 @xcite and rxj0852.04622 @xcite have been published by the h.e.s.s . collaboration , the main focus of this review will lie on similarities and differences between the two objects with the addition of comparisons to sn1006 as the most prominent non - detected snr where appropriate . figure [ fig::shell_morphology ] shows @xmath0-ray excess maps for rxj1713.73946 , rxj0852.04622 , and sn1006 . both @xmath0-ray emitting objects show a shell - like structure with a surprising resemblance of their respective x - ray morphology ( the correlation coefficients between @xmath0-ray and x - ray counts are @xmath16 ) . for both objects the x - ray emission is completely dominated by non - thermal x - ray emission without traces of line emission , exhibiting small filamentary structures that are interpreted as zones where the magnetic field is high ( @xmath17 g ) such that electrons rapidly lose energy through synchrotron emission in these areas @xcite . both objects appear rather faint in radio with typical fluxes below or in the several tenth of jansky - regime for the whole shell , certainly lower than what would be expected from equipartition arguments @xcite . the distance to both objects is somewhat uncertain , for rxj1713.73946 it seems that a distance of @xmath18kpc is preferred from the column density inferred from x - ray data . this distance would make rxj1713.73946 most likely the remnant of the historical supernova event of ad393 . for rxj0852.04622 distance estimates range from as close as the vela pulsar ( @xmath19pc ) to as far as the vela molecular ridge ( @xmath18kpc ) . the age ranges from @xmath20 years in the close case to @xmath21 years in the far case . morphologically their @xmath0-ray emission , in particular the width of the shells is rather different . the apparent width of the shell for rxj1713.73946 comprises 45% of the radius of the snr , while the for rxj0852.04622 it approximates to 20% of the radius . there is no apparent correlation between the dense molecular material surrounding rxj1713.73946 as measured by the nanten telescope and the vhe @xmath0-ray emission as measured by h.e.s.s . , but in fact , assuming a typical energy of @xmath22 ergs in accelerated protons , the density needed to explain the @xmath0-ray flux through hadronic interactions is only 1 @xmath23 . sn1006 is somewhat distinct in its multi - frequency picture in that its surface brightness is higher in radio ( @xmath24 jy ) @xcite showing a pronounced shell - like structure . the x - ray emission , especially in the shell is dominated by non - thermal emission up to @xmath25kev . sn1006 was not detected in a deep ( 1000 ksec ) integral exposure above 20 kev @xcite and was also not detected in sensitive h.e.s.s . observations @xcite . the density surrounding the source was estimated from x - ray as well as optical observations and values as low as n@xmath26 @xmath23 have been invoked to explain the apparent absence of @xmath0-ray emission in a hadronic scenario . from the h.e.s.s . non - detection assuming a leptonic @xmath0-ray emission scenario on the cmbr a lower limit on the post - shock magnetic field of @xmath27 g can be derived @xcite . higher values of the magnetic field in excess of 40@xmath8 g have been derived from x - ray observation and application of diffusive shock acceleration scenarios , so the lower limit on the magnetic field is not in contradiction to these values . comparing the energy spectra of the two @xmath0-ray detected snrs , strong similarities can be made out . the spectral energy distribution ( sed ) for the three shell - type snrs discussed here is shown in figure [ fig::shell_spectra ] along with model spectra , showing typical leptonic and hadronic @xmath0-ray emission models . as can be seen from this plot rxj1713.73946 and rxj0852.04622 show a remarkably similar @xmath0-ray energy spectrum with a rather flat @xmath28-type distribution at lower energies with a deviation from this power - law at higher energies . the flat spectrum at lower energies has advocated claims that the @xmath0-rays might be generated by pionic decays rather than inverse compton scattering @xcite . however , @xcite claim that the data can be well fitted in terms of a leptonic model when applying an unbroken electron spectrum along with the galactic radiation fields . therefore , at the moment , no strong conclusions can be drawn from the spectral shape on the particle population responsible for the @xmath0-ray emission . the upcoming glast satellite , measuring in the energy range between 30 mev and 300 gev might be able to distinguish between hadronic and leptonic @xmath0-ray production mechanisms . also interesting to note is that the h.e.s.s . @xmath0-ray upper limit for sn1006 is more than an order of magnitude below these @xmath0-ray detections and therefore starts to be rather constraining for the values of the magnetic field ( in a leptonic scenario ) or the ambient matter density ( in a hadronic scenario ) . these first unambiguous detections of individual shell - type snrs allowed for important advances in the understanding @xmath0-ray emission from these objects . however , the open question remains what differentiates non - detected snrs such as sn1006 from prominent @xmath0-ray emitters such as rxj1713.73946 . pulsar wind nebulae ( pwn ) or plerions are objects powered by a relativistic particle outflow ( electrons and positrons ) from a central source a pulsar . this pulsar is a rapidly rotating neutron star generated in the supernova event . the wind of relativistic particles flows freely out until the outflow pressure is balanced by that of the surrounding medium . at that point a standing termination shock is formed at which particles are accelerated @xcite . the existence of electrons accelerated to energies @xmath29 tev in such pwn has been established by x - ray observations of synchrotron emission , e.g. in the crab nebula @xcite . vhe @xmath0-rays are generated in pwn from the high - energy electrons by non - thermal bremsstrahlung or inverse compton ( ic ) scattering on photon target fields , such as the cosmic microwave background ( cmbr ) or star - light . apart from the crab nebula , no individual pwne have been unambiguously associated with egret sources , although several unidentified egret sources are located in close proximity to prominent pwn , such as in the kookaburra region , or msh155__2__. pwn are however one candidate for the population of slowly varying low - latitude unidentified sources . glast will shed more light on this population and possibly establish pwn as emitters in the mev to gev range . in vhe @xmath0-rays pwn make up the majority of the identified galactic sources detected so far @xcite . apart from the crab nebula ( the brightest steady vhe @xmath0-ray source ) several prominent pwn were identified in vhe @xmath0-rays in the last two years . these detections include msh155__2 _ _ @xcite , vela x @xcite , the two sources in the kookaburra region @xcite and lately hessj1825137 @xcite . vhe @xmath0-ray emission from pwn comes in various disguises as shown in figure [ fig::pwn_morphology ] : these include a ) point - like emission such as from the crab nebula @xcite and from the composite snr g0.9 + 0.1 , where the @xmath0-ray emission was shown to originate from the central pwn @xcite , b ) emission tracing the x - ray contours around a central pulsar such as in msh15 - 5__2 _ _ or c ) asymmetrically extending to one side and tracing the x - ray contours such as in vela x and finally d ) the emerging new class of offset pwn exemplified by hessj1825137 where the @xmath0-ray emission shows a similar morphology to the x - ray emission but on a much larger scale @xcite . calculating the efficiency of the energetic pulsars powering the pwn that is necessary to account for the vhe @xmath0-ray luminosity , values between 0.02% ( crab nebula ) and 7.5% ( hessj1825137 ) of the spin - down luminosity are found . the broadband seds of the vhe @xmath0-ray pwn can typically be well described by leptonic models , although claims have been made for a hadronic component at the high - energy end of the spectrum @xcite . vela x is the first vhe @xmath0-ray source in which the peak in the inverse compton energy flux has been detected within the h.e.s.s . energy range @xcite . figure [ fig::pwn_spectra ] shows , that while the ranges of @xmath0-ray fluxes for the detected pwn is rather small , the differences in the x - ray energy fluxes span a large range . this might be alluded to largely different magnetic fields , to different angular scales on which the x - ray emission has been measured or simply to the fact , that different populations of electrons are responsible for the x - ray and the @xmath0-ray emission . the most prominent example of the new class of _ offset pwn _ is hessj1825137 . this object can serve as a template for a whole new class of @xmath0-ray pwn in which a ) the @xmath0-ray emission is shifted away from the pulsar , possibly due to dense material on one side that prevents an isotropic expansion of the pwn and b ) the size of the vhe @xmath0-ray pwn is on a much larger scale ( @xmath30 ) than the x - ray pwn ( @xmath31 ) @xcite . concerning the offset morphology , asymmetric reverse shock interactions were first proposed to explain the offset morphology of the vela x pwn based on hydro - dynamical simulations by @xcite . the different sizes for the @xmath0-ray and x - ray pwne can be explained by the difference in the synchrotron cooling lifetimes of the ( higher energy ) x - ray emitting and the ( lower energy ) ic-@xmath0-ray emitting electrons . the @xmath0-ray sources that can be explained in this framework are typically extended , their emission region overlaps with energetic pulsars ( energetic enough to explain the @xmath0-ray flux by their spindown power ) and very importantly also show evidence for an x - ray pwn . so far only vela x and hessj1825137 match this picture , several other unidentified vhe @xmath0-ray sources have been proposed to be offset pwn , but all these cases lack the detection of an x - ray pwn . originally supernova remnants have been detected by means of radio observation sensitive to synchrotron emission in magnetic fields . radio observations are particularly suited to detect supernova remnants in the inner galaxy since they are insensitive to the prevailing dust emission within the galactic plane . the same holds for hard x - ray emission and is particularly true for @xmath0-rays . a survey of the galaxy in the @xmath0-ray regime proved to be a good means to detect new @xmath0-ray snrs , that are inconspicuous in other wavebands . h.e.s.s . observations of the inner part of the galactic plane revealed @xmath32 new @xmath0-ray sources @xcite . while some could be identified at other wavebands , such as the supernova remnants described in the previous sections or the microquasar ls5039 @xcite , most objects were left unidentified following their detection . a programme of detailed mwl studies using existing radio and x - ray facilities is underway to establish positional counterparts of the @xmath0-ray sources . a particularly interesting object found in the survey of the galactic plane is hessj1813178 . at first flagged unidentified @xcite , it was quickly found to be positionally coincident with : a ) a previously unpublished archival faint radio ( vla ) source ( 20 cm ) showing a shell - like morphology @xcite b ) a previously unpublished archival bright x - ray asca source ( 210 kev ) @xcite , and c ) a hard x - ray integral source 10100 kev @xcite . , asca as well as integral lacked the spatial resolution to resolve the object . the vla radio source showed a shell - like morphology , suggesting that the vhe @xmath0-ray emission originates in the shell of a supernova remnant . however , in a 30 ksec xmm - newton x - ray observation non - thermal synchrotron emission was found not from a shell , but rather from an object embedded within the shell with a faint tail towards the north - east ( see figure [ fig::new_1813 ] ) resembling in its shape a pwn @xcite . this detection reveals that hessj1813178 is connected to a composite snr similar to e.g. the @xmath0-ray source g0.9 + 0.1 @xcite . in hessj1813178 for the first time , an snr initially detected in vhe @xmath0-rays and subsequently confirmed with superior angular resolution radio and x - ray data . this detection shows , that @xmath0-ray observations especially in the galactic plane are well suited to detect snr , that are otherwise hard to detect due to obscuration and dust absorption . other objects tentatively connected to shell - type or composite snrs are hessj1640465 @xcite , and hessj1834087 @xcite . the upcoming glast satellite , in its energy range between 30 mev and 300 gev the successor to egret will provide a unique tool for studying supernova remnants . especially in crowded regions in the galactic plane dominated by diffuse emission , the improved angular resolution of the instrument in comparison to egret will be important for disentangling source confusion and identification of counterparts . as obvious from figure [ fig::shell_spectra ] , glast will provide spectral measurements in an energy regime in which differences between hadronic and leptonic production @xmath0-ray production mechanisms are significant . thus glast might finally disentangle whether the @xmath0-ray emission detected by h.e.s.s . is in fact the first direct evidence of accelerated hadrons in shell - type snrs , a question that directly relates to the origin of cosmic rays . glast will provide a highly even sky - coverage on long timescales and therefore population studies of snrs will be possible . since egret is thought to have just not been sensitive enough to single out individual pwne and enable us to conclude about their population , it is expected that these objects will appear as a source classin the glast sky . population studies of snrs provides information not only on individual objects but on spatial - statistical aspects that can be used to understand the transition of source populations through the regime of gev - cutoffs as already evident in numerous egret sources . already now snrs are an established source class in vhe @xmath0-ray astronomy . observations in this wavebands provide an important tool to understand the acceleration processes within these galactic accelerators . the author would like to acknowledge the support of their host institutions , and additionally support from the german ministry for education and research ( bmbf ) and from the department of energy ( doe ) . the author would like to thank the whole h.e.s.s . collaboration for their support , cospar for the financial support , the unknown referee(s ) for their valuable comments , and finally the organisers of the session e1.4 j.vink and p. o. slane for the invitation to the conference . aharonian , f. , atoyan , a. , & kifune , t. , `` inverse compton gamma radiation of faint synchrotron x - ray nebulae around pulsars '' , 1997 , mnras 291 , 162 aharonian , f. , et al . , `` evidence for tev gamma ray emission from cassiopeia a '' , 2001 , a&a 370 , 112 aharonian , f. , et al . , `` high - energy particle acceleration in the shell of a supernova remnant '' , 2004 , nature , 432 , 75 aharonian , f. , et al . , `` upper limits to the sn1006 multi - tev gamma - ray flux from hess observations '' , 2005a , a&a 437 , 135 aharonian , f. , et al . , `` detection of tev @xmath0-ray emission from the shell - type supernova remnant rx j0852.0 - 4622 with hess '' , 2005b , a&a 437 , 7 aharonian , f. , et al . , `` discovery of extended vhe gamma - ray emission from the asymmetric pulsar wind nebula in msh 15 - 52 with hess '' , 2005c , a&a 435 , l17 aharonian , f. , et al . , `` a possible association of the new vhe @xmath0-ray source hessj1825137 with the pulsar wind nebula g18.00.7 '' , 2005d , a&a 442 , l25 aharonian , f. , et al . , `` very high energy gamma rays from the composite snr g0.9 + 0.1 '' , 2005e , a&a 432 , 25 aharonian , f. , et al . , `` a new population of very high energy gamma - ray sources in the milky way '' , 2005f , science 307 , 1938 aharonian , f. , et al . , `` discovery of vhe gamma - ray emission from the x - ray binary ls 5039 '' , 2005 g , science 309 , 746 aharonian , f. , et al . , `` a detailed spectral and morphological study of the gamma - ray supernova remnant rxj1713.73946 with hess '' , 2006a , a&a 449 , 223 aharonian , f. , et al . , `` primary particle acceleration above 100 tev in the shell - type supernova remnant rx j1713.7 - 3946 with deep h.e.s.s . observations '' , 2006b , a&a in press ( astro - ph/0611813 ) aharonian , f. a. , et al . `` discovery of the two wings of the kookaburra complex in vhe gamma - rays with h.e.s.s '' , 2006c , a&a 456 , 245 aharonian , f. , et al . , `` h.e.s.s . observations of the supernova remnant rxj0852.04622 '' , 2006d , apj in press ( astro - ph/0612495 ) aharonian , f. , et al . , `` first detection of a vhe gamma - ray spectral maximum from a cosmic source : hess discovery of the vela x nebula '' , 2006e , a&a 448 , 43 aharonian , f. , et al . , `` discovery of the two `` wings '' of the kookaburra complex in vhe @xmath0-rays with hess '' , 2006f , a&a 456 , 245 aharonian , f. , et al.,``energy dependent @xmath0-ray morphology in the pulsar wind nebula hessj1825137 '' , 2006 g , a&a 460 , 365 aharonian , f. , et al.,``observations of the crab nebula with hess '' , 2006h , a&a 457 , 899 aharonian , f. , et al . , `` the h.e.s.s . survey of the inner galaxy in very high - energy gamma - rays '' , 2006i , apj 636 , 777 aharonian , f. , et al.,``3.9 day orbital modulation in the tev gamma - ray flux and spectrum from the x - ray binary ls 5039 '' , 2006j , a&a in press ( astro - ph/0607192 ) allen , g. e. , petre , r. , & gotthelf , e. v. `` x - ray synchrotron emission from 10 - 100 tev cosmic - ray electrons in the supernova remnant sn 1006 '' , 2001 , apj 558 , 739 aschenbach , b. , `` discovery of a young nearby supernova remnant . '' , 1998 , nature 396 , 141 bamba , a. , yamazaki , r. , ueno , m. , & koyama , k. , `` small - scale structure of the sn 1006 shock with chandra observations '' , 2003 , apj 589 ( 2 ) , 827 bell , a. r. , `` the acceleration of cosmic rays in shock fronts . ii '' , 1978 , mnras 182 , 443 blandford , r. d. , & ostriker , j. p. , `` particle acceleration by astrophysical shocks '' , 1978 , apj 221 , l29 blandford , r. , & eichler , d. , `` particle acceleration at astrophysical shocks - a theory of cosmic - ray origin '' , 1987 , phr , 154 , 1 blondin , j.m . , chevalier , r.a . , & frierson , d.m . , `` pulsar wind nebulae in evolved supernova remnants '' , 2001 , apj 563 , 806 brogan , c. l. , et al . , `` discovery of a radio supernova remnant and nonthermal x - rays coincident with the tev source hessj1813178 '' , 2005 , apj 629 , l105 buckley , j. h. , et al . , `` constraints on cosmic - ray origin from tev gamma - ray observations of supernova remnants '' , 1998 , a&a 329 , 639 cassam - chena , g. , decourchelle , a. , ballet , j. , sauvageot , j. l. , dubner , g. , & giacani , e. , `` xmm - newton observations of the supernova remnant rxj1713.73946 and its central source '' , 2004 , a&a 427 , 199 clark , g. w. , `` balloon observation of the x - ray spectrum of the crab nebula above 15 kev '' , 1965 , phys.rev.lett . , 14 , 91 drury , l. o. , `` an introduction to the theory of diffusive shock acceleration of energetic particles in tenuous plasmas '' , 1983 , rep.prog.phys . 46 , 973 drury , l. oc . , aharonian , f. , voelk , h. j. , the gamma - ray visibility of supernova remnants . a test of cosmic ray origin , 1994 , a&a 287 , 959 enomoto , r. , et al . , `` the acceleration of cosmic - ray protons in the supernova remnant rx j1713.7 - 3946 '' , 2002 , nature 416 , 823 esposito , j. a. , hunter , s. d. , kanbach , g. , & sreekumar , p. , `` egret observations of radio - bright supernova remnants '' , 1996 , apj 461 , 820 forot , m. , hermsen , w. , renaud , m. , laurent , p. , grenier , i. , goret , p. , khelifi , b. , & kuiper , l. , `` high - energy particles in the wind nebula of pulsar b1509 - 58 as seen by integral '' , 2006 , apj 651 , l45 funk , s. , `` status of identification of vhe @xmath0-ray sources '' , 2006 , apss , in press ( astro - ph/0609586 ) funk , s. , et al . , `` xmm - newton observations of hess j1813 - 178 reveal a composite supernova remnant '' , 2006 , submitted to a&a ( astro - ph/0611646 ) funk , s. , et al . , `` xmm - newton observations reveal the x - ray counterpart of the very - high - energy gamma - ray source hessj1640 - 465 '' , 2007 , submitted to apj ( astro - ph/0701166 ) gallant , y. , `` associations of very high energy gamma - ray sources discovered by h.e.s.s . with pulsar wind nebulae '' , 2006 , apss , in press ( astro - ph/0611720 ) gardner , f. f. , & milne , d. k. , `` the supernova of a.d . 1006 '' , aj 70 , 754 green , d. a. , `` galactic supernova remnants : an updated catalogue and some statistics '' , 2004 , basi , 32 , 335 gaensler , b. m. , schulz , n. s. , kaspi , v. m. pivovaroff , m. j. , & becker , w. e. , `` xmm - newton observations of psr b1823 - 13 : an asymmetric synchrotron nebula around a vela - like pulsar '' , 2003 , apj 588 , 441 gaensler , b. m. , & slane , p. o. , `` the evolution and structure of pulsar wind nebulae '' , 2006 , ara&a 44 ( 1 ) , 17 ginzburg , v. l. , & syrovatskii , s. i. 1964 , `` the origin of cosmic rays '' , macmillan hillas , m. a. , `` can diffusive shock acceleration in supernova remnants account for high - energy galactic cosmic rays ? '' , 2005 , j.phys.g : nucl . 31 , r95 horns , d. , aharonian , f. , santangelo , a. , hoffmann , a. i. d. , & masterson , c. , `` nucleonic gamma - ray production in velax '' , 2006 , a&a 451 51 iyudin , a. f. , et al . , `` comptel observations of ti-44 gamma - ray line emission from cas a '' , 1994 , a&a 284 , l1 iyudin , a. f. , et al . , `` emission from @xmath10ti associated with a previously unknown galactic supernova '' , 1998 , nature , 296 , 142 iyudin , a. f. , aschenbach , w. becker , k. dennerl and f. haberl , `` xmm - newton observations of the supernova remnant rxj0852.04622/groj08524642 '' , 2005 , a&a 429 , 225 - 234 jones , f. c. , & ellison , d. c. , `` the plasma physics of shock acceleration '' , 1991 , ssrv 58 , 259 kalemci , e. , et al . , `` x - ray observations of sn 1006 with _ integral _ '' , apj , 644 , 274 kennel , c. f. , coroniti , f. v. , `` magnetohydrodynamic model of crab nebula radiation '' , 1984 , apj 283 , 710 koyama , k. , petre , r. , gotthelf , e. v. , hwang , u. , matsuura , m. , ozaki , m. , & holt , s. s. , `` evidence for shock acceleration of high - energy electrons in the supernova remnant sn1006 '' , 1995 , nature , 378 , 255 koyama , k. , et al . , `` discovery of non - thermal x - rays from the northwest shell of the new snr rxj1713.73946 : the second sn 1006 ? '' , 1997 , pasj 49 , l7 lazendic , j. s. , slane , p. o. , gaensler , b. m. , reynolds , s. p. , plucinsky , p. p. , & hughes , j. p. , `` a high - resolution study of nonthermal radio and x - ray emission from supernova remnant g347.30.5 '' , 2004 , apj 602 , 271 lemoine - gourmard , m. , et al . , `` morphological and spectral study of rxj1713.73946 & rxj0852.04622 '' , 2006 , apss , in press malkov , m. a. , oc drury , l. , `` nonlinear theory of diffusive acceleration of particles by shock waves '' , 2001 , rep.prog.phys . , 64 , 429 mangano , v. , massaro , e. , bocchino , f. , mineo , t , & cusumano , g. , `` the extended hard x - ray emission from the vela plerion '' , 2005 , a&a 436 , 917 marsden , d. , et al . , `` the x - ray spectrum of the plerionic system psrb150958/msh 1552 '' , apj 491 , l39 mineo , t. , cusumano , g. , maccarone , m. c. , massaglia , s. , massaro , e. , & trussoni , e. , `` the hard x - ray emission from the complex snr msh 15 - 52 observed by bepposax '' , 2001 , a&a 380 , 695 porter , t. a. , moskalenko , i. v. , & strong , a. w. , `` inverse compton emission from galactic supernova remnants : effect of the interstellar radiation field '' , 2006 , apj 648 ( 2 ) , l29 reimer , o. , & funk , s. , `` demystifying an unidentified egret source by vhe @xmath0-ray observations '' , 2006 , in press , ap&ss ( astro - ph/0611653 ) roberts , m. s. e. , romani , r. w. , kawai , n. , `` the asca catalog of potential x - ray counterparts of gev sources '' , 2001 , apjs 133 , 451 roberts , m. s. e. , brogan , c. l. , gaensler , b. m. , hessels , j. w. t. , ng , c .- y . ; romani , r. w. , `` pulsar wind nebulae in egret error boxes '' , 2005 , ap&ss 297 , 93 roger , r. s. , milne , d. k. , kesteven , m. j. , wellington , k. j. , & haynes , r. f. , `` symmetry of the radio emission from two high - latitude supernova remnants , g296.5 + 10.0 and g327.6 + 14.6 ( sn 1006 ) '' , 1988 , apj 332 , 940 romero , g. e. , benaglia , p. , & torres , d. f. , `` unidentified 3eg gamma - ray sources at low galactic latitudes '' , 1999 , a&a , 348 , 868 slane , p. , et al . , `` rxj0852.04622 : another nonthermal shell - type supernova remnant ( g266.21.2 ) '' , 2001 , apj 548 , 814 sturner , s. j. , & dermer , c. d. , `` association of unidentified , low altitude egret sources with supernova remnants '' , 1995 , a&a 293 ( 1 ) , l17 tamura , k. , kawai , n. , yoshida , a. , & brinkmann , w. , `` evidence for a pulsar jet producing a hot nebula in the supernova remnant msh 1552 '' , 1996 , pasj 48 , l33 tanimori , t. , et al . , `` discovery of tev gamma rays from sn 1006 : further evidence for the supernova remnant origin of cosmic rays '' , 1998 , apj , 497 , l25 tanimori , t. , et al . , `` recent status of the analyses for stereoscopic observations with the cangaroo - iii telescopes '' , 2005 , in proc . of the 29th icrc , vol . 4 , ed . b. s. acharya et al . , 215 torres , d. f. , romero , g. e. , dame , t. m. , combi , j. a. , & butt , y. m. , `` supernova remnants and @xmath0-ray sources '' , 2003 , phr , 382 , 303 ubertini , p. et al . , `` integral igr j18135 - 1751 = hessj1813 - 178 : a new cosmic high - energy accelerator from kev to tev energies '' , 2005 , apj 629 , l109 uchiyama , y. , aharonian , f .. , takahashi , t. , `` fine - structure in the nonthermal x - ray emission of snr rxj1713.73946 revealed by chandra '' , 2003 , a&a 400 , 567 vink , j. , kaastra , j. s. , bleeker , j. a. m. , preite - martinez , a. , `` the bepposax x - ray spectrum of the remnant of sn 1006 '' , 2000 , a&a , 354 , 931 vink , j. , laming , j. m. , kaastra , j. s. , bleeker , j. a. m. , bloemen , h. , & oberlack , u. , `` detection of the 67.9 and 78.4 kev lines associated with the radioactive decay of @xmath10ti in cassiopeia a '' , 2001 , apj , 560 , l79 vink , j. , laming , j. m.,``on the magnetic fields and particle acceleration in cassiopeia a '' , 2003 , aj 584 , 758 weeks , t. c. , et al . , `` observation of tev gamma rays from the crab nebula using the atmospheric cerenkov imaging technique '' , 1989 , apj 342 , 379 weisskopf , m. c. , et al . , `` discovery of spatial and spectral structure in the x - ray emission from the crab nebula '' , 2000 , apj 536 , l81

the dynamics of evolution depend an organism s fitness landscape " , the mapping from genotypes to reproductive capacity . knowledge of the fitness landscape can help resolve questions such as how quickly a pathogen will acquire drug resistance , or by what pattern of mutations . but direct measurement of a fitness landscape is impossible because of the vast number of genotypes . here we critically examine regression techniques used to approximate fitness landscapes from data . we find that such regressions are subject to two inherent biases that distort the biological quantities of greatest interest , often making evolution appear less predictable than it actually is . we discuss methods that may mitigate these biases in some cases . an organism s fitness , or expected reproductive output , is determined by its genotype , environment , and possibly the frequencies of other genotypes in the population . in the simplified setting of a fixed environment , and disregarding frequency - dependent effects , which is typical in many experimental populations @xcite , fitnesses are described by a map from genotypes to reproductive rates , called the _ fitness landscape_. the dynamics of an adapting population fundamentally depend on characteristics of the organism s fitness landscape @xcite . however , mapping out an organism s fitness landscape is virtually impossible in practice because of the coarse resolution of fitness measurements , and because of _ epistasis _ : the fitness contribution of one locus may depend on the states of other loci . to account for all possible forms of epistasis , a fitness landscape must assign a potentially different fitness to each genotype , and the number of genotypes increases exponentially with the number of loci . as a result of these practical difficulties , fitness landscapes have been directly measured in only very limited cases , such as for individual proteins , rna molecules , or viruses . even in these limited cases genetic variation was restricted to a handful of genetic sites @xcite . alternatively , one might try to infer properties of a fitness landscape from a time - series of samples from a reproducing population . despite considerable effort along these lines @xcite , this approach is difficult and such inferences from times - series can be subject to systematic biases @xcite . as a result , very little is known about fitness landscapes in nature , despite their overwhelming importance in shaping the course of evolution . technological developments now allow researchers to assay growth rates of microbes or enzymatic activities of individual proteins and rnas for millions of variants @xcite . as a result , researchers are now beginning to sample and measure larger portions of the fitness landscapes than previously possible . nonetheless , even in these cases , the set of sampled genotypes still represents a tiny proportion of all genotypes , and likely also a tiny proportion of all viable genotypes . in order to draw conclusions from the limited number of genotypes whose fitnesses can be assayed , researchers fit statistical models , notably by penalized regression , that approximate the fitness landscape based on the data available . this situation is perhaps best illustrated by recent studies of fitness for the hiv-1 virus , based on the measured reproductive capacity of hiv - derived amplicons inserted into a resistance test vector @xcite . these hiv genotypes were sampled from infected patients . ( an alternative approach , often used for measuring activities of an individual enzyme , is to introduce mutations randomly into a wild - type sequence @xcite ) . whereas the entire fitness landscape of hiv-1 consists of reproductive values for roughly @xmath0 genotypes , only @xmath1 genotypes were assayed in the experiment @xcite . researchers therefore approximated the fitness landscape by penalized regression , based on the measured data , using an expansion in terms of main effects of loci and epistatic interactions between loci . the principal goal of estimating the underlying fitness landscape was to assess the extent and form of epistasis @xcite , and , more generally , to understand how adaptation would proceed on such a landscape @xcite . these @xcite and other high - throughput fitness measurement studies @xcite produce massive amounts of data , but not nearly enough to determine an entire fitness landscape . this presents the field with several pressing questions : do statistical approximations based on available data faithfully reproduce the relevant aspects of the true fitness landscape and accurately predict the dynamics of adaptation ? or , do biases arising from statistical fits or measurement noise influence the conclusions we draw from such data ? here , we begin to address these fundamental questions about empirical fitness measurements and how they inform our understanding of the underlying fitness landscape and evolution on the landscape . we study the effects of approximating a fitness landscape from data in terms of main and epistatic effects of loci . we demonstrate that such approximations , which are required to draw any general conclusions from a limited sample of genotypes , are subject to two distinct sources of biases . although these biases are known features of linear regressions , they have important consequences for the biological quantities inferred from such fitness landscapes . these biases systematically alter the form of epistasis in the inferred fitness landscape compared to the true underlying landscape . in particular , the inferred fitness landscape will typically exhibit less local ruggedness than the true landscape , and it will suggest that evolutionary trajectories are less predictable than they actually are in the true landscape . most of our analysis is based on samples from mathematically constructed fitness landscapes . but we argue that the types of biases we identify apply generally , and in more biologically realistic situations . indeed , we show that the same types of biases occur in rna - folding landscapes as well as empirically measured regulatory sequence binding landscapes . although it may be impossible to completely remove these biases , we conclude by suggesting steps to mitigate the biases in some cases . a function that maps genotype to fitness may be written as an expansion in terms of main effects and interactions between loci @xcite : @xmath2 where @xmath3 represents a genetic variant at locus @xmath4 , @xmath5 is the fitness ( typically the logarithm of growth rate ) , and @xmath6 is the number of loci . the nucleotides atgc , or any number of categorical variables , may be encoded by dummy variables , represented by @xmath3 , which equal either @xmath7 or @xmath8 in this study ( see methods ) . the term @xmath9 represents a sum over all pairs of interactions between loci , and the elipses represent higher - order terms , such as three - way interactions . since the statistical model is linear in the coefficients @xmath10 and @xmath11 , etc , the best - fit coefficients can be inferred by linear regression . experimental data are now sufficiently extensive that both the additive and pairwise epistatic coefficients , @xmath10 and @xmath11 , can often be estimated , whereas three - way and higher - order interactions are typically omitted from the statistical model of the fitness landscape . we refer to a statistical model with only additive and pairwise interactions as a quadratic model . even in the quadratic case , the statistical model may involve more free coefficients than empirical observations , so that over - fitting could become a problem . techniques to accommodate this problem , and the biases they introduce , are discussed below . the first type of bias we study arises from the use of penalized regressions which are required when a large number of parameters must be inferred from a limited amount of data . under standard linear regression with limited data , overfitting can cause the magnitudes of inferred coefficients to be large , resulting in positive and negative effects that cancel out to fit the observed fitness measurements . the standard remedy for overfitting is a so - called penalized least - squares regression " , such as ridge or lasso regression @xcite , which constrains the complexity of the inferred model by limiting the magnitudes of the inferred coefficients . for example , in fitting a quadratic landscape to sampled hiv-1 fitness measurements , hinkley et al . employed a form of penalized linear regression in order to avoid overfitting their data @xcite . although often required when fitting complex fitness landscapes to data , the penalized least square regression has some drawbacks . in general , the mean square error of any regression can be decomposed into a bias and a variance . the gauss - markov theorem guarantees that the standard least - squares linear regression produces the lowest possible mean squared error ( mse ) that has no bias , whereas penalized least squares can reduce the mse further by adding bias in exchange for a reduction in variance @xcite . thus , in order to provide predictive power for the fitnesses of un - observed genotypes , these regressions necessarily produce biased fits . while the accuracy of predicting unobserved fitnesses may be improved by such a biased fit , other quantities of biological interest derived from these predictions , such as measures of epistasis , may be distorted by the bias . in order to quantify the biases introduced by penalized least square regression , we compared mathematically constructed fitness landscapes to the landscapes inferred from a quadratic model fit by ridge regression ( similar results hold for lasso regressions , see discussion ) . our analyses are based on two types of mathematical fitness landscapes . the widely used @xmath12-landscapes of kauffman et al @xcite comprise a family of landscapes that range from additive to highly epistatic , depending upon the parameter @xmath13 , which determines the number of ( typically sparse ) interactions between sites . we also study polynomial " landscapes , which consist of additive effects and all possible pairwise and three - way interactions . in these landscapes , the amount of epistasis can be tuned by controlling the proportion of fitness variation that arises from the additive contributions , pairwise interactions , and three - way interactions ( see methods ) . we constructed @xmath12 and polynomial landscapes with only additive and pairwise effects , we sampled genotypes and fitness from these landscapes , and we fit a quadratic model of the landscape based on the sampled data " using penalized least square regression . for both @xmath12 and polynomial landscapes , we found that the inferred landscape tends to overestimate the fitnesses of low - fitness genotypes and underestimate the fitnesses of high - fitness genotypes ( polynomial landscape fig . [ fig : regbias ] , @xmath12 landscape fig . [ fig : regbias_nk ] ) . thus , there is a fitness - dependent bias in the inferred landscape compared to the true underlying landscape . the extent of this bias depends on the amount of penalization used , which in turn depends on the amount of data sampled relative the number of free coefficients in the statistical model . when the number of independent samples equals the number of free coefficients these biases disappear ( fig . [ fig : regbias ] , red curve ) , but whenever data are in short supply these biases arise and they can be substantial in magnitude ( fig . [ fig : regbias ] ) . the only way to avoid this bias entirely is to obtain at least as many independent observations as model parameters , which is typically unfeasible for realistic protein lengths or genome sizes . furthermore , as we will discuss below , these inherent biases have important consequences for our understanding of epistasis in the fitness landscape and for our ability to predict the dynamics of adaptation . , towards the average value , so that the high fitnesses are underestimated and low fitnesses are overestimated . the plot shows the mean ( solid line ) and standard deviation ( shaded area ) , of the distribution of residuals at a given true fitness value @xmath5 , smoothed by a gaussian moving window ( see methods ) . the fewer the observations ( i.e. the smaller the number of genotypes sampled for fitting the statistical model ) , the stronger the effect of this bias , as seen by comparing fits to training datasets of different sizes : 250 sampled genotypes ( red ) , 200 sampled genotypes ( green ) , and 150 sampled genotypes ( blue ) . genotypes were sampled randomly from a quadratic polynomial fitness landscape , which lacks any three - way interactions ( parameters @xmath14 , @xmath15 , @xmath16 , @xmath17 , and @xmath18 sites ; see methods ) . the training data were fit to a quadratic model , which has @xmath19 parameters , so that the statistical model is well - specified . a test set of 5,000 random genotypes was used to compare the predicted ( @xmath20 ) and true ( @xmath5 ) fitnesses of genotypes . with sufficient sampled data , no penalization is required and the resulting statistical fit contains no bias ( red ) . [ fig : regbias ] ] even when there is sufficient data so that a penalized regression is unnecessary , there is another source of potential bias in the inferred fitness landscape due to variables that are omitted from the statistical model but present in the true landscape , e.g. higher - order interactions between loci @xcite . in this case , the estimated coefficients of the statistical model will be biased in proportion to the amount of correlation between the omitted variables and the included variables @xcite . uncorrelated omitted variables , by contrast , may be regarded as noise , as we discuss below . interactions of different orders , e.g. three - way and pairwise interactions , are generally correlated with each other , unless the genotypes are sampled randomly and encoded as @xmath21 forming an orthogonal basis @xcite . in this case , which rarely applies to samples drawn from an evolving population , the omitted interactions may be regarded as noise and the estimated coefficients are guaranteed to be unbiased . however , even in this case the inferred @xmath5 values may still be biased . fig . [ fig : modelbias ] illustrates the biases arising from model mis - specification . to produce this figure we fit quadratic models to fitness landscapes that contain higher - order interactions . in both the cubic polynomial ( fig . [ fig : modelbias ] ) and @xmath12 ( fig . [ fig : modelbias_nk ] ) landscapes , fitnesses that are very high or very low are likely to contain contributions from higher - order interactions with positive or negative effects , respectively . but these higher - order interactions are not estimated by the statistical model , and so the inferred model overestimates low fitnesses and underestimates high fitnesses . bias arising from model mis - specification is qualitatively similar to bias arising form penalized regression , discussed above . bias from a mis - specified model can be large , but it would be not be visible in a plot of residuals versus _ inferred _ fitnesses . the mis - specified model bias shown in figure [ fig : modelbias ] is a form of _ regression towards the mean _ , and it is present even in a simple univariate regression with a large amount of noise @xcite . the slope in figure [ fig : modelbias ] , which plots true fitness @xmath5 against the residual @xmath22 , arises because the quadratic statistical model can not estimate the higher - order ( cubic ) interactions , which effectively act as noise in the regression . in fact , the slope in the figure equals @xmath23 , where @xmath24 denotes the coefficient of determination of the original regression ( see material and methods for a derivation ) . whether regression towards the mean is viewed as bias depends on the interpretation of the statistical model . if one assumes that the model can not be improved by adding any more predictor variables , i.e. that the noise is caused by purely random factors , as opposed to unknown systematic factors , then the regression results are unbiased and the observed negative slope between @xmath22 and @xmath5 simply reflects the fact that the regression can not estimate the noise . however , in situations when there is a systematic signal that is missing from the statistical model , such as when fitting a quadratic model to a fitness landscape that contains higher - order interactions , then the regression is biased towards the mean in proportion to the amount of variance that is not explained by the model . this phenomenon is not caused by measurement noise but by the omission of relevant variables . in the experiments summarized in fig . [ fig : modelbias ] there is no measurement noise in the fitnesses , and so the negative slope shown in figure [ fig : modelbias ] reflects a true bias : the mis - specified model over - estimates low fitnesses and under - estimates high fitnesses . with carefully tuned parameters , other forms of mis - specified model bias are also possible , see fig . [ fig : modelbias2 ] . in any case , whatever form it takes , mis - specified model bias has consequences for how accurately the landscape inferred from an experiment will reflect the amount of epistasis in the true landscape or predict the dynamics of adaptation , as we will demonstrate below . parameters ) to 5,000 randomly sampled individuals from three different cubic - polynomial landscapes each with with @xmath18 sites and @xmath17 : @xmath25 , @xmath15 , @xmath16 ( red ) , @xmath26 , @xmath27 , @xmath28 ( green ) , and @xmath25 , @xmath29 , @xmath30 ( blue ) . the larger the value of @xmath31 , the greater the amount of model mis - specification and the stronger the bias . a test set of 5,000 random genotypes was used to compare the predicted ( @xmath20 ) and true ( @xmath5 ) fitnesses of individuals . [ fig : modelbias ] ] a third difficulty that arises when fitting a statistical model to measured fitnesses is the presence of correlations between observed states of loci in sampled genotypes . an adapting population does not explore sequence space randomly , but rather is guided by selection towards higher - fitness genotypes . sequences sampled from a population under selection will thus tend to have correlated loci , due either to shared ancestry or due to epistasis . correlated variables do not themselves bias inferred coefficients ( at least , when the model is specified correctly ) , but they can inflate the variance of those estimates @xcite . predictions from the inferred model are not affected , in expectation , provided the new data have the same correlations as in the original training data . however , in the context of the expansion in eq . [ eq : quad ] , if there are correlations between the included variables , then there are also correlations between the omitted higher - order interactions and the included variables . thus , sampling from a wright - fisher ( wf ) population will exacerbate the mis - specified model bias . as a result , the inferred @xmath10 s and @xmath11 s will be further biased and so too will the inferred fitnesses . to illustrate biases that arise from sampling a population under selection , we simulated wright - fisher populations for 100 generations on cubic polynomial and @xmath12 landscapes . we used large population sizes and mutation rates to produce a large amount of standing genetic diversity for sampling genotypes ( see methods ) . the resulting fits exhibit very high @xmath24 values , in many cases even larger than fits to randomly sampled genotypes . but the large @xmath24 values and apparent lack of bias in the training data are very misleading when the model is misspecified , i.e. when the true landscape contains higher - order interactions . when predicting fitnesses of genotypes just one or two mutations away from the training data we find again large biases and large variance ( polynomial landscape fig . [ fig : wfsample2 ] and @xmath12 landscape fig . [ fig : wfsample_nk ] ) . as before , the resulting bias tends towards intermediate fitness values . ) were computed from cross - validated training data ( red ) , for genotypes sampled one mutation away from the training data ( purple ) , and for genotypes sampled two mutations away from the training data ( cyan ) . the true fitnesses ( @xmath5 ) are determined by a cubic polynomial fitness landscape on @xmath18 sites with @xmath26 , @xmath27 , @xmath28 , and @xmath32 . genotypes for fitting the quadratic statistical model were sampled from the population after 100 generations of wright - fisher evolution , with mutation rate @xmath33 and population size @xmath34 ( see methods ) . [ fig : wfsample2 ] ] one of the motivations for fitting a statistical model of a fitness landscape is to predict the fitnesses of genotypes that were not sampled or assayed in the original experiment . this immediately raises the question , how much predictive power do such statistical fits have , and how does their power depend upon the form of the underlying landscape from which genotypes have been sampled , as well as the form of the fitting procedure ? although extrapolation is easy to visualize in a linear regression with one component of @xmath35 , it can not be plotted as easily in high dimensions , where it is sometimes called _ hidden extrapolation _ @xcite . in the discrete , high - dimensional space of genotypes , no genotype is between any two other genotypes , so that every prediction is in some sense an extrapolation rather than an interpolation . given experimental data , it may be hard to determine if a model is extrapolating accurately or not @xcite . here , we quantify the accuracy of extrapolation explicitly using mathematical fitness landscapes . [ fig : wfsample3 ] illustrates the ability of statistical fits to predict fitnesses of genotypes that were not sampled in the training data , for a range of models and for regressions with varying degrees of mis - specification . this figure quantifies the amount of error when predicting the fitnesses of genotypes that are one or two mutations from the training data , as well as for predicting fitness of random , unsampled genotypes . away from the training data , the bias and variance increase with each mutation , as reflected by the lower squared correlation coefficients between true and inferred values . the predictions are progressively worse as the amount of model mis - specification increases . a statistical model that has a good fit to the training data , i.e. a high @xmath24 , does not necessarily imply that the model can make accurate predictions , especially if there is over - fitting . in fact , fig . [ fig : wfsample3 ] shows that even a high _ cross - validated _ @xmath24 can be misleading in the context of predicting unobserved fitnesses when the model is mis - specified . it is interesting to compare the extrapolative power of landscapes fitted to genotypes sampled from a wright - fisher population , versus genotypes sampled randomly . the dashed line in fig . [ fig : wfsample3 ] indicates the expected @xmath24 for regressions fitted to randomly sampled genotypes . on the one hand , predictions that are local , i.e. within a few mutations from the training data , typically have a higher @xmath24 for a model trained on wf - sampled genotypes compared to a model trained on random genotypes . on the other hand , predictions that are far from the training data ( i.e. predictions for random genotypes ) , typically have much lower @xmath24 for a model trained on wf - sampled genotypes compared to a model trained on random genotypes . thus , samples from a wright - fisher population produce a more biased model , even of the training data , but may nonetheless produce better predictions for local unsampled genotypes , compared to a model fitted to random genotypes @xcite . ) and true fitnesses ( @xmath5 ) , for sequences that are one mutation away from the training data ( purple ) , two mutations away from the training data ( cyan ) , and for random sequences ( black ) . genotypes used as training data were sampled from a wf population after 100 generations of evolution with mutation rates @xmath33 and population size @xmath34 ( see methods ) . landscapes were each instances of a cubic polynomial form , with @xmath31 values ranging from zero to one ( @xmath35-axis ) , with @xmath36 drawn uniformly in range @xmath37 , and with @xmath38 . the number of unique sequences sampled from each wf population varied from 34 to 603 ( not shown ) . [ fig : wfsample3 ] ] the dynamics of an adapting population depend fundamentally on the form of epistasis , that is , the way in which fitness contributions from one locus depend upon the status of other loci . indeed , one of the primary goals in fitting a fitness landscape to empirical data is to quantify the amount and form of its epistasis , in order to understand how adaptation will proceed . given the two sources of biases discussed above , which are inherent to fitting fitness landscapes to empirical data and exacerbated by sampling from populations under selection , the question arises : how do these inferential biases influence the apparent form of epistasis in the fitted landscape ? in this section we address this question by comparing the form of epistasis in the true , underlying fitness landscape to the form of epistasis in the inferred landscape obtained from fitting a quadratic model to sampled genotypes . there are several measures of epistasis known to influence the dynamics of adaptation . we will focus on three measures commonly used in the experimental literature on epistasis . one measure of epistasis , which reflects the degree of predictability in adaptation , is to reconstruct all the possible genetic paths between a low - fitness ancestral genotype and a high - fitness derived genotype sampled from an experimental population @xcite the proportion of such paths that are `` accessible '' , or monotonically increasing in fitness , is then a natural measure of epistasis . when this proportion is high , many possible routes of adaptation are allowable , suggesting that the evolutionary trajectory can not be easily predicted in advance . whereas when this proportion is small , it suggests that the evolutionary trajectory is more predictable , at least in a large population . to generate data similar to what would arise in an evolution experiment , we ran wright - fisher simulations on mathematical fitness landscapes ( see methods ) . each population began monomorphic and the most populated genotype at the end of the simulation was taken as the derived genotype , which typically contained between 5 to 7 mutations compared to the ancestral genotype . genotypes and their associated fitnesses were sampled from the population after 100 generations and used to fit a quadratic model of the landscape . figure [ fig : paths2]a shows an example of all the mutational paths between the ancestral and derived genotypes separated by 5 mutations , for both the true and the inferred fitnesses . since low fitnesses are likely to be overestimated , and high fitnesses are likely to be underestimated , the bias in the inferred landscape tends to eliminate fitness valleys . as a result , the number of accessible paths is higher in the inferred landscape than it is in the true underlying landscape ( fig . [ fig : paths2]a ) . epistasis appears to be less severe , and adaptation appears able to take more paths , than it actually is . this effect occurs systematically , as we have observed it over many realizations of different underlying fitness landscapes ( fig . [ fig : paths2]b ) . -axis ) and inferred ( @xmath5-axis ) landscapes . the inferred landscapes tend to over - estimate the proportion of accessible paths , compared to the true landscape : the proportion of accessible paths was overestimated 3.9-times more often than it was underestimated . in all cases the true landscape was cubic polynomial with @xmath25 , @xmath29 , @xmath30 , and @xmath39 . wf simulation parameters : @xmath33 , @xmath40 , with 500 generated landscapes and simulations . [ fig : paths2 ] ] we also investigated two other measures of epistasis : the number of local maxima in the fitness landscape @xcite , and the prevalence of sign epistasis between pairs of mutations @xcite . both of these quantities are global measures of epistasis , which depend upon the entire landscape , as opposed to the local measure of accessible paths between an ancestral and a derived genotype . generally speaking , local maxima tend to slow adaptation towards very high fitnesses , even though valley - crossing can occur in large populations @xcite . sign epistasis occurs when the fitness effect of a mutation at one site changes sign depending upon the status of a second site . reciprocal sign epistasis is a subset of sign epistasis , and it occurs when the second site also has sign epistasis on the first site . [ fig : max_signep ] compares the true and inferred amounts of these two global measures of epistasis . both quantities can be either underestimated or overestimated by the inferred landscape , depending on the circumstances . when the model is miss - specified ( fig . [ fig : max_signep ] bottom row ) , and without penalization or local sampling of genotypes , both measures of epistasis are heavily underestimated , since the model is unable to capture the local maxima and sign epistasis caused by three - way interactions . but when genotypes are sampled locally ( fig . [ fig : max_signep ] top row ) , i.e. sampled within a few mutations around a focal genotype , and penalized regression is applied , then the inferred landscape is influenced both by penalization bias and extrapolation error . the penalization bias tends to smooth the inferred landscape and eliminate local maxima ; wherease extrapolation adds noise to the estimated fitnesses and it may create spurious local maxima . which of these two effects dominates depends on the amount of data sampled and how it is distributed . sign epistasis appears to be less sensitive to extrapolation and bias when the model is well - specified , presumably because it depends only on the signs of effects and not magnitudes . in all cases considered , however , the inferred landscapes exhibit systematically biased global measures of epistasis . to complement the studies above , which are based on mathematically constructed fitness landscapes , we also investigated more realistic landscapes : computational rna - folding landscapes and empirical data relating regulatory sequences and expression levels ( i.e. a regulatory sequence binding landscape , see methods ) . the rna - folding landscape and the regulatory sequence binding landscape ( fig . [ fig : rnakinney ] ) both exhibit the same form of bias that we observed in the mathematical fitness landscapes . in the case of the rna - folding landscape ( see fig . [ fig : rnakinney]a ) there is no measurement error and sufficient data to avoid the need for penalized regression . thus , the bias towards the mean fitness seen in fig . [ fig : rnakinney]a with @xmath41 is due entirely to model misspecification : the quadratic model does not capture some higher - order interactions that influence rna folding . the regulatory sequence expression level data ( fig . [ fig : rnakinney]b ) , on the other hand , contain some measurement noise which comprises about 10 - 24% of the variance @xcite , whereas the @xmath24 for statistical model is @xmath42 . these numbers suggest that higher - order interactions bias the predictions made from the statistical model of the regulatory sequence binding landscape , as well , at least to some extent . the form of the biases observed when fitting a quadratic model to these realistic fitness landscapes ( fig . [ fig : rnakinney ] ) are similar to the biases observed when fitting quadratic models to @xmath12 landscapes or to polynomial landscapes . therefore we expect that these biases will have similar consequences for measures of epistasis . bias arising in a quadratic fit to ( a ) computational rna - folding landscape ( see methods ) , and ( b ) regulatory sequence binding landscape from kinney et al . @xcite ( see methods ) . the data are discrete in true fitnesses , @xmath5 . circles indicate means of distributions of residuals within each bin , and error bars indicate standard deviations . the quadratic fit exhibits the same type of bias towards the mean fitness as observed in @xmath12 and polynomial fitness landscapes . ] bias arising from a mis - specified model may be reduced by adding relevant missing variables . in the context of the expansion in eq . [ eq : quad ] this requires adding higher - order interactions such as triplets of sites , quartets , etc . in practice , this approach is often infeasible because the number of such interactions is extremely large and relatively few of them may be present in the data . in fig . [ fig : unbias]a we show that the bias in the statistical model can be reduced by fitting a model that includes three - way interactions . however , there is still some residual bias , due to penalized regression . thus , incorporating additional predictor variables effectively trades bias due to model mis - specification for bias due to a more severe penalized regression . it is not always necessary to use penalized regression , especially when higher - order interactions are sparse . in such cases it may be appropriate first to select a limited number of relevant variables , and then to use standard regression that avoids the bias associated with penalization . selecting the variables to use in the statistical model may be done `` by hand '' based on prior knowledge , or by statistical methods such as lasso @xcite ( see methods ) . this approach is expected to perform well only when a relatively small number of variables / interactions are in fact present in the true fitness landscape . as a proof of principle , we have shown how lasso followed by standard unpenalized regression reduces bias in fits to the cubic polynomial and @xmath12 landscapes . notably , lasso is a form of penalized regression that favors sparse solutions . we do not use lasso in this procedure to make predictions , but rather to select the variables to retain for the eventual unpenalized regression . the cubic polynomial landscape contains interactions between all triplets of sites , and so the variables retained by lasso are expected to omit some important variables , leaving some bias . by contrast , the @xmath12 landscape has sparse interactions and the resulting fit of this two - step procedure is therefore expected to be far less biased . both of these expectations are confirmed in fig . [ fig : unbias ] . although this approach may have the benefit of reducing bias in the inferred fitnesses , it will not improve the overall @xmath24 of the fit , or the extrapolative power . reducing bias when fitting a statistical model to ( a ) cubic polynomial landscape , and ( b ) @xmath12 landscape . bias in the inferred fitnesses can be reduced by adding third - order interactions to the statistical model ( quadratic fit in red , compared to cubic fit in green ) . although the green model is correctly specified , some bias still remains due to the penalized regression ( @xmath43 parameters fit with @xmath44 data points ) . further reduction of bias can be achieved by selecting model variables with lasso @xcite , and then performing an unpenalized regression only the selected variables ( blue ) . the variable selection step may omit some important variables , especially when the true landscapes includes a large number of higher - order interactions , as in the cubic polynomial case ( a ) . if the true landscape has a sparse set of interactions ( e.g. the @xmath12 landscape , panel b ) , then bias can be removed almost entirely by this two - step procudure . ( a ) : cubic polynomial landscape with @xmath25 , @xmath29 , @xmath30 . ( b ) : @xmath12 landscape with @xmath45 . a test set of 5,000 random genotypes was used to compare the predicted ( @xmath20 ) and true ( @xmath5 ) fitnesses . ] our ability to measure the genotype - fitness relationship directly in experimental populations is advancing at a dramatic pace . and yet we can never hope to measure but a tiny fraction of an entire fitness landscape , even for individual proteins . as a result , there is increasing need to fit statistical models of landscapes to sampled data . here we have shown that such statistical fits can warp our view of epistasis in the landscape and , in turn , our expectations for the dynamics of an evolving population . we have identified two distinct sources of biases : penalized regression and model mis - specification . our analysis of the effects of penalized regression have been performed using ridge regression , but the same qualitative results hold for lasso regressions ( figs . [ fig : regbias_poly_lasso ] , [ fig : regbias_nk_lasso ] , [ fig : wf_poly_lasso ] , [ fig : wf_nk_lasso ] , [ fig : paths_all_lasso ] analogous to figs . [ fig : regbias ] , [ fig : regbias_nk ] , [ fig : wfsample3 ] , [ fig : wfsample_nk ] , [ fig : paths2]b ) , and also , we expect , for other forms of penalization such as elastic net or the generalized kernel ridge regression in @xcite . notably , the bias arising from penalized regression has a slightly different form depending upon whether genotypes are encoded in the @xmath46 basis relative to the wild - type , in which case inferences will be biased towards the wild - type fitness , or the @xmath47 basis , used here , in which case inferences will be biased towards the average fitness . aside from these biases , we have also shown that statistical fits have poor predictive power . even when a fitted landscape exhibits a large cross - validated @xmath24 value , the fitted landscape generally has poor power to predict the fitnesses of unsampled genotypes , including genotypes within only a single mutation of the genotypes used to fit the landscape . common measures of epistasis may be grossly distorted by statistical fits to sampled data . interestingly , different measures of epistasis can be affected differently . for example , the number of paths accessible between an ancestal and derived genotype will be systematically over estimated in such fits , suggesting that evolution is less predictable than it actually is . the number of local maxima in the inferred landscape can be severly over- or under - estimated , whereas the prevalence of sign epistasis is less prone to bias ( fig . [ fig : max_signep ] ) . estimates of pairwise sign epistasis may be more robust because quadratic fits capture pairwise interactions , whereas the number of local maxima can depend on higher - order interactions . the problems of bias towards smoothness due to penalized regression and ruggedness due to extrapolation error can co - exist in the same dataset , as observed in a recent study of regulatory sequence binding data @xcite . which of these two effects will dominate is unclear , in general , as it depends upon the underlying landscape and the form of sampling . as a result , it is difficult to interpret the predictions made by quadratic fits to fitness landscapes , such as fits made to hiv data @xcite . nevertheless , it is often possible to at least deduce the presence of epistatic interactions from sampled data @xcite . in some cases , such as when the true landscape has sparse high - order interactions between loci , a combination of variable selection followed by unpenalized regression may ameliorate the biases we have identified . however , the degree to which this approach will reduce bias will surely depend upon the biological context . thus researchers should incorporate as much prior biological knowledge as possible when choosing a statistical model and fitting procedure . at the very least , it is important that researchers be aware of the biases inherent in fitting statistical of fitness landscapes to data . in order to build a statistical model with linear and interacting terms ( eq [ eq : quad ] ) , genetic sequences must be encoded as dummy variables , @xmath3 . if there is a well - defined wild - type sequence , then a natural parameterization is using zeros and ones , with the wild - type denoted as all zeros . the coefficients @xmath10 are the effect of single mutations , @xmath11 are the effects of pairs of mutations , and the constant term @xmath48 is the inferred wild - type fitness . in the case of a population with large diversity and no well - defined reference sequence , the reference - free parameterization with @xmath49 may be more appropriate , with @xmath48 then denoting the inferred average fitness @xcite . in this work we used the @xmath49 basis , because a wild - type was not defined , and the polynomial landscapes ( see below ) are defined in the @xmath49 basis . we constructed `` polynomial landscapes '' by terminating the expansion ( eq . [ eq : quad ] ) at the third - order and specifying the coefficients in such a way as to control the contributions to the total variation in log fitness that arise from interactions of each order . in particular , the variance of @xmath5 is @xmath50 where @xmath51 denotes an average over all genotypes in the orthogonal ( @xmath49 ) basis , and the sums are taken overal all sites , pairs of sites , and triplets . the coefficients @xmath10 , @xmath11 , and @xmath52 are chosen from normal distributions with mean zero and variances : @xmath53 @xmath54 @xmath55 where @xmath56 is the fraction of total variance determined by the @xmath4th order , @xmath6 is the number of loci , and the total variance is @xmath57 . for our numerical investigations using the cubic polynomial landscapes we chose @xmath18 . @xmath58 makes no contribution to the variance or the evolutionary dynamics , and was set to zero . we followed a standard @xcite procedure for constructing @xmath12 fitness landscapes . the parameter @xmath59 denotes the length of the binary string defining the genotype ( @xmath60 throughout our analyses ) the logarithm of the fitness of a genotype is calculated as the mean of contributions from each site , which are themselves determined by a table of values each drawn independently from a uniform probability distribution . when @xmath61 , the contribution of a site depends only on its own state : 0 or 1 , and not on the state of other sites . when @xmath62 , the contribution of a site depends on its own ( binary ) allele as well as the states of k other sites , yielding a lookup table with @xmath63 entries thus , there are in general @xmath59 lookup tables each with @xmath63 independently drawn entries , which together determine the contribution of each locus , based on the status of all other loci . under such @xmath12 models , the fitness effect of a substitution depends strongly and randomly on some fraction of the genetic background , determined by @xmath13 . @xmath13 is constant across sites and genotypes for a particular landscape , and the identifies of the @xmath13 sites upon which a given locus depends are drawn uniformly from the @xmath64 possibilities . notably , @xmath61 is an additive landscape , and @xmath65 is additive with sparse pairwise interactions . the amount of total variance in fitness due to the @xmath66th - order interactions is proportional to @xmath67 @xcite . the rna - folding landscape was generated by the vienna rna software @xcite . the target secondary structure was chosen as the most common structure observed in a sample of 10,000 structures generated from random genotypes of length 15 . the fitness function was defined as @xmath68 , where @xmath69 is hamming distance to the target in the tree - edit metric . training data consisted of @xmath70 random genotypes . the data consisted of 129,000 sequences of the _ e. coli _ lac promoter and associated gene expression levels . each sequence contained 75 nucleotides and contained roughly 7 mutations relative to the wild type . _ e. coli _ were facs sorted by expression levels into 9 bins , and the bin numbers serves as the phenotype for fitting the quadratic model . in this case , lasso was used for regression . for more details see @xcite . a quadratic fit with ridge regression was used ( unless otherwise stated ) , which identifies the coefficients that minimize @xmath71 where the last sum is taken over all coefficients denoted as @xmath72 . the first term is the mean squared error , and the last term is the penalization which biases coefficients towards zero . the free parameter @xmath73 was determined by choosing the largest @xmath73 within one standard deviation of the smallest ten - fold cross - validated mean squared error . an alternative form of penalization is lasso @xcite , which has a penalization term of the form @xmath74 . lasso favors sparse solutions of coefficients , and is useful for picking out important variables . ridge and lasso regression were done by matlab version r2013b . monte - carlo simulations of adaptation were based on standard wright - fisher dynamics @xcite . a population consists of @xmath59 individuals , each with a genotype consisting of a bit string of length 20 . the population replaces itself in discrete generations , such that each individual has a random number of offspring in proportion to its fitness , which is determined by its genotype via the fitness landscape . in practice , this is done efficiently with a multinomial random number generator . mutations are defined as bit flips , and they are introduced in every individual at each generation with probability @xmath75 per genome . the number of individuals receiving a mutation is thus binomially distributed , as double mutations are not allowed . the populations were initialized as monomorphic for a low - fitness genotype , chosen by generating 100 random genotypes and picking the one with the lowest fitness . simulations were run for 100 generations , and the resulting population was reduced to unique genotypes , and those genotypes , with the corresponding fitnesses , were used as the training data for regressions . the number of unique genotypes in the population is sensitive to @xmath75 , @xmath59 , the number of generations , and @xmath76 . these parameters were chosen such that there were at least a few hundred unique genotypes , representing substantial diversity in the population . the figures plotting true fitness versus residuals were produced using a gaussian moving window applied to the raw data . for each value of true fitness , @xmath5 , a mean and standard deviation was calculated by weighting all the data points by a gaussian with a width @xmath77 , and normalized by the sum of all the weights for the given @xmath5 value . this procedure provides a sense of the distribution at a given @xmath5 , without regard to the density of points . areas on the extremes of @xmath5 had few points to estimate a mean and variance , and they were excluded if the sum of weights was smaller than the 10% percentile of the distribution of all normalization factors . we used the smoothing parameter @xmath78 for figures [ fig : regbias ] , [ fig : modelbias ] , [ fig : unbias]a , [ fig : modelbias2 ] , and [ fig : regbias_poly_lasso ] ; @xmath79 for figures [ fig : unbias]b , [ fig : regbias_nk ] , [ fig : modelbias_nk ] , and [ fig : regbias_nk_lasso ] ; @xmath80 for figures [ fig : wfsample2 ] and [ fig : wf_poly_lasso ] ; and @xmath81 for figures [ fig : wfsample_nk ] and [ fig : wf_nk_lasso ] . a scatter plot of true fitnesses , @xmath5 , versus estimated fitnesses inferred by regression , @xmath20 , reflects the quality of the statistical fit . one can calculate the slope of @xmath5 versus @xmath20 by using a second regression : @xmath82 where @xmath83 denotes the slope , found by minimizing the mean squared error @xmath84 @xmath85 @xmath86 recall that @xmath87 , where @xmath88 is the residual from the initial regression , and @xmath89 because residuals are uncorrelated with @xmath20 . as a result , we conclude that @xmath90 , which simply reflects the properties of the original linear regression . this result is analogous to plotting residuals on the @xmath5 axis and inferred values on the @xmath35-axis , and observing no relationship . in the main text , by contrast , we show plots of the true values @xmath5 versus the residuals @xmath22 . in this case we observe a `` bias '' , in that genotypes with large @xmath5 are underestimated , and genotypes with small @xmath5 are overestimated . this type of bias is a form of regression towards the mean . we can calculate the slope of @xmath5 versus @xmath20 as follows : @xmath91 and with a similar calculation we find @xmath92 if we have mean - centered data ( @xmath93 ) , then this slope equals the coefficient of determination of the initial regression , denoted @xmath24 . equivalently , the slope in plots of @xmath5 versus @xmath22 equals @xmath94 . we thank a. feder , j. draghi , d. mccandlish for constructive input ; and p. chesson for clarifying the usage of the word between . j.b.p . acknowledges funding from the burroughs wellcome fund , the david and lucile packard foundation , the u.s . department of the interior ( d12ap00025 ) , the u.s . army research office ( w911nf-12 - 1 - 0552 ) , and the foundational questions in evolutionary biology fund ( rfp-12 - 16 ) . lenski re , travisano m ( 1994 ) dynamics of adaptation and diversification : a 10,000-generation experiment with bacterial populations . _ proceedings of the national academy of sciences of the united states of america _ 91:68086814 . blount zd , borland cz , lenski re ( 2008 ) historical contingency and the evolution of a key innovation in an experimental population of escherichia coli . _ proceedings of the national academy of sciences of the united states of america _ 105:7899906 . lee yh , dsouza l , fox ge ( 1993 ) experimental investigation of an rna sequence space . _ origins of life and evolution of the biosphere : the journal of the international society for the study of the origin of life _ 23:36572 . schlosser k , li y ( 2005 ) diverse evolutionary trajectories characterize a community of rna - cleaving deoxyribozymes : a case study into the population dynamics of in vitro selection . _ journal of molecular evolution _ 61:192206 . jimnez ji , xulvi - brunet r , campbell gw , turk - macleod r , chen ia ( 2013 ) comprehensive experimental fitness landscape and evolutionary network for small rna . _ proceedings of the national academy of sciences of the united states of america _ 110:149849 . kinney jb , murugan a , callan cg , cox ec ( 2010 ) using deep sequencing to characterize the biophysical mechanism of a transcriptional regulatory sequence . _ proceedings of the national academy of sciences of the united states of america _ 107:915863 . maharjan rp , ferenci t ( 2013 ) epistatic interactions determine the mutational pathways and coexistence of lineages in clonal escherichia coli populations . _ evolution ; international journal of organic evolution _ 67:27628 . covert aw , lenski re , wilke co , ofria c ( 2013 ) experiments on the role of deleterious mutations as stepping stones in adaptive evolution . _ proceedings of the national academy of sciences of the united states of america _ 110:e31718 . a quadratic model was fit to a cubic polynomial landscape with large negative additive effects , small mean - zero pairwise effects , and positive 3-way effects . the result is a complicated bias in the inferred landscape , with very low fitnesses being underestimated and very high fitnesses being overestimated . nevertheless , the bias from missing , higher - order interactions as in fig . [ fig : modelbias ] may be a generic feature . when the distribution of interaction effects is centered around zero , the model tends to be biased towards the average fitness . ] the inferred number of local maxima and the prevalence of sign epistasis may be under- or over - estimated by a statistical model of the fitness landscape fit to sampled data . in the top row the genotypes were sampled locally " , so that they are close together in hamming distance . the sampled genotypes consist of a focal genotype plus all sequences within one and two mutations , totaling 211 sampled sequences . the true landscape was a quadratic polynomial ( @xmath14 , @xmath15 , @xmath16 ) , and it was fit with a quadratic model by ridge regression ( 211 coefficients ) . this fit did not require a penalized regression , but we imposed some arbitrary values of the penalization parameter , @xmath73 , to demonstrate its effects . when @xmath98 , there is no bias ( not shown ) , but when the penalization is modest @xmath99 , purple dots ) , the number of local maxima is over - estimated . the spurious local maxima are likely due to large extrapolation errors , away from the sampled genotypes . when the penalization is yet larger ( @xmath100 , pink ) , the penalization bias smooths out the inferred landscape and thus under - estimates the number of local maxima . the fraction of inferred sign epistasis ( pluses , upper right cluster ) and reciprocal sign epistasis ( crosses , cluster in the lower left ) , is typically overestimated for locally sampled genotypes . in the bottom row , 5000 genotypes are sampled randomly so that no penalization is required . however , the true landscape in these cases is cubic polynomial ( @xmath25 , @xmath29 , @xmath30 ) , whereas the statistical model is quadratic , resulting in bias from model mis - specification . in these cases , the number of local maxima and the prevalence of sign epistasis are both under - estimated . in all cases , @xmath17 and dashed line indicates the line @xmath101 . ]

while the probability density function ( pdf ) of the sum of two independent random variables is easily described as the convolution of their pdfs , the expression for the pdf of the product is significantly more complicated . given two independent random variables @xmath0 and @xmath1 with pdfs @xmath2 and @xmath3 , where @xmath4 and @xmath5 , the pdf @xmath6 of their product , @xmath7 , can be succinctly written as @xmath8 where @xmath9 is the delta function defined as @xmath10 for a class of test functions @xmath11.when written in a more traditional form , @xmath12 the difficulty of computing @xmath6 becomes apparent since the integral in diverges at @xmath13 as long as e.g. @xmath14 and @xmath15 . indeed , in the simplest case when independent random variables are selected from two normal distributions with zero means , the pdf @xmath6 has a logarithmic singularity at @xmath13 ( see below ) . as far as we know , the only universal method currently available for computing the pdf of the product of two independent random variables relies on a monte - carlo approach , where one samples the individual pdfs , computes the products , and collects enough samples to achieve certain accuracy in the computation of @xmath6 in . however , due to the slow convergence of such method ( typically @xmath16 , where @xmath17 is the number of samples ) achieving high accuracy is not feasible . in this paper we show how to accurately and efficiently compute the pdf @xmath6 in using an approximate multiresolution analysis ( mra ) , where the scaling function is a gaussian . we demonstrate that , in contrast with the sampling method , our new algorithm is both fast and accurate . moreover , while the sampling method only provides a histogram of @xmath6 , we obtain the result in a functional form that can be used in further computations . a representation of a function in gaussian - based mra ( gmra ) can also be viewed as a multiresolution gaussian mixture . approximations by the gaussian mixtures have been used early on , see e.g. @xcite . we prove that the integral ( [ eq : integral - to - compute-1 ] ) can be efficiently evaluated as a multiresolution gaussian mixture and provide a tight estimate for the accuracy of the result . our approach to constructing a gmra is based on the observation that , for any finite user - selected accuracy @xmath18 , the function @xmath19 satisfies an approximate two scale relation for an appropriately selected parameter @xmath20 . as a consequence , the scaled and shifted functions @xmath21 form a gaussian - based approximate mra . the approximate nature of the gmra does not limit its applicability since any finite accuracy can be selected . for example , in our computations we roughly match the double precision arithmetic of a typical computer by selecting @xmath18 in the range @xmath22 . the integral ( [ eq : integral - to - compute-1 ] ) can , in principle , be computed using the mellin transform . it has been observed that , for non - negative random variables , the mellin transform of @xmath6 in ( [ eq : integral - to - compute-1 ] ) is equal to the product of the mellin transforms of @xmath2 and @xmath3 ( equivalently , @xmath6 is the so - called mellin convolution of @xmath2 and @xmath3 ) . in e.g. @xcite , this approach has been extended to random variables taking both positive and negative values . while analytically appealing , a numerical implementation of mellin convolution is problematic and has not resulted in a reliable numerical method . instead of using the mellin transform to represent the pdfs of non - negative random variables , we can use linear combinations of decaying exponentials with complex exponents constructed via algorithms developed in @xcite . such representation of pdfs of non - negative random variables leads to a universal method for computing their product . we plan to address such approach in a separate paper . however , for random variables that can take both positive and negative values , it is more natural to use multiresolution gaussian mixtures as described in this paper . moreover , a generalization to higher dimensions becomes problematic if one were to split the treatment of positive and negative values of such random variables . computing the pdf of the product of independent random variables using gmra is just the first of several interesting applications of this approximate basis . the idea of using gaussians to approximate functions appeared earlier in @xcite , where the coefficients of the shifted gaussians were associated with function values . it turns out that much more accurate approximations can be achieved if the coefficients are computed differently , as described later in this paper . since many classical operators of mathematical physics can be accurately approximated via a linear combination of gaussians ( see e.g. @xcite ) , using gmra to apply these operators can be reduced to the analytic evaluation of integrals , and hence , suggests a new class of fast multiresolution algorithms . however , constructing such algorithms is not straightforward since integer translates of gaussians are far from being orthogonal ( after all , they are positive functions so that no cancellations are possible ) . therefore , the use of gmra differs from the usual mra with orthogonal or nearly orthogonal scaling functions and we briefly discuss some of these issues in the appendix . we present our basis construction for functions of one variable and note that the generalization to two and three variables is fairly straightforward and we will describe it in appropriate contexts . extending our approach to bases for multivariate functions in higher dimensions is more complicated and we plan to address it later . finally , we note that the quotient of two independent random variables can be evaluated in the same fashion as the product so that gmra allows us to develop a numerical calculus of pdfs . we start by proving a key approximation result and considering its consequences in section [ sec : approximation - of - shifted ] . in section [ sec : distribution - of - the ] we prove a relative error estimate for the representation of the pdf of the product via gmra when one of the distributions is gaussian . we also present two approaches to approximating pdfs via a linear combination of gaussians . we then turn to numerical examples in section [ sec : numerical - examples ] . conclusions and further work is discussed in section [ sec : conclusions - and - further ] . finally , we present a brief discussion of certain features of gmra in the appendix . we start by proving a key estimate ( theorem [ thm : projection of a gaussian ] below ) that allows us to replace any shifted gaussian by an explicit linear combination of gaussians ( with a larger , fixed exponent ) shifted at integer values . this estimate yields an approximate two - scale relation for gaussians that makes possible to construct an approximate multiresolution basis with gaussians as the scaling functions . since the basis of shifted gaussians is far from being orthogonal , one would expect difficulties in computing projections of functions on such basis . remarkably , theorem [ thm : projection of a gaussian ] and theorem [ thm : the - gaussian - on arbitrary scale ] allow us to compute such projections accurately for any gaussian with an arbitrary exponent and shift . since we are interested in applications where all functions and the results of computing integrals can be approximated via linear combinations of gaussians , the following result is the key to using gmra as a numerical tool . [ thm : projection of a gaussian]if @xmath23 then , for @xmath24 , we have @xmath25 where @xmath26 and @xmath27 is the jacobi theta function . we have @xmath28 and need to estimate @xmath29 where @xmath30 using poissons summation formula , we have @xmath31 where @xmath32 we obtain @xmath33 yielding the estimate . see also ( * ? ? ? * section 2 , eq.5 ) for this type of analysis . [ cor : app_two_scale_eq]defining @xmath19 and assuming that @xmath34 , we obtain from @xmath35 where @xmath36 also , setting @xmath37 , @xmath38 and replacing @xmath39 by @xmath40 in , we obtain @xmath41 an approximate two - scale relation for gaussians . consequently , we can use the function @xmath42 as a scaling function for an approximate multiresolution basis . by selecting an appropriate @xmath43 we can reduce the value of @xmath18 in and achieve arbitrary finite precision in . in the examples presented later in the paper , we use @xmath44 so that @xmath45 ( see figures [ fig : example0101 - 2 ] and [ fig : example010101 - 2 ] ) . we note that selecting @xmath46 yields @xmath47 achieving double precision accuracy . if @xmath48 in , we recover the approximation described in @xcite . however , if @xmath49 , our approximation is substantially different and significantly more accurate than that obtained in @xcite . the key difference , of course , is in the approach to the selection of coefficients of the shifted gaussians . while in ( * ? ? ? * section 2 ) we use the step size of the discretization to control the accuracy of approximating integrals by a sum , here we fix a unit step size and select the exponent of the gaussian @xmath42 in order to achieve the desired accuracy . both , in theorem [ thm : projection of a gaussian ] and corollary [ cor : app_two_scale_eq ] , the error is described in terms of the quantity @xmath50 for some appropriate @xmath51 . in the next lemma we show that , as a function of @xmath52 , @xmath50 decreases exponentially fast . thus , slightly increasing the value of @xmath52 provides a much improved error . also , using the lemma , for a given @xmath43 and target accuracy @xmath53 we can accurately estimate the maximum value of @xmath54 to use in corollary [ cor : app_two_scale_eq ] . [ lem : estimate_theta_at_zero ] for @xmath55 , @xmath56 in particular , for @xmath57 , we have @xmath58 and @xmath59 where @xmath60{\pi}}{\gamma\left(\frac{3}{4}\right)}-1\right)<2.002.\label{constant_c_pi}\ ] ] consider @xmath61 since @xmath62 , @xmath2 is a decreasing function which implies the first claim . for , use @xcite to write @xmath63{\pi}}{\gamma\left(\frac{3}{4}\right)}-1\right)<2.002 $ ] . for , note that @xmath64 since for @xmath65 we have @xmath66 , we obtain the result using with @xmath52 replaced by @xmath67 . we consider the family of functions @xmath68 normalized so that the @xmath69-norm @xmath70 does not depend on the scale @xmath71 ( _ n.b_. , setting the norm to @xmath72 would have resulted in bulkier formulas ) . in the next theorem , we show that a gaussian with an arbitrary exponent can be projected ( approximated ) at an appropriate scale of gmra . it is remarkable that the projection coefficients are obtained explicitly without solving any equations . [ thm : the - gaussian - on arbitrary scale]the gaussian @xmath73 with exponent @xmath54 and shift @xmath74 @xmath75 where @xmath76 given @xmath54 , we select the scale @xmath71 so that @xmath77 . once @xmath71 is selected , we rescale the exponent @xmath54 and the shift @xmath74 as @xmath78 and @xmath79 and compute coefficients using for the rescaled gaussian @xmath80 , @xmath81 replacing the variable @xmath39 by @xmath82 in the expression above , we arrive at . given two independent gaussian variables @xmath83 , @xmath84 , with @xmath85 , the distribution of their product variable , @xmath86 is given by @xmath87 where @xmath88 is modified bessel function of the second kind . the pdf @xmath89 has an integrable logarithmic singularity at zero even though the normal distributions do not have any singularities . in general , the contribution of such singularity to the cumulative distribution function ( cdf ) may or may not be significant but it is always present as it follows from the representation of @xmath6 in . specifically , for @xmath90 in , as @xmath91 approaches @xmath92 we have @xmath93 where @xmath52 is the euler constant . for three or more gaussian pdfs with zero means , their product can be expressed via meijer g - functions . however , when the gaussian pdfs have non - zero means , we are not aware of any explicit expression for the pdf of their product in terms of special functions . in this section , we compute the pdf of the product of two independent random variables , @xmath0 with pdf @xmath94 , @xmath95 , and the gaussian variable @xmath84 with pdf @xmath96 the pdf of their product , @xmath7 , is expressed via the integral @xmath97 we assume that the pdf @xmath2 is a smooth function except at zero and a finite number of points where it may have integrable singularities . specifically , @xmath2 may have an integrable logarithmic singularity at @xmath98 , @xmath99 , for some integer @xmath100 . we also assume a mild decay of @xmath2 as @xmath101 , @xmath102 where @xmath103 and @xmath104 are positive constants . in particular , @xmath2 can be the cauchy distribution . in the next theorem we obtain a multiresolution approximation of @xmath105 , where the coefficients at all scales are given explicitly by integrals of well - behaved functions over a unit interval . these integrals are evaluated numerically via adaptive integration , see section [ sub : adaptive - method - for ] below . [ thm : the - pdf - of the product]the pdf of the product of two independent random variables @xmath0 and the gaussian variable @xmath84 in can be approximated as @xmath106 where @xmath18 is given in , @xmath107 @xmath108 and @xmath109 splitting the integration over @xmath39 in , we have @xmath110 and changing @xmath111 in the second integral , obtain @xmath112 since @xmath113 for @xmath114 , we have @xmath115 changing variables @xmath116 , we obtain @xmath117 where @xmath118 . splitting the integration in into integrals over unit intervals , we rewrite @xmath119 and changing the variable of integration , @xmath120 , we arrive at @xmath121d\tau.\label{eq : intermediate formula}\end{aligned}\ ] ] for @xmath122 $ ] we have @xmath123 , so that @xmath124 satisfies @xmath125 . consequently , using and , we obtain @xmath126 where @xmath127 and use these expressions to replace the corresponding gaussians in . setting @xmath128 as in , we now estimate @xmath129 . we have @xmath130d\tau.\nonumber \end{aligned}\ ] ] since @xmath2 is non - negative , we obtain @xmath131d\tau.\label{eq : intermediate formula-2}\end{aligned}\ ] ] using and , we have @xmath132 yielding the estimate in . we compute the integrals @xmath133 using adaptive integration . the integrand may have only integrable singularities in the interior of the interval @xmath134 $ ] due to our assumption on @xmath2 . since @xmath2 may have a logarithmic singularity at @xmath98 , i.e. @xmath135 near @xmath98 , we have @xmath136 so that , if @xmath71 is large , @xmath2 behaves like a polynomial as a function of @xmath137 . such behavior of the integrand explains our choice of a simple and efficient method for adaptive integration . specifically , we subdivide the integral @xmath134 $ ] adaptively in a hierarchical fashion while using a fixed number of legendre nodes to compute the integral on each subinterval . the subdivision terminates when the integral computed on two subintervals matches that computed on the parent interval within an appropriately chosen accuracy . in our examples we used the gauss - legendre quadrature with 10 nodes and set the accuracy parameter of the adaptive integrator to @xmath138 ; the local maximum depth of subdivision remained small in all cases . we consider two approaches to approximating pdfs via a linear combination of gaussians ( a gaussian mixture ) . the first approach is applicable to distributions that can be expressed via integrals involving gaussians . we illustrate this case using the laplace distribution . the second approach relies on a rapidly decaying fourier transform of a pdf so that it can be approximated by a bandlimited function . we illustrate this case using the gumbel distribution . we also refer to @xcite for a way to approximate a smooth pdf via a linear combination of gaussians provided that the accuracy requirements are not too stringent . although the pdf of the laplace distribution has two parameters , it is sufficient to approximate the function in for the particular case @xmath139 and @xmath140 , @xmath141 following ( * ? ? 37 ) , we have @xmath142 a linear combination of gaussians is then obtained by discretizing this integral . it is sufficient to use the trapezoidal rule due to the rapid decay of the integrand as @xmath143 . we limit integration to the interval @xmath144 $ ] and use @xmath145 as the step size . setting @xmath146 , and using @xmath147 terms , we arrive at the gaussian mixture approximation , @xmath148 we use this approximation in example [ sub : the - laplace - and - gumbel ] to compute the pdf of the product of two independent random variables , one with laplace distribution and the other with gumbel distribution . approximation of the latter via a gaussian mixture is described next . an interpolating gaussian - based scaling function spans the same subspace as the function @xmath42 , since it is constructed as @xmath149 using the interpolating property @xmath150 , @xmath151 , we also have @xmath152 multiplying this identity by @xmath153 , and adding over @xmath154 , we obtain @xmath155 by evaluating the fourier transform , @xmath156 and using , we arrive at @xmath157 by poisson summation formula , @xmath158 we finally obtain @xmath159 to illustrate the use of the interpolating scaling functions in the construction of gaussian mixtures , we consider the gumbel distribution in with parameters @xmath160 and @xmath140 , @xmath161 gaussian mixtures for general parameters are then obtained from this case by rescaling . first we select an interval to approximate @xmath3 . for example , outside the interval @xmath162 $ ] the function @xmath3 is less than @xmath163 . we treat @xmath3 on this interval as approximately periodic and , changing variables , @xmath164 , rescale it to the interval @xmath134 $ ] , @xmath165 sampling @xmath166 at @xmath17 points , where @xmath17 is sufficiently large , we write its approximation using the interpolating scaling function @xmath167 as @xmath168 using the scaling function @xmath169 , we seek the coefficients @xmath170 such that @xmath171 computing the fourier transform , we obtain @xmath172 and @xmath173 sampling at @xmath17 points and using , we obtain @xmath174 and compute the coefficients @xmath170 via the fast fourier transform ( fft ) . in this construction , @xmath17 should have small prime factors ( since we use fft ) but @xmath17 does not have to be a power of @xmath175 . we illustrate the result in figure [ fig : error - of - approximating gumbel ] , where we set @xmath176 . we use this gaussian mixture as an approximation of the gumbel distribution in example [ sub : the - laplace - and - gumbel ] . a more efficient representation ( with fewer terms ) can be obtained by a multiresolution version of the approach presented in this section . of approximating the rescaled gumbel pdf @xmath166 in as a gaussian mixture with @xmath176 terms . ] for computing the pdf of the product of independent random variables where the original pdfs are in gmra representation , we repeat the derivation in theorem [ thm : the - pdf - of the product ] using two representative gaussians from our approximate basis , namely @xmath177 we have @xmath178 and employing the same changes of variables as in the proof of theorem [ thm : the - pdf - of the product ] , arrive at @xmath179 changing variables @xmath180 , @xmath181 we define @xmath182 changing variable @xmath183 , @xmath184 , and splitting the real axis into unit subintervals ( see proof of theorem [ thm : the - pdf - of the product ] ) , we obtain @xmath185 changing variables one more time , @xmath120 , we arrive at @xmath186 finally , we use theorem [ thm : the - gaussian - on arbitrary scale ] to obtain @xmath187 where @xmath188 since the integrand is comprised of well - behaved functions , we can discretize the integral using the gauss - legendre quadrature with @xmath189 nodes directly ( rather than using an adaptive integrator ) . it follows that @xmath190 where @xmath191 , @xmath192 are the weights of the gauss - legendre quadrature , @xmath193 and @xmath194 we note that matrices @xmath195 and @xmath196 can be precomputed and that they are sparse ( banded ) since small entries can be removed . also , these matrices do not depend on the scales @xmath154 and @xmath197 of the original gaussians @xmath2 and @xmath3 in . [ ordering remark]the integral describing the pdf of the product does not depend on the order of the factors @xmath2 and @xmath3 . however , the computation of relies on the evaluation of the coefficients in and these coefficients do depend on the order of the factors . while the resulting pdf does not depend on the chosen order , one of the two possible representations may require significantly more coefficients . to explain the reason for such behavior , let us assume that both random variables in the product have normal distributions but different means and deviations , @xmath198 , @xmath199 and @xmath200 , @xmath201 . according to theorem [ thm : the - pdf - of the product ] , we need to evaluate @xmath202 as @xmath71 becomes large , we have @xmath203 thus , in order to reduce the number of scales , given the ratios , @xmath204 we pick @xmath2 in as the distribution with the largest ratio . the impact of ordering the factors in this way is illustrated in example [ sub : example6121 ] . this remark implies that the ordering of the gaussians @xmath2 and @xmath3 in - should be such that @xmath205 . we also note that in this particular case the ratios in do not depend on the scales @xmath154 and @xmath197 . [ moments remark]once a pdf @xmath2 is represented via gmra , @xmath206 where @xmath71 and @xmath207 only take a finite number of integer values , it is immediate to compute its moments since the moments of the basis functions @xmath208 , are readily available . to be precise , for any nonnegative integer @xmath154 , we compute @xmath209 where @xmath210 thus we have @xmath211 , @xmath212 , and higher moments of the basis functions @xmath208 can be evaluated using the recurrence , @xmath213 obtained by rewriting as @xmath214 and integrating by parts . for the moments of @xmath2 we arrive at the recurrence @xmath215 where @xmath216 and @xmath217 in order to compute the variance , we have @xmath218=\mathcal{m}^{\left(2\right)}-\left(\mathcal{m}^{\left(1\right)}\right)^{2},\ ] ] with similar formulas for skewness and kurtosis . unless noted otherwise , in all our numerical examples we set the parameter @xmath44 , so that @xmath219 in . for representing the pdf of the product of random variables , we use a gmra with a very large number of scales ( @xmath220 scales ) to demonstrate that we can accurately resolve the singularity at zero . obviously , there is no need to use so many scales in practical applications ; we use them to illustrate that the accuracy of our approach is well beyond the computational feasibility of any monte - carlo method . the test code was written in fortran 90 and run on a laptop with @xmath221 ghz intel processor . while no speed comparison with a monte - carlo method is meaningful , we note the actual time required for our algorithm . computing the coefficients of the product of two normal distributions and using @xmath220 scales takes @xmath222 seconds and evaluation of the resulting function at @xmath223 points takes @xmath224 second . while the computation of the coefficients @xmath128 in is trivially parallelizable , we did not take advantage of it in our code . in all our examples , we display significant coefficients of the gmra representation , @xmath128 in , as a matrix where the horizontal direction corresponds to the shift @xmath207 and the vertical direction to the scale @xmath71 . in interpreting these images ( see figures [ fig : example0101 - 1 ] , [ fig : example010101 - 2 ] , [ fig : example2111 - 2 ] , [ fig : example6121 - 3 ] , [ fig : cauchy-2 ] , [ fig : laplace2 ] , [ fig : gumbel2 ] and [ fig : the - laplace - and - the - gumbel - figure-2 ] below ) , it is important to remember that the shift @xmath207 of the basis functions @xmath225 indicates the distance to the singularity at @xmath13 , which is measured in scale dependent units , i.e. @xmath226 at the scale @xmath71 . just as an extra check , in the examples where explicit answers are not available , we verify the computed distributions by evaluating zero , first and second moments . in all cases these moments were within the accuracy of computation . the moments of pdfs in gmra representation were computed as described in remark [ moments remark ] . the pdfs of the product of two and three normal random variables with zero means are available analytically and we use them to ascertain accuracy of our algorithm . the exact pdf of the product of two random variables @xmath83 , @xmath84 , with @xmath85 , is given in . setting @xmath227 , we compute the pdf of the product @xmath7 via our algorithm and compare it with the exact pdf . the computed pdf @xmath228 and its gmra coefficients are displayed in figure [ fig : example0101 - 1 ] and the relative errors @xmath229 for different values of the parameter @xmath43 of the gmra , are displayed in figure [ fig : example0101 - 2 ] . the exact pdf in has a logarithmic singularity at the origin and we use @xmath220 scales to recover the pdf up to the interval @xmath230 $ ] with accuracy @xmath231 . the limit of resolution of this singularity is illustrated in figure [ fig : example0101 - 2 ] ( again , well beyond what may be needed in applications ) . pdf @xmath228 of the product in example [ sub : example0101 ] ( left ) and logarithm ( base 10 ) of gmra coefficients of @xmath228 ( right).,title="fig : " ] pdf @xmath228 of the product in example [ sub : example0101 ] ( left ) and logarithm ( base 10 ) of gmra coefficients of @xmath228 ( right).,title="fig : " ] the limit of resolution of singularity of the pdf @xmath228 in example [ sub : example0101 ] using gmra with @xmath220 scales and @xmath232 ( left ) and the relative error curves shown over the interval @xmath233 $ ] for different parameters @xmath43 in the gmra . the bottom curve corresponds to @xmath232 , followed by @xmath234 , @xmath235 and @xmath236 , as we go up in the display ( right).,title="fig : " ] the limit of resolution of singularity of the pdf @xmath228 in example [ sub : example0101 ] using gmra with @xmath220 scales and @xmath232 ( left ) and the relative error curves shown over the interval @xmath233 $ ] for different parameters @xmath43 in the gmra . the bottom curve corresponds to @xmath232 , followed by @xmath234 , @xmath235 and @xmath236 , as we go up in the display ( right).,title="fig : " ] we compute the pdf of the product of three zero - mean gaussian random variables , starting with the gmra representation of the product of two of them computed in the previous example . given three zero - mean gaussian random variables , @xmath237 , @xmath238 , and @xmath239 , the pdf of their product , @xmath240 , is available analytically as @xmath241 where @xmath242 is a special case of meijer g - function , see ( * ? ? ? * definition 9.301 ) . the computed product pdf @xmath243 is obtained by multiplying @xmath244 by the gmra representation of the pdf of @xmath245 , as computed in example [ sub : example0101 ] . the pdfs of the random variables @xmath246 , @xmath247 , and @xmath248 are illustrated in figure [ fig : example010101 - 1 ] . as in example [ sub : example0101 ] , we use @xmath220 scales and @xmath232 to recover the pdf @xmath243 with accuracy @xmath231 . in figure [ fig : example010101 - 2 ] we display the gmra coefficients of @xmath243 and its logarithmic singularity resolved to the interval @xmath249 $ ] . we compute the relative error , @xmath250 for different values of the parameter @xmath43 in the gmra and display them in figure [ fig : example010101 - 2 ] . pdfs of the random variables @xmath245 and @xmath247 in example [ sub : example010101 ] ( left ) and pdf of the product random variable @xmath251 in example [ sub : example010101 ] ( right).,title="fig : " ] pdfs of the random variables @xmath245 and @xmath247 in example [ sub : example010101 ] ( left ) and pdf of the product random variable @xmath251 in example [ sub : example010101 ] ( right).,title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath243 in example [ sub : example010101 ] ( left ) and the relative error curves shown over the interval @xmath233 $ ] for different parameters @xmath43 in the gmra . the bottom curve corresponds to @xmath232 , followed by @xmath234 , @xmath235 and @xmath236 , as we go up in the display , demonstrating resolution up to the interval @xmath249$].,title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath243 in example [ sub : example010101 ] ( left ) and the relative error curves shown over the interval @xmath233 $ ] for different parameters @xmath43 in the gmra . the bottom curve corresponds to @xmath232 , followed by @xmath234 , @xmath235 and @xmath236 , as we go up in the display , demonstrating resolution up to the interval @xmath249$].,title="fig : " ] in the next five examples we use our algorithm to compute the pdfs of the product of two independent normal random variables with non - zero means as well as the product of independent normal random variable and independent random variables with laplace , cauchy and gumbel distributions . we also compute the product of independent random variables with laplace and gumbel distributions where both distributions are approximated via gaussian mixtures as described in section [ sub : representing - pdfs - in - gmra ] . for the random variables @xmath252 , @xmath253 with @xmath254 , @xmath255 and @xmath256 , @xmath257 , we compute the pdf @xmath228 of their product , @xmath7 . figure [ fig : example2111 - 1 ] shows the pdfs of the random variables @xmath0 , @xmath1 and @xmath246 . as far as we know , there is no analytic expression for the result which we compute with accuracy @xmath258 . using @xmath220 scales , we resolve the logarithmic singularity at zero within an interval as small as @xmath259 $ ] ( see figure [ fig : example2111 - 2 ] ) . we observe that to compute the pdf of @xmath246 to such accuracy and resolution via a monte - carlo method is not feasible . moreover , we obtain the pdf of @xmath246 in a functional form which can be used for further computations . in contrast , a monte - carlo method using @xmath260 samples and @xmath261 bins of approximate size @xmath262 yields just the histogram shown in figure [ fig : example2111 - 3 ] . pdfs of the random variables @xmath0 and @xmath1 of example [ sub : example2111 ] ( left ) and the computed pdf @xmath228 of example [ sub : example2111 ] ( right).,title="fig : " ] pdfs of the random variables @xmath0 and @xmath1 of example [ sub : example2111 ] ( left ) and the computed pdf @xmath228 of example [ sub : example2111 ] ( right).,title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath228 in example [ sub : example2111 ] ( left ) and logarithmic singularity of @xmath228 resolved to the interval @xmath263$].,title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath228 in example [ sub : example2111 ] ( left ) and logarithmic singularity of @xmath228 resolved to the interval @xmath263$].,title="fig : " ] a histogram of the distribution of example [ sub : example2111 ] computed with @xmath260 samples using @xmath261 bins of approximate size @xmath262 . the logarithmic singularity at the origin is poorly resolved ( cf . the resolution of our method illustrated in figure [ fig : example2111 - 2 ] ) . ] we demonstrate the effect of the order of factors on the coefficients of the representation of the product as described in remark [ ordering remark ] . we compute the pdfs of @xmath264 and @xmath265 , where @xmath266 and @xmath267 , and display the results in figure [ fig : example6121 - 1 ] and [ fig : example6121 - 3 ] . while the values of the pdfs @xmath268 and @xmath269 are the same within the accuracy of computation , the coefficients of @xmath268 and @xmath269 do depend on the order of the factors as illustrated in figure [ fig : example6121 - 3 ] . pdfs of random variables @xmath0 and @xmath1 in example [ sub : example6121 ] ( left ) and computed product pdf @xmath268 ( right).,title="fig : " ] pdfs of random variables @xmath0 and @xmath1 in example [ sub : example6121 ] ( left ) and computed product pdf @xmath268 ( right).,title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath268 ( left ) and logarithm ( base 10 ) of gmra coefficients of the pdf @xmath269 ( right).,title="fig:"]logarithm ( base 10 ) of gmra coefficients of the pdf @xmath268 ( left ) and logarithm ( base 10 ) of gmra coefficients of the pdf @xmath269 ( right).,title="fig : " ] we now turn to computing the pdf of the product of random variables where one of them has a non - gaussian distribution . first , we consider the pdf of the cauchy distribution @xmath270 where @xmath271 is the location parameter and @xmath52 is the scale parameter . the cauchy distribution is defined on the real line and is an example of a heavy - tailed distribution for which the moments are infinite . in this example , we compute the pdf @xmath228 of @xmath7 , where @xmath272 and @xmath273 . we used @xmath274 `` negative scales '' , i.e. @xmath275 in order to capture the heavy - tail behavior of the product distribution ( which also has no finite moments ) . by using so many coarse scales , the integral of the resulting pdf over the real line differs from @xmath72 by approximately @xmath276 . pdfs of the factors @xmath0 and @xmath1 and the product @xmath246 are illustrated in figure [ fig : cauchy-1 ] . in figure [ fig : cauchy-2 ] , we display the gmra coefficients of @xmath228 and resolve the singularity at zero of @xmath228 within the interval @xmath249 $ ] . the pdfs of the random variables @xmath0 and @xmath1 in example [ sub : examplecauchy ] ( left ) and computed product pdf @xmath228 ( right).,title="fig : " ] the pdfs of the random variables @xmath0 and @xmath1 in example [ sub : examplecauchy ] ( left ) and computed product pdf @xmath228 ( right).,title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath228 in example [ sub : examplecauchy ] ( left ) and logarithmic singularity of @xmath228 resolved to the interval @xmath249 $ ] ( right ) . , title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath228 in example [ sub : examplecauchy ] ( left ) and logarithmic singularity of @xmath228 resolved to the interval @xmath249 $ ] ( right ) . , title="fig : " ] we consider the pdf of the laplace distribution @xmath277 where @xmath278 is the location parameter and @xmath279 is the scale parameter . the density function , @xmath280 , has a cusp at @xmath281 . in this example , we compute the pdf @xmath228 of the product @xmath7 , where @xmath282 and @xmath283 and display these pdfs in figure [ fig : laplace1 ] . in figure [ fig : laplace2 ] we display the gmra coefficients of @xmath228 and resolve , within the interval @xmath249 $ ] , the singularity of @xmath228 at zero . note that the resulting distribution does not have a singularity at @xmath281 but has a singularity at zero as in the other examples . the pdfs of the random variables @xmath0 and @xmath1 in example [ sub : examplelaplace ] ( left ) and computed product pdf @xmath228 ( right ) . , title="fig : " ] the pdfs of the random variables @xmath0 and @xmath1 in example [ sub : examplelaplace ] ( left ) and computed product pdf @xmath228 ( right ) . , title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath228 in example [ sub : examplelaplace ] and ( b ) logarithmic singularity of @xmath228 resolved to the interval @xmath249 $ ] . , title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath228 in example [ sub : examplelaplace ] and ( b ) logarithmic singularity of @xmath228 resolved to the interval @xmath249 $ ] . , title="fig : " ] we also consider the gumbel distribution which is a special case of the generalized extreme value distribution . its pdf is not symmetric and is given by @xmath284 where @xmath278 is the location parameter and @xmath285 is the scale parameter . here , we compute the pdf @xmath228 of the product random variable @xmath7 , where @xmath286 and @xmath287 and display these pdfs in figure [ fig : gumbel1 ] . in figure [ fig : gumbel2 ] we display the gmra coefficients of @xmath228 and resolve , within the interval @xmath249 $ ] , the singularity of @xmath228 at zero . pdfs of random variables @xmath0 and @xmath1 in example [ sub : examplegumbel ] ( left ) and computed product pdf @xmath228 ( right).,title="fig : " ] pdfs of random variables @xmath0 and @xmath1 in example [ sub : examplegumbel ] ( left ) and computed product pdf @xmath228 ( right).,title="fig : " ] logarithm ( base 10 ) of gmra coefficients of @xmath228 in example [ sub : examplegumbel ] ( left ) and logarithmic singularity of @xmath228 resolved to the interval @xmath249 $ ] ( right ) . , title="fig : " ] logarithm ( base 10 ) of gmra coefficients of @xmath228 in example [ sub : examplegumbel ] ( left ) and logarithmic singularity of @xmath228 resolved to the interval @xmath249 $ ] ( right ) . , title="fig : " ] in this example we compute the pdf @xmath228 of the product random variable @xmath7 , where @xmath288 is the laplace distribution and @xmath289 is the gumbel distribution . both distributions are approximated as gaussian mixtures as described in section [ sub : representing - pdfs - in - gmra ] . in computing @xmath228 we ordered the parameters of gaussians to reduce the total number of coefficients as explained in remark [ ordering remark ] . the laplace distribution was approximated by a gaussian mixture with @xmath290 terms and the gumbel distribution with 300 terms , see section [ sub : representing - pdfs - in - gmra ] for details . we display the results in figure [ fig : the - laplace - and - the - gumbel - figure-1 ] and figure [ fig : the - laplace - and - the - gumbel - figure-2 ] . the accuracy of the result is @xmath291 as indicated by computing moments of @xmath228 . the pdfs of the random variables @xmath0 and @xmath1 in example [ sub : the - laplace - and - gumbel](left ) and computed product pdf @xmath228 ( right).,title="fig:"]the pdfs of the random variables @xmath0 and @xmath1 in example [ sub : the - laplace - and - gumbel](left ) and computed product pdf @xmath228 ( right).,title="fig : " ] logarithm ( base 10 ) of gmra coefficients of the pdf @xmath228 in example [ sub : the - laplace - and - gumbel ] . ] gmra allows us to develop a numerical calculus of pdfs since we can now compute sums and products of independent random variables ( as well as quotients , _ _ since the integral for the quotient of two independent random variables can be evaluated in a manner similar to that of the product ) . indeed , if the initial pdfs are expressed in the gmra , the resulting pdf is also expressed in the gmra , i.e. as a multiresolution gaussian mixture . similar results hold for cdfs of random variables . we expect that generalizations of the gmra to two and three dimensions can be accomplished in a straightforward manner , perhaps supplemented by employing the singular value decomposition ( svd ) to simplify the representation of coefficients on each scale . we plan to address such generalizations within relevant applications . a practical generalization to dimensions higher than three is more challenging and we plan to work towards it . computing the pdf of the product of independent random variables does not require all of the features of mra since we only compute projections on the scaling functions at appropriate scales . in this appendix we briefly present some of the usual mra constructs . we assume that the reader is familiar with the basic concepts of orthogonal and bi - orthogonal wavelet bases , see e.g. @xcite . we note that the approximate gmra is in many respects similar to mra based on high order splines , see e.g. @xcite . however , there are several special properties of gmra that are not shared by spline based mra . for example , as shown in theorem [ thm : the - gaussian - on arbitrary scale ] , a gaussian with an arbitrary exponent and an arbitrary shift can be represented at an appropriate subspace which does not depend on the shift . another advantage of using gmra is that it becomes possible to apply operators to functions in a semi - analytic fashion which is important in a number of applications we plan to address separately . finally , we would like to emphasize that the fact that the gmra is approximate is not an obstacle in practice since we can adjust the accuracy of the gmra as needed . @xmath293 more precisely , using the poisson summation formula applied to @xmath294 , we have @xmath295 for @xmath296 , we obtain the error estimate @xmath297 where the last equality can be found in ( * ? ? ? 8.180.3 ) . since @xmath298 lemma [ lem : estimate_theta_at_zero ] shows that the error decays exponentially fast as a function of the parameter @xmath43 . for example , if @xmath46 , the estimate provides an error bound which is approximately @xmath299 . this leads us to define the filter @xmath300 for gmra as @xmath301 and observe that ( for any @xmath302 ) implies @xmath303 where @xmath304 , see . therefore , for the values of @xmath43 of interest , @xmath305 is extremely close to one . for example , for @xmath46 , yields @xmath306 next , we compute @xmath307 to define the fourier transform of the approximate , orthogonal gaussian - based scaling function as @xmath308 the corresponding approximate quadrature mirror filter ( qmf ) , determined by the identity @xmath309 , is then given by @xmath310 we note that @xmath311 so that , for @xmath46 , @xmath312 as well . the real valued filter @xmath313 satisfies an approximate qmf equation , @xmath314 for @xmath46 , we numerically verify that @xmath315 in order to obtain the filter that is dual to @xmath300 , we rewrite as @xmath316 so that the dual filter @xmath317 is given by @xmath318 we use @xmath317 to obtain a formula for the projection of coefficients onto a coarser scale . it is sufficient to consider a projection from zero scale into the next coarser scale . therefore , starting with @xmath2 of the form @xmath319 so that @xmath320 the coefficients of @xmath2 in the next coarser scale are computed via ( see e.g. ( * ? ? ? * section 8.3.1 ) ) , @xmath321e^{2\pi inp}dp.\nonumber \end{aligned}\ ] ] we would like to point out a numerical problem associated with using the filter @xmath317 . the plot of @xmath317 for @xmath322 in figure [ fig : filter for m00 ] shows that one can expect a loss of significant digits due to its large dynamic range . to avoid the possible loss of significant digits we have several options . first , as is done in this paper , we can avoid using @xmath317 altogether . the second option is to accept a lower ( e.g. single precision ) accuracy in computations with gmra . finally , one can look for an alternative ( nonlinear ) algorithm for computing coefficients @xmath324 in . to illustrate the second option , we choose @xmath236 and observe the relatively small dynamic range of @xmath317 for this value of @xmath43 in figure [ fig : filter for m00 ] . using @xmath236 in computing the pdfs results in roughly single precision accuracy as illustrated in figures [ fig : example0101 - 2 ] and[fig : example010101 - 2 ] . an alternative approach motivated by the construction of gmra is to develop an exact mra such that the resulting scaling function is closely approximated by a single gaussian . although we do not use this approach in this paper , we briefly indicate how to attain such construction . we can demonstrate that the nonnegative filter is an excellent approximation of the filter @xmath313 in . in fact , for @xmath322 , we verify numerically that @xmath326.\label{diff_m_a_and_m_0}\ ] ] in the next theorem we show that the filter @xmath327 has the necessary properties to generate an exact orthogonal mra . since @xmath332 satisfies the well - known property @xmath333 and @xmath334 , we obtain @xmath335 by construction , we have @xmath331 and , since @xmath336 is an even function , @xmath327 is even . using and observing that @xmath337 is strictly decreasing for @xmath51 , we have @xmath338 hence , denoting by @xmath339 the constant in , we know that @xmath340 . for @xmath341 , we denote by @xmath2 the function @xmath342 and notice that , due to , we have @xmath343 and @xmath344 . since @xmath345 and @xmath346 is strictly decreasing on @xmath347 , to show that @xmath2 is strictly decreasing in this interval , it is enough to show that @xmath3 has the same property . by , we have @xmath348 and , taking derivatives , @xmath349 is equivalent to the inequality @xmath350 from @xcite , we have , for any @xmath51 , @xmath351 and , hence , @xmath352 which implies and , thus , that @xmath2 is strictly decreasing on @xmath347 . therefore , @xmath2 is also positive in that interval . it follows that @xmath353 is strictly decreasing on @xmath347 . since @xmath327 has no zeros in @xmath354 $ ] , using corollary 6.3.2 in @xcite , we know that the qmf @xmath327 generates an orthogonal mra . the fourier transform of the resulting scaling function @xmath355 can now be obtained as @xmath356 due to , if we use a finite number of scales , the exact mra and approximate gmra generated by the function @xmath357 in are exchangeable for an appropriately selected parameter @xmath43 tuned to the user - selected accuracy . g. beylkin , g. fann , r. j. harrison , c. kurcz , and l. monzn . multiresolution representation of operators with boundary conditions on simple domains . , 33:109139 , 2012 .

efforts to gain a quantitative microscopic understanding of melting have spanned more than a century . the lindemann criterion developed in 1910 @xcite describes melting in terms of the root mean square ( rms ) deviation from the atomic equilibrium positions . since long - range positional order stems from the periodic arrangement of atoms in crystalline solid , atomic deviations that are comparable to the separation between atomic species could obscure the regularity of the underlying crystal lattice with a concomitant loss of positional order . the lindemann criterion specifies that melting has occurred if the rms deviations reach on the order of a tenth of a lattice constant , and has proved to be a reasonably effective theory for three dimensional systems . the lindemann analysis does not take into account correlations of the motions of neighboring atoms . correlations are more important at lower dimensions , and the process of melting is hence strongly dimensionally dependent . while three dimensional crystals exhibit long - range order below certain temperatures , statistical fluctuations play a significant role in one dimensional systems , precluding all but short - ranged local ordering for @xmath0 . the process of melting in two dimensions is more subtle , and is understood in the modern context to occur in more than one stage . an initial continuous loss of positional order precedes the proliferation of lattice defects , which accumulate and eventually complete the melting process at sufficiently high temperatures by destroying even orientational order , where each atom has a fixed number of neighbors . thermally induced fluctuations in atomic positions can have an important effect on nano - engineered systems where structures may be on the atomic scale . atomic clusters or quantum `` dots '' are mesoscopic assemblies of atoms where the scale is confined in all directions . linear structures such as carbon nanotubes are essentially one dimensional objects ( although having cross sections on the atomic scale ) where the tube length may approach macroscopic scales . finally , two dimensional systems with nanoscale thickness such as covalently bonded graphene sheets are genuine monolayers with thicknesses on the atomic scale , but spanning macroscopic areas . the novel charge transport properties of graphene have been of intrinsic fundamental interest , and have also inspired scenarios for the use of graphene in semiconductor microprocessor applications . technological uses for graphene will need a stable planar substrate for the implementation of nano - circuitry , and fundamental scientific research will also benefit from the minimization of the amplitude of random undulations in graphene layers . we examine two dimensional crystals with properties that would generically be found in two dimensional covalently bonded crystals , including stiffness with respect to displacements perpendicular to the plane of the sheet . although we do not consider temperature regimes capable of disrupting the lattice topology or number of neighbors for each atom ( e.g. by thermal rupture of bonds between neighboring atomic species ) , we examine the loss of order caused by fluctuations of atomic positions about their equilibrium positions which nonetheless leave the bonding pattern intact . if the motion of particles comprising the crystal is confined to the plane of the lattice , the gradual loss of long - range crystalline order with increasing system size has been understood as being in some respects similar to the destruction of ferromagnetic ordering in the @xmath1-@xmath2 model ( the motion of the spins are confined to the plane with a ferromagnetic coupling between them ) by thermally excited spin waves . nevertheless , on a detailed level the two systems differ . in the case of the @xmath1-@xmath2 model , spin - spin correlation functions decay algebraically with the spatial separation between spins below the kosterlitz - thouless temperature for vortex unbinding . on the other hand , the rms deviation in atomic positions in two dimensional crystals has been described as logarithmically divergent ( i.e. varying as @xmath3 $ ] where @xmath4 is a temperature dependent parameter ) for any finite temperature @xcite . in section ii , we discuss theoretical techniques and the system geometries under consideration . then , in section iii we examine three dimensional lattices where we show directly for suitably rigid lattice geometries that the rms deviations from equilibrium converge to a finite value in the thermodynamic limit , an anticipated property of three dimensional systems . moreover , we determine a reference temperature threshold @xmath5 where mean square fluctuations about equilibrium reach one tenth of a lattice constant , corresponding to the melting point according to the lindemann criterion . this temperature will serve as a point of reference in the examination of two dimensional systems where thermal fluctuations disrupt long - range order for any finite temperature . however , although we find stable crystalline order in three dimensional geometries , we also discuss a significant caveat which applies for simple cubic lattices and other geometries which lack local stiffness . to a great extent , the lattice geometries we report on are based on the two dimensional examples shown in fig . [ fig : fig12 ] . a square lattice pattern , and a triangular lattice structure are shown . the former lacks inherent rigidity , but the square lattice gains local rigidity through the activation of an extended coupling scheme in which both nearest neighbors and next - nearest neighbors interact . in the same way , a simple cubic lattice requires interactions between next - nearest as well as nearest neighbors to resist thermal fluctuations and maintain long - range crystalline order . by considering two geometries and appropriate three dimensional generalizations which differ in significant ways ( i.e. one base on a square pattern and the other assembled of triangles or tetrahedra joined at their corners ) , we identify generic thermodynamic characteristics common to both . square lattice extended coupling geometry with interactions with nearest and next - nearest neighbors ( left ) and the periodic triangular lattice and labeling scheme ( right ) ] in section iv , we examine two dimensional lattices such as those shown in fig . [ fig : fig12 ] with motion confined to the crystal plane , finding the very slow ( i.e. logarithmic ) loss of crystalline order anticipated for two dimensional crystals . on the other hand , for a two dimensional crystal embedded in three dimensions , it is important to consider transverse perturbations tending to push atoms out of the plane . we find that in the absence of binding to a substrate , two dimensional crystals are much less able to resist extraplanar distortions than fluctuations which are confined to the lattice plane . in section v , we examine dual - layer systems where coupling between the crystal planes imparts local stability with respect to extra - planar variations of atomic positions in a caricature of physical systems ( e.g. sheets of graphene ) which have a finite thickness , and would be imbued with local stiffness . we find for two distinct locally rigid dual - layer geometries similar rapid divergences of mean square displacements as the crystal is made larger , corresponding to thermally induced rippling of the crystal , and scaling linearly with the size of the system . analysis of the density of vibrational states reveals that the length scale of the random undulations increases with the size of the system with strong long wavelength contributions . on the other hand coupling to a flat substrate , however weak , places an asymptotic upper bound on the ripple amplitudes and also limits the average wavelength of thermally induced undulations . we examine thermodynamic properties ( e.g. the mean square deviations of atoms about equilibrium positions ) for crystals with short range bonding in the regime where bonds remain intact and thermally induced lengthening and shortening of bonds is small relative to the unperturbed , or equilibrium , bond length . with individual bonds varying only slightly in length , it is appropriate to model the bonds as harmonic potentials so the couplings between neighboring atoms are effectively treated as springs connecting the two particles . it is important to note that although we neglect anharmonic effects from the bonds , the noncollinearity of bonds in the crystal geometry may in principle introduce anharmonic terms in the hamiltonian . nevertheless , at temperatures near and below the melting point , many scenarios are amenable to the harmonic approximation where the neglect of anharmonicities ( whether intrinsic or geometric ) has a small impact on the accuracy of the calculation . analytical results obtained in the context of the harmonic approximation are validated in the cases we consider by good agreement with monte - carlo calculations where the anharmonic characteristics of the bonding stemming from peculiarities of the lattice geometry are rigorously taken into account . the hamiltonian is given by @xmath6 where @xmath7 is the equilibrium energetically favored bond length and @xmath8 is the instantaneous separation between atoms @xmath9 and @xmath10 . the outer sum is over the atoms in the ( finite ) crystal , and the inner sum is over the neighbors associated with the atom indexed by the label @xmath9 . the additional factor of 1/2 is included to compensate for double counting of bonds . the constant @xmath11 is the second derivative of the interatomic potential @xmath12 at the equilibrium separation @xmath7 . we develop the harmonic approximation directly from the bond length @xmath13 for the @xmath14 coordinates , it is convenient to write , for example , @xmath15 where @xmath16 is the equilibrium coordinate and @xmath17 is the shift about equilibrium . we operate in the same way for the @xmath18 and @xmath19 coordinates , finding @xmath20 where @xmath21 , @xmath22 , and @xmath23 . one may develop the harmonic approximation by expanding terms such as @xmath24 to quadratic order in the shift differences @xmath25 , @xmath26 , and @xmath27 . the result will be @xmath28^{2}$ ] , where @xmath29 is a unit vector formed from @xmath30.the terms @xmath31 and @xmath32 are vector atomic displacements such that , e.g. , @xmath33 . a salient characteristic of the bond energy is its dependence on the differences of the coordinate shifts ( e.g. @xmath34 for the @xmath14 direction ) instead of @xmath17 , @xmath35 , and @xmath36 by themselves , a condition which under many circumstances permits the neglect of anharmonicities due to bond non - collinearity . in the harmonic approximation , the lattice energy due to deviations from equilibrium positions will be @xmath37^{2}\end{aligned}\ ] ] on expanding , one obtains a quadratic expression mixing the displacements @xmath38 \ ! \ ! \ ! \left [ \ ! \begin{array}{ccc } a_{xx } & a_{xy } & a_{xz } \\ a_{yx } & a_{yy } & a_{yz } \\ a_{zx } & a_{zy } & a_{zz } \end{array } \ ! \right ] \ ! \ ! \ ! \ ! \left [ \ ! \begin{array}{c } \delta_{j}^{x } \\ \delta_{j}^{y } \\ \delta_{j}^{z } \end{array } \ ! \right ] \end{aligned}\ ] ] diagonalizing the appropriate matrix yields 3@xmath39 eigenvectors , taken to be normalized . each of the set of 3n eigenvectors has a component for the individual degrees of freedom in the crystal lattice , permitting the lattice hamiltonian to be written in decoupled form as @xmath40 with eigenvector expansion coefficients @xmath41 and eigenvalues @xmath42 ; the parameter @xmath43 is the `` primary '' harmonic constant , which is taken to be the nearest neighbor intra - planar coupling constant in schemes , such as extended models with multiple coupling constants . the eigen - modes are independently excited by thermal fluctuations , and thermodynamic equilibrium observables may be calculated by evaluating gaussian integrals . as an example , the thermally averaged mean square fluctuation per atomic species @xmath44 is ( first moments of the coordinate shifts such as @xmath45 vanish in the thermal average and do not appear in the expression below ) @xmath46 indexing the eigenvectors with the label @xmath47 and noting , e.g. , that @xmath48 , we see that the total square of the instantaneous fluctuations per particle is @xmath49 \end{aligned}\ ] ] in calculating the thermal average the term @xmath50 will be as often negative as positive when @xmath51 , and there will only be a non - zero contribution to @xmath52 if @xmath53 . hence , the double sum enclosed in square brackets will collapse to a single sum , and the calculation is reduced to the thermal average @xmath54 \end{aligned}\ ] ] the eigenvector normalization condition gives @xmath55 = 1,\end{aligned}\ ] ] and hence @xmath56 appears simply as @xmath57 the partition function @xmath58 may be calculated with the aid of @xmath59 , and one has a product of decoupled gaussian integrals , which may be written as @xmath60 with @xmath61 , @xmath62 the boltzmann constant , and the temperature @xmath63 is given in kelvins . for the sake of convenience , units are chosen such that the lattice constant @xmath64 is equal to unity , and a reduced temperature is defined with @xmath65 . evaluating the integrals in the product given in eq . [ eq : eq100 ] yields for @xmath58 @xmath66 the thermally averages mean square displacement may be written in terms of a thermal logarithmic derivative of @xmath58 , and in particular , one finds @xmath67 hence , the thermally averaged mean square deviation from equilibrium may be written as the square root of a sum over eigenvalue reciprocals . @xmath68 zero eigenvalues would lead to a diverging expression , but eigenvalues which are strictly equal to zero are artifacts of periodic boundary conditions , correspond to global translations of the crystal lattice , and are excluded from the sum . the dependence on reduced temperature consists of a @xmath69 factor . to concentrate on characteristics specific to a lattice geometry and its coupling scheme , as well as trends with respect to system size @xmath70 , the normalized mean square displacement @xmath71 will often be discussed in lieu of the full temperature dependent quantity . in the case of a periodic regular crystal lattice , it is useful exploit translational invariance , which will lead to exact expressions for the vibrational mode eigenstates and frequencies for periodic crystals ( or at the very least yielding a small matrix which may be diagonalized analytically or by numerical means if necessary ) if atomic displacements are written in terms of the corresponding fourier components . using monte carlo calculations to sample thermodynamic quantities incorporates anharmonic effects in a rigorous manner , providing a means of assessing the validity of the harmonic approximation . we employ the metropolis technique @xcite to introduce random displacements and sample the distribution corresponding to thermal equilibrium . we follow the standard metropolis prescription , where an attempted random displacement with an associated energy shift @xmath72 is accepted with probability @xmath73 if @xmath74 and the monte carlo move is invariably accepted for cases in which @xmath75 . in calculating thermodynamic quantities , we operate in terms of monte carlo sweeps where a sweep , on average , consists of an attempt to move each atom in the crystal with the acceptance of the move subject to the metropolis condition . in the calculations , the sampling of thermodynamic quantities is postponed until the completion of the first 25% of the total number of sweeps to eliminate bias from the initial conditions , which are not typical thermal equilibrium configurations for the system . to reduce errors due to statistical fluctuations in the monte carlo simulation and obtain several digits of accuracy in the results , we conduct at least @xmath76 sweeps . figure [ fig : fig1 ] ( for the square lattice with an extended coupling scheme ) and figure [ fig : fig2 ] ( for the triangular lattice geometry ) show mean square deviation curves for various temperatures ranging from temperatures an order of magnitude smaller than @xmath77 to temperatures on par with the lindemann criterion result for the melting temperature of the corresponding three dimensional system . the solid lines correspond to analytical results , while the symbols are rms values obtained with monte carlo calculations . the curves show very good agreement between the monte carlo data and analytical results over a wide range of temperatures and system sizes , and deviations are primarily mild statistical errors ( on the order of one part in @xmath78 ) in the monte carlo calculations . graph of the rms fluctuation about equilibrium versus systems size @xmath70 for various values of the reduced temperature @xmath79 for the square lattice with next - nearest neighbor couplings . the solid lines are analytic results obtained in the harmonic approximation , and symbols are results from monte carlo calculations . ] graph of the rms fluctuation about equilibrium versus systems size @xmath70 for various values of the reduced temperature @xmath79 for the triangular lattice . the solid lines are analytic results obtained in the harmonic approximation , and symbols are results from monte carlo calculations . ] illustration of the simple cubic , nonrigid structure and rigidity gained by incorporating next nearest - neighbor couplings as shown in the image to the right . ] to establish a temperature scale for the two dimensional systems , where long - range crystalline order is not expected to exist at temperatures above 0k , we first examine three dimensional lattices , which may exhibit long - range positional order at finite temperature if the lattice is suitably rigid . as a preliminary step , we perform an analysis similar to the lindemann treatment where an atom in a simple cubic geometry is coupled to six nearest neighbors . since we do not take into account the motion of neighboring atoms , we take their displacements to be zero ; certainly the excursions of neighboring atoms would average to zero , although to be precise , one would need to take into account cooperative effects of the atomic motions of the neighbors . the lattice energy has the form @xmath80 where , for example , @xmath81 is the shift in length of the bond to the nearest neighbor in the positive @xmath82 direction . applying the harmonic approximation and taking the atomic shifts to be @xmath83 , the energy becomes @xmath84\end{aligned}\ ] ] in the calculation of @xmath85 , the partition function has the form @xmath86 from the gaussian integration , we find @xmath87 . the rms displacement will be @xmath88 where @xmath89 hence , the thermally averaged mean square shift is @xmath90 . the lindemann criterion places the melting temperature at a temperature high enough that the mean square deviation @xmath85 reaches a tenth of a lattice constant , which corresponds to @xmath91 , a reduced temperature on the order of @xmath92 . if correlations among atoms are taken into account , next - nearest neighbor couplings become crucial to imparting local stiffness and maintaining long - range crystalline order . to see how rigidity is an important factor , we calculate the rms displacements for a simple cubic lattice where only couplings between nearest neighbors are taken into account . the energy stored in the lattice will be @xmath93\end{aligned}\ ] ] where a periodic geometry is assumed , and the counting factor of @xmath94 does not appear since the sum has been constructed to avoid redundancies . if we use the transformations @xmath95 where @xmath96 is the imaginary unit , and similar expressions are used for @xmath97 and @xmath98 . in terms of the fourier components , the lattice energy may be written as @xmath99 \end{aligned}\ ] ] the @xmath14 , @xmath18 , and @xmath19 degrees of freedom @xmath100 , @xmath101 , and @xmath102 automatically decouple . the normalized mean square deviation is @xmath103 , where the sum is restricted to non - zero eigenvalues . we identify three eigenvalues , @xmath104 , @xmath105 , and @xmath106 for each wave vector @xmath107 as can be seen in figure , the mean square fluctuation about equilibrium positions grows very rapidly with increasing system size . the divergence in the rms displacements is a consequence of the lack of rigidity of the simple cubic geometry , which facilitates the destruction of long range crystalline order by thermal fluctuations . however , next - nearest neighbor couplings make the lattice rigid , and are very effective in suppressing fluctuations about equilibrium and establishing long - range crystalline order for the simple cubic lattice . rms displacements graphed with system size @xmath70 for various values of @xmath108 , expressed in units of the nearest - neighbor coupling @xmath109 . panel ( a ) shows a closer view of the @xmath85 curves over a smaller range of system sizes , and panel ( b ) is a graph with a broader range of system sizes included in the plot . ] rms displacements graphed as a function of the ratio of next nearest neighbor to nearest neighbor coupling , @xmath110 . inset ( a ) is a graph of the normalized rms deviation with respect to @xmath111 , and inset ( b ) shows the rms fluctuations raised to the 1/4 power versus @xmath111 . ] the structure of the eigenvalue density states profile has informative characteristics particular to the lattice geometry from which it is obtained , and the density of states is calculated for many of the systems we report on . we achieve the thermodynamic in a genuine sense by not restricting @xmath112 , @xmath113 , and @xmath114 to discrete values as is done for finite systems . the density of states is built up by monte carlo sampling in which the wave - vector components are each generated independently from a uniform random distribution . to obtain good statistics , at least on the order of @xmath115 eigenvalues are sampled in constructing the dos . the same monte carlo sampling procedure is used to calculate the @xmath85 values shown in fig . [ fig : fig5 ] , and thereby completely remove any bias from finite size effects . the density of states corresponding to the simple cubic system ( shown in the graph in fig . [ fig : fig6 ] ) is consistent with the divergence of the rms fluctuations with increasing system size . the bimodal structure is sharply peaked in the low and high eigenvalue regimes , with the former contributing to the steady rise of @xmath85 with increasing system size @xmath70 . normalized eigenvalue density of states for the simple cubic system for an extended coupling scheme with @xmath116 with a sampling of @xmath117 eigenvalues . ] normalized eigenvalue density of states for the simple cubic system for an extended coupling scheme with @xmath118 with a sampling of @xmath117 eigenvalues . ] normalized eigenvalue density of states for the simple cubic system with coupling for an extended coupling scheme with @xmath119 with a sampling of @xmath117 eigenvalues . ] in the extended coupling scheme in the simple cubic geometry , the energy stored in the lattice is @xmath120^{2 } \ ! + \ ! [ \hat{y } \ ! \cdot \ ! ( \vec{\delta}_{ij+1k } ! \vec{\delta}_{ijk})]^{2 } \\ \ ! + \ ! [ \hat{z } \ ! \cdot \ ! ( \vec{\delta}_{ijk+1 } \ ! - \ ! \vec{\delta}_{ijk } ) ] ^{2 } \end{array } \ ! \ ! \right \ } \\ \ ! \ ! \ ! + \kappa_{2 } \ ! \left \ { \ ! \ ! \ ! \begin{array}{l } \left [ \frac{1}{\sqrt{2}}(\hat{x } + \hat{y } ) \cdot ( \vec{\delta}_{i+1j+1k } - \vec{\delta}_{ijk } ) \right]^{2 } \ ! \left [ \frac{1}{\sqrt{2 } } ( \hat{x } \!- \!\hat{y } ) \ ! \cdot \ ! ( \vec{\delta}_{i+1j-1k } ! \vec{\delta}_{ijk } ) \right]^{2 } \ ! \ ! + \ ! \\ \left [ \frac{1}{\sqrt{2 } } ( \hat{y } \!+ \!\hat{z } ) \ ! ( \vec{\delta}_{ij+1k+1 } \!- \ ! \vec{\delta}_{ijk } ) \right]^{2 } \ ! \ ! + \ ! \\ \left [ \frac{1}{\sqrt{2 } } ( \hat{y } \ ! - \ ! \hat{z } ) \ ! \cdot \ ! ( \vec{\delta}_{ij+1k-1 } \!- \ ! \vec{\delta}_{ijk } ) \right]^{2 } \ ! \ ! + \ ! \\ \left [ \frac{1}{\sqrt{2 } } ( \hat{x } \ ! + \ ! \hat{z } ) \ ! \cdot \ ! ( \vec{\delta}_{i+1jk+1 } \!- \ ! \vec{\delta}_{ijk } ) \right]^{2 } \ ! \ ! + \ ! \\ \left [ \frac{1}{\sqrt{2 } } ( \hat{x } \ ! - \ ! \hat{z } ) \ ! \cdot \ ! ( \vec{\delta}_{i+1jk-1 } \!- \ ! \vec{\delta}_{ijk } ) \right]^{2 } \end{array } % [ ( \frac{\hat{x}}{\sqrt{2 } } - \frac{\hat{y}}{\sqrt{2 } } ) \cdot % ( \vec{delta}_{i+1j-1k } - \vec{\delta}_{ijk})]^{2 } \end{array } \ ! \ ! \ ! \ ! \right \ } % e = \displaystyle { \sum_{i=0}^{n-1 } \sum_{j=0}^{n-1 } \sum_{k=0}^{n-1 } } \\ % \begin{array}{l } \frac{k_{1}}{2 } \left \ { \begin{array}{c } ( \delta_{i+1jk}^{x } - \delta_{ijk}^{x})^{2 } \\ + % ( \delta_{ij+1k}^{y } - \delta_{ijk}^{y})^{2 } + ( \delta_{ijk+1}^{z } - \delta_{ijk})^{2 } \end{array } \right \ } + \\ % \\ % \frac{k_{2}}{4 } \left \ { \begin{array}{l } % \left [ ( \delta_{i+1j+1k}^{x } - % \delta_{ijk}^{x } ) + ( \delta_{i+1j+1k}^{y } - \delta_{ijk}^{y } ) \right ] ^{2 } % \\ + \left [ ( \delta_{i+1j-1k}^{x } - % \delta_{ijk}^{x } ) - ( \delta_{i+1j-1k}^{y } - \delta_{ijk}^{y } ) \right ] ^{2 } % \\ + \left [ ( \delta_{ij+1k+1}^{y } - % \delta_{ijk}^{y } ) + ( \delta_{ij+1k+1}^{z } - \delta_{ijk}^{z } ) \end{array } \right ] ^{2 } % \\ + \left [ ( \delta_{ij+1k-1}^{y } - % \delta_{ijk}^{y } ) - ( \delta_{ij+1k-1}^{z } - \delta_{ijk}^{z } ) \right ] ^{2 } % \\ + \left [ ( \delta_{i+1jk+1}^{x } - % \delta_{ijk}^{x } ) + ( \delta_{i+1jk+1}^{z } - \delta_{ijk}^{z } ) \right ] ^{2 } \\ + % \left [ ( \delta_{i+1jk-1}^{y } - % \delta_{ijk}^{y } ) - ( \delta_{i+1jk-1}^{z } - \delta_{ijk}^{z } ) \right ] ^{2 } \end{array } % \right \ } \end{array } \ ! \ ! \ ! \ ! \right)\end{aligned}\ ] ] where @xmath109 is the coupling to nearest neighbors , @xmath108 is the coupling to next - nearest neighbors , and @xmath121 is the ratio of the next - nearest and nearest neighbor coupling constants . in terms of fourier components , one has @xmath122 \ ! \lvert \delta_{\bf{k}}^{x } \rvert^{2 } \ ! + \!\\ \left [ 2 \ ! - \ ! \cos k_{x } \cos k_{y } \ ! - \ ! \cos k_{y } \cos k_{z } \right ] \ ! \lvert \delta_{\bf{k}}^{y}\rvert^{2 } \ ! \left [ 2 \ ! - \ ! \cos k_{y } \cos k_{z } \ ! - \ ! \cos k_{x } \cos k_{z } \right ] \ ! \lvert \delta_{\bf{k}}^{z } \rvert^{2 } \ ! + \ ! \\ \sin k_{x } \sin k_{y } ( \delta_{\bf{k}}^{x } \delta_{\bf{k}}^{*y } + \delta_{\bf{k}}^{*x } \delta_{\bf{k}}^{y } ) + \\ \sin k_{y } \sin k_{z } ( \delta_{\bf{k}}^{y } \delta_{\bf{k}}^{*z } + \delta_{\bf{k}}^{*y } \delta_{\bf{k}}^{z } ) + \\ \sin k_{z } \sin k_{x } ( \delta_{\bf{k}}^{z } \delta_{\bf{k}}^{*x } + \delta_{\bf{k}}^{*z } \delta_{\bf{k}}^{x } ) \end{array } \ ! \ ! \ ! \right \ } \end{array } \ ! \ ! \ ! \ ! \ ! \right ) \end{aligned}\ ] ] with @xmath123 indicating the wave - vector with components @xmath112 , @xmath113 , and @xmath114 , and again @xmath124 . the eigenvalues are hence obtained by diagonalizing the 3@xmath1253 matrix @xmath126\end{aligned}\ ] ] although solving the cubic characteristic equation will yield analytical expressions for the eigenvalues , the result is cumbersome , and we instead use standard algorithms for the diagonalization of a symmetric matrix to efficiently obtain the eigenvalues numerically . the eigenvalues determined in this manner are used to calculate the means square atomic fluctuations , and the results are shown in figure [ fig : fig4 ] , where @xmath85 is graphed with respected to @xmath70 for a range of the next to nearest neighbor coupling strength ratio @xmath127 . whereas the mean square displacement steadily rises with system size when @xmath128 ( i.e with only nearest - neighbor couplings active ) , the curves behave very differently for nonzero @xmath129 , ultimately saturating with increasing @xmath70 . the stabilization of @xmath85 in the thermodynamic limit indicates the presence of intact long - range crystalline order . in the case of @xmath128 , the mean square deviation steadily diverges with increasing @xmath70 . the same divergence with only nearest - neighbor interactions taken into a account occurs whether one is considering the simple cubic structure , a square lattice , or a linear chain . hence , in some lattice geometries , having a three dimensional structure may be insufficient to stabilize long - range order if an extended coupling scheme is not taken into consideration . by switching on and varying the strength of the next - nearest coupling @xmath108 , one sees the appearance of long - range crystalline order as the cubic system is made increasingly rigid . in fig . [ fig : fig5 ] the mean square displacement is shown graphed versus the coupling ratio @xmath110 . the tendency for atoms to be driven from their positions in the lattice does increase as @xmath108 is shut off , but the divergence occurs at a slow rate . inset ( a ) is a graph of @xmath130 versus the logarithm of @xmath110 . while the concavity of the curve indicates a somewhat more rapid than logarithmic divergence , a semi - logarithmic plot of @xmath85 ( i.e. as shown in inset ( b ) of fig . [ fig : fig5 ] ) shows an approximately linear scaling of @xmath131 with the logarithm of the system size , still a relatively slow divergence , albeit somewhat more rapid than a simple logarithmic divergence . hence , the next - nearest neighbor couplings in the extended coupling simple cubic model are very effective in restoring long - range crystalline order . trends in the eigenvalue density of states profile with increasing @xmath110 to next - nearest neighbors are shown in in fig . [ fig : fig6 ] , fig . [ fig : fig7 ] , fig . [ fig : fig8 ] . the almost immediate retreat of the low and high frequency peaks toward the center is consistent with the effectiveness of an extended coupling scheme in stabilizing long - range crystalline order even for very small values of the ratio @xmath127 . the dos profile has a simple structure for small @xmath132 , while intermediate @xmath129 values are associated with a richer density of states curve which changes rapidly as the coupling ratio is increased further . illustration of the tetrahedral lattice geometry ] as in the case of the cubic lattice with the extended coupling scheme , we may calculate the lattice energy in the harmonic approximation for the tetrahedral lattice , which is @xmath133^{2 } + \\ \left [ \ ! \left ( \ ! \frac{1}{2 } \hat{x } \ ! + \ ! \frac{\sqrt{3}}{2 } \hat{y } \ ! \right ) \ ! \cdot \ ! ( \vec{\delta}_{ij+1k } \ ! - \ ! \vec{\delta}_{ijk } ) \ ! \right]^{2 } \ ! + \ ! \\ \left [ \ ! \left ( \ ! \frac{1}{2 } \hat{x } \ ! - \ ! \frac{\sqrt{3}}{2 } \hat{y } \ ! \right ) \ ! \cdot \ ! ( \vec{\delta}_{i+1j-1k } \ ! - \ ! \vec{\delta}_{ijk } ) \ ! \right]^{2 } \ ! + \ ! \\ \left [ \ ! \left ( \ ! \frac{1}{2 } \hat{x } \ ! + \ ! \frac{1}{2\sqrt{3 } } \hat{y } \ ! + \ ! \sqrt{\frac{2}{3 } } \hat{z } \ ! \right ) \ ! \cdot \ ! ( \vec{\delta}_{ijk+1 } \vec{\delta}_{ijk } ) \ ! \right]^{2 } \ \left [ \ ! \left ( \ ! -\frac{1}{2 } \hat{x } \ ! + \ ! \sqrt{3 } } \hat{y } \ ! + \ ! \sqrt{\frac{2}{3 } } \hat{z } \ ! \right ) \ ! \cdot \ ! ( \vec{\delta}_{i-1jk+1 } ! \vec{\delta}_{ijk } ) \ ! \right]^{2 } \\ \ ! + \ ! \left [ \left ( -\frac{1}{\sqrt{3 } } \hat{y } \ ! + \ ! \sqrt{\frac{2}{3 } } \hat{z } \ ! \right ) \cdot ( \vec{\delta}_{ij-1k+1 } \ ! - \ ! \vec{\delta}_{ijk } ) \ ! \right]^{2 } \end{array } \ ! \ ! \ ! \right ) \end{aligned}\ ] ] where there is only one coupling constant @xmath43 since bonds are considered between nearest neighbors only , the tedrahedral geometry being intrinsically rigid , and we have used the fact that the altitude of a tetrahedron is @xmath134 times the lattice constant . the energy may be expressed in terms of fourier components , and one has the task of diagonalizing the @xmath135 matrix @xmath136\end{aligned}\ ] ] normalized root mean square ( rms ) deviation shown versus @xmath137 for the three dimensional tetrahedral crystal . the inset is a graph of the normalized rms deviations , again plotted with respect to @xmath137 , with the horizontal line indicating the extrapolated @xmath85 in the thermodynamic limit . ] the three dimensional tetrahedral lattice is locally stiff even with only nearest - neighbor couplings taken into account , and the rigidity inherent in the tetrahedral lattice geometry is sufficient to preserve long - range crystalline order , as may be seen in figure [ fig : fig9 ] which displays the normalized mean square deviation versus the system size @xmath70 . the inset is a semi - logarithmic plot with the horizontal axis extending over three decades of system sizes . the saturation of the normalized rms displacement with increasing @xmath70 is evident in both of the graphs , and in the thermodynamic limit , is in the vicinity of @xmath138 . with temperature dependence included , one will have @xmath139 . hence , the lindemann criterion would give @xmath140 , compatible with the previous estimate which neglected correlations of the atomic displacements from equilibrium . in inset ( a ) of fig . [ fig : fig10 ] , the density of states is shown for the simple cubic lattice geometry with the extended coupling scheme , and for the tetrahedral lattice in inset ( b ) . for both systems , while other details of the density of states profiles differ , the curves tend swiftly to zero in the small eigenvalue regime , a hallmark of intact long range crystalline order in rigid three dimensional lattices . normalized eigenvalue density of states for the cubic crystal with nearest and next - nearest neighbor couplings ( a ) , and the density of states profile for the tetrahedral lattice ( b ) . ] we first examine the case where motion perpendicular to the plan is forbidden , and atomic deviations from equilibrium are confined to the lattice plane . we consider various geometries , but first we consider a square ( effectively a face - centered system ) , illustrated in figure [ fig : fig1 ] , where coupling to the four next nearest neighbors is taken into account . we then consider a locally rigid triangular lattice where each atom interacts with six nearest neighbors . in both the face - centered square and triangular systems , we find a logarithmic divergence with increasing system size in the mean square fluctuations about equilibrium . for the periodic square geometry with the coupling scheme extended to next - nearest neighbors , the lattice energy to quadratic order is @xmath141^{2 } \ ! + \ ! \left [ \hat{y } \ ! \left ( \vec{\delta}_{ij+1}^{y } \ ! - \ ! \vec{\delta}_{ij}^{y } \right ) \ ! \right]^{2 } \\ \ ! + \ ! \left [ \ ! \tfrac{1}{\sqrt{2 } } ( \hat{x } \ ! + \ ! \hat{y } ) \ ! \left ( \vec{\delta}_{i+1j+1}^{x } \ ! - \ ! \vec{\delta}_{ij}^{x } \right ) \ ! \right]^{2 } \\ \ ! + \ ! \left [ \ ! \tfrac{1}{\sqrt{2 } } ( \hat{x } \ ! - \ ! \hat{y } ) \ ! \left ( \vec{\delta}_{i+1j-1}^{x } \ ! - \ ! \vec{\delta}_{ij}^{x } \right ) \ ! \right ] ^{2 } \end{array } \ ! \ ! \right ) \end{aligned}\ ] ] operating in reciprocal space , one diagonalizes the @xmath142 matrix @xmath143\end{aligned}\ ] ] yielding the eigenvalues @xmath144 in the case of the triangular lattice with six fold coordination , one may also obtain analytical expressions for the mean square deviations . in real space , the harmonic approximation for the energy stored in the lattice is @xmath145^{2 } + \\ \left [ ( \frac{1}{2}\hat{x } + \frac{\sqrt{3}}{2 } \hat{y } ) \ ! \left ( \vec{\delta}_{ij+1 } - \vec{\delta}_{ij } \right ) \right]^{2 } + \\ \left [ ( \frac{1}{2}\hat{x } - \frac{\sqrt{3}}{2 } \hat{y } ) \ ! \left ( \vec{\delta}_{i+1j-1 } - \vec{\delta}_{ij } \right ) \right]^{2 } \end{array } \ ! \ ! \right ) \end{aligned}\ ] ] expressing the displacements in terms of fourier components , one decouples the @xmath14 and @xmath18 degrees of freedom by diagonalizing the matrix @xmath146 \end{array } \ ! \ ! \right ) \ ! \ ! & \tfrac{\sqrt{3}}{2 } \left ( \cos [ k_{y } \ ! - \ ! k_{x } ] \ ! - \ ! \cos k_{y } \right ) \\ \\ \tfrac{\sqrt{3}}{2 } \left ( \cos [ k_{y } \ ! - \ ! k_{x } ] \ ! - \ ! \cos k_{y } \right ) \ ! \ ! & \left ( \ ! \ ! \begin{array}{c } 3 \ ! - \ ! \cos k_{y } \\ \ ! - \ ! \tfrac{3}{2 } \cos [ k_{y } \ ! - \ ! k_{x } ] \end{array } \ ! \ ! \right ) \end{array } \ ! \ ! \ ! \right]\end{aligned}\ ] ] yields the eigenvalues @xmath147 \\ \nonumber \pm \sqrt{\begin{array}{c } \cos^{2 } k_{x } \ ! + \ ! \cos^{2 } k_{y } \ ! + \ } ( k_{y } \ ! - \ ! k_{x } ) \ ! - \ ! \cos k_{x } \cos k_{y } \\ \ ! - \ ! \cos k_{y } \cos ( k_{y } \ ! - \ ! k_{x } ) \ ! - \ ! \cos ( k_{y } \ ! - \ ! k_{x } ) \cos k_{x } \end{array}}\end{aligned}\ ] ] for convenience in comparison with the analytical results in the harmonic approximation , we consider periodic boundary conditions in the crystal plane . we have also examined anchored lattices , where atoms at the periphery are prevented from moving , while those in the interior are free to move . for both the free and fixed boundary conditions , as in the three dimensional case , we obtain qualitatively similar results , and the same physical phenomena . in fig . [ fig : fig13 ] and fig . [ fig : fig14 ] , the normalized mean square deviation @xmath148 is graphed with respect to the system size @xmath70 for the square lattice in the extended scheme and the triangular lattice , respectively . the overall behavior of the mean square deviations from equilibrium is qualitatively the same for both lattice geometries . in both cases , the main graph is semi - logarithmic with @xmath149 on the ordinate . the traces are linear to a very good approximation for all regimes of @xmath70 ( i.e. for small , moderate , and large ) shown , and the linearity is maintained for four decades of system sizes ranging from several to on the order of a few times @xmath150 lattice constants . in fig . [ fig : fig13 ] and fig . [ fig : fig14 ] , inset ( a ) is a standard plot , and the apparent saturation of the @xmath151 curve is a hallmark of the slow loss of long - range crystalline order best seen on a semi - logarithmic graph . inset ( b ) in fig . [ fig : fig13 ] and fig . [ fig : fig4 ] contains as semi - logarithmic plot with @xmath71 [ instead of @xmath149 ] on the ordinate axis . the curves plotted in this manner are not linear , and it is evident that the divergence of the fluctuations about equilibrium is actually somewhat slower than logarithmic ; instead , it is @xmath149 which scales as @xmath152 . square of the normalized root mean square ( rms ) deviation shown versus @xmath137 for the square lattice system with extended couplings . the solid line encompassing the open circular symbols is a strictly linear fit . inset ( b ) is a semi - logarithmic graph of the normalized rms deviations , plotted with respect to @xmath137 . ] square of the normalized root mean square ( rms ) deviation shown versus @xmath137 for the triangular lattice . the solid line encompassing the open circular symbols is a strictly linear fit . inset ( a ) is a standard plot of the rms deviation with respect to system size @xmath70 , while inset ( b ) is a semi - logarithmic graph of the normalized mean square deviations , plotted with respect to @xmath137 . ] as in the case of the three dimensional systems , it is informative to examine the density of states , shown in the graph of fig . [ fig : fig15 ] for the square lattice in the extended coupling scheme in panel ( a ) and the triangular lattice in panel ( b ) of fig . [ fig : fig15 ] . again , while details of the density of states profiles shown are peculiar to the lattice under consideration , the behavior in the regime of low eigenvalues is quite similar , and both curves tend to a finite value instead of dropping swiftly to zero as in the density of states for the rigid three dimensional lattices . the failure of the density of states to vanish in the small eigenvalue limit contributes to the slow divergence of @xmath85 in @xmath70 . normalized eigenvalue density of states for the face centered square lattice with motion confined to the lattice plane , depicted in panel ( a ) , and for the triangular lattice in panel ( b ) . ] the locally stiff face - centered square and triangular lattices show the anticipated slow logarithmic divergence in system size . however since laboratory systems often are not vertically constrained , it is important to examine a scenario where motion perpendicular to the plane of the lattice may be considered . there is an important difficulty with single layer systems , in that motion perpendicular to the plane is not hindered since there are no restraining bonds with a directional component transverse to the plane of the layer . however , by considering dual - layer geometries , it is possible to incorporate local stiffness with respect to perturbations that would push atoms above or below the lattice . we examine analogs of the simple cubic lattice , where we again use an extended coupling scheme to create local stiffness . on the other hand , we also consider a dual - layer tetrahedral lattice . although the two lattice geometries achieve local stiffness in different ways , the similarities we find in thermodynamic behavior of the mean square atomic fluctuations suggest these characteristics would appear in the generic case as well . [ fig : fig16 ] illustrates the structure of the dual - layer square lattice with an extended coupling scheme ; the additional couplings between next nearest neighbors impart local stiffness to the system with respect to perturbations perpendicular to the planes of the square lattices . [ fig : fig17 ] shows how the dual - layer systems is constructed as a caricature of the graphene lattice . the image labeled ( a ) is a schematic illustration of a single hexagonal cell in a graphene monolayer . the bonding shares similarities with that in a benzene ring with delocalized @xmath153 orbitals forming honeycomb networks of charge density above and below the plane occupied by the carbon atomic nuclei . the superimposed lattice work is a rigid network compatible with the symmetries of the graphene layer and set up to capture the rigidity of the hexagonal cells making up a sheet of graphene . with the honeycomb graphene pattern removed , the remaining lattice geometry and the labeling scheme for the crystal members is shown in fig . [ fig : fig18 ] . illustration of the periodic dual - layer square lattice with nearest and next - nearest neighbor coupling and labeling scheme ; blue indices refer to the upper layer , while red indices pertain to the lower plane . ] schematic representation of a coarse - grained super - structure for a graphene sheet , title="fig : " ] schematic representation of a coarse - grained super - structure for a graphene sheet , title="fig : " ] illustration of the periodic dual - layer triangular lattice and labeling scheme ; blue indices refer to the upper layer , while red indices pertain to the lower plane . ] with the superscript i representing the lower plane and ii indicating the upper plane , the lattice energy in the harmonic approximation is @xmath154^{2 } \ ! \ ! + \ ! \left [ \hat{y } \ ! \cdot \ ! ( \vec{\delta}_{ij+1}^{\alpha } -\vec{\delta}_{ij}^{\alpha } ) \ ! \right]^{2 } \\ + \left [ \frac{1}{\sqrt{2}}(\hat{x } + \hat{y } ) \ ! \cdot \ ! ( \vec{\delta}_{i+1j+1}^{\alpha } - \vec{\delta}_{ij}^{\alpha } ) \ ! \right]^{2 } + \\ \left [ \frac{1}{\sqrt{2 } } ( \hat{x } - \hat{y } ) \cdot ( \vec{\delta}_{i+1j-1}^{\alpha } \ ! - \ ! \vec{\delta}_{ij}^{\alpha } ) \ ! \right ] ^{2 } \end{array } \ ! \ ! \right \ } \\ + \kappa_{z } \left \ { \ ! \ ! \begin{array}{l } \left [ \hat{z } \cdot \left ( \vec{\delta}_{ij}^{\mathrm{ii } } - \vec{\delta}_{ij}^{\mathrm{i } } \right ) \right ] ^{2 } \\ \ ! + \ ! \left [ \tfrac{1}{\sqrt{2}}(\hat{x } \ ! + \ ! \hat{z } ) \ ! \cdot \ ! \left ( \vec{\delta}_{i+1j}^{\mathrm{ii } } \ ! -\ ! \vec{\delta}_{ij}^{\mathrm{i } } \right ) \right]^{2 } \\ \ ! + \ ! \left [ \tfrac{1}{\sqrt{2}}(-\hat{x } \ ! + \ ! \hat{z } ) \ ! \cdot \ ! \left ( \vec{\delta}_{i-1j}^{\mathrm{ii } } \ ! -\ ! \vec{\delta}_{ij}^{\mathrm{i } } \right ) \ ! \right]^{2 } \\ \ ! + \ ! \left [ \tfrac{1}{\sqrt{2}}(\hat{y } \ \hat{z } ) \ ! \cdot \ ! \left ( \vec{\delta}_{ij+1}^{\mathrm{ii } } \ ! -\ ! \vec{\delta}_{ij}^{\mathrm{i } } \right ) \ ! \right]^{2 } \\ \ ! + \ ! \left [ \tfrac{1}{\sqrt{2}}(-\hat{y } \ ! + \ ! \hat{z } ) \ ! \cdot \ ! \left ( \vec{\delta}_{ij-1}^{\mathrm{ii } } \ ! -\ ! \vec{\delta}_{ij}^{\mathrm{i } } \right ) \ ! \right]^{2 } \end{array } \ ! \ ! \right \ } \end{array } \ ! \ ! \ ! \ ! \right)\end{aligned}\ ] ] the corresponding complex hermitian 6@xmath1256 matrix to be diagonalized has the form @xmath155 = \left [ \ ! \begin{array}{cc } \hat{a } & \hat{b } \\ \hat{b}^{\dagger } & \hat{a } \end{array } \ ! \right]\end{aligned}\ ] ] where @xmath156 and @xmath157 are @xmath135 matrices and @xmath158 is the hermitian conjugate of @xmath157 . the sub - matrices @xmath156 and @xmath157 are given by @xmath159\end{aligned}\ ] ] where @xmath160 for @xmath156 and @xmath161\end{aligned}\ ] ] for @xmath157 where the six eigenvalues for each wave - number pair are calculated numerically with code available in the eispack linear algebraic library for diagonalizing complex hermitian matrices . on the other hand a locally stiff dual - layer system which may be regarded as a section of the three dimensional tetrahedral lattice such that the upper and lower layers are triangular lattices with connections between the layers . the dual - layer lattice structure based on the tetrahedral geometry is illustrated in figure [ fig : fig18 ] . the vertices of the upper layer are positioned above the centers of the triangles in the lower layer with bonds extending from atoms in the upper layer to each of the corners of the triangle below such that each atom in the dual - layer system is a member of a rigid tetrahedron ; the result is a locally stiff layer , as in the dual - layer square lattice extended model , but with a very different geometric structure . the lattice energy for the dual - layer tetrahedral system to quadratic order in the displacements @xmath162 and @xmath163 has the form @xmath164^{2 } + \\ \left [ \left ( \tfrac{\hat{x}}{2 } \ ! + \ ! \tfrac { \sqrt{3 } \hat{y}}{2 } \right ) \ ! \cdot \ ! ( \vec{\delta}_{ij+1}^{\alpha } - \vec{\delta}_{ij}^{\alpha } ) \right ] ^{2 } + \\ \left [ \left ( \tfrac{\hat{x}}{2 } - \tfrac { \sqrt{3 } \hat{y}}{2 } \right ) \cdot ( \vec{\delta}_{i+1j-1}^{\alpha } \ ! - \ ! \vec{\delta}_{ij}^{\alpha } ) \right]^{2 } \end{array } \ ! \ ! \right \ } \!+ \ ! \\ \\ \ ! \kappa_{z } \ ! \left \ { \ ! \ ! \ ! \begin{array}{c } \left [ \left ( \tfrac{\hat{x}}{2 } \ ! + \ ! \tfrac{\hat{y}}{2 \sqrt{3 } } \ ! + \ ! \sqrt{\tfrac{2}{3 } } \hat{z } \right ) \cdot ( \vec{\delta}_{ij}^{\mathrm{ii } } - \vec{\delta}_{ij}^{\mathrm{i } } ) \right]^{2 } + \\ \left [ \ ! \left(\ ! - \tfrac{\hat{x}}{2 } \ ! + \ ! \tfrac{\hat{y}}{2 \sqrt{3 } } \ ! + \ ! \sqrt{\tfrac{2}{3 } } \hat{z } \right ) \ ! \cdot \ ! ( \vec{\delta}_{ij}^{\mathrm{ii } } \ ! - \ ! \vec{\delta}_{i+1j}^{\mathrm{i } } ) \ ! \right]^{2 } \ ! + \ ! \\ \left [ \ ! \left ( \ ! -\tfrac{\hat{y}}{\sqrt{3 } } \ ! + \ ! \sqrt{\tfrac{2}{3 } } \hat{z } \right ) \ ! \cdot \ ! ( \vec{\delta}_{ij}^{\mathrm{ii } } \ ! - \ ! \vec{\delta}_{ij+1}^{\mathrm{i } } ) \ ! \right]^{2 } \end{array } \ ! \ ! \ ! \right \ } \end{array } \ ! \ ! \ ! \ ! \right)\end{aligned}\ ] ] expressing the lattice energy in terms of fourier components leads to a @xmath165 matrix to be diagonalized , which may again be written in terms of @xmath135 submatrices as @xmath166 $ ] , where @xmath167\end{aligned}\ ] ] where again @xmath160 for the sub - matrix @xmath156 and @xmath168\end{aligned}\ ] ] for the sub - matrix @xmath157 . the use of a dual - layer lattice geometry to provide resistance to transverse deviations is not sufficient to prevent a rapid divergence in @xmath85 with increasing system size @xmath70 . whereas thermally averaged mean square fluctuations grew very slowly ( i.e. logarithmically ) when the atomic motions are confined to the lattice plane , @xmath85 for the dual - layer systems increases linearly with @xmath70 ; ultimately , it is not difficult for the sheet to bend and flex in the presence of thermal fluctuations in spite of its locally stiff characteristics . the diverging mean square deviations from equilibrium and other thermodynamic characteristics of the the dual - layer square lattice in the extended coupling scheme and its counterpart based on a tetrahedral geometry are examined , with consideration given to the effects of increasing @xmath70 and variations in the inter - layer coupling strength . the graph of the normalized rms displacement , shown in fig . [ fig : fig20 ] for the dual - layer square lattice with an extended coupling pattern shows a dependence on systems size which is an asymptotically linear growth in @xmath70 . @xmath85 curves for several values of the interlayer coupling @xmath169 are shown ; the intra - layer coupling is taken to be unity , so @xmath169 is effectively expressed in units of @xmath43 . although the relative interlayer couplings range over three orders magnitude , there is little variation of the curves , especially for @xmath170 . similarly , the @xmath85 curves ultimately vary linearly in the system size @xmath70 with little dependence on the relative magnitude of @xmath169 , which again is expressed in units of @xmath43 . again , it is useful to examine the density of states for the eigenvalues in the case of the locally rigid dual - layer systems , which are richer than the density of states profiles corresponding to rigid three dimensional lattices or those of the single layer geometries with atomic fluctuations confined to intra - planar motion . although details of the density of states profiles for the two geometries differ , the both curves show a divergence of the density of states with decreasing eigenvalue magnitude , whereas the density states remained constant in the case of the planar systems with exclusively intra - planar motion and vanished altogether for the rigid three dimensional systems . inset ( a ) of fig . [ fig : fig19 ] show the dos for the dual - layer square lattice , while inset ( b ) is a graph of the density of states for the locally stiff tetrahedral system . the dos cusp for both lattices at the zero eigenvalue point is responsible for the rapid divergence of @xmath85 with increasing @xmath70 . normalized mean square displacements for @xmath169 = 0.01 , 0.03 , 0.1 , 1.0 , and 10.0 for the dual - layer square lattice with an extended coupling scheme , where @xmath169 is in units of the intralayer coupling @xmath43 . the inset is a closer view of the rms curves . ] normalized mean square displacements for @xmath169 = 0.01 , 0.1 , 1.0 , and 10.0 for the dual - layer tetrahedral lattice where @xmath169 is in units of the triangular lattice intra - layer coupling @xmath43 . the inset is a closer view of the rms curves . ] normalized eigenvalue density of states for the dual - layer cubic system with with nearest and next - nearest neighbor interactions ( a ) and for the dual - layer locally stiff lattice based on a tetrahedral lattice geometry ( b ) for system size @xmath171 . ] while adjusting the interaction between the layers to enhance the resistance to local transverse perturbations has little effect on the mean square fluctuations for large values of @xmath70 , the eigenvalue density of states evolves as the interplanar to intraplanar coupling ratio @xmath172 is modified . density of states profiles for @xmath173 values ranging from @xmath174 to @xmath175 are shown for the dual - layer square system with and extended coupling pattern in fig . [ fig : fig22 ] and for the tetrahedral counterpart in fig . [ fig : fig23 ] . density of states profiles are shown for strong ( @xmath175 ) and moderate ( @xmath176 ) values of the the coupling ratio in insets ( a ) and ( b ) of fig . [ fig : fig22 ] and fig . [ fig : fig23 ] , and there is little change in the dos curve in the low eigenvalue regime . on the other had , as @xmath173 decreases further and the interplanar coupling begins to fall below parity with that in the plane , the eigenvalue density of states profiles begin to change more drastically , as may be seen in panels ( b ) and ( c ) of fig . [ fig : fig22 ] for the dual - layer square system and fig . [ fig : fig23 ] for the dual - layer tetrahedrally based geometry ; the distribution in both cases rapidly grows narrower with decreasing @xmath173 . although the two lattice geometries are very distinct , similar ( and likely generic to locally rigid dual - layer lattices ) trends may be seen in the dos profiles in the regime of low eigenvalues as @xmath173 is reduced . normalized density of states profiles for @xmath169 = 3.0 , 1.0 , 0.3 , and 0.1 for the dual - layer square lattice with an extended coupling scheme where @xmath169 is in units of the triangular lattice intra - layer coupling @xmath43 . ] normalized density of states profiles for @xmath169 = 3.0 , 1.0 , 0.3 , and 0.1 for the dual - layer tetrahedral lattice where @xmath169 is in units of the triangular lattice intra - layer coupling @xmath43 . ] we incorporate an attractive interaction with a flat substrate by including an additional harmonic potential acting on the lower members of the tetrahedral and square extended coupling dual - layer systems . we take the attraction to depend only on the atomic shift @xmath177 above the planar system , and the additional term hence has the form @xmath178 figure [ fig : fig24 ] shows the effect of the substrate coupling on @xmath85 in the case of the dual - layer square system with an extended coupling scheme . the graph , which shows mean square deviation curves for a wide range of @xmath179 values , indicates the capacity of even a very mild substrate coupling to suppress thermally induced undulations in the dual - layer sheet . similarly , for the tetrahedrally based dual - layer lattice geometry , an attractive interaction with a substrate considerably reduces fluctuations transverse to the lattice planes , preventing a rapid divergence of @xmath85 . the mean square deviation curves are shown in figure [ fig : fig25 ] . we also examine the effect of an attractive substrate coupling on the density of states profiles , and results are displayed in fig . [ fig : fig26 ] for a range of substrate coupling constants @xmath179 . with increasing @xmath179 , a salient trend is the opening of a separation between the sharp cusp and the zero eigenvalue mark on the abscissa . the migration of the maximum formerly at the zero eigenvalue point to a peak at a larger eigenvalue is associated with a sharp reduction of the mean square fluctuations about equilibrium , and the lattice is better able to withstand transverse fluctuations . the presence of a flat substrate plays a very important role in dictating the overall structure and amplitude of ripples in the dual - layer geometries we report on here . this result is in accord with recent experiments on graphene sheets deposited on cleaved mica substrates @xcite , where the careful preparation of flat substrates significantly dampens the ripple amplitude , whereas much larger undulations are seen with sheets attached to substrates with poorer control over the flatness . to determine which length scales are associated with the strongest contributions to the thermally averaged mean square deviations about equilibrium , we have prepared histograms showing the relative contribution to @xmath85 versus inverse wave - vector magnitude , with the latter providing a length scale . apart from a significant diminution in the height of thermally excited undulations in the dual - layered sheets , we also find a considerable reduction in their typical wavelength . in figure [ fig : fig27 ] , for the dual - layer tetrahedral lattice in the absence of a substrate couplings , the dominant contribution to @xmath85 comes from large length scales comparable to the scale of the lattice . however the picture changes with the activation of a finite substrate coupling as may be seen in the inset with the peak height at the minimal wave - number decreasing with increasing @xmath179 . moreover , as may be seen in figure [ fig : fig28 ] , introducing even a weak anchoring to the foundation below immediately creates a strong peak in the short wavelength regime , skewing the size of thermally induced ripples toward smaller length scales . for the dual - layer square lattice in the extended scheme , the main figure and the inset are graphs of mean square fluctuations versus system size @xmath70 for a range of substrate couplings @xmath179 , with @xmath179 given in units of the inter - planar and intra - planar coupling constant ( both equal to @xmath43 ) . the symbol legend on the main plot also pertains to the inset ] for the dual - layer tetrahedral lattice , the main figure and the inset are graphs of mean square fluctuations versus system size @xmath70 for a range of substrate couplings @xmath179 , with @xmath179 given in units of the inter - planar and intra - planar coupling constant ( both equal to @xmath43 ) . the symbol legend on the main plot also pertains to the inset ] normalized density of states for the dual layer system for various substrate coupling strengths @xmath179 . panels ( a ) , ( b ) , ( c ) , and ( d ) show the density of states for @xmath179 equal to @xmath180 , @xmath181 , @xmath182 , and @xmath183 respectively . ] in conclusion , we have examined thermally induced fluctuations about equilibrium in two and three dimensional crystalline solids with a local bonding scheme . while long - range crystalline order may exist in three dimensional crystal lattices , some geometries ( e.g. the simple cubic lattice ) are not rigid when only nearest - neighbor couplings are taken into account , and an extended coupling scheme is needed to prevent the divergence of mean square fluctuations with increasing system size @xmath70 . in two dimensional lattices , we find rms fluctuations to increase at a very slow ( logarithmic ) rate when motion is confined to the lattice plane . on the other hand , when transverse motion is permitted , thermal fluctuations are very effective in bringing about significant vertical displacements of particles which contribute to rapidly growing deviations from equilibrium , and @xmath85 ultimately diverges at a linear rate in @xmath70 . the asymptotically linear divergence in the mean square deviations from equilibrium is insensitive to the strength of the interlayer coupling ; @xmath85 values appear to converge and eventually show identical behavior with increasing system size whether the coupling @xmath169 established between the layers to provide local stiffness is quite weak or very strong relative to the bonding @xmath43 between atoms in the same layer . introducing a coupling @xmath179 to a flat substrate very effectively hinders transverse fluctuations in two dimensional crystal lattices , even in the coupling is very weak , and reflects the importance of a substrate in shaping the characteristics of ripples set up by thermal fluctuations by inhibiting transverse deviations . an attractive coupling to a fixed substrate also reduces the typical lateral length scale or wavelength of thermally excited undulations in lattices bound to a substrate . these tendencies are consistent with recent experimental observations that control over the flatness of the underlying surface is directly related to the amplitude and length scale of thermally induced ripples .

investigations of dynamical properties of fluid fluid dispersions ( e.g. , emulsions @xcite and polymer blends @xcite ) require detailed understanding of the behavior of viscous drops in creeping flows . such understanding is also crucial in development of new drop - based microfluidic systems @xcite . therefore , drop dynamics at small reynolds numbers have been extensively studied experimentally @xcite , computationally @xcite and theoretically @xcite . these investigations revealed complex nonlinear drop dynamics resulting from the coupling between the drop shape and fluid flow . examples of nonlinear phenomena that occur under creeping - flow conditions include formation of self - similar neck regions during a drop breakup process @xcite , universal slow evolution of drops near the critical flow strength above which there are no stationary drop shapes @xcite , and existence of two branches of stable stationary shapes of highly viscous drops in two - dimensional stokes flows with nonzero vorticity @xcite . as revealed by the analysis presented by @xcite , there exists a flow - parameter range where a high - viscosity drop can either adopt a nearly spherical shape stabilized primarily by the rotational flow component or an elongated shape stabilized primarily by capillary forces . abrupt changes of the drop shape from one state to the other can be used in manipulation of emulsion micro - structure and for controlling the behavior of highly viscous drops in microfluidic devices . due to discontinuous changes of emulsion microstructure , the bistable drop behavior may also significantly affect emulsion rheology . moreover , the mechanism of the bistability is of fundamental interest because of close analogies to the dynamics of vesicles @xcite and macromolecules @xcite in external flows . due to the fundamental significance and because of potential applications ( such as those mentioned above ) it is important to explore the dynamics of highly viscous drops in linear flows with nonzero vorticity . however , the investigations presented by @xcite were limited to stationary drop shapes and stationary external flows . in the present study we focus on drop behavior in time - dependent flows . we elucidate the physical mechanism that give rise to the bistable drop behavior and examine the consequence of these mechanism for drop response to time variation of the fluid vorticity . the system dynamics is investigated via direct boundary - integral simulations @xcite and by using a small - deformation approach @xcite . in particular we show that the small - deformation equations with only several essential terms retained reproduce complex dynamical features of drop evolution that are associated with drop bistability . to emphasize important aspects of the drop dynamics we consider three flow variation protocols . in the first protocol , the vorticity is slowly increased and then decreased . we find that such quasistatic vorticity ramping gives access to both the elongated and compact , nearly spherical stationary drop shapes . the drop exhibits a hysteretic behavior , with transitions between the compact shape ( rotationally stabilized ) and elongated shape ( stabilized by capillary forces ) occurring at different values of the vorticity when it is slowly ramped up or down . in the second protocol , the vorticity undergoes finite - frequency periodic oscillations . as expected , at low frequencies the drop behavior is quasistatic , with a hysteresis loop analogous to the one observed for linear ramping . at high frequencies the oscillations average out , and the drop undergoes small oscillations around the stationary shape corresponding to the average flow . however , at intermediate frequencies we find a much more complex behavior . in particular we show that there exists a frequency and amplitude domain where the drop response to the periodic forcing is chaotic . since in the creeping flow regime fluid motion is governed by the linear stokes equations , the nonlinear chaotic behavior of the drop stems entirely from the coupling of the drop shape to the fluid velocity . an analysis of drop motion in the chaotic domain indicates that the transition to chaos is associated with the existence of two stationary states observed in steady flow . in our third flow - variation protocol , the vorticity of the imposed flow undergoes random changes . we observe , that the resulting statistical distribution of the drop length is bimodal in a certain regime of flow parameters , with two peaks around the drop length corresponding to the short and long stationary solutions . hence , we find that also in this problem the existence of two stationary states underpins drop behavior in a time - dependent flow . in [ viscous drop in linear flows ] the system considered in our paper is defined . the quasistatic hysteretic drop behavior is analyzed in [ hysteretic behavior ] , our results for chaotic drop dynamics are presented in [ sec : sinusoidal_chaos ] , and drop motion in linear flows with randomly varying vorticity is discussed in [ drop statistics in a linear flow with stochastic vorticity ] . our findings are summarized in [ sec : conclusion ] . we consider a viscous drop suspended in an incompressible newtonian fluid of a constant viscosity @xmath0 . the viscosity of the drop fluid is @xmath1 , and the interfacial tension between the two phases is @xmath2 . the drop is surfactant free , and no marangoni stresses are present . there are also no buoyancy forces . we focus here on nonlinear effects that stem entirely from the coupling of the fluid flow to the drop shape ( but not from the fluid inertia ) . therefore , the creeping - flow conditions are assumed . in the creeping - flow regime the fluid motion in the regions inside ( @xmath3 ) and outside ( @xmath4 ) the drop is governed by the stokes equations @xmath5 the fluid velocity @xmath6 is continuous at the drop interface @xmath7 . due to the absence of the marangoni stresses the tangential viscous traction is also continuous . the jump in the normal viscous traction across @xmath7 is equal to the capillary pressure @xmath8 = 2\kappa\sigma,\ ] ] where @xmath9 is the viscous stress tensor , @xmath10 is the outward normal unit vector , and @xmath11 is the local curvature . ( 300,8)(0,0 ) ( 75,0)@xmath12 ( 176,0)@xmath13 ( 290,0)@xmath14 the drop is subject to two - dimensional linear incident flow @xmath15 where @xmath16 is the strain rate , @xmath17 is the dimensionless vorticity parameter , @xmath18 is the position , and @xmath19 and @xmath20 are the symmetric and antisymmetric parts of the normalized velocity - gradient tensor . in an appropriately adopted coordinate system we have @xmath21 without a loss of generality . accordingly , @xmath22 corresponds to a purely straining flow with the extensional axis @xmath23 and the compressional axis @xmath24 , and @xmath25 corresponds to shear flow in the @xmath26 direction with the velocity gradient in the @xmath27 direction . the tensor @xmath20 in equation describes the rigid - body rotation in the anti - clockwise direction with the angular velocity @xmath28 the decomposition of the flow field into the straining and rotational components is sketched in figure [ flow decomposition ] . the dynamics of our system is characterized by three dimensionless parameters : there is the viscosity ratio @xmath29 that characterizes dissipative forces in the drop- and continuous - phase fluids . the capillary number @xmath30 ( where @xmath31 is the radius of an undeformed drop ) characterizes the ratio between the deforming viscous forces produced by the imposed flow and the capillary forces that resist drop deformation and drive the drop towards the equilibrium spherical shape . finally the vorticity parameter @xmath17 characterizes the magnitude of the rotational component of the external flow relative to the extensional component . in this paper we focus on the parameter regime where the drop deformation may be significant , which requires that @xmath32 . ( however , the flow is not strong enough to cause drop breakup . ) we also assume that the drop - phase fluid is much more viscous than the continuous - phase fluid , @xmath33 drop deformation process is hindered at large drop - phase viscosities , while drop rotation is only weakly affected by the viscous stresses inside the drops . therefore the relative effect of the drop rotation is amplified in the regime , and the rotational component of the external flow produces nontrivial qualitative effects . and capillary number @xmath34 . normalized drop length @xmath35 and drop angle @xmath36 versus vorticity parameter @xmath17 . arrows indicate the direction of increasing time , and inset shows dependence of @xmath17 on time . ( results from boundary - integral simulations . ) , title="fig:",width=240 ] and capillary number @xmath34 . normalized drop length @xmath35 and drop angle @xmath36 versus vorticity parameter @xmath17 . arrows indicate the direction of increasing time , and inset shows dependence of @xmath17 on time . ( results from boundary - integral simulations . ) , title="fig:",width=240 ] to illustrate the effect of the rotational component of the flow on the dynamics of a highly viscous drop we consider a system where the parameter @xmath17 is first slowly increased and then slowly decreased . we adopt here a linear ramping protocol where the vorticity is slowly ramped up from @xmath22 to @xmath37 and then ramped down back to zero . before the ramping occurs the flow is maintained at @xmath22 ( for 5% of the total ramping time ) to allow the drop to relax to the stationary shape in purely straining flow . this vorticity variation protocol is represented in the inset of figure [ fig:1 ] . the evolution of the drop shape in this time - dependent flow is depicted in figure [ fig:1 ] . figure [ fig:1 ] represents the drop length @xmath35 , and figure [ fig:1 ] shows the drop angle @xmath38 ( measured anticlockwise from the axis @xmath26 ) ; both quantities are plotted versus the vorticity parameter @xmath17 . in our example , the drop viscosity is @xmath39 . the capillary number @xmath40 is below the critical value for drop breakup ( @xmath41 in 2d straining flow ) but it is sufficiently large to allow for a significant flow - induced drop deformation . the total ramping time is @xmath42 , so that the drop response to the flow variation is nearly quasistatic . the calculations were performed using the adaptive boundary - integral procedure developed by @xcite . the results shown in figure [ fig:1 ] indicate that the drop response to the vorticity variation is hysteretic . at @xmath22 drop is elongated , and it is aligned with the extensional axis of the straining component of the flow ( @xmath43 ) . with increasing @xmath17 , the drop orientation slowly changes towards the symmetry axis @xmath26 ( @xmath44 ) , and the drop length slowly decreases . at a critical value of the vorticity parameter , @xmath45 , a discontinuous change occurs : the drop length and the angle suddenly decrease . afterwards the drop is almost spherical and nearly aligned with the axis @xmath26 . when the direction of the vorticity change is reversed , the drop initially retraces its trajectory . however , the drop does not jump back to the elongated shape until the vorticity reaches the lower critical value @xmath46 . the bistable drop behavior and the associated hysteretic shape evolution stem from the existence of two mechanisms that can stabilize a viscous drop in linear flows with rotation . namely , the drop can be stabilized by the capillary stresses ( which drive the drop towards the equilibrium spherical shape ) or by the vorticity flow component ( which rotates the drop out of the extensional axis of the straining component of the flow ) . in a purely straining flow the drop assumes the interfacial - tension - stabilized elongated shape @xcite . the form of the drop results from the balance between drop deformation by the flow and drop relaxation due to the capillary forces . the deformation and relaxation occur on the respective time scales @xmath47 @xmath48 both of which are proportional the viscosity ratio @xmath49 . the drop deformation @xmath50 is determined by the time scale ratio @xmath51 and therefore it is independent of the viscosity ratio in the limit @xmath52 . in purely staining flows , the drop is oriented along the extensional axis @xmath23 . for small values of @xmath17 , the vorticity flow component produces an @xmath53 perturbation of the drop orientation . the corresponding decrease of the drop length is @xmath54 . however , a further drop rotation is arrested because the straining component of the flow produces hydrodynamic stresses that pull the elongated drop back towards the straining axis ( as illustrated in figure [ mechanism ] ) . since the @xmath55 internal circulation inside an elongated high - viscosity drop is weak , the drop in its stationary state behaves analogously to a rigid body whose equilibrium orientation results from the balance of the torques produced by the straining and rotational components of the external flow . the transition to the compact drop shape occurs when the vorticity flow component becomes too strong to be balanced . under such conditions , a rigid body would undergo a transition to a periodic motion with continuous rotation in the clockwise direction . similarly , a drop also starts to continuously rotate when @xmath17 achieves the upper critical value @xmath56 . however , during the rotation the drop length decreases because the drop becomes misaligned with the extensional axis of the flow . as a result , the drop relaxes to a nearly spherical shape . in this new , compact shape the fluid inside the drop circulates with the angular velocity @xmath57 that is nearly equal to the angular velocity of the external flow . within each period of rotation the drop undergoes a small deformation produced by the straining component of the external flow ( as schematically illustrated in figure [ mechanism ] ) . however , the deformation does not accumulate because it is constantly convected away by the rotational component of the flow . since the rotation occurs on the time scale @xmath58 and the drop deforms on the much longer timescale , we find that the drop deformation @xmath59 is small for @xmath49 , consistent with the results shown in figure [ fig:1 ] . relation indicates that the deformation of the rotationally stabilized drop increases with the decreasing parameter @xmath17 . when @xmath17 falls below the lower critical value @xmath60 , the hydrodynamic torque associated with the straining component of the flow acting on a slightly elongated drop becomes strong enough to reorient the drop along the straining axis and arrest further drop rotation . deformation thus starts to accumulate , the drop is stretched , and a transition to the interfacial - tension stabilized elongated state takes place . as shown in figure [ fig:1 ] , a drop in the compact , rotationally stabilized stationary state is nearly aligned with the symmetry axis of the applied flow @xmath26 . this behavior stems from the flow - reflection symmetry of stokes equations and the fact that the drop is stabilized by rotation rather than the capillary forces . in the absence of the capillary forces ( or in the limit of infinitely strong flow ) the symmetry of stokes equations implies that the stationary drop shape is invariant with respect to flow reflection . hence , the shape is also invariant with respect to the corresponding transformation @xmath61 of the spatial coordinates ( and this symmetry corresponds to drop alignment in the @xmath26 direction ) . a perturbation due to the presence of the capillary stresses produces only a small asymmetry because the effect of capillary forces is insignificant for a nearly spherical drop . with vorticity parameter @xmath17 for @xmath39 , 100 , 50 , and different values of capillary number ( as labeled ) . ( results from boundary - integral simulations . ) , title="fig:",width=307 ] + with vorticity parameter @xmath17 for @xmath39 , 100 , 50 , and different values of capillary number ( as labeled ) . ( results from boundary - integral simulations . ) , title="fig:",width=240 ] with vorticity parameter @xmath17 for @xmath39 , 100 , 50 , and different values of capillary number ( as labeled ) . ( results from boundary - integral simulations . ) , title="fig:",width=240 ] the quasistatic response of the drop length to the variation of the vorticity is illustrated in figure [ drop response for different ca and lambda ] for different values of the capillary number @xmath62 and viscosity ratio @xmath29 . the results indicate that the size of the hysteresis loop is the largest for large values of @xmath29 and @xmath62 . when the capillary number is decreased , the upper critical vorticity parameter @xmath56 ( corresponding to the transition from the elongated to the compact drop ) decreases but the lower critical parameter @xmath63 remains nearly unaffected . in contrast , the viscosity ratio @xmath29 affects primarily the position of the lower critical parameter @xmath60 ( corresponding to the transition from the compact to the elongated drop ) . this behavior is consistent with the scaling relations and and the mechanism of drop bistability explained in [ capillary and rotational stabilizing mechanisms ] . according to these relations the deformation of an elongated drop scales with @xmath62 and is independent of @xmath29 , while the deformation of a compact drop is independent of @xmath62 and scales with @xmath64 . we recall that the critical states corresponding to the transitions between the elongated and compact drop shapes correspond to the points where the maximal torque @xmath65 exerted by the extensional component of the flow on the drop just before the transition marginally balances the torque @xmath66 exerted by the vorticity flow component . since @xmath67 whereas @xmath66 is approximately independent of @xmath68 , we conclude that @xmath60 varies with @xmath64 and @xmath56 with the @xmath62 . this conclusion is consistent with our simulation results . the plots shown in figure [ drop response for different ca and lambda ] indicate that for a given viscosity ratio , there exists a critical value of the capillary number @xmath69 below which the drop response to the changes of the flow vorticity does not exhibit a hysteretic loop . the bifurcation point occurs at the critical value of the vorticity parameter @xmath70 that corresponds to the position of the infinitesimal hysteresis loop for @xmath62 slightly above the critical value @xmath69 . it can be shown from the results derived by @xcite that @xmath71 for @xmath49 . the plots shown in figure [ drop response for different ca and lambda ] are consistent with the above expressions . we note that for @xmath72 ( see figure [ drop response for different ca and lambda ] ) the hysteretic drop behavior is not observed . the reason is that in this viscosity range @xmath69 exceeds the critical value of the capillary parameter for drop breakup . hence , the elongated stationary drop shape does not exist for @xmath73 . crucial features of the evolution of a highly viscous drop in two - dimensional flows with nonzero vorticity are captured by small - deformation equations . in our approach @xcite , the position @xmath74 of the drop interface is expanded in spherical harmonics @xmath75,\ ] ] where @xmath76 and @xmath77 denote the spherical - harmonic order , @xmath78 and @xmath79 are the expansion coefficients , and the parameter @xmath80 is given by the drop - volume constraint . since for @xmath81 all spherical harmonics are real , we set @xmath82 in expansion . moreover , flow - induced drop deformation preserves the symmetry of the incident flow . therefore only even values of @xmath83 and @xmath84 need to be included in the analysis . evolution equation for the expansion coefficients @xmath78 and @xmath79 are obtained by inserting the series into the boundary - value problem , performing a boundary - perturbation analysis , and reexpanding resulting products of spherical harmonics using appropriate clebsch gordan coupling coefficients . the detailed analysis and explicit expressions for the evolution equations at different truncation levels are presented elsewhere @xcite . for simplicity , our small - deformation calculations are performed with the expansion truncated at the lowest spherical - harmonic order @xmath85 , which leaves us with three independent drop - shape components : @xmath86 , and @xmath87 . noting that @xmath88 we find that the shape parameters @xmath89 and @xmath90 correspond to the drop deformation along the symmetry axis @xmath26 and the straining axis @xmath23 , respectively . the parameter @xmath87 described an axisymmetric deformation along the axis @xmath91 . the evolution equations for the shape parameters @xmath86 , and @xmath87 , truncated at the second - order in the drop deformation , can be represented in the following form [ small - deformation equations ] @xmath92,\\ \label{equation f'22 } \dot{f'_{22 } } & = & -2\omega f''_{22}+\lambda^{-1}\left[d_{21}f'_{22}f''_{22 } -{{\rm ca}}^{-1}(d_1 + 2d_2f'_{20})f'_{22}\right],\\ \label{equation f''22 } \dot{f''_{22 } } & = & 2\omega f'_{22}+\lambda^{-1 } \left[(d_{31}+d_{32}f'_{20 } + d_{33}f^{\prime\hspace{0.5pt}2}_{20 } + d_{34}f^{\prime\hspace{0.5pt}2}_{22 } + d_{35}f^{\prime\hspace{0.5pt}2}_{22})- { { \rm ca}}^{-1}(d_1 + 2d_2f'_{20})f''_{22}\right],\nonumber\\\end{aligned}\ ] ] where the dot denotes the time derivative ( normalized by @xmath93 ) . the terms involving the coefficients @xmath94 correspond to drop deformation by the external flow , and the terms involving @xmath95 describe the capillary relaxation . all these terms are @xmath55 in the large - viscosity - ratio regime . explicit expressions for the coefficients @xmath94 and @xmath95 are given in @xcite ; here we only note that these coefficients are functions of the viscosity ratio @xmath29 and have finite limits for @xmath52 . the two remaining terms on the right - hand side of equations and ( the terms proportional to @xmath96 ) are viscosity independent . these terms represent the rigid - body rotation of the drop , with the angular velocity @xmath97 where @xmath98 . consistent with our qualitative physical picture described in [ capillary and rotational stabilizing mechanisms ] ( and illustrated in figure [ mechanism ] ) , the rotational velocity involves two terms . the first term corresponds to the rotation of the drop by the vorticity component of the flow . the second term , which is proportional to the shape parameter @xmath89 that described deformation in the @xmath26 direction , corresponds to the rotation of a deformed drop by the straining component of the external flow towards the straining axis @xmath23 . it has been shown by @xcite that for @xmath49 the drop behavior near the bifurcation point can be described by simplified asymptotic equations [ simplified evolution equation ] @xmath99 where @xmath100 and @xmath101 are the high - viscosity limits of @xmath102 and @xmath103 . the asymptotic result is obtained form on assumption that near the bifurcation point there is a balance between drop deformation and rotation ( which corresponds to @xmath104 ) and the balance between capillary relaxation and rotation ( which yields @xmath105 and @xmath106 ) . moreover , the two contributions to the angular velocity are of the same order but do not cancel ( i.e. , @xmath107 ) . equations are rescaled accordingly , and only the leading order terms are retained . equations have all necessary ingredients that are needed to describe the hysteretic drop behavior . we have terms representing drop rotation by the straining and vorticity components of the external flow ( i.e. , the terms proportional to @xmath96 ) , drop relaxation due to the presence of the capillary forces ( the terms proportional to @xmath108 ) , and stretching of the drop along the straining axis of the straining component of the external flow ( the last term in equation ) . neglecting any of these terms would qualitatively alter the solution structure , which no longer would manifest the key features of the drop evolution in the parameter range considered herein . in the regime @xmath109 and @xmath110 the stationary solutions of equations can be obtained by a singular - perturbation analysis . to the leading order in the small parameters , we find that the elongated drop is described by the relations [ equations for elongated drop ] @xmath111 @xmath112 the first of the above relations correspond to the fact that an elongated drop does not rotate ( there is only a weak fluid circulation inside it , as predicted by our qualitative analysis ) . the second relation describes the balance between drop deformation by the external flow and relaxation due to the capillary forces . recalling the definition of the angular velocity @xmath96 we find that relations yield [ solution for elongated drop ] @xmath113 @xmath114 by inserting the above relations back into one can verify that they constitute a consistent leading - order asymptotic stationary solution . according to equation drop elongation along the straining axis @xmath23 ( i.e. , @xmath43 ) scales with the capillary number , which is consistent with the scaling result . the drop angle @xmath115 only slightly deviates from @xmath43 because @xmath116 ( assuming that @xmath117 and @xmath110 ) . the leading - order stationary solution corresponding to the compact drop is obtained from the following relations [ equations for compact drop ] @xmath118 @xmath119 which are obtained by dropping from the evolution equations the @xmath55 capillary - relaxation terms . taking into account the definition of @xmath96 we thus obtain [ solution for compact drop ] @xmath120 and @xmath121 ( the solution with the plus sign in front of the square root is unstable ) . since the shape parameter @xmath90 vanishes according to equation , the drop is oriented in the @xmath26 direction . for @xmath122 we find @xmath123 which is consistent with our scaling estimate . for @xmath124 , where @xmath125 the solution does not exist ; the drop thus undergoes a transition to the long solution . the critical vorticity parameter decreases with increasing @xmath29 , in agreement with our numerical results presented in figure [ drop response for different ca and lambda ] . the solution of the simplified small - deformation equations is perturbative . we note , however , the exact stationary solution can also be found @xcite . our analytical solution quantitatively agrees with the results of numerical simulations , provided that the drop deformation is not too large . with vorticity parameter @xmath17 for @xmath39 and different values of capillary number ( as labeled ) . solution of small - deformation equations ( [ small - deformation equations ] ) ; comparison of small - deformation results ( dashed lines ) with boundary integral simulations ( solid lines ) . , title="fig:",width=240,height=259 ] with vorticity parameter @xmath17 for @xmath39 and different values of capillary number ( as labeled ) . solution of small - deformation equations ( [ small - deformation equations ] ) ; comparison of small - deformation results ( dashed lines ) with boundary integral simulations ( solid lines ) . , title="fig:",width=240,height=259 ] predictions of the small - deformation equations for drop behavior in a two - dimensional linear flow with slowly varying vorticity are depicted in figure [ small deformation results for lambda=200 ] for a system with the viscosity ratio @xmath39 . figure [ small deformation results for lambda=200 ] shows the dependence of the drop length @xmath126 on the vorticity parameter @xmath17 for the same set of capillary numbers as those represented in figure [ drop response for different ca and lambda ] . figure [ small deformation results for lambda=200 ] compares the small deformation results directly with the results of the boundary - integral simulations . the small - deformation calculations were performed using the second - order equations because for an elongated drop they are more accurate than the simplified equations . the results shown in figures [ drop response for different ca and lambda ] and [ small deformation results for lambda=200 ] indicate that for small and moderate capillary numbers the small - deformation theory yields accurate quantitative predictions . at high values of @xmath62 drop behavior is also captured quantitatively , except for the upper portion of the hysteresis loop ( i.e. when the drop is in the elongated state ) . for all values of the capillary number the lower and upper critical vorticity parameters @xmath63 and @xmath127 are obtained within the numerical error of the boundary - integral simulations . our additional calculations ( not shown ) indicate that a similar accuracy is obtained for @xmath72 and @xmath128 . @xmath129 { figs / fig6a.eps } } & { \includegraphics[width=2.5in , height=4.8in , trim=20 0 0 0 ] { figs / fig6b.eps } } \end{array } $ ] ( solid lines ) , drop deformation @xmath68 ( dashed ) , and drop angle @xmath36 ( dash - dotted ) are shown versus time @xmath130 normalized by the oscillation period @xmath131 . viscosity ratio @xmath132 , capillary number @xmath40 , period @xmath133 , mean vorticity @xmath134 , and vorticity oscillation amplitude @xmath135 and 0.8 . ( results from small - deformation theory . ) , title="fig:",width=432 ] ( solid lines ) , drop deformation @xmath68 ( dashed ) , and drop angle @xmath36 ( dash - dotted ) are shown versus time @xmath130 normalized by the oscillation period @xmath131 . viscosity ratio @xmath132 , capillary number @xmath40 , period @xmath133 , mean vorticity @xmath134 , and vorticity oscillation amplitude @xmath135 and 0.8 . ( results from small - deformation theory . ) , title="fig:",width=432 ] dynamical systems with multiple equilibrium states often exhibit novel dynamics when driven by simple forcing @xcite . thus , despite the laminar nature of the stokes flow , we expect to find interesting nonlinear dynamics of a viscous drop in a time - varying linear flow with rotation . to explore this dynamics we will now investigate the drop response to harmonic variation of the vorticity @xmath136 where @xmath137 is the average vorticity value , @xmath138 is the oscillation amplitude , and @xmath131 is the oscillation period . we have performed a series of small deformation calculations ( [ subsec : chaos_sdt ] ) and boundary - integral simulations ( [ subsec : chaos_bis ] ) for different values of the flow parameters @xmath137 , @xmath138 , and @xmath131 . if the oscillation period @xmath131 is much shorter than the drop deformation and oscillation times and , we find that the drop undergoes small oscillations about a stationary shape corresponding to the mean value of @xmath17 ( which is an expected behavior ) . in the opposite regime @xmath139 , the quasistatic drop behavior described in [ hysteretic behavior ] is recovered . in what follows we focus on the most interesting parameter domain @xmath140 and @xmath141 , in which an interaction of different timescales as well as an interplay between the short and elongated drop shapes is anticipated . and period @xmath133 . viscosity ratio and capillary number are the same as in figures [ fig : chaos01 ] and [ example of period doubling ] . ( results from small - deformation theory . ) , title="fig:",width=336 ] figure [ fig : chaos01 ] illustrates the dependence of the drop evolution in linear flow with the oscillatory vorticity on the oscillation period @xmath131 . the viscosity ratio in this example is @xmath132 , and the capillary number is @xmath40 . the mean value of the vorticity @xmath134 is close to the lower critical value @xmath142 , and the oscillation amplitude is @xmath143 . figure [ fig : chaos01 ] represent our results for the shortest oscillation period of the flow vorticity @xmath144 . the drop oscillates about the compact stationary shape in this case . both the drop lengths and the drop angle vary periodically , with the period @xmath145 equal to the period @xmath131 of the external forcing . the drop length decreases when the drop is in the compressional quadrant @xmath146 and increases for @xmath147 . upon an increase of the period of the external forcing the amplitude of the angular drop oscillations increases . when the oscillation amplitude reaches @xmath148 the drop starts to tumble , as illustrated in figure [ fig : chaos01 ] . @xmath129 { figs / fig9a.eps } } & { \includegraphics[width=2.5in , height=4.8 in ] { figs / fig9b.eps } } \end{array } $ ] a further increase of the period of the external forcing results in a qualitative change of drop response . we find that the drop still undergoes a tumbling motion ; however , the evolution is not periodic but it becomes _ chaotic _ , as shown in figure [ fig : chaos01 ] . the chaotic motion continues up to @xmath133 , and then the drop reverts to periodic motion . for @xmath149 ( figure [ fig : chaos01 ] ) the drop oscillates about the elongated stationary shape , and in the regime @xmath150 ( figure [ fig : chaos01 ] ) the system approaches the quasistatic behavior discussed in [ hysteretic behavior ] . occurs right after the drop jumps from the long to the compact shape when @xmath17 increases above the upper critical value @xmath56 . namely , before the drop settles down to the compact stationary shape it undergoes a tumbling motion with the amplitude decaying on the timescale . ] the transition to the chaotic drop motion occurs through a cascade of period doubling events , as illustrated in figures [ example of period doubling ] and [ fig : period_doubling ] . figure [ example of period doubling ] depicts the drop evolution at three values of the amplitude of vorticity oscillations @xmath138 ( the remaining system parameters are the same as those that yield the chaotic motion depicted in figure [ fig : chaos01 ] ) . the results shown in figure [ example of period doubling ] indicate that for a sufficiently small oscillation amplitude ( @xmath151 in our example ) the drop evolves with the period @xmath152 ( i.e. , the period equal to that of the external forcing ) . at a larger amplitude @xmath153 the drop oscillation period is @xmath154 and for @xmath155 we find @xmath156 . the period - doubling scenario of the transition to chaos in our system is further supported by the bifurcation diagram shown in figure [ fig : period_doubling ] , where the drop length @xmath35 at times @xmath157 ( @xmath158 ) is plotted versus the flow - oscillation amplitude @xmath138 . an analysis of the results shown in figure [ example of period doubling ] indicates that the period doubling occurs as a result of a resonance between the drop tumbling motion and the vorticity oscillations . namely , if the drop is relatively long and approximately aligned with the straining axis of the external flow when the vorticity parameter @xmath159 reaches a minimum , the drop rotation may be significantly slowed down or even arrested ( as in the long - drop stationary state discussed in [ hysteretic behavior ] ) . such a temporary arrest of drop rotation corresponds to the shoulders in the plots of the angular evolution depicted in figure [ example of period doubling ] . on the other hand , if the drop angle exceeds @xmath160 when the vorticity goes through a minimum , the arrest of drop rotation does not occur . as seen in figure [ example of period doubling ] this interplay of drop tumbling with oscillations of the external forcing produces the period - doubling bifurcation that leads to alternating accelerated and retarded drop - rotation cycles . in our numerical calculations depicted in figures [ fig : chaos01]-[fig : period_doubling ] we have used the full set of the small - deformation equations but we find that the simplified asymptotic equations yield similar results . in particular these equations correctly reproduce the cascade of the period doubling bifurcations and the chaotic - evolution domain . however , a further simplification of the evolution equations is not possible : if any of the terms in equations is removed the solutions qualitatively change and the chaotic domain disappears . we also find that chaotic drop dynamics occurs only for highly viscous drops with @xmath161 . , @xmath162 , @xmath132 , and @xmath40 . ( results from boundary - integral simulations . ) , title="fig:",width=336 ] figure [ fig : chaos03 ] shows examples of our simulation results for the periodic and chaotic drop evolution . the flow parameters are similar to those used in our small - deformation calculations described in [ subsec : chaos_sdt ] . as with the small - deformation calculations , for short periods of the external forcing @xmath131 the drop oscillates around the compact stationary shape , for moderate periods the system undergoes a transition to chaotic evolution , and for long periods the drop motion approaches the quasistatic behavior . consistent with the small - deformation results , the chaotic drop dynamics revealed by the boundary - integral simulations is due to a cascade of period doubling bifurcations . a bifurcation diagram illustrating this behavior is presented in figure [ fig : period_doubling_dns ] . the domain of chaotic motion found in the direct boundary - integral simulations somewhat differs from the corresponding domain obtained from the small - deformation theory . also , the magnitude of the chaotic fluctuations in the drop length is larger in the boundary - integral runs . we have tested the convergence of the boundary - integral simulations , and we believe that the differences between the drop behavior obtained by the two different methods stem from the approximations involved in the small - deformation theory . we note , however , that the evolution in the chaotic and period - doubling regimes is very sensitive to the to small changes of the system parameters ( hence , also sensitive to the approximations involved in our calculations ) . and @xmath40 , in linear flow with stochastic vorticity . mean value of the vorticity parameter @xmath163 , variance @xmath164 , and correlation time @xmath165 . inset shows vorticity probability distribution . ( results from boundary - integral simulations.),title="fig:",width=336,height=336 ] and @xmath40 , in linear flow with stochastic vorticity . mean and variance of the vorticity parameter are the smae as in figure [ fig : cgaussdns ] ; the correlation times are @xmath166 for lines marked 15 , respectively . ( results from small - deformation theory . ) , title="fig:",width=249,height=211 ] and @xmath40 , in linear flow with stochastic vorticity . mean and variance of the vorticity parameter are the smae as in figure [ fig : cgaussdns ] ; the correlation times are @xmath166 for lines marked 15 , respectively . ( results from small - deformation theory . ) , title="fig:",width=249,height=211 ] in some systems ( e.g. emulsion flows through a packed bed of fibers @xcite or turbulent emulsion flows with drops that are much smaller then the kolmogorov scale @xcite ) a viscous drop undergoes deformation in a random external creeping flow . to gain some understanding of the role that drop bistability may play in such systems we consider the drop behavior in a flow with stochastic vorticity . we assume that the time variation of the vorticity parameter @xmath17 is described by a stationary markovian gaussian process ( i.e , the ornstein uhlenbeck process ) with the mean @xmath167 , variance @xmath168 , and correlation time @xmath169 . a standard numerical scheme @xcite for generating such a time - correlated gaussian process is used to model the time variation of @xmath17 for a given drop trajectory . an example of drop behavior in a stochastic flow with a gaussian variation of the vorticity is presented in figure [ fig : cgaussdns ] . the mean value of the vorticity is @xmath163 and the variance is @xmath164 . the correlation time of the vorticity distribution @xmath170 is several times shorter than the drop relaxation time . the figure depicts the probability distribution of the drop length , obtained using the boundary - integral simulations . the capillary number is @xmath40 , and the drop viscosity is @xmath39 . the results indicate that the drop - length distribution is bimodal for the above parameter values . this behavior is expected since the vorticity undergoes random variation in the domain that includes both the lower and upper critical values @xmath60 , and @xmath56 of the vorticity parameter @xmath17 . since drop response to a flow with slow variation of vorticity is hysteretic , a drop in the random flow tends to stay in the neighborhood of the compact and the elongated stationary states . we note that the peak of the length probability distribution at @xmath171 is shifted towards the shorter drop lengths compared to the length of a drop in the elongated stationary shape . this is because @xmath40 is close to the critical capillary number for drop breakup . thus , due to the slow time scale in the drop dynamics near the critical capillary number @xcite , the drop does not have sufficient time to fully extend before the vorticity significantly changes . figure [ fig : cgauss01 ] shows the probability density distribution for the drop length @xmath35 and drop angle @xmath36 for different values of the flow correlation time @xmath169 . other system parameters are the same as in figure [ fig : cgaussdns ] . the calculations were performed using the small - deformation equations . the results indicate that at short flow correlation times the drop - length probability distribution is peaked around small values corresponding to the short - drop stationary solution , and the has a moderate - hight peak at @xmath172 . as the flow correlation time @xmath169 increases , the length probability distribution becomes bimodal : one of its peaks corresponds to compact and the other to elongated drops . a corresponding change occurs in the angle distribution , i.e. , its peak becomes more pronounced , and shifts towards the straining axis @xmath43 . the shift of the typical drop length and orientation from the compact to elongated state when the flow correlation time is increased resembles the analogous shift for a system with harmonic vorticity oscillations ( see figures [ fig : chaos01 ] and figures [ fig : chaos01 ] ) . this behavior is further illustrated in figure [ fig : cgauss02 ] which shows the average values and the variance of @xmath35 and @xmath36 versus the correlation time @xmath169 . normalized by drop deformation time , for @xmath39 , @xmath40 , @xmath163 , and @xmath164 . , title="fig:",width=249,height=211 ] normalized by drop deformation time , for @xmath39 , @xmath40 , @xmath163 , and @xmath164 . , title="fig:",width=249,height=211 ] we have presented results of numerical and theoretical investigations of the dynamics of highly viscous drops in two - dimensional linear creeping flows with time - dependent vorticity . in our earlier publication @xcite we predicted that in stationary flows such drops exhibit bistable behavior : there is a range of system parameters where the drop may assume either an elongated shape approximately aligned with the straining axis of the external flow or a nearly spherical shape approximately aligned in the flow direction . here we analyze the consequences of this behavior for the system dynamics . we also elucidate the physical mechanism that leads to drop bistability . a direct consequence of the existence of two stationary states is hysteretic drop response to a flow with slowly varying vorticity . we have explained that the rapid transition from an elongated non - rotating drop shape to the nearly spherical compact shape occurs when the vorticity becomes strong enough to overcome the effect of the straining flow component that aligns the drop with the straining axis . this transition is thus analogous to the behavior of an elongated rigid particle which starts to tumble when the vorticity grows above a critical value @xcite . a viscous drop also begins to tumble at a critical vorticity magnitude @xmath56 . however , when the drop becomes misaligned with the extensional axis it relaxes under the action of capillary forces towards a nearly spherical rotationally stabilized stationary shape . if , in turn , the vorticity is slowly decreased , the drop returns to the elongated shape only after the vorticity magnitude reaches a lower critical value @xmath173 . this hysteretic drop response occurs at high drop viscosities because the rotational stabilizing mechanism is more efficient in the high - viscosity regime . a highly viscous drop deforms less within each drop revolution , so the compact shape remains stable even for small vorticity magnitudes . the existence of two stationary states affects drop dynamics not only in the quasistatic regime but also at finite frequencies of the external forcing . at small amplitudes of harmonic vorticity oscillations the drop simply oscillates ( with the same frequency as the external forcing ) about one of the stationary states . however , if the vorticity - variation range includes both critical values @xmath60 and @xmath56 the dynamics of the system is much richer . we find that with an increasing magnitude of the vorticity oscillations the system undergoes a cascade of period - doubling bifurcations resulting in chaotic drop dynamics . the period doubling stems from the resonance between the periodicity of the external forcing and the tumbling motion of the drop when it jumps from a ( partially ) elongated shape towards the compact rotationally stabilized state . chaos in our system emerges despite linearity of stokes equations the system dynamics is nonlinear because of the coupling of the flow to the evolving fluid interface . a detailed analysis of small - deformation equations describing drop dynamics reveals that in addition to chaos associated with the period - doubling mechanism there also exists in our system a different kind of chaotic evolution that results from manifold tangling @xcite . our analysis of different types of chaos will be presented in a separate publication . to our knowledge , chaotic drop dynamics in stokes - flow regime has never been observed before . we note , however , that in a recent independent study @xcite has reported chaotic dynamics in co - annular stokes flow with insoluble surfactant adsorbed on the fluid interface . chaos in their system also appears as a result of period doubling . understanding of drop bistability and the associated dynamical phenomena is relevant for many practical problems . for example , interpretation of rheological response of emulsions of highly viscous drops to time - varying flows requires insight into drop dynamics . our results may also be useful in design new methods for manipulating emulsion microstructure in material processing and controlling drop behavior in microfluidic flows . drop bistability could , e.g. , be used to construct microfluidic switches , and chaotic drop dynamics may be relevant for microfluidic mixing . in particular , our analysis suggests that macromolecules with high degree of internal dissipation may undergo a transition between a nearly spherical and moderately elongated states ( in addition to the standard coil - stretch transition predicted by * ? ? ? * ) . in fact , the dynamics of macromolecules can be modeled using equations analogous to , supplemented with terms representing random thermal forces . such a simplified description correctly captures the most important features of power spectra of dna molecules evolving in linear flows with nonzero rotational component @xcite . there are also close analogies between drop and vesicle motion . the main difference between these two systems is that vesicles satisfy a constant - area constraint whereas the drop area can vary . this constraint gives rise to periodic vesicle motion ( such as tank treading and tumbling ) even in stationary flows @xcite . it would be interesting to determine if a coupling of vesicle oscillations to a harmonic variation of the external flow can lead to chaotic dynamics . it would also be of significant interest to experimentally explore the bistable and chaotic drop dynamics ( as well as related phenomena for other deformable particles ) . in such experiments a four - roll mill could be used to produce a linear flow with a controlled magnitude of vorticity . the experiments could also be performed using recently developed microfluidic analogues of a four - roll mill device @xcite . we would like to acknowledge helpful discussions with paul steen , demetrios papageorgiou , and michael loewenberg . we thank michael loewenberg for permission to use his improved boundary - integral code , and also petia vlahovska for use of her mathematica codes for the coefficients in the small - deformation theory . jb was supported by nsf career grant cts-0348175 , yny acknowledges a nsf / dms grant ( dms-0708977 ) and a sbr grant from njit . the simulations were conducted on the njit computer cluster supported by nsf / mri grant dms-0420590 . 1983 nonlinear oscillations , dynamical systems , and bifurcations of vector fields . in _ nonlinear oscillations , dynamical systems , and bifurcations of vector fields _ f. john , j. e. marsden & l. sirovich ) . new york : springer - verlag .

we begin by recalling two closely related operads . the first is the _ little 2-discs _ operad of boardman and vogt , denoted @xmath0 . the arity @xmath1 space , @xmath2 , is the space of embeddings of the disjoint union of @xmath1 discs into a standard disc , where each disc is embedded by a map which is a translation composed with a dilation . at the level of spaces , group complete algebras over this operad are 2-fold loop spaces , and at the level of homology an algebra over @xmath3 is precisely a gerstenhaber algebra . the second operad is a variant of the little 2-discs ; it is the _ framed little 2-discs _ operad , denoted @xmath4 , first introduced by getzler @xcite . here each little disc is allowed to be embedded by a composition of a _ rotation _ in addition to a dilation and translation , so in particular , @xmath5 . getzler observed that algebras over the homology operad @xmath6 are precisely batalin - vilkovisky algebras , and at the space level the second author and wahl @xcite proved that a group complete algebra over @xmath4 is a 2-fold loop space on a based space with a circle action . tamarkin @xcite first proved that the operad @xmath0 is _ formal _ over @xmath7 , meaning that the singular chain operad @xmath8 is quasi - isomorphic to the homology operad @xmath9 . his proof is algebraic in nature , using braid groups and drinfeld associators . kontsevich @xcite sketched a different proof of this formality theorem over @xmath10 which generalizes to show formality of the little @xmath11-discs for all @xmath12 . roughly speaking , kontsevich works dually , constructing a cooperad of cochain complexes of graphs which maps to both the cochains and the cohomology of @xmath0 by quasi - isomorphisms . the full details of kontsevich s proof have been explained nicely by lambrechts and volic @xcite . the kontsevich proof method has the advantage over tamarkin s proof of simultaneously proving the formality in the sullivan dga sense of the individual spaces of the operad . recently severa and willwacher @xcite have shown that the quasi - isomorphisms of kontsevich and tamarkin are homotopic for a particular choice of drinfeld associator . tamarkin s proof of formality for @xmath0 has been adapted to the operad @xmath4 by severa @xcite , and kontsevich s proof has been adapted to @xmath4 by the present authors in @xcite . both adaptations are relatively straightforward . operad formality theorems have significant applications . the formality of @xmath0 plays an important role in tamarkin s proof @xcite of kontsevich s celebrated deformation quantization theorem , and the formality of @xmath13 for @xmath14 is a primary ingredient in the computation of the homology of spaces of long knots in high dimensions @xcite . formality of the operad @xmath4 is used in @xcite to construct chain level homotopy bv - algebra structures for topological conformal field theories and for the 2-fold loop space on a space with a circle action , as well as to give a particular solution to the cyclic deligne conjecture . unlike @xmath13 , the operad @xmath15 has the additional structure of being homotopy equivalent to a _ cyclic operad _ ( see @xcite for an explicit cyclic model ) ; being a cyclic operad means roughly that the roles of inputs and outputs can be interchanged . the present paper was motivated by the question of whether the formality of the operad can be made compatible with the cyclic structure . we answer this question in the affirmative when @xmath16 . in section [ cyclic - model ] we will recall a cyclic model @xmath17 for @xmath4 constructed in terms of compactified moduli spaces of genus zero curves with marked points . the elements of @xmath17 are genus zero stable curves decorated with real tangent rays at the marked points and nodes . the operad composition is obtained by gluing stable curves at marked points and tensoring the corresponding rays . this model was first introduced by kimura - stasheff - voronov @xcite , and it is closely related to the modular operad of deligne - mumford compactified moduli spaces of curves whose formality is proved in @xcite . moreover @xmath17 is homotopy equivalent , as a cyclic operad , to the operad of genus zero riemann surfaces with boundary , where the operad composition is defined by gluing along boundary components . the main result of this paper is : [ main ] the cyclic operad @xmath18 of singular chains on @xmath17 with real coefficients is quasi - isomorphic as a cyclic operad to its homology , @xmath19 , which is the batalin - vilkovisky operad . the new and important point of this theorem is that the chain of quasi - isomorphisms is compatible with the _ cyclic _ operad structure . a second point , coming out of the proof , is that this operad formality simultaneously realises the formality of the individual spaces of the operad in the sullivan commutative dga sense , as was the case for the kontsevich proof in the unframed case . this cyclic formality has a consequence in computing the cohomology of diffeomorphism groups of 3-dimensional handlebodies , as discussed further below . it is not immediately clear that the previous operad formality proofs given in @xcite and @xcite are compatible with the cyclic structure , nor is it clear that they are compatible with the formality of the individual spaces of the operad . however , since the completion of this work severa has indicated how to show that his formality proof is in fact compatible with the cyclic structure . as for the other proof , the kontsevich graph complex , as modified for the framed 2-discs operad in @xcite , in arity @xmath1 has an action of the enlarged symmetric group @xmath20 ( this action would be part of the cyclic operad structure ) by linear automorphisms , but it is not clear if this action respects the differential . it seems likely that the action is in fact compatible , but proving this appears to be a very difficult combinatorial problem . instead , here we take a more conceptual approach and build a new manifestly cyclic operad of `` projective graph complexes '' as a replacement for kontsevich s graph complexes . the differential is now a combination of both edge deletions and edge contractions . the projective graph complexes do form a cyclic cooperad . our proof of theorem [ main ] then follows in outline the method used by kontsevich in the unframed case . the projective graph complexes map onto the cohomology of @xmath17 , and they also map to the ( semi - algebraic ) forms on @xmath17 by a variation of kontsevich s map defined by certain configuration space integrals . we prove that both of these maps are quasi - isomorphisms and are compatible with the cyclic cooperad structures . the theorem then follows by dualising . for @xmath21 the question of formality of @xmath15 as a cyclic operad , or even just as an ordinary operad , remains open . in future work we hope to address this . in the forthcoming paper @xcite , the first author proves that the modular operad generated by the cyclic operad @xmath17 is homotopy equivalent to the modular operad made from @xmath22s of handlebodies . this leads to a bousfield - kan spectral sequence computing the cohomology of @xmath23 for @xmath24 a handlebody of genus @xmath25 . theorem [ main ] implies that this spectral sequence degenerates at the @xmath26 page . ( note that formality as an operad rather than a cyclic operad is not enough to imply degeneration of the spectral sequence . ) the framed little 2-discs operad @xmath4 is homotopy equivalent to a cyclic operad , ( although it is not itself cyclic on the nose ) . a convenient cyclic model was introduced in @xcite . we recall that model and discuss some of its properties . recall that a _ cyclic operad _ is an extension of the concept of operad in which inputs and outputs are on equal footing ; this can be formalised by asking that the @xmath27 action on the arity @xmath1 space extends to an action of @xmath20 in a way compatible with the operad composition maps . the notion of a cyclic operad was first introduced by getzler and kapranov @xcite . in more detail , a cyclic operad in a symmetric monoidal category @xmath28 is defined by the following data : a functor @xmath29 from the category of non - empty finite sets and bijections to @xmath28 , and for each pair of finite sets @xmath30 with elements @xmath31 a composition morphism , @xmath32 natural in @xmath33 and @xmath34 and satisfying the following axioms . * * ( associativity ) * given finite sets @xmath35 , and elements @xmath36 , the following diagram commutes : @xmath37 * * ( commutativity ) * given finite sets @xmath30 and elements @xmath38 and @xmath39 , the following diagram commutes : @xmath40 where @xmath41 is the canonical bijection @xmath42 . * * ( unit axiom ) * for each set @xmath43 of cardinality 2 there is a morphism @xmath44 , where @xmath45 is the monoidal unit , that is natural in @xmath46 and such that for any finite set @xmath33 and an element @xmath38 the following diagram commutes : @xmath47{p(a\sqcup i \smallsetminus \{a , i\ } ) , } \end{diagram}\ ] ] where @xmath48 is the canonical bijection sending @xmath49 to @xmath50 . dually one defines a _ cyclic cooperad _ by reversing the arrows in the definition . concretely , a cyclic cooperad is specified by the data of a functor @xmath29 from finite sets and bijections to @xmath28 , and a co - composition map @xmath51 ( where @xmath52 and @xmath53 are auxiliary points ) for each partition @xmath54 . these co - composition maps must satisfy dual versions of the naturality , symmetry , and associativity conditions . a cyclic ( co)operad in the category of chain complexes is called a _ differential graded cyclic ( co)operad_. we now recall the moduli space operad introduced by kimura , stasheff and voronov in @xcite . in their notation , the operad is denoted @xmath55 , but we prefer @xmath17 as being more consistent with other notations . this operad is manifestly a cyclic operad , and we will show in proposition [ operad - equivalence ] that its underlying operad is homotopy equivalent to the framed little 2-discs operad . given a non - empty finite set @xmath56 , let @xmath57 denote the moduli space of genus zero riemann surfaces with marked points labelled by @xmath56 . explicitly , @xmath58 where @xmath59 is the space of injective maps from @xmath56 to the complex projective line , and the automorphism group @xmath60 acts on the target in the standard way . we will write @xmath61 when @xmath56 is the set @xmath62 . there is an identification between @xmath61 and the complement in @xmath63 of the union of the hyperplanes of the form @xmath64 if @xmath65 then the moduli space @xmath61 is noncompact and has a deligne - mumford - knudsen compactification @xmath66 which can be constructed blowing up in a suitable order all proper intersections of hyperplanes of the form in @xmath63 @xcite . the points of @xmath67 correspond to stable nodal riemann surfaces , ( _ nodal _ means that the surfaces are allowed to have double - point singularities , and _ stable _ in this case means that no irreducible component is a sphere with strictly fewer than 3 points that are either nodes or marked points ) . for @xmath68 there is a complex line bundle @xmath69 over @xmath67 whose fibre over a given riemann surface @xmath70 is the tangent space at the @xmath71 marked point . the boundary , @xmath72 , is a normal crossing divisor . it has an irreducible component for each unordered partition @xmath73 into two subsets of cardinality at least 2 , and the corresponding component is canonically isomorphic to @xmath74 . the inclusions @xmath75 define a cyclic operad structure on the deligne - mumford - knudsen moduli spaces . the space @xmath76 is defined as the real oriented blowup of the boundary locus of @xmath77 . we pull the complex line bundles @xmath78 back from @xmath77 ( and denote them by the same symbol ) . the space @xmath76 can be interpreted as the moduli space of genus zero stable nodal riemann surfaces with points marked by @xmath56 and equipped with the additional data of a ray in the tensor product of the two tangent spaces at either side of each node . it can be constructed as the iterated _ real oriented _ blowup of @xmath63 along _ all _ intersections of hyperplanes , by blowing up first the proper intersections in the same order as for the deligne - mumford space , and finally the hyperplanes themselves . the space @xmath77 is a compact smooth manifold with _ faces _ ; i.e. it is a manifold with corners , stratified with the interior as the unique codimension 0 stratum and all of the positive codimension strata lying on the boundary , each of which is a transversal intersection of faces ( codimension 1 strata ) . faces correspond to unordered partitions of @xmath56 into two subsets @xmath79 of cardinality at least 2 ; given such a partition , the corresponding face is canonically isomorphic to the unit circle bundle of the exterior tensor product @xmath80 over @xmath81 let @xmath82 denote the ( coarse ) moduli space of genus zero smooth surfaces equipped with points marked by @xmath56 with real tangent rays . explicitly , @xmath83 where @xmath84 is the circle bundle associated with the tangent bundle of @xmath85 . observe that there are diffeomorphisms @xmath86 in particular , @xmath87 is compact for @xmath88 . when @xmath65 , @xmath87 is noncompact . forgetting the tangent rays at the marked points gives a trivial principal @xmath89-bundle @xmath90 which extends uniquely to a trivial torus bundle over @xmath91 . we call its total space @xmath92 . note that the trivialisation of the bundle is not canonical . more explicitly , @xmath93 one sees that @xmath94 is a compactification of @xmath87 obtained by adding boundary and corners , and hence it is homotopy equivalent to its interior . when @xmath88 we simply define @xmath95 . the functor @xmath96 forms a cyclic operad . composition with @xmath97 acts by rotating tangent rays , and composition with @xmath98 removes a marked point . all other composition maps are defined by gluing marked points together to form a node and tensoring their tangent rays to produce the decoration at the node ; in more detail , if @xmath99 and @xmath100 both have cardinality at least 2 then the composition map @xmath101 is a trivial circle bundle over the face associated with the partition @xmath102 and it is induced by the multiplication map @xmath103 . recall that a homotopy equivalence of topological operads is a morphism of operads ( a natural transformation commuting with the composition and unit maps ) that is level - wise an equivariant homotopy equivalence . [ operad - equivalence ] there is homotopy equivalences of operads @xmath104 there is also a homotopy equivalence of spaces , @xmath105 , where the torus acts by rotating all but the zeroth tangent direction . in @xcite the second author constructed a zigzag of homotopy equivalences between the operad @xmath0 and the _ fulton - macpherson operad _ , @xmath106 , introduced by getzler and jones in @xcite . the space @xmath107 is the real oriented fulton - macpherson compactification of the space of configurations of @xmath1 points in @xmath108 modulo dilation and translation . it can be constructed by iterated real oriented blowup from the @xmath109-sphere @xmath110 , where @xmath111 is the thin diagonal defined by @xmath112 and @xmath113 is the affine group generated by positive dilations and translations . the blowups are performed along all diagonals . the space @xmath107 is a manifold with faces ; a face corresponds to an unordered partition of @xmath114 into two subsets of cardinality at least 2 , and it is canonically diffeomorphic to @xmath115 . the inclusion of a face with @xmath116 represents a ( non - cyclic ) @xmath117 operad composition . from @xcite , @xmath0 is an operad in the category of spaces with @xmath118 action and so one can form the semidirect product operad @xmath119 which is isomorphic to @xmath4 . similarly , @xmath106 is an operad in @xmath118-spaces , and one can form the _ framed fulton - macpherson operad _ @xmath120 . the homotopy equivalence between @xmath0 and @xmath106 is realised by a zigzag of @xmath118-equivariant maps and thus induces a homotopy equivalence of operads , @xmath121 it turns out that @xmath122 and @xmath17 are isomorphic as operads . this is immediate in arity @xmath123 in higher arity , note that blowing up the sphere @xmath124 at the diagonal @xmath125 is the same as blowing up @xmath126 at the hypersurfaces @xmath127 . blowing up all other diagonals identifies @xmath128 , and thus we have diffeomorphisms @xmath129 compatible with the operad composition , by definition . for the proof of theorem [ main ] , we will need to work in the category of semi - algebraic spaces . the theory of semi - algebraic spaces and its corresponding de rham theory has been developed by hardt , lambrechts , turchin and volic in @xcite . the functor @xmath96 forms a cyclic operad in the category of semi - algebraic manifolds . sinha constructs in @xcite a @xmath27-equivariant semi - algebraic embedding @xmath130^{\binom{n}{3}}$ ] by evaluating the direction between each pair of points and ratio of distances for each triple of points . thus the symmetric group acts semi - algebraically . moreover , lambrechts and volic check in ( * ? ? ? * prop 5.5 ) that the operad composition @xmath131 is the restriction of a semi - algebraic map . passing to the semi - direct product with @xmath118 , it follows that @xmath122 is an operad in semi - algebraic manifolds . finally , we must verify that the action of @xmath27 on @xmath132 is semi - algebraic . for this it is convenient instead to @xmath27-equivariantly embed @xmath133 by evaluating angles between rays for pair of points and the cross ratio for 4-tuples of points . this shows that @xmath134 acts semi - algebraically on @xmath94 . forgetting subsets of the marked points defines a collection of important projections between the various moduli spaces . given a subset @xmath135 , let @xmath136 be the map induced by forgetting all marked points labelled by @xmath46 . note that @xmath137 is the iterated operad composition with the point of @xmath138 for each element of @xmath46 . the map @xmath139 , forgetting the points in @xmath46 , is a semi - algebraic fibre bundle . the projection @xmath140 is equivalent to the composition @xmath141 . the first map is a semi - algebraic bundle by ( * ? ? ? * appendix a ) , the second map is a trivial bundle , and the composite of bundle projections is a bundle projection by ( * ? ? ? the fibres of the bundle @xmath140 are manifolds with faces , so the fibrewise boundary is stratified into pieces of codimension 1 ( in the total space ) and greater . we shall refer to the codimension 1 strata as the _ fibrewise boundary components_. [ fibrewise - boundary ] for @xmath142 , the fibrewise boundary components of the bundle @xmath143 correspond bijectively with the set of unordered partitions of @xmath56 into two subsets @xmath144 of cardinality at least 2 such that @xmath145 . corresponding to a partition @xmath146 is the closure of the locus in @xmath147 in which a node separates the sphere into two lobes , one containing the points of @xmath148 and the other containing the points of @xmath149 . this closed stratum fibres over @xmath150 if and only if one of the @xmath151 contains at most one point of @xmath152 ( a priori it is an unordered partition , but we can always choose to call the set with this property @xmath148 ) . given a partition @xmath153 as above , we will write @xmath154 for the corresponding fibrewise boundary component of @xmath137 . fibrewise orientations of the canonical projections play an important role in the formality proof in this paper . here we establish some general definitions and conventions for working with orientations . given a finite dimensional real vector space @xmath155 , let @xmath156 denote the top exterior power of @xmath155 ; a choice of a ray in this line is equivalent to a choice of an orientation of @xmath155 . if @xmath70 is a finite set then we write @xmath157 ( and @xmath158 when @xmath70 is empty ) . if @xmath70 has cardinality at least 2 then choosing a ray in @xmath159 is equivalent to choosing an ordering of @xmath70 up to even permutation . an orientation form on an @xmath1-dimensional manifold @xmath160 is a nowhere vanishing @xmath1-form @xmath161 . clearly an orientation form determines an orientation of @xmath160 , and two orientation forms determine the same orientation if and only if they agree up to multiplication by a positive scalar function . given a fibre bundle @xmath162 , a fibrewise orientation form is a form on @xmath163 whose restriction to each fibre is an orientation form . we will use the following general orientation conventions . * given @xmath164 , we define an isomorphism @xmath165 by the rule @xmath166 ( inverse to the map @xmath167 ) . * if @xmath160 is a manifold with boundary with orientation form @xmath161 and @xmath168 is an outward pointing normal vector field on @xmath169 , then @xmath169 is given the induced orientation @xmath170 . thus the stokes formula is @xmath171 . * if @xmath172 and @xmath173 are fibre bundles with fibrewise orientation forms @xmath174 and @xmath175 respectively , then we give the the composite bundle @xmath176 the fibrewise orientation form @xmath177 . given a finite set @xmath56 and an element @xmath178 , the canonical projection @xmath179 has a standard fibrewise orientation defined as follows . it factors as @xmath180 where @xmath181 forgets the tangent ray decoration at the marked point @xmath182 , and @xmath183 forgets the position of @xmath182 , thus forgetting it entirely . the principal @xmath184-bundle @xmath181 inherits a fibrewise orientation @xmath185 from the counterclockwise orientation of the circle . the map @xmath183 is a family of riemann surfaces and so the complex structures of the fibres determine a fibrewise orientation @xmath186 . using the above orientation conventions , we then give @xmath187 the fibrewise orientation determined by the form @xmath188 . given @xmath189 , we define a bijection between rays in @xmath190 and fibrewise orientations of @xmath191 as follows . the ray spanned by @xmath192 ( where @xmath193 are the elements of @xmath46 ) corresponds to the fibrewise orientation determined by the above conventions and the factorisation @xmath194 . we will first recall arnold s well - known presentation of the cohomology of the ordered configuration spaces of @xmath108 , or equivalently of the spaces of the unframed little discs operad @xmath0 . then we will use this result to give a new presentation of the cohomology of the framed little discs operad which will be convenient for working with the cyclic operad structure . consider the little @xmath195-discs operad @xmath0 . for @xmath196 , consider the map @xmath197 which forgets all but the @xmath71 and @xmath198 discs . define a collection of 1-forms @xmath199 where @xmath200 is the standard volume form of @xmath184 . ( we will often abuse notation by using the same symbol for both a differential form and the cohomology class it represents . ) ( arnold)[arnold - pres ] the real cohomology of @xmath2 is generated by the degree 1 classes @xmath201 for @xmath202 , subject to only to the symmetry relation , @xmath203 and the arnold relation , @xmath204 for each triple of distinct indices @xmath205 . in fact , this gives a presentation of the integral cohomology , but we shall only be concerned with real cohomology in this paper . since @xmath206 , where the @xmath71 @xmath184 factor measures the angle by which the @xmath71 disc is rotated , one has @xmath207 where @xmath208 is the class represented by the volume form @xmath209 of the @xmath49-th circle . thus the cohomology is generated by the classes @xmath201 together with the classes @xmath208 , and subject only to the symmetry and arnold relations . the symmetric group @xmath20 acts on @xmath210 by permuting the labels of the marked points . the subgroup @xmath211 , consisting of permutations which fix 0 , permutes the generators and relations among themselves in the obvious way . however , the permutations which do not fix zero act in a more complicated way . we will now give a different presentation in which the generators and relations of @xmath212 are both permuted by the full symmetric group @xmath20 . let @xmath213 the projection forgetting all but the @xmath49-th and the @xmath214-th points with their tangent rays ( see earlier section [ canonical - projections ] ) . since @xmath97 , we define closed 1-forms @xmath215 which represent generators of the cohomology . to describe the relations among these generators we must relate them to the generators @xmath201 , @xmath208 described above . there are two ways of identifying the space @xmath216 with @xmath217 . the first way is to apply the unique conformal automorphism of @xmath218 which puts the marked points @xmath219 labelled by @xmath220 at the points @xmath221 respectively . one then reads off the angle @xmath222 , in the counter - clockwise sense , from the equator ( going in the direction from @xmath223 to @xmath224 to @xmath225 ) to the tangent ray at the @xmath49 marked point . in terms of these coordinates , @xmath226 . the other way of identifying @xmath227 with @xmath217 is to put the points at @xmath228 and then rotate the sphere around the axis through @xmath223 and @xmath225 , so that the tangent ray of the point at @xmath225 is parallel to the equator ( pointing in the direction from @xmath225 to @xmath223 to @xmath224 ) . we define @xmath229 to be the angle between the ray at @xmath223 and the equator . parallel transport along the geodesic from @xmath230 to @xmath231 of the tangent ray to the equator at @xmath230 gives a reference ray at @xmath231 and @xmath232 is the angle from this ray to the tangent ray of @xmath231 . the angle @xmath233 is measured at @xmath230 from the equator to the geodesic from @xmath230 and @xmath231 . in terms of these coordinates , @xmath234 and @xmath235 . these two different coordinate systems are illustrated below . one easily sees that these two coordinate systems are related as follows : @xmath236 hence , in the cohomology of @xmath94 the two sets of generators are related by : @xmath237 in terms of the @xmath238 classes , the arnold relation becomes : @xmath239 the action of the symmetric group must send relations to relations , so one can write the relations in a more invariant form . let @xmath240 be four distinct elements of @xmath241 . cyclic arnold relations _ are @xmath242 where @xmath243 is the alternating group on 4 letters . clearly the symmetric group @xmath20 sends each relation of this type to another relation of this type . [ cyclic - pres ] the real cohomology of @xmath210 is generated by the 1-dimensional classes @xmath244 ( @xmath245 ) subject only to the cyclic arnold relations , one for each 4-tuple of distinct indices @xmath240 between @xmath223 and @xmath1 , and to the symmetry relation @xmath246 . the main ingredient in kontsevich s proof of formality of the unframed little discs operad is the construction of a certain differential graded algebra of graphs , the _ kontsevich graph complex _ @xmath247 , which resolves the usual arnold presentation ( theorem [ arnold - pres ] ) of the cohomology of the little discs @xmath2 . as @xmath1 varies these dgas of graphs fit together to form a cooperad , and the dual operad is quasi - isomorphic to the operad of chains on the little discs . tensoring kontsevich s resolution @xmath247 of @xmath248 with @xmath249 gives a resolution of the presentation of the cohomology ring of @xmath250 we shall refer to this dga as the _ affine graph complex _ , denoted @xmath251 . the collection of dgas @xmath252 forms a cooperad , but it is _ not _ compatible with the _ cyclic _ operad structure of @xmath17 . to deal with this we will pass to the _ projective graph complex _ @xmath253 . the projective graph complexes do form a cyclic cooperad and we prove that @xmath253 is indeed a resolution of @xmath212 by a comparison with @xmath251 . summarising this : * @xmath247 is kontsevich s graph complex ; it resolves @xmath248 over @xmath254 $ ] . * @xmath255 is the affine graph complex ; it resolves @xmath212 over @xmath254 $ ] . * @xmath253 is the projective graph complex , which resolves @xmath212 over @xmath256 $ ] . each of these complexes will be defined in detail below . by a graph @xmath257 we shall mean a finite set @xmath258 of vertices and a set @xmath163 of unordered pairs of vertices representing the edges . note that our graphs can not have loops or double edges . those vertices in @xmath259 are called _ internal vertices _ , and those in @xmath56 are called _ external vertices _ ; there is an induced partition of the set of edges @xmath260 where the set of _ internal edges _ , @xmath261 , consists of those edges for which both endpoints are internal , the set of _ external edges _ , @xmath262 , consists of those for which both endpoints are external vertices , and the set of boundary edges , @xmath263 , consists of those which have one internal endpoint and one external endpoint . an _ affine orientation _ of a graph @xmath264 is a choice of a ray in the real line @xmath265 . a _ projective orientation _ of @xmath266 is a choice of a ray in @xmath267 . if @xmath266 is equipped with an affine or projective orientation then we shall write @xmath268 for the same underlying graph equipped with the opposite orientation . given a graph @xmath269 and an edge @xmath270 , we define a new graph @xmath271 by deleting the edge @xmath270 . if @xmath266 has a projective or affine orientation , then @xmath272 inherits an orientation of the same type by the isomorphism @xmath273 from section [ orientations ] . given a graph @xmath269 and an edge @xmath270 that is not part of a triangle of edges and has at least one internal endpoint , let @xmath274 denote the graph constructed from @xmath266 by contracting the edge @xmath270 , i.e. deleting @xmath270 from the set of edges and identifying its two endpoints together . if @xmath270 is a boundary edge then the resulting vertex is external ; if @xmath275 then the resulting vertex is internal . thus the set of external vertices of @xmath274 is identified with @xmath56 . if @xmath266 has an affine orientation then @xmath274 inherits an affine orientation by the isomorphism @xmath273 . inducing projective orientations is slightly more complicated and depends on whether @xmath270 is boundary , or internal . * if @xmath276 then one endpoint , @xmath182 , is external and one endpoint , @xmath277 , is internal . in this case there is a canonical identification of the edges of @xmath274 with @xmath278 ; we identify the external vertices of @xmath274 with @xmath56 , and the internal vertices with @xmath279 , by identifying the vertex in @xmath274 at which @xmath270 was contracted with @xmath182 . using these identifications , the induced projective orientation of @xmath274 is the image of the projective orientation of @xmath266 under the isomorphism @xmath280 . * if @xmath275 then inducing a projective orientation on @xmath274 is equivalent to choosing an orientation of @xmath270 , meaning a choice of one end , @xmath277 , as _ head _ and the other , @xmath182 as _ tail_. as in the boundary edge case , we contract @xmath270 onto its tail and identify the internal vertices of @xmath274 with @xmath281 , and then we induce a projective orientation via @xmath280 . suppose that @xmath282 is a pair of edges in a graph @xmath266 meeting at an internal vertex @xmath182 and such that @xmath270 is not a side of a triangle in @xmath283 . then we shall write @xmath284 if @xmath266 has a projective orientation and @xmath270 is either a boundary edge or an oriented internal edge , then @xmath285 inherits a projective orientation by the above discussion . first recall kontsevich s complex of `` admissible graphs '' ( * ? ? ? * definition 13 ) , denoted @xmath286 . the vector space @xmath287 is spanned by isomorphism classes of affine oriented graphs with external vertex set @xmath56 modulo the relations 1 . @xmath288 if there is a connected component of @xmath266 containing no external vertices , 2 . @xmath288 if there is an internal vertex of valence @xmath289 , 3 . @xmath290 . there is a grading defined by @xmath291 . the differential of degree 1 is given by the formula @xmath292 where @xmath270 runs among those boundary and internal edges such that @xmath274 is defined . there is an associative and graded commutative product defined by gluing the external vertices together according to the labelling , and by tensoring orientation rays . note that the differential is a derivation of the product , and so @xmath287 forms a commutative dga . we will write @xmath293 when @xmath294 . [ agc ] the affine graph complex is the cdga defined as @xmath295 we now come to principal object of this paper : the cdga of _ projective graphs_. [ defproj ] the vector space of projective graphs @xmath296 , for a finite set @xmath56 , is spanned by isomorphism classes of projectively oriented graphs @xmath266 with @xmath56 as set of external vertices , modulo the relations * if @xmath266 has an internal vertex of valence @xmath297 then @xmath288 , * if @xmath266 has a component with at most one external vertex and at least one internal vertex then @xmath288 , * @xmath290 , * the `` pinwheel relation '' : for @xmath277 an internal vertex of @xmath266 and @xmath298 the set of edges incident at @xmath277 , @xmath299 there is a grading defined by @xmath300 . there is an associative and graded commutative algebra structure on @xmath296 given by gluing at the external vertices , and tensoring orientation rays . in the case @xmath301 , we write @xmath302 . we next define a differential @xmath303 on @xmath296 . the formula will involve choices , but we will show that because of the pinwheel relation it is in fact independent of these choices . to define @xmath304 , with @xmath305 , we first arbitrarily choose an orientation of each internal edge @xmath270 of @xmath266 . we orient the boundary edges in each case so that their heads are at their internal endpoints . the differential is now given by the formula @xmath306 where the sum runs over all ordered pairs of edges @xmath282 for which the head of @xmath270 is internal and incident to @xmath307 ( so in particular , they must lie in @xmath308 ) and such that @xmath309 is defined ( i.e. deleting @xmath307 and contracting @xmath270 does not result in a loop or double - edge ) . @xmath310 is well - defined , independent of the choices of orientations for the edges between internal vertices . suppose @xmath270 is an edge between internal vertices @xmath277 , @xmath311 . the difference between @xmath310 with @xmath270 oriented one way or the other is precisely the pinwheel relation in @xmath274 at the vertex where @xmath270 is contracted . thus @xmath310 is independent of the choices of orienting the internal edges . it remains to check that @xmath303 descends to the quotient by the pinwheel relation . given a graph @xmath266 with an internal vertex @xmath277 , orient all of the ( internal ) edges of @xmath298 with their heads at @xmath277 . then , @xmath312 the first of the two above summations can be grouped as a sum of terms , each of which is a pinwheel sum over edges @xmath25 incident at @xmath277 in @xmath285 . the terms of the second summation cancel in pairs since @xmath313 oriented with head at @xmath277 means that @xmath314 , and @xmath315 . @xmath316 . to compute @xmath317 we first choose an orientation of each internal edge of @xmath266 ; @xmath310 is a sum of terms of the form @xmath318 and the internal edges of each of these terms inherit orientations . now , @xmath317 is a sum of graphs of the form @xmath319 ( the sign depends only on @xmath266 ) , where @xmath320 are internal or boundary edges in @xmath266 with @xmath307 incident at the head of @xmath270 , and @xmath321 are internal or boundary edges in @xmath318 with @xmath322 incident at the head of @xmath323 . this means that , as edges in @xmath266 , the possible configurations are enumerated as follows : type ( 0 ) : the head of @xmath323 is incident at @xmath322 in @xmath266 and the heads of @xmath270 and @xmath323 are disjoint . if the head of @xmath323 is incident at @xmath322 in @xmath266 but the configuration is not of type ( 0 ) then it must be type ( 1 ) : @xmath307 , @xmath322 , @xmath270 and @xmath323 all meet at a vertex . if the head of @xmath323 is not incident at @xmath322 in @xmath266 then it must be incident at @xmath322 in @xmath285 . this can happen only if , in @xmath266 , either @xmath323 and @xmath322 are incident at opposite ends of @xmath270 ( types ( 2 ) , ( 4 ) , and ( 5 ) below ) , or if @xmath322 is incident at the tail of @xmath323 and contracting @xmath270 causes @xmath323 to reverse orientation ( type ( 3 ) below ) . so the configuration must be one of the following types . we will show case by case that all terms can be grouped to exactly cancel . for type ( 0 ) , the way that orientations are induced yields @xmath324 and so these two terms in @xmath325 cancel . for types ( 1 ) and ( 2 ) , @xmath326 cancels with @xmath327 . in the configuration of type ( 3 ) , @xmath328 cancels with @xmath329 which is a term of type ( 4 ) , and hence the terms of type ( 3 ) cancel in pairs with the terms of type ( 4 ) . type ( 5 ) is slightly more involved . by the pinwheel relation , @xmath330 consider the sum of the expression over all @xmath331 ; the sum over @xmath332 is zero because of the symmetry of swapping @xmath307 and @xmath25 ( just as for type ( 0 ) ) , and the sum over @xmath333 exactly cancels with the sum over @xmath333 of the type ( 5 ) terms @xmath334 occurring in @xmath325 . here we introduce maps of graph complexes that are combinatorial analogues of certain configuration space integrals . fix finite sets @xmath335 . let @xmath336 denote the subalgebra of graphs having no internal vertices . note that for such a graph @xmath337 , a projective orientation is simply a ray in @xmath265 . given a ray @xmath338 , there is a linear map @xmath339 it is defined by turning the @xmath259 vertices into internal vertices and mapping the orientations by @xmath340 note that @xmath341 geometrically , this map corresponds to integrating out the @xmath259 vertices as we shall see in section [ int - section ] . a general projectively oriented graph @xmath305 with set of internal vertices @xmath259 can be written as @xmath342 , where @xmath343 and @xmath344 is a ray in @xmath345 . we now define a cyclic cooperad structure on the collection of projective graph complexes . for any partition @xmath346 we need to define a co - composition map @xmath347 let @xmath348 be the graph consisting of a single edge between external vertices @xmath349 and @xmath50 , oriented from @xmath349 to @xmath50 . then , @xmath350 a graph @xmath351 can be written as a product of graphs each having only a single edge . the order of the factors corresponds to the chosen orientation . the multiplication of the formulae above for the edges of @xmath266 defines @xmath352 . given a partition @xmath353 , choose rays @xmath354 and @xmath355 such that @xmath356 . now consider the composition @xmath357 we can now define the cooperad co - composition map on an arbitrary graph @xmath358 by the formula @xmath359 there is a useful alternative description of this co - composition that will make the verification of certain properties easier . an @xmath360-_splitting _ of a graph @xmath266 is a partition of the vertices into two sets @xmath361 with @xmath362 and @xmath363 and a partition of the edges into two sets @xmath364 such that if both of the endpoints of an edge @xmath270 are in @xmath365 ( @xmath366 ) then @xmath367 . given a splitting @xmath41 , we can construct two new graphs @xmath368 with vertices @xmath369 and edges @xmath370 ; if @xmath367 , as an edge in @xmath266 , has only one endpoint in @xmath365 then as an edge in @xmath371 it s endpoint that is not in @xmath365 is replaced by the auxiliary vertex @xmath372 . if @xmath266 has a projective orientation @xmath373 then we give @xmath374 and @xmath375 projective orientations @xmath376 and @xmath377 such that @xmath378 . by inspection one finds the following . the co - composition map defined above in is given by the formula @xmath379 where the sum runs over the set @xmath380 of all @xmath360-splittings of @xmath266 . the co - composition maps given above are well - defined and they make @xmath296 into a cyclic cooperad in the category of commutative differential graded algebras . there are three facts to verify . * @xmath381 is compatible with the relations @xmath382 of definition [ defproj ] . * @xmath381 commutes with the differential . * these maps satisfy the cyclic cooperad axioms given in section [ cyclic - operad - definitions ] . for ( i ) , it is easy to verify compatibility with @xmath383 . for the pinwheel relation @xmath384 , suppose @xmath266 is a projective graph with an internal vertex @xmath277 and external vertices @xmath385 . then @xmath386 where @xmath41 runs over the set @xmath387 of @xmath360-splittings of @xmath388 . then can be written as @xmath389 this verifies statement ( i ) . we turn to ( ii ) now . given an oriented edge @xmath270 , let @xmath390 denote the head and @xmath391 the tail . observe that choosing a splitting of @xmath318 is the same as choosing a splitting of @xmath266 such that @xmath392 are either all on the left or all on the right . hence , @xmath393 applying the differential to a co - composition and expanding similarly , one sees that @xmath394 is given by the above sum but without the requirement @xmath391 be on the same side as @xmath390 . we will show that these additional terms cancel out . suppose we are given a pair @xmath282 in @xmath266 meeting at an internal vertex @xmath395 and a splitting @xmath396 such that @xmath397 , @xmath398 , and @xmath399 . define a new splitting @xmath400 by switching @xmath320 and @xmath390 from the right to the left . then @xmath401 and so these two terms exactly cancel . all of the additional terms in @xmath402 can thus be paired in this way to cancel . finally , the verification of ( iii ) is entirely straightforward . kontsevich showed that the map @xmath405 , defined by sending a graph with no internal vertices to the corresponding product of arnold classes @xmath201 , one for each edge @xmath406 , and sending all graphs with internal vertices to zero , is a quasi - isomorphism of cooperads ; see ( * ? ? ? * section 3.3.4 ) or ( * ? ? ? * theorem 9.1 ) . it follows immediately that the analogous map @xmath407 is also a quasi - isomorphism of cooperads . the cohomology @xmath408 is a cyclic cooperad , but @xmath409 is _ not _ a cyclic cooperad . hence we must pass from affine to projective graphs . [ projective - graphs - resolution ] the map @xmath410 defined by sending @xmath266 to zero if it has any internal vertices , and otherwise sending it to the corresponding product of @xmath244 classes , one for each edge @xmath411 , is a quasi - isomorphism of cyclic cooperads . we will derive this theorem from the corresponding statement for @xmath412 by a spectral sequence comparison of @xmath412 and @xmath403 . the first step is to construct a map between @xmath412 and @xmath403 , and for this we must introduce a couple of auxiliary graded vector spaces . let @xmath413 denote the graded vector space covering @xmath412 in which each graph is equipped with an injective map from the set of its internal vertices to @xmath414 . let @xmath415 denote the graded vector space covering @xmath403 in which the pinwheel relation is _ not _ imposed and each graph is equipped with an injective map from the set of its internal vertices to @xmath414 . there is an injective linear map @xmath416 defined by sending @xmath417 to the graph constructed by adding a disjoint external vertex @xmath223 , denoted @xmath418 , and then successively adding an external edge @xmath419 between @xmath418 and the external vertex labelled by @xmath420 for @xmath421 , and then finally by adding a boundary edge @xmath422 from @xmath418 to each internal vertex @xmath277 . the affine orientation @xmath52 of @xmath266 determines a projective orientation of @xmath423 by the expression @xmath424 where @xmath277 runs over the internal vertices in some arbitrary order . the image of @xmath425 is precisely the vector space spanned by graphs for which each internal vertex is connected to @xmath418 by an edge . one checks that @xmath425 descends to a linear map @xmath426 this is a morphism of graded algebras , but it does _ not _ commute with the differentials . however , we have the following . [ linear - iso ] the map @xmath427 is a linear isomorphism . the idea of the proof of the above lemma is to explicitly construct an inverse to @xmath233 by using the pinwheel relation to rewrite any projective graph as a sum of projective graphs in which each internal vertex is connected by an edge to the external vertex @xmath418 . we now formalise this procedure . for any @xmath428 , we define two linear operators @xmath429 by specifying what they do on generators . given a projective graph @xmath430 , if there is no internal vertex labelled by @xmath11 then @xmath431 ; if there is such an internal vertex then @xmath432 simply adds an edge @xmath270 from that vertex to @xmath418 or outputs zero if there is already such an edge ; the new projective orientation is given by @xmath433 . observe that the operators @xmath432 anticommute . let @xmath434 denote the set of edges in @xmath266 which are incident at the internal vertex with label @xmath11 , where @xmath435 if there is no such vertex . if @xmath434 contains an edge to the external vertex @xmath418 then @xmath436 ; otherwise @xmath437 one can think of @xmath438 as using the pinwheel relation to write @xmath266 as a sum of graphs that are in the image of @xmath233 , although we have not yet passed from @xmath439 to @xmath440 . we can now proceed to define an inverse to @xmath233 . consider the infinite composition @xmath441 since any graph @xmath266 has only finitely many internal vertices , all but finitely many of the @xmath438 act by identity on @xmath266 , and so this is well defined . [ p - ops ] * @xmath442 is equivariant with respect to the action of the group @xmath443 of permutations of @xmath414 by relabelling the internal vertices . * @xmath266 and @xmath444 project to the same element of @xmath403 . * any two operators @xmath445 commute . * the pinwheel ideal is contained in the kernel of @xmath442 . part ( i ) is trivial . part ( ii ) is immediate from the pinwheel relation . for part ( iii ) we need only consider the case where @xmath266 is a graph with internal vertices @xmath214 and @xmath11 , neither of which has an edge to @xmath418 since in all other cases at least one of @xmath446 or @xmath438 acts by identity on @xmath266 . in this situation , @xmath447 the operations of deleting edges @xmath270 and @xmath307 anticommute , deleting either @xmath270 or @xmath307 anticommutes with @xmath448 and @xmath432 , and @xmath448 and @xmath432 anticommute with one another . hence @xmath449 we turn to part ( iv ) . it suffices to check that for any graph @xmath266 with internal vertex @xmath11 , @xmath450 since the operators @xmath438 all commute , we need only show that @xmath438 sends @xmath451 to zero . first suppose that there is no edge in @xmath266 between the vertices @xmath11 and @xmath418 . then @xmath452 and this is zero since the @xmath282 term cancels exactly with the @xmath453 term . now suppose that there is an edge @xmath454 in @xmath266 between the internal vertex labelled @xmath11 and the external vertex @xmath418 . then @xmath455 since @xmath456 so all terms cancel . observe that @xmath457 lies in the image of @xmath425 . since @xmath425 is injective we can thus regard @xmath442 as a linear map @xmath458 which clearly satisfies @xmath459 and , by lemma [ p - ops ] , descends to a map @xmath460 such that @xmath461 . on the other hand , lemma [ p - ops ] implies that @xmath442 descends to the identity on @xmath403 . but the induced map on @xmath403 is equal to @xmath462 . therefore @xmath233 is bijective . the complex @xmath403 has a descending filtration @xmath463 in which the @xmath11-th filtration level , @xmath464 , is the subcomplex spanned by projective graphs having at least @xmath11 edges connecting @xmath418 to other external vertices . the filtration is well defined because the pinwheel relation does not affect edges between external vertices , and similarly , the differential is compatible with the filtration . there is a spectral sequence @xmath465 associated with this filtration . [ iso ] the map @xmath233 induces an isomorphism of complexes @xmath466 . let us filter @xmath467 so that @xmath468 is spanned by elements @xmath469 where @xmath470 is a product of at least @xmath11 distinct generators @xmath208 . the linear isomorphism @xmath427 respects the filtrations . we show that if @xmath471 then @xmath472 where the error term , @xmath473 , is in strictly higher filtration , i.e. @xmath474 . this will prove the lemma . recall that @xmath475 has an edge from the external vertex @xmath418 to each internal vertex @xmath277 ; let @xmath422 denote this edge . the differential of @xmath476 is a sum of terms of the form @xmath477 we now consider the various possible positions of the pair of adjacent edges @xmath282 in @xmath476 . case 1 : @xmath270 is an internal edge from @xmath478 to @xmath479 . if @xmath307 connects @xmath479 to the external vertex @xmath418 , then @xmath480 . to see that the orientations agree , suppose @xmath266 has affine orientation @xmath52 . then @xmath475 has projective orientation @xmath481 where @xmath482 here stands for a wedge product of the boundary edges incident at @xmath418 corresponding to the class @xmath470 , and @xmath477 has projective orientation @xmath483 which is precisely the projective orientation of @xmath484 . if @xmath307 connects @xmath479 to a vertex that is not @xmath418 then @xmath485 because this graph has a double edge coming from the edges @xmath486 and @xmath487 in @xmath488 . case 2 : @xmath270 goes from an external vertex @xmath489 to an internal vertex @xmath277 . if @xmath490 is the edge from @xmath418 to @xmath277 , then @xmath491 as in case 1 . the remaining possibilities for the position of @xmath307 will contribute only to the error term @xmath473 . if @xmath492 then @xmath493 is in strictly higher filtration because it has an additional external edge between @xmath52 and @xmath418 . case 3 : @xmath270 goes from @xmath418 to an internal vertex @xmath277 . the internal vertices are at least 4-valent , so there are at least 3 edges connecting @xmath277 to vertices @xmath494 other than @xmath418 . we may assume that @xmath307 is not the edge from @xmath277 to @xmath349 . if @xmath349 is internal then @xmath495 because it has a double edge coming from the edges from @xmath349 to @xmath277 and from @xmath349 to @xmath418 in @xmath475 . if @xmath349 is external then @xmath496 has one more edge than @xmath475 connecting @xmath349 to @xmath418 and thus lives in higher filtration . by lemma [ iso ] the @xmath497-term is @xmath498 . on the other hand by definition of the affine graph complex ( definition [ agc ] ) , and since kontsevich s graph complex @xmath499 is a resolution of @xmath500 , @xmath501 we claim that the spectral sequence collapses at the @xmath497-term . the projection @xmath502 is surjective in cohomology because the forms @xmath244 generate the cohomology ring of @xmath94 and @xmath244 is the image of the graph with a single edge between external vertices labelled @xmath49 and @xmath214 . thus the dimension of @xmath503 is not smaller than the dimension @xmath404 . the isomorphism @xmath504 and finite dimensionality imply that @xmath505 and @xmath506 is an isomorphism of cyclic cooperads . in this section we construct a quasi - isomorphism @xmath507 where @xmath508 is the complex of `` piecewise semi - algebraic forms '' . the idea is as follows . a graph @xmath264 determines a differential form on @xmath509 given by multiplying @xmath244 1-forms corresponding to the edges , and one then integrates this form along the fibres of the bundle @xmath510 to obtain a form on @xmath147 . both the integrand and the integration map have a sign ambiguity , but if @xmath266 has a projective orientation then these sign ambiguities are resolved . the differential form that results is generally _ not _ smooth , and so we must pass to the semi - algebraic setting in which the kontsevich integral is well - defined . given a semi - algebraic manifold , the complex of _ minimal _ forms , @xmath511 , is spanned by elements @xmath512 where the @xmath513 are semi - algebraic functions . the complex of @xmath514 forms , @xmath515 , consists of those forms which can be written as a pushforward of a minimal form . when @xmath168 is smooth , the commutative differential graded algebra of @xmath514 forms is quasi - isomorphic to the ordinary smooth forms . we first recall some facts about semi - algebraic fibrewise integration . suppose @xmath516 is a semi - algebraic fibre bundle with @xmath1-dimensional compact fibres and equipped with a fibrewise orientation . integration of forms fibrewise defines a map ( see ( * ? ? ? * section 8) ) @xmath517 in particular , if @xmath518 restricts to a volume form on each fibre then @xmath519 is the semi - algebraic function on @xmath520 which sends @xmath521 to the volume of the fibre over @xmath50 . we shall need _ iterated _ fibrewise integrals at several points in this section . as the theory is developed in @xcite , the pushforward of a minimal form is a @xmath514 form by definition , and the pushforward of a @xmath514 form is not defined . note that the minimal forms constitute a sub - cdga of all @xmath514 forms . [ twosteps ] if @xmath522 and @xmath523 are semi - algebraic bundles with compact oriented fibres , and @xmath524 is a minimal form on @xmath163 such that @xmath525 is a minimal form on @xmath520 , then @xmath526 is defined and @xmath527 this is a slight generalisation of ( * ? ? ? * prop 8.11 ) and has a similar proof . given any form @xmath528 , one has the push - pull formula ( ( * ? ? ? * prop . 8.13 ) ) , @xmath529 now recall the fibrewise stokes formula . 8.10 ) [ stokes ] let @xmath516 be a semi - algebraic bundle of oriented @xmath11-manifolds with boundary , with @xmath530 denoting the associated fibrewise boundary bundle . then , giving the boundary the induced orientation , @xmath531 for any form @xmath532 . we first define @xmath533 on the subcomplex @xmath534 of graphs with no internal vertices . let @xmath535 be such a graph with edge set @xmath163 and orientation ray in @xmath265 spanned by @xmath536 . an edge @xmath270 between vertices @xmath182 and @xmath277 determines a semi - algebraic minimal 1-form @xmath537 where @xmath538 forgets all marked points except @xmath182 and @xmath277 . define @xmath539 recall that ( the equivalence class of ) a projectively oriented graph @xmath266 with internal vertex set @xmath259 and external vertex set @xmath56 can be written as @xmath540 for some ray @xmath338 and graph @xmath541 . we then define @xmath542 where the fibrewise orientation of @xmath543 is determined by the ray @xmath344 as described in section [ orientations ] . we prove in lemma [ kont - integral - properties ] below that @xmath533 respects the relations @xmath382 of definition [ defproj ] , and hence it descends to a map from @xmath296 , that we denote by the same name . [ kont - integral - properties ] the kontsevich integral satisfies * @xmath544 . * @xmath533 vanishes on graphs having an internal vertex of valence @xmath297 . * @xmath533 vanishes on graphs which have a connected component containing at most one vertex in @xmath56 and at least one internal vertex . * the pinwheel ideal is contained in the kernel of @xmath533 . hence we have a well - defined morphism @xmath545 of graded algebras . statement ( i ) is trivial . for ( ii ) , suppose @xmath266 has a trivalent internal vertex @xmath52 adjacent to vertices @xmath182 , @xmath277 and @xmath311 as below . by proposition [ twosteps ] one may integrate out the internal vertices by first integrating out @xmath52 and then integrating out the remaining internal vertices , i.e. @xmath546 where if @xmath547 has fibrewise orientation @xmath548 then @xmath549 has fibrewise orientation @xmath550 and @xmath551 has its standard fibrewise orientation . this works because the integrand @xmath552 is the product of @xmath553 with a minimal form @xmath554 that is independent of @xmath52 , so by formula , @xmath555 is equal to the product of @xmath554 with the @xmath514 form @xmath556 and since the fibres of @xmath557 have dimension 3 , is a semi - algebraic function on @xmath558 , and so @xmath555 is minimal . putting @xmath182 , @xmath277 and @xmath311 at @xmath223 , @xmath224 and @xmath225 identifies each fibre with the complement of a codimension 1 semi - algebraic subset of the unit tangent circle bundle of the riemann sphere punctured at @xmath223 , @xmath224 and @xmath225 , @xmath559 , and the differential form @xmath553 extends to @xmath559 . complex conjugation is an orientation - preserving diffeomorphism of the tangent circle bundle , and it reverses the sign of @xmath560 since it reverses the sign of each of the three factors . hence the integral over each fibre is zero , and the fibrewise integral is identically zero . statement ( iii ) follows from kontsevich s lemma on vanishing of integrals over configuration spaces ( * ? ? ? * lemma 6.4 ) . we choose an external vertex as the point at infinity , and write a form @xmath552 as polynomial in the forms @xmath201 and @xmath561 . integrate out the internal vertices by first integrating out the tangent ray decorations at the internal vertices , resulting in a @xmath514 and in fact minimal form on a configuration space ; what remains is a configuration space integral of the form that kontsevich s lemma refers to . for ( iv ) we will in show that if @xmath277 is an internal vertex then integrating out the tangent ray decoration at @xmath277 sends the element @xmath562 to zero . write the orientation of @xmath563 as @xmath564 where @xmath565 is the set of edges incident at @xmath277 . let @xmath566 denote the opposite endpoint of @xmath567 . then @xmath568 and integrating out the decoration at @xmath277 extracts the coefficient of @xmath569 . it is not hard to see that the coefficient of @xmath569 is in fact zero . here we shall prove the following . [ differential - of - graph ] the linear map @xmath545 commutes with the differentials and is thus a morphism of differential graded algebras let @xmath305 be a projectively oriented graph . we will prove this theorem by using the stokes formula to compute the exterior derivative of the pa form @xmath570 . write @xmath571 . by the fibrewise stokes formula ( theorem [ stokes ] ) , @xmath572 since @xmath552 is closed . we can factor @xmath573 into the composition of first forgetting the points in @xmath574 and then forgetting the remaining internal vertices : @xmath575 by lemma [ fibrewise - boundary ] , the fibrewise boundary component @xmath576 is the ( closure of the ) locus where a node separates the sphere into two lobes @xmath577 , with the points labelled by @xmath151 on the lobe @xmath578 . the map @xmath579 forgets the first lobe if @xmath580 , and collapses it to a single marked point if @xmath148 contains an external vertex ; hence it is a semi - algebraic fibre bundle , and in fact it is trivial . if @xmath581 then the corresponding term in the stokes formula vanishes : @xmath582 consider performing the fibrewise integration by first integrating along the fibres of @xmath583 and then integrating out the remaining internal vertices . furthermore , consider performing the integral over the fibres of @xmath584 by first integrating out the tangent ray decorations of the points in @xmath574 and then integrating out the positions of these points : @xmath585 since iterated integration of @xmath514 forms is not defined in general , we must justify as above both of these iterated integrations . we will see that each successive integration results again in a minimal form , so the iteration is allowed . relative to a choice of a point @xmath586 as point at infinity , the form @xmath552 can be expanded in terms of the affine presentation generators as in the proof of lemma [ kont - integral - properties ] ( iii),(iv ) . integrating out the @xmath587 decorations then yields a polynomial @xmath588 in @xmath201 forms and @xmath561 forms for @xmath589 , and hence a minimal form . on the locus @xmath576 , let @xmath590 denote the node between the two lobes . on this locus , if @xmath591 and @xmath592 then @xmath593 . hence each monomial in @xmath29 factors as a product @xmath594 , where @xmath595 is a monomial of @xmath470 and @xmath596 forms on the second lobe ( relative to @xmath52 and including @xmath590 as a marked point ) and @xmath597 is a monomial of only @xmath470 forms on the first lobe ( relative to @xmath590 ) . now , integrating out the @xmath587 positions , @xmath598 and if @xmath581 then @xmath599 by kontsevich s vanishing lemma ( * ? ? ? * lemma 6.4 ) . thus the only fibrewise boundary components which contribute are those for which @xmath600 . if @xmath601 then we will write @xmath602 . we will now compute the stokes boundary term @xmath603 . factor @xmath604 as @xmath605 the first map is given its standard fibrewise orientation and the second map is given the fibrewise orientation @xmath606 , where @xmath338 is the fibrewise orientation of @xmath547 . note that with these choices of orientation , @xmath607 the stokes formula then becomes @xmath608 where the sum runs over unordered pairs @xmath609 at most one of which is external ( if one is internal then we call it @xmath277 , and if both are internal then we choose one to call @xmath277 ) . [ key - relation ] suppose we are given four distinct vertices @xmath610 . then on @xmath611 , @xmath612 we can consider @xmath52 to be at infinity and in the corresponding affine presentation , @xmath613 . hence , on @xmath611 we have the relation @xmath614 . the result follows from multiplying both sides with @xmath615 since 1-forms anticommute . [ volume - forms ] for any triple of vertices @xmath616 , the 2-form @xmath617 is a fibrewise negative unit volume form for @xmath618 . in terms of the affine presentation ( relative to @xmath311 ) , one sees that @xmath619 is a fibrewise negative unit volume form for @xmath620 as follows . we factor @xmath621 as @xmath622 , and each of these maps has a fibrewise orientation as described in section [ orientations ] . the form @xmath569 is a fibrewise positive unit volume form for @xmath623 . the form @xmath624 is a negative fibrewise unit volume form for the restriction of @xmath625 to the face @xmath626 . translating to the cyclic presentation , @xmath627 since @xmath628 is constant on the fibres of @xmath629 , the first term in the final line integrates to zero on each fibre , and so we obtain the result . [ boundary - expansion ] consider a vertex set @xmath630 . on the boundary component @xmath631 the following identity holds . @xmath632 use induction on @xmath11 and the relation . suppose @xmath266 is a projective graph . write @xmath633 . given an edge @xmath270 between vertices @xmath634 with @xmath277 internal ( if @xmath182 is internal as well then we orient @xmath270 with head at @xmath277 ) and a second edge @xmath307 meeting @xmath277 , @xmath635 where the sign comes from sliding @xmath277 to the left past the edges @xmath636 . by lemmas [ boundary - expansion ] and [ volume - forms ] , @xmath637 by combining these we find that @xmath638 summing this over all unordered pairs @xmath639 and using then gives the desired result . as an illustration of theorem [ differential - of - graph ] , let us consider the product @xmath640 corresponding to the graph @xmath266 shown below , with internal vertex @xmath277 and projective orientation @xmath641 . we can express @xmath642 as a linear combination of edge forms as follows . on @xmath643 , @xmath644 by lemma [ volume - forms ] , @xmath645 this corresponds to the sum of graphs where each is given the orientation induced from that of @xmath266 by first deleting an edge and then contracting the @xmath646 edge . from this one finds that @xmath647 the @xmath514-forms on @xmath17 do not constitute a cyclic cooperad because the exterior product of forms , @xmath648 is invertible up to quasi - isomorphism but not on the nose . however , lemma [ cooperad - diagram ] below shows compatibility between the cyclic operad structure on @xmath17 and the cyclic cooperad structure on @xmath440 . if @xmath182 and @xmath277 are both in @xmath33 then commutativity of the diagram @xmath653 implies the formula in this case ; the reasoning is the same when @xmath654 . suppose now that @xmath655 and @xmath656 . choosing an identification @xmath97 , the map @xmath657 corresponds with the product map @xmath658 , and @xmath659 . commutativity of the diagram @xmath660 then implies the formula . [ cooperad - diagram ] ( c.f . * proposition 8.19 ) . ) the following diagram commutes : @xmath661{e , t}{_i\bullet_j } \node[2]{{\mathcal{pg}}_{i \sqcup \{x\ } } \otimes { \mathcal{pg}}_{j\sqcup\{y\ } } } \arrow{s , r}{{\mathscr{i}}\otimes { \mathscr{i } } } \\ \node { \omega_{pa}^*({f\underline{\mathcal{m}}}_{i\sqcup j } ) } \arrow{se , t}{({_x\circ_y})^ * } \node[2]{\omega_{pa}^*({f\underline{\mathcal{m}}}_{i\sqcup\{x\ } } ) \otimes \omega_{pa}^*({f\underline{\mathcal{m}}}_{j\sqcup\{y\ } } ) . } \arrow{sw } \\ \node[2]{\omega_{pa}^*({f\underline{\mathcal{m}}}_{i\sqcup \{x\ } } \times { f\underline{\mathcal{m}}}_{j\sqcup \{y\ } } ) } \end{diagram}\ ] ] the composition map @xmath662 is a principal @xmath184-bundle over its image @xmath663 , where the circle action is by rotating the rays at the @xmath52 and @xmath53 points in opposite directions . consider a graph @xmath664 with set of internal vertices @xmath259 . there are pullback squares @xmath665 and @xmath666 on each fibre over @xmath663 the quotient map @xmath667 is a diffeomorphism in the interior and finite to1 on the boundary . with this observation and the above pullback squares , one has @xmath668 where we choose compatible fibrewise orientations for @xmath137 and @xmath669 . the lemma now follows directly from this , lemma [ pullback - lemma ] , and the definition of the cooperad structure on @xmath440 from section [ cooperad - structure ] . the projective graph complexes @xmath670 collectively form a cyclic cooperad in the category of commutative dgas , and they are degree - wise finite dimensional . hence the graded duals @xmath671 form a cyclic operad in the category of cocommutative differential graded coalgebras . the projections @xmath672 determine a quasi - isomorphism of cyclic cooperads , so the dual morphisms , @xmath673 constitute a quasi - isomorphism of cyclic operads . let @xmath674 denote the symmetric monoidal functor of semi - algebraic chains with real coefficients @xcite . this functor is weakly equivalent to the functor of real singular chains @xmath675 , so the cyclic operads @xmath18 and @xmath676 are weakly equivalent . by ( 7.3 ) , if @xmath168 is compact then the natural evaluation map @xmath677 is a quasi - isomorphism . although the complexes @xmath678 do not form a cyclic operad , the compositions @xmath679 give a quasi - isomorphism of cyclic operads by lemma [ cooperad - diagram ] . this completes the proof . j. giansiracusa and p. salvatore , formality of the framed little 2-discs and semidirect products . to appear in proceedings of `` homotopy theory of function spaces and related topics , '' oberwolfach , april 5 - 11 , 2009 .

cluster categories were defined in @xcite as tools for categorification of fomin - zelevinsky cluster algebras @xcite . they are defined as the orbit categories of the derived category @xmath0 of hereditary abelian categories @xmath1 by a certain autoequivalence . in the situation where @xmath1 is the category of finite dimensional representations of a finite acyclic quiver , the cluster category has been subject to intense investigation . in this case it has been shown that the cluster category and the set of exceptional objects form a good model for the cluster algebra associated with the same quiver . in this paper we work with a cluster category @xmath2 defined from a different hereditary abelian category , namely the tube @xmath3 . this category is called the cluster tube and has recently been studied in @xcite and @xcite . although this category is also a hom - finite triangulated 2-calabi - yau category , it does not admit all of the nice properties of cluster categories from quivers . in particular , the maximal rigid ( also called maximal exceptional ) objects do not satisfy the more restrictive definition of cluster - tilting objects . moreover , the gabriel quivers of the endomorphism rings of maximal rigid objects in the cluster tube have loops . consequently , @xmath2 with its maximal rigid objects does not carry a cluster structure in the sense of @xcite . the axioms for cluster structures can be modified , however , to apply also to cluster tubes , see @xcite . the aim of the present paper is to study the endomorphism rings of the maximal rigid objects . we will find a description in terms of quivers with relations . like cluster - tilted algebras , the algebras we consider here are gorenstein of gorenstein dimension 1 , unless @xmath4 , in which case they are self - injective . however , the proof ( from @xcite ) for cluster - tilted algebras has no analogy in our setting . instead , we use the fact that our algebras are gentle , and apply the technique from @xcite to our quivers with relations in order to prove the result . the properties of the algebras we study in this paper are thus reminiscent of those of the algebras recently studied in @xcite . since the endomorphism rings are gentle , they are string algebras . we use the theory of string- and band - modules to show that the endomorphism rings are of finite type . one of the main results about cluster - tilted algebras , which was proved in @xcite , is the close connection between the module category of the cluster - tilted algebra and the cluster category it arises from . this connection is provided by the hom - functor . in our situation , the hom - functor is not full , and therefore there is no analogous theorem . we will nevertheless study the action of the hom - functor on the objects , and in particular show that it is dense . indeed , when @xmath5 is maximal rigid , we find an explicit description of @xmath6 for every indecomposable @xmath7 in @xmath2 . the paper is organised as follows : section [ sec : objects ] contains the definition of the cluster tube and a description of maximal rigid objects recalled from @xcite . in section [ sec : endoring ] we give the description of the endomorphism rings , before we in section [ sec : gentle ] study the gentleness and gorenstein dimension and give some facts about indecomposable representations . finally , in section [ sec : homfunctor ] we describe the action of the hom - functor . throughout the paper we will work over some field @xmath8 , which is assumed to be algebraically closed . modules over an algebra will always mean left modules , and we will read paths in quivers from right to left . we start off by reviewing some properties of cluster tubes . these categories have recently been studied in @xcite and @xcite , and more details can be found in these papers . for any integer @xmath9 , let @xmath3 be a tube of rank @xmath10 , that is , the category of nilpotent representations of a cyclically oriented @xmath11-quiver . it can also be realised as the thick subcategory generated by a tube in the regular component in the ar - quiver of a suitable tame hereditary algebra . all maps in this category are linear combinations of finite compositions of irreducible maps , and are subject to mesh relations in the ar - quiver . @xmath3 is a hereditary abelian category , and following the construction introduced in @xcite , we form its cluster category , called the _ cluster tube of rank @xmath10 _ and denoted @xmath2 . this is by definition the orbit category obtained from the bounded derived category @xmath12 by the action of the self - equivalence @xmath13 $ ] . here , @xmath14 $ ] denotes the suspension functor of @xmath15 , while @xmath16 is the auslander - reiten translation . unless the actual value of @xmath10 is important , we will usually suppress the subscript @xmath10 in the notation , and write @xmath17 , @xmath18 and @xmath19 . for a finite - dimensional hereditary algebra @xmath20 , a theorem due to keller @xcite guarantees that the associated cluster category @xmath21 is triangulated with a canonical triangulated structure inherited from the derived category . keller s result is not directly applicable in our situation , since @xmath17 has no tilting objects . nevertheless , @xmath2 also inherits a triangulated structure from @xmath15 , see @xcite for a rigorous treatment of this . the indecomposable objects of the cluster tube @xmath19 are in bijection with the indecomposables in @xmath17 itself , and we will sometimes use the same symbol to denote both an object in the tube @xmath17 and its image in the cluster tube @xmath19 . the irreducible maps in @xmath19 are the images of the irreducible maps in @xmath18 , which again are the shifts of the irreducible maps in the tube @xmath17 . so the ar - quiver of @xmath19 is isomorphic to the ar - quiver of @xmath17 . for a given rank @xmath10 , we will use a coordinate system on the indecomposable objects . choose once and for all a quasisimple object and give it coordinates @xmath22 . now give the other quasisimples coordinates @xmath23 such that @xmath24 , where @xmath25 is reduced modulo the rank @xmath10 . then give the remaining indecomposables coordinates @xmath26 in such a way that there are irreducible morphisms @xmath27 for @xmath28 and @xmath29 for @xmath30 . throughout , the first coordinates will be reduced modulo @xmath10 . see figure [ figure : coordinates ] . and @xmath2.,width=453 ] the infinite sequence of irreducible maps @xmath31 is called a _ ray_. similarly , the infinite sequence @xmath32 is called a _ coray_. note that the sum of the coordinates is constant , modulo @xmath10 , for indecomposables located on the same coray . for an indecomposable object @xmath33 where @xmath34 we will also need the notion of the _ wing _ @xmath35 determined by @xmath7 . this is by definition the set of indecomposables whose position in the tube is in the triangle with @xmath7 on top . @xmath7 will be called the _ summit _ of @xmath35 . in terms of coordinates , the objects in the wing @xmath36 are @xmath37 such that @xmath38 and @xmath39 . the _ height _ of @xmath35 is the quasi - length @xmath40 . the dimensions of the hom - spaces in @xmath19 are given by the following lemma , proved in @xcite . [ lem : homspaces ] for @xmath7 and @xmath41 indecomposable in @xmath19 , we have @xmath42 where @xmath43 is the usual @xmath8-vector space duality @xmath44 . when @xmath7 and @xmath41 are indecomposable , the maps in @xmath45 which are images of maps in @xmath46)$ ] for @xmath47 in @xmath17 will be called @xmath18-maps , and those which are images of maps in @xmath17 itself will be called @xmath17-maps . since @xmath17 is hereditary , all maps in @xmath19 are linear combinations of maps of these two kinds . the hom - hammock of an indecomposable object ( that is , the support of @xmath48 ) is illustrated in figure [ figure : homhammock ] . note that the two components in the figure wrap around the tube and intersect . moreover , if @xmath49 , then each component intersects itself , possibly with several layers , and therefore there exist hom - spaces of arbitrary finite dimension between indecomposables . . there are @xmath17-maps to the indecomposables in the right component , and @xmath18-maps to the indecomposables in the left component.,width=340 ] so for indecomposable @xmath7 and @xmath41 , the existence of a @xmath18-map @xmath50 is equivalent to the existence of a @xmath17-map @xmath51 . the following lemma is then easily verified : [ lem : dendo ] let @xmath7 be an indecomposable object of @xmath2 . then there exists a @xmath18-endomorphism of @xmath7 if and only if @xmath52 . we will need the following lemma on the relationship between @xmath17-maps and @xmath18-maps : [ lem : rayfactoring ] for @xmath7 , @xmath41 and @xmath53 indecomposable objects in @xmath2 , we have the following : * assume that @xmath54 and @xmath55 . if there are non - zero @xmath18-maps @xmath56 and @xmath57 , and an irreducible map @xmath58 , then @xmath59 up to multiplication by a non - zero scalar . * assume that @xmath54 . if there are non - zero @xmath18-maps @xmath60 and @xmath56 , and an irreducible map @xmath61 , then @xmath62 , up to multiplication by a non - zero scalar . [ rem : rayfactoring ] note that by repeated application , the same applies to compositions of irreducible maps , i. e. to all @xmath17-maps , under the assumption that the required hom - spaces are non - zero for each indecomposable that the composition factors through . * ( i ) : * we lift the maps to the derived category @xmath18 , and denote by @xmath63 , @xmath64 and @xmath65 the preimages of the objects in @xmath17 . since @xmath7 and @xmath41 have quasilength @xmath66 , we have that the space @xmath67)$ ] of @xmath18-maps is at most one - dimensional for any indecomposable @xmath68 , and similarly for @xmath41 . the aim is to show that the map @xmath69 ) : \hom_{\d}(\hat{y},\tau^{-1}\hat{z}[1 ] ) \to \hom_{\d}(\hat{x},\tau^{-1}\hat{z}[1])\ ] ] is surjective . we can view this as a map @xmath70 or , by duality and the ar - formula , @xmath71 which we now wish to show is injective . but this is clear from the structure of the tube when the hom - spaces are non - zero . * ( ii ) : * we need to show that the map @xmath72 ) : \hom_{\d}(\hat{x},\tau^{-1}\hat{y}[1])\to \hom_{\d}(\hat{x},\tau^{-1}\hat{z}[1])\ ] ] is surjective . similarly as above , by duality this is equivalent to the map @xmath73 being injective . but by the combinatorics of the tube , this is clearly an isomorphism , since by assumption both spaces are 1-dimensional . let @xmath74 , and choose some indecomposable @xmath7 in @xmath3 with quasilength @xmath75 . let @xmath76 be a linearly oriented quiver with underlying graph the dynkin diagram @xmath77 . then the category @xmath78 of finitely generated modules over the path algebra @xmath79 is naturally equivalent with the subcategory @xmath80 of @xmath3 . embedding to @xmath15 and projecting to @xmath2 we find that @xmath78 embeds into the subcategory @xmath81 of @xmath2 . the image is the subcategory @xmath82 obtained by deleting the @xmath18-maps from @xmath83 . from now on , we will drop the subscript when we speak of an additive hull as a set of objects , since there is a bijection between the objects of @xmath17 and those of @xmath19 . the triangulated category @xmath19 is a 2-calabi - yau category , which in particular means that for any two objects @xmath7 and @xmath41 , we have symmetric ext - spaces : @xmath84 two indecomposable objects @xmath7 and @xmath41 will be called _ compatible _ if @xmath85 . it is worth noticing that @xmath7 and @xmath41 are compatible if and only if @xmath86 . in an abelian or triangulated category @xmath87 , an object @xmath5 is called _ rigid _ if it satisfies @xmath88 . if it is maximal with respect to this property , that is if @xmath89 implies that @xmath90 , then it is called _ maximal rigid_. the maximal rigid objects in the cluster tube @xmath19 do not satisfy the stronger condition of _ cluster tilting _ , see @xcite . the following description of the maximal rigid objects was given in @xcite : [ prop : tiltedbijection ] there is a natural bijection between the set of maximal rigid objects in @xmath2 and the set @xmath91 where @xmath92 is a linearly oriented quiver with the dynkin diagram @xmath93 as its underlying graph . the proposition follows from the following considerations , which will be needed for the rest of the paper : all summands of a maximal rigid object in @xmath2 are concentrated in the wing @xmath94 determined by a _ top summand _ @xmath95 with @xmath96 . now the claim follows from the embedding of @xmath97 into @xmath98 , since it is easily seen that @xmath99 for two indecomposables @xmath7 and @xmath41 in @xmath94 vanishes if and only if both @xmath100 and @xmath101 vanish , where @xmath102 and @xmath103 are the corresponding @xmath104-modules . since there are @xmath10 choices for the top summand , this provides the bijection . with the description of the maximal rigid objects of @xmath19 presented in section [ sec : objects ] , we now proceed to determine their endomorphism rings in terms of quivers and relations . let @xmath5 be a maximal rigid object in the cluster tube @xmath2 , and let @xmath105 denote the tilting @xmath104-module associated with @xmath5 according to proposition [ prop : tiltedbijection ] . since the module category of a hereditary algebra @xmath20 sits naturally embedded in the cluster category @xmath21 , we can think of the module @xmath105 as a cluster - tilting object in @xmath106 . the endomorphism ring , or cluster - tilted algebra , @xmath107 can easily be found from the tilted algebra @xmath108 , by the results in @xcite , or more generally in @xcite . every minimal relation on the quiver of a tilted algebra of type @xmath109 is a zero relation of length two . the quiver of the cluster - tilted algebra is then obtained by inserting an arrow @xmath110 from the end vertex to the start vertex of each defining relation path @xmath111 . the relations for the cluster - tilted algebra are , as prescribed by @xcite , the compositions of any two arrows in any of the 3-cycles formed by adding the new arrows . we can now formulate the main theorem of this section . [ theorem : endoring ] let @xmath5 be a maximal rigid object in @xmath2 . then the endomorphism ring @xmath112 is isomorphic to the algebra @xmath113 where * @xmath114 is the quiver obtained from the quiver of @xmath115 by adjoining a loop @xmath116 to the vertex corresponding to the projective - injective @xmath104-module . * @xmath117 is the ideal generated by the relations in @xmath115 and in addition @xmath118 . before we can present the proof of the theorem , we need some considerations on the combinatorial structure of the maximal rigid objects , so we postpone the proof until we have established a few lemmas . we define a _ non - degenerate subwing triple _ @xmath119 to be a triple @xmath120 of indecomposables in @xmath19 with @xmath121 such that if @xmath122 , then @xmath123 and @xmath124 for some @xmath125 . this means that @xmath7 is on the ray @xmath126 and on the coray @xmath127 , so in particular @xmath128 and @xmath129 are contained in @xmath35 . moreover , there is exactly one quasisimple @xmath130 which is in @xmath35 but not in @xmath131 . see figure [ figure : subwingtriple ] . a _ degenerate subwing triple _ @xmath119 is a triple with @xmath132 such that if @xmath122 , then either @xmath133 and @xmath134 or @xmath135 and @xmath136 . note that any subwing triple ( degenerate or non - degenerate ) is determined by the top indecomposable @xmath7 and the unique quasisimple which is not in any of the two subwings @xmath128 or @xmath129 . . if @xmath122 and @xmath123 with @xmath137 , then @xmath124.,width=377 ] [ lem : subwingmaps ] let @xmath119 be a non - degenerate subwing triple . let @xmath138 and @xmath139 . * there are no @xmath17-maps @xmath140 . * there are no @xmath17-maps @xmath141 . * there is a @xmath18-map @xmath140 if and only if @xmath142 is on the left edge of @xmath129 and @xmath143 is on the right edge of @xmath129 . in this case , this map factors through the @xmath18-map @xmath144 . * there is a @xmath18-map @xmath141 if and only if @xmath142 is on the right edge of @xmath129 and @xmath143 is on the left edge of @xmath128 , and @xmath145 . in this case , this map factors through the @xmath18-endomorphism of @xmath7 . claims ( i ) and ( ii ) are easily verified ; one must keep in mind that @xmath146 by the definition of subwing triples . since the existence of a @xmath18-map @xmath140 is equivalent to the existence of a @xmath17-map @xmath147 , we see that the only way such a map can arise is when @xmath142 is on the left edge of @xmath129 , and @xmath143 is on the right edge of @xmath128 . now by lemma [ lem : rayfactoring ] and remark [ rem : rayfactoring ] , this @xmath18-map factors through the ray @xmath148 . in particular , it factors @xmath149 where @xmath150 is the @xmath17-map from @xmath142 to @xmath53 , and @xmath151 is the unique ( up to multiplication with scalars ) @xmath18-map @xmath152 . applying lemma [ lem : rayfactoring ] to @xmath151 , we find that it factors @xmath153 , where @xmath154 is the @xmath18-map from @xmath53 to @xmath41 and @xmath155 is the @xmath17-map from @xmath41 to @xmath143 . so claim ( iii ) holds . for claim ( iv ) , observe that since @xmath156 , a necessary condition for the existence of a @xmath17-map @xmath157 is that @xmath158 . moreover , we have that @xmath142 must be on the right edge of @xmath35 and @xmath143 must be on the left edge of @xmath35 . now the claim is proved using a similar argument as for ( iii ) and the fact from lemma [ lem : dendo ] that if @xmath145 then @xmath7 has a @xmath18-endomorphism . [ lem : subwingext ] let @xmath119 be a subwing triple , and let @xmath159 . * @xmath41 and @xmath160 are compatible if and only if @xmath161 or @xmath162 . * @xmath53 and @xmath160 are compatible if and only if @xmath161 or @xmath163 . by the 2-calabi - yau property , we have symmetric ext - groups , so it is enough to check vanishing of @xmath164 and @xmath165 . for this , we use the ar - formula @xmath166 and similarly for @xmath53 . then consider the intersection of the hom - hammock of @xmath41 with @xmath167 . [ lem : maxcompatible ] let @xmath168 be a wing in @xmath2 of height @xmath169 , and let @xmath170 be a set of pairwise compatible indecomposable objects in @xmath171 . * @xmath170 has at most @xmath172 elements . * if @xmath170 has less than @xmath172 elements , there exists a set @xmath173 of @xmath172 pairwise compatible indecomposable objects , containing @xmath170 . * if @xmath170 has @xmath172 elements , then @xmath7 is an element of @xmath170 . the argument is essentially the same as for proposition [ prop : tiltedbijection ] , and the result follows from the theory of tilting modules applied to @xmath79-modules , noting that the projective - injective indecomposable is a summand of every tilting module . [ lem : subwingstructure ] let @xmath174 be some indecomposable summand of a maximal rigid object @xmath175 in @xmath2 . * there are @xmath176 indecomposable summands of @xmath5 in @xmath177 . * assume @xmath178 . then there is a subwing triple @xmath179 such that @xmath180 and @xmath181 are either summands of @xmath5 or zero , and all summands of @xmath182 which are in @xmath177 are in @xmath183 , with @xmath184 summands in @xmath185 and @xmath186 summands in @xmath187 . given a rank @xmath10 , we let @xmath188 be a maximal rigid object in @xmath2 . we proceed by reverse induction on the quasilength of the summands of @xmath5 . ( that is , from summands of larger quasilength to summands of smaller quasilength . ) this is ok , since there is a ( unique ) summand of maximal quasilength . in accordance with proposition [ prop : tiltedbijection ] , consider the top summand @xmath95 of @xmath5 , which has quasilength @xmath189 . without loss of generality assume that this has coordinates @xmath190 . the claim ( i ) holds for @xmath95 by the proof of proposition [ prop : tiltedbijection ] . assume first that there are no summands of @xmath5 among the objects @xmath191 , where @xmath192 . then all @xmath193 summands of @xmath194 are in @xmath195 . so by lemma [ lem : maxcompatible ] , the object @xmath196 must be a summand of @xmath5 . hence in this situation claim ( ii ) also holds , with the triple @xmath197 . a similar argument shows that none of the objects @xmath198 , where @xmath199 , are summands if and only if @xmath200 is a summand , and thus ( ii ) holds with @xmath201 . suppose therefore that @xmath200 is not a summand , and that there is at least one summand of @xmath5 with coordinates @xmath191 , where @xmath202 . let @xmath203 be the one of these with highest quasi - length ( that is , maximal @xmath204 ) . consider the subwing triple @xmath205 where @xmath206 . by lemma [ lem : subwingext ] and the maximality of @xmath207 , all @xmath193 summands of @xmath194 must be in @xmath208 . then it follows from lemma [ lem : maxcompatible ] that there must be @xmath207 summands in @xmath94 and @xmath209 summands in @xmath35 and moreover that @xmath7 is indeed a summand of @xmath5 . we conclude that both claims ( i ) and ( ii ) hold for the top summand . assume now that @xmath174 is some summand of @xmath5 with @xmath178 , but not the top summand . let @xmath210 be a summand of @xmath5 of smallest quasilength with @xmath211 such that @xmath212 . ( such a summand exists , since all summands are in @xmath94 . ) then by induction , the claims hold for @xmath213 , and by the minimality in the choice of @xmath213 , the subwing triple corresponding to @xmath213 is @xmath214 where either @xmath215 or @xmath216 is @xmath174 , and the other one is also a summand . in any case , since by the induction hypothesis claim ( ii ) holds for @xmath213 , there are @xmath176 summands of @xmath5 in @xmath177 , and so ( i ) holds for @xmath174 . now that we know that there are @xmath217 summands of @xmath182 in @xmath177 , we can prove that ( ii ) holds by the same arguments as for @xmath95 above . a subwing triple @xmath179 such that @xmath218 and @xmath181 are summands of a maximal rigid object @xmath5 will be called a @xmath5-_subwing triple_. with lemma [ lem : subwingstructure ] we have obtained a combinatorial description of the maximal rigid objects as a system of subwing triples partially ordered by inclusion . note that for a maximal rigid @xmath5 with top summand @xmath95 there is a natural map from the objects in @xmath94 to the summands of @xmath5 given by sending an object @xmath7 to the summand @xmath219 with smallest quasilength such that @xmath220 . the restriction of this map to the set of quasisimples is a bijection . see figure [ fig : example ] a ) for an example . also , we have the following : [ lem : subwingintersection ] let @xmath221 be summands of a maximal rigid @xmath5 . then either @xmath222 , or @xmath223 , or @xmath224 and @xmath225 have empty intersection . suppose @xmath226 and @xmath227 . let @xmath174 be a summand of @xmath5 such that both @xmath228 and @xmath229 , and which has minimal quasilength among summands with this property . there is some non - degenerate @xmath5-subwing triple @xmath179 , and by lemma [ lem : subwingstructure ] , the summands @xmath230 and @xmath213 must be in @xmath183 . now by the minimality in the choice of @xmath174 , we know that @xmath230 must be in @xmath185 and @xmath213 in @xmath187 or vice versa . it follows that @xmath224 and @xmath225 have zero intersection , since @xmath185 and @xmath187 have empty intersection . [ lem : tmapsonrays ] let @xmath230 and @xmath213 be summands of a maximal rigid @xmath5 . then the following are equivalent . * there exists a non - zero @xmath17-map @xmath231 . * either @xmath213 is on the ray @xmath232 or @xmath230 is on the coray @xmath233 . this follows from the observation that since @xmath230 and @xmath213 are ext - orthogonal , there is no map @xmath234 . [ lem : tsubwingtmaps ] let @xmath5 be a maximal rigid object . * if @xmath235 is a non - degenerate @xmath5-subwing triple , there is a @xmath17-map @xmath236 and a @xmath17-map @xmath237 , and these maps are irreducible in @xmath238 . * if @xmath239 a degenerate @xmath5-subwing triple , there is a @xmath17-map @xmath236 , and this is irreducible in @xmath240 . a similar statement holds for a degenerate @xmath5-subwing triple @xmath241 . * there are no other irreducible @xmath17-maps in @xmath238 than those described in ( i ) and ( ii ) . * if @xmath235 is non - degenerate , the composition @xmath242 is zero . claims ( i ) and ( ii ) are obvious , so consider summands @xmath243 and assume that there is a @xmath17-map @xmath244 . by lemma [ lem : tmapsonrays ] , either the summand @xmath245 must be on the ray @xmath246 or @xmath219 must be on the coray @xmath247 . assume the former case , so @xmath219 is on the left edge of @xmath248 . we know from lemma [ lem : subwingstructure ] that there is some @xmath5-subwing triple @xmath249 , where @xmath219 must necessarily be on the left edge of @xmath250 . ( if it was on the right , it would violate lemma [ lem : subwingintersection ] . ) then clearly @xmath251 factors through the @xmath17-map @xmath252 . a similar argument can be given in the case where @xmath219 is on the coray @xmath247 by considering a @xmath5-subwing triple @xmath253 , so ( iii ) holds as well . the claim ( iv ) also follows from lemma [ lem : tmapsonrays ] . [ lem : tsubwingdmaps ] let @xmath221 be summands of a maximal rigid @xmath5 . then the following are equivalent . * there is a non - zero @xmath18-map @xmath254 . * one of the following is satisfied * * there is some non - degenerate @xmath5-subwing triple @xmath255 such that @xmath230 is on the left edge of @xmath250 and @xmath213 is on the right edge of @xmath248 . * * @xmath230 is on the left edge of @xmath94 and @xmath213 is on the right edge of @xmath94 , where @xmath95 is the top summand . assume first that @xmath256 , the top summand . then there is a @xmath18-map @xmath257 if and only if @xmath213 is on the right edge of @xmath94 , which is a special case of b ) . similarly , if @xmath258 , then there is a @xmath18-map @xmath254 if and only if @xmath230 is on the left edge of @xmath94 . so under the assumption that at least one of @xmath221 is the top summand , the claim holds . consider therefore the case where neither @xmath230 nor @xmath213 is the top summand . it is easily seen that if @xmath223 or @xmath259 , there can be no @xmath18-map @xmath254 . so consider the summand @xmath219 of minimal quasilength such that both @xmath230 and @xmath213 are in @xmath260 . by lemma [ lem : subwingstructure ] there is a @xmath5-subwing triple @xmath255 , and by the minimality of @xmath219 , either @xmath261 and @xmath262 or the other way around . the claim now follows from lemma [ lem : subwingmaps ] . we can now prove the main theorem . in the argument , we will implicitly use the fact that for indecomposable objects @xmath7 and @xmath41 with @xmath263 , there is up to multiplication by scalars at most one @xmath17-map and one @xmath18-map from @xmath7 to @xmath41 . our task is to determine the quiver @xmath114 and defining relations of @xmath264 . first recall that the functor @xmath265 induces an equivalence between @xmath238 and the category @xmath266 of projective @xmath267-modules . the vertices in @xmath114 are therefore in bijection with the indecomposable summands of @xmath5 , and the arrows correspond to maps which are irreducible in @xmath238 . no @xmath17-map can factor through a @xmath18-map , so let us first consider the arrows in @xmath114 coming from @xmath17-maps and their relations , that is the endomorphism ring of @xmath5 as an object in the subcategory @xmath268 , where @xmath95 is the top summand of @xmath5 . by virtue of the equivalence of this category with the module category @xmath97 , we know that this will be the quiver with relations for the tilted algebra @xmath269 . by lemma [ lem : tsubwingtmaps ] part ( i ) , for each non - degenerate @xmath5-subwing triple @xmath235 there exist @xmath17-maps @xmath270 and @xmath271 which are irreducible in @xmath238 and hence correspond to arrows @xmath272 and @xmath273 in @xmath114 . similarly , by part ( ii ) , for each degenerate @xmath5-subwing triple @xmath239 there is an arrow @xmath272 , and for a @xmath5-subwing triple @xmath241 there is an arrow @xmath273 . moreover , by part ( iii ) of the same lemma , these are the only arrows in @xmath114 coming from @xmath17-maps . assuming that the triple is non - degenerate , by part ( iv ) , the composition @xmath274 is zero , and hence @xmath275 is a zero relation for the quiver . it follows from lemma [ lem : tmapsonrays ] that there are no other minimal relations on the arrows coming from @xmath17-maps , since a path @xmath276 such that no @xmath277 comes from a composition @xmath278 from a non - degenerate @xmath5-subwing triple @xmath279 corresponds to a map following a ray or a coray . so the above gives a description of the tilted algebra @xmath269 . see figure [ fig : example ] b ) for an example . now consider the @xmath18-maps , and postpone for a moment the situation with maps to or from the top summand @xmath95 . by lemma [ lem : subwingmaps ] part ( iii ) , for each non - degenerate @xmath5-subwing triple @xmath235 there is a @xmath18-map @xmath280 . we claim that this map is irreducible in @xmath238 . so assume that there exists a summand @xmath219 , not isomorphic to @xmath174 or @xmath213 , and a @xmath18-map @xmath281 such that @xmath282 where @xmath283 . since the composition of two @xmath18-maps is zero , @xmath284 must be a @xmath17-map . so by lemma [ lem : tmapsonrays ] we either have that @xmath219 is on @xmath285 , or that @xmath174 is on @xmath286 . the former case contradicts lemma [ lem : subwingintersection ] , since @xmath174 is on the right side of @xmath224 and on the left side of @xmath260 . the latter case contradicts lemma [ lem : subwingmaps ] , since @xmath219 is on the right edge of @xmath177 . a similar argument shows that we can not have a factorisation @xmath287 where @xmath288 is a @xmath17-map and @xmath289 is a @xmath18-map . thus @xmath290 is irreducible and corresponds to an arrow @xmath291 in @xmath114 . for the non - degenerate @xmath5-subwing triple @xmath235 , the composition @xmath292 is a @xmath18-map . we claim that it must be zero . assume therefore that there is a @xmath17-map from @xmath213 to @xmath293 . then , since @xmath230 sits on @xmath294 and there is no @xmath17-map @xmath295 by the ext - orthogonality of @xmath230 and @xmath213 , there must be @xmath17-maps from @xmath213 to _ all _ indecomposables of quasilength @xmath296 , _ except _ @xmath297 . but this would require @xmath298 , and this is impossible , since @xmath299 . we conclude that there is no @xmath18-map @xmath231 , and the composition is zero . similarly , the composition @xmath300 is also zero , since there is no @xmath17-map from @xmath230 to @xmath301 . it follows that the paths @xmath302 and @xmath303 are zero relations on the quiver @xmath114 . by lemma [ lem : dendo ] , there is a @xmath18-map @xmath304 which is an endomorphism of the top summand @xmath95 . this map must be irreducible in @xmath238 , for the only objects in @xmath94 to which there are maps from @xmath95 are the ones on the right edge of @xmath94 , but there are no maps from any of these to @xmath95 . thus there is a loop @xmath116 on the vertex corresponding to the top summand . the composition @xmath305 is zero , since any composition of two @xmath18-maps is the image of a map @xmath306 $ ] in @xmath18 , which must necessarily be zero since @xmath17 is hereditary . so @xmath118 is a zero relation on the quiver of @xmath267 . by lemmas [ lem : subwingmaps ] and [ lem : tsubwingdmaps ] , there are no other irreducible @xmath18-maps between summands of @xmath5 , and also no other minimal relations involving arrows from @xmath18-maps : let @xmath307 be an arrow corresponding to a @xmath18-map @xmath308 where @xmath235 is a non - degenerate subwing triangle . then if @xmath309 is a path in @xmath114 such that @xmath310 and @xmath311 do not traverse any relations , then necessarily @xmath312 are arrows coming from maps on the left edge of @xmath177 and @xmath313 are arrows coming from the right edge of @xmath225 . then this path corresponds to a non - zero map by lemma [ lem : tsubwingdmaps ] . we now see that the arrows @xmath314 are in bijection with the zero relations for the quiver of @xmath315 , and complete the relation paths to oriented cycles , so by @xcite or @xcite the arrows @xmath316 and @xmath314 form the quiver of @xmath317 . furthermore , the relations imposed on this quiver coincide with the relations defining @xmath115 . so the quiver of @xmath267 is obtained from the quiver of @xmath115 by adjoining the loop @xmath116 at the loop corresponding to the top summand , which again corresponds to the projective - injective @xmath315-module . also , the relations for @xmath267 are the relations for @xmath115 and in addition @xmath318 . see figure [ fig : example ] for an example of a maximal rigid object @xmath5 and the tilted algebra @xmath315 , the cluster - tilted algebra @xmath115 and the endomorphism ring @xmath267 . example of a maximal rigid object @xmath5 and the associated algebras @xmath315 , @xmath115 and @xmath267.,width=529 ] an explicit description of the quivers for cluster - tilted algebras of type @xmath109 was given in @xcite , and also in @xcite . it can be deduced from the type @xmath109 cluster category model from @xcite . they are exactly the quivers satisfying the following : * all non - trivial minimal cycles are oriented and of length 3 * any vertex has valency at most four * if a vertex has valency four , then two of its adjacent arrows belong to one 3-cycle , and the other two belong to another 3-cycle * if a vertex has valency three , then two of its adjacent arrows belong to a 3-cycle , and the third does not belong to any 3-cycle in the first condition , a _ cycle _ means a cycle in the underlying graph . a _ connecting vertex _ for such a quiver , as defined in @xcite , is a vertex which either has valency one , or has valency two and is traversed by a 3-cycle . note that for the endomorphism ring @xmath267 of a maximal rigid @xmath5 , the loop vertex is connecting for the quiver of @xmath115 . there is a sort of converse to theorem [ theorem : endoring ] , so we have the full description of the algebras which can arise : let @xmath319 be a cluster - tilted algebra of type @xmath93 , and let @xmath320 be a connecting vertex for @xmath321 . then there exists a maximal rigid object @xmath5 of @xmath2 such that @xmath322 is obtained from @xmath321 by adjoining a loop at @xmath320 . we do induction on @xmath10 . the claim is easily verified for small values . given a cluster - tilted algebra @xmath319 of type @xmath93 , let @xmath114 be its quiver and let @xmath320 be some connecting vertex of @xmath114 . then @xmath320 has valency 1 or 2 in @xmath114 . consider first the case where @xmath320 has valency 2 . then @xmath320 is traversed by a 3-cycle @xmath323 in @xmath114 . the quiver @xmath324 has two disconnected components @xmath325 and @xmath326 , where @xmath327 is connecting for @xmath328 . also , @xmath328 is the quiver of some cluster - tilted algebra @xmath329 of type @xmath330 , where @xmath331 . by induction we can assume that for each @xmath332 there exists a maximal rigid object @xmath230 in @xmath333 such that the endomorphism ring of @xmath230 is obtained from the quiver of @xmath329 by adjoining a loop at @xmath327 . let @xmath334 be the corresponding tilting @xmath335-module . we have a natural embedding of the module category @xmath336 into the wing @xmath337 , and of @xmath338 into @xmath339 . it is now easily seen that the images of the indecomposable summands of @xmath340 and @xmath341 under these embeddings are all compatible , and that the direct sum of these can be completed to a maximal rigid object @xmath5 by adding the object @xmath190 . then the quiver of @xmath5 is obtained from @xmath114 by adding a loop at @xmath320 . the case where @xmath320 has valency one in @xmath114 is easier ; one considers @xmath342 which is cluster - tilted of type @xmath343 , and uses induction similarly as above . there are in fact exactly @xmath10 maximal rigid objects in @xmath2 with the prescribed endomorphism algebra , and these form a @xmath16-orbit . in this section we show that the endomorphism rings under discussion are gentle . we use this to determine their gorenstein dimension . a finite - dimensional algebra @xmath113 where @xmath114 is a finite quiver is called _ special biserial _ @xcite if * for all vertices @xmath344 in @xmath114 , there are at most two arrows starting in @xmath344 and at most two arrows ending in @xmath344 * for every arrow @xmath345 in @xmath114 , there is at most one arrow @xmath346 in @xmath114 with @xmath347 and at most one arrow @xmath348 with @xmath349 . a special biserial algebra @xmath113 is _ gentle _ @xcite if moreover * @xmath117 is generated by paths of length 2 * for every arrow @xmath345 in @xmath114 there is at most one arrow @xmath350 such that @xmath351 is a path and @xmath352 , and at most one arrow @xmath353 such that @xmath354 is a path and @xmath355 . [ theorem : gentle ] if @xmath5 is a maximal rigid object in the cluster tube @xmath19 , then the endomorphism ring @xmath356 is gentle . it follows from the description of cluster - tilted algebras of type @xmath109 , and theorem [ theorem : endoring ] , that condition ( i ) is satisfied whenever @xmath344 is not the loop vertex . if @xmath344 is the loop vertex , the quiver generally looks locally like this : @xmath357{loopvertex.eps}\ ] ] where the relations are indicated by dashed lines and it may also happen that one of the two other vertices pictured and consequently the adjacent arrows are not there . we see that also for this vertex ( i ) is fulfilled . as proved in @xcite , any cluster - tilted algebra of type @xmath109 is gentle , and therefore if @xmath345 is not an arrow incident with the loop vertex , ( ii ) and ( iv ) are satisfied . if @xmath345 is incident with the loop vertex , ( ii ) and ( iv ) follow from the local description pictured above , with the observation that by the description of the cluster - tilted algebras there are no minimal relations involving both an arrow in the picture and an arrow outside the picture . moreover , by theorem [ theorem : endoring ] , the ideal @xmath117 is generated by paths of length two , so ( iii ) is fulfilled , and the algebra is gentle . it is known from @xcite that all gentle algebras are gorenstein , that is , a gentle algebra @xmath358 has finite injective dimension as both a left and a right @xmath358-module . this dimension is then called the gorenstein dimension of @xmath358 . in order to prove the next result , we need to recall the main result in @xcite in more detail . let @xmath359 be a gentle algebra . an arrow @xmath360 in @xmath114 is said to be _ gentle _ if there is no arrow @xmath361 such that @xmath362 is a non - zero element of @xmath117 . a _ critical path _ in @xmath114 is a path @xmath363 such that @xmath364 for all @xmath365 . [ theorem : gr ] let @xmath359 be a gentle algebra , and let @xmath366 be the supremum of the lengths of critical paths starting with a gentle arrow . ( @xmath366 is taken to be zero if there are no gentle arrows . ) * @xmath366 is bounded by the number of arrows in @xmath114 . * if @xmath367 , then @xmath358 is gorenstein of gorenstein dimension @xmath366 . * if @xmath368 , then @xmath358 is gorenstein of gorenstein dimension at most one . we can use this to find the gorenstein dimension of our algebras : [ prop : gorenstein ] let @xmath5 be a maximal rigid object in @xmath19 . if @xmath4 , the gorenstein dimension of @xmath369 is zero , and if @xmath370 , the gorenstein dimension is one . by theorem [ theorem : gentle ] , the algebra @xmath267 is gentle , and we can apply theorem [ theorem : gr ] . if @xmath4 , then @xmath5 has only one summand , and the endomorphism algebra @xmath267 is isomorphic to the self - injective algebra @xmath371/(x^2)$ ] , so in this case the gorenstein dimension is zero . assume therefore that @xmath370 . the gentle arrows are exactly the arrows which are not traversed by any minimal oriented cycle . ( in particular , the loop is not gentle . ) but if @xmath360 is any such arrow , then @xmath372 is a path for at most one arrow @xmath345 , and @xmath372 can never be a zero relation , since it is not a part of a 3-cycle . so if gentle arrows exist , the maximal length @xmath373 of critical paths starting in gentle arrows is 1 , and therefore the gorenstein dimension of @xmath267 is also 1 . if gentle arrows do not exist , we have @xmath374 , and the gorenstein dimension is at most 1 . it remains to be shown that @xmath267 can not be self - injective for @xmath375 . for this , consider the indecomposable projective associated with the loop vertex . it is easily seen that it is not injective . since @xmath376 is gentle , it is in particular a string algebra . we will use this to show that @xmath267 has finite representation type . for this , we will recall some basic facts about representations of string algebras . more details on this can be found e.g. in @xcite . let @xmath113 be a string algebra . we consider words from the alphabet formed by arrows in @xmath114 and their formal inverses . inverse words are defined in the obvious way . for an inverted arrow @xmath377 we set the end vertex @xmath378 equal to the start vertex @xmath379 of the original arrow and vice versa . a word @xmath380 , where no two consecutive @xmath381 are inverses of each other , is called a _ string _ if @xmath382 for @xmath383 and moreover no subword of @xmath384 or its inverse is a zero relation . in addition , there is a trivial string of length zero for each vertex of @xmath114 . we say that the start- and end vertices of the @xmath381 which appear in the strings are the vertices _ traversed _ by the string . for technical purposes , we also consider a unique zero ( or empty ) string of length @xmath385 . to any string @xmath386 of length @xmath387 in @xmath114 there is an associated indecomposable @xmath388-dimensional representation @xmath389 of @xmath113 given by one - dimensional vector spaces in each vertex traversed by the string ( with multiplicity ) and one - dimensional identity maps for each of the arrows ( and inverted arrows ) appearing in @xmath390 . we have that @xmath391 if and only if @xmath392 or @xmath393 . the @xmath113-modules given by such representations are called _ string modules_. if there exist closed strings @xmath394 ( i.e. strings starting and ending in the same vertex ) such that @xmath395 and powers of the string are strings as well , there will also be infinite families of indecomposables called _ band modules_. the string modules and band modules constitute a complete set of representatives of isoclasses of indecomposable modules over @xmath113 . now some remarks on the strings in the quiver of @xmath267 , which will be useful in the proof of theorem [ thm : finitetype ] and in section [ sec : homfunctor ] . in what follows , an arrow @xmath360 that arises from an irreducible @xmath17-map in @xmath238 will be called a @xmath17-arrow , and similarly an arrow that arises from a @xmath18-map will be called a @xmath18-arrow . since the vertices of the quiver are in a natural bijection with the indecomposable summands of @xmath5 , we will transfer some terminology about the summands to the vertices . so the quasilength of a vertex is the quasilength of the corresponding summand , and vertices ( also arrows , strings ) are said to be in a wing if the corresponding summands are in the wing . [ lem : stringsandcycles ] let @xmath235 be a non - degenerate @xmath5-subwing triple , and @xmath396 and @xmath307 the corresponding arrows . then * if @xmath397 is a string , then @xmath398 is not of the form @xmath399 . * if @xmath400 is a string , then @xmath401 is not of the form @xmath402 . the analogous statements hold for the inverses of the arrows . the assertions follow from the fact that @xmath302 and @xmath303 are both zero relations . * if @xmath403 is a @xmath18-arrow , and @xmath404 is a string , then @xmath401 is in the wing of @xmath405 and @xmath398 is in the wing of @xmath204 , and similarly for inverses of @xmath18-arrows . * a string in the quiver of @xmath267 contains at most one @xmath18-arrow or inverse of such . * ( i ) : * let @xmath404 be a string , where @xmath403 is a @xmath18-arrow corresponding to a @xmath18-map @xmath406 . suppose first that @xmath407 . then there is a subwing triple @xmath408 . now if the string @xmath401 which starts in @xmath405 is the trivial string for vertex @xmath405 , then there is nothing to show , since @xmath405 is definitely in the wing of itself . assume therefore that @xmath401 has length at least one . then @xmath409 , where @xmath410 is either an arrow starting in @xmath405 or the inverse of an arrow ending in @xmath405 . by lemma [ lem : stringsandcycles ] we know that @xmath410 can not be the arrow @xmath411 which connects @xmath405 to the vertex @xmath8 associated with @xmath174 . now the remaining possibilities for @xmath410 are contained in the wing of @xmath405 . by lemma [ lem : stringsandcycles ] again , it follows that none of the @xmath412 can be @xmath18-arrows ( or inverse @xmath18-arrows ) . so @xmath401 traverses vertices of successively smaller quasilength and can not return to @xmath405 . so @xmath401 is in the wing of @xmath405 . if @xmath413 , then @xmath230 is the top summand and @xmath345 is the loop , in which case the claim follows by the fact that @xmath414 is a zero relation , and a similar argument as above . the statement for @xmath398 and @xmath204 is proved analogously , and also the statements for inverses of @xmath18-arrows . * ( ii ) : * this follows from ( i ) . if @xmath415 and @xmath416 are two @xmath18-arrows , then clearly @xmath417 can not be in the wing of @xmath204 or @xmath405 when at the same time @xmath418 is in the wing of @xmath8 or @xmath419 . so they can not both appear in the same string . the same goes if one ( or both ) of @xmath418 and @xmath417 is the inverse of a @xmath18-arrow . [ rem : closed ] it follows from this lemma that if @xmath390 is a closed string , then either @xmath390 is a trivial string , or @xmath390 contains the loop as the only @xmath18-arrow . [ lem : stringsubwing ] in the following , we consider strings only up to orientation . * the strings of length @xmath420 in the quiver of @xmath267 which do not contain a @xmath18-arrow or inverse of such are in bijection with sequences @xmath421 such that there are subwing triples @xmath422 or @xmath423 for @xmath424 . * the strings in the quiver of @xmath267 that do not contain a @xmath18-arrow or inverse of such are in bijection with pairs @xmath221 of summands of @xmath5 such that @xmath259 . * ( i ) : * let @xmath390 be a string without a @xmath18-arrow or inverse @xmath18-arrow . if @xmath390 is trivial , the claim is obviously true , so assume @xmath390 has length @xmath425 . choose @xmath426 to be an indecomposable summand of @xmath5 which corresponds to a vertex @xmath427 traversed by @xmath390 such that no of the other vertices traversed by @xmath390 have higher quasilength . assume @xmath428 . then there is some @xmath5-subwing triple @xmath429 . let @xmath430 and @xmath431 be the corresponding arrows . now since there are no @xmath18-arrows , and @xmath427 has maximal quasilength among the vertices traversed by @xmath390 , the string @xmath390 must be of the form @xmath432 where @xmath398 , if it is non - trivial , is of the form @xmath433 or @xmath434 , and similarly @xmath435 or @xmath436 if it is non - trivial . so for @xmath437 to be a string , one of these has to be the trivial string associated with @xmath427 , since the composition @xmath438 is zero . so @xmath427 is the start- or end vertex of @xmath390 , and the first arrow ( or inverse arrow ) connects @xmath427 to one of the vertices from the @xmath5-subwing triple with @xmath427 on top . by repeating the process with the string @xmath439 or @xmath440 , we get the desired chain of subwing triples . * ( ii ) : * follows directly from ( i ) . see figure [ fig : stringexample ] for an example of strings in the quiver of @xmath267 . , where @xmath441 is the only @xmath18-arrow , starts in the vertex corresponding to @xmath219 and ends in the vertex corresponding to @xmath245 . note how a path like @xmath442 ( or its inverse ) can not be in a string , since it is a zero relation . also , the path @xmath443 is not a string , while both @xmath444 and @xmath445 are.,width=415 ] [ thm : finitetype ] for a maximal rigid object @xmath5 in @xmath2 , the endomorphism ring @xmath267 is of finite type , and the number of indecomposable representations is @xmath446 let @xmath114 be the quiver of @xmath267 and @xmath117 the relation ideal . moreover , let @xmath419 denote the loop vertex , and @xmath116 the loop itself . for a string @xmath390 we will denote the associated indecomposable representation by @xmath389 . first consider strings which do not involve the loop . these are in bijection with ordered pairs of vertices : for each pair @xmath447 from @xmath448 , let @xmath221 be the associated summands of @xmath5 . if @xmath259 or vice versa , there is , as in lemma [ lem : stringsubwing ] , a string without a @xmath18-arrow connecting @xmath204 and @xmath405 . so suppose this is not the case . consider the ( unique ) summand @xmath174 which is of minimal quasilength such that @xmath230 and @xmath213 are both in @xmath177 . then there is some @xmath5-subwing triple @xmath449 where @xmath450 and @xmath451 or the other way around . now use lemma [ lem : stringsubwing ] again to find strings connecting @xmath204 with the vertex @xmath452 associated with @xmath453 and @xmath405 with the vertex @xmath454 associated with @xmath455 . now there is a @xmath18-arrow @xmath456 . connecting the two string by way of @xmath457 yields the desired string , and we can choose the orientation , so this is up to ordered pairs of vertices . in particular , by remark [ rem : closed ] , any non - loop string starting and ending in the same vertex is a trivial string . denote by @xmath458 the unique non - loop string starting in @xmath204 and ending in @xmath405 , so @xmath459 . for the corresponding indecomposable representations we have isomorphisms @xmath460 . the simple representations are the @xmath461 . the total number of representations corresponding to non - loop strings is therefore @xmath462 where the first term is the number of simple representations and the second is the number of strings of length @xmath425 up to orientation . now for the strings passing through the loop . for each pair @xmath447 of vertices , there are two strings from @xmath204 to @xmath405 passing through the loop @xmath116 : @xmath463 and these are all possible loop strings . we see that @xmath464 , so @xmath465 for any choice of @xmath447 , in particular for @xmath466 . we deduce that the indecomposable representations associated with loop strings are in bijection with ordered pairs of vertices , and the number of such representations is @xmath467 . these are all the strings there are . since ( by remark [ rem : closed ] and lemma [ lem : stringsubwing ] ) the only closed strings are either trivial , @xmath116 or of the form @xmath468 for some arrow ( or inverse arrow ) @xmath360 and some string @xmath390 , it follows that there are no band modules . so the string modules we have presented form a complete set of isomorphism classes of @xmath267-modules . summarising , we find that the total number of representations is @xmath469 for a cluster - tilted algebra @xmath470 arising from the cluster category @xmath21 of some hereditary algebra @xmath20 , there is a close connection between the module category of @xmath471 and the cluster category itself . the main theorem from @xcite says that the functor @xmath472 induces an equivalence @xmath473 in particular , the cluster - tilted algebra is of finite representation type if and only if @xmath21 has finitely many objects ( which again happens if and only if @xmath20 is of finite type ) . by theorem [ thm : finitetype ] , a similar result can not hold for the cluster tubes . the analogous argument fails because the hom - functor is not full . in this section , we will study some properties of this functor . we introduce some notation . for any indecomposable object @xmath7 in @xmath19 , let @xmath474 be the hom - hammock of @xmath7 , where @xmath475 is the set of indecomposables to which @xmath7 has @xmath17-maps , and similarly for @xmath476 . also , consider the _ reverse hom - hammock _ @xmath477 , that is , the support of @xmath478 among the indecomposables . like the ordinary hom - hammock , this has a natural structure as the union of two components , one denoted by @xmath479 containing the indecomposables that have non - zero @xmath17-maps to @xmath7 , and another one denoted by @xmath480 containing those that have @xmath18-maps to @xmath7 . note that by the description of the hom - hammocks , @xmath481 and @xmath482 . so the shape of @xmath483 is similar to the shape of @xmath484 ( figure [ figure : homhammock ] ) . for the remainder of this chapter , @xmath485 will be a maximal rigid object in @xmath2 , and we assume that the top summand of @xmath5 is @xmath486 . clearly , by redefining the coordinates we can use the results for all maximal rigid objects of @xmath19 . as in the preceeding sections , we will denote by @xmath267 the endomorphism ring @xmath376 . we define @xmath487 to be a certain set of indecomposable objects in @xmath19 : @xmath488 see figure [ fig : fundamentaldomain ] . the region @xmath487 in the tube consists of the rigid part and in addition a triangle of height @xmath189 in the non - rigid part . ( we have defined wings only for rigid indecomposables , but we can think of @xmath487 as the wing of the object @xmath489 . ) in @xmath490 , below the dashed curve . @xmath5 is concentrated in the indicated wing.,width=264 ] the following claims are easily verified : [ lem : onecomponent ] if @xmath7 is an indecomposable in @xmath94 , then @xmath491 forms one rectangle - shaped subgraph of the tube , and similarly for @xmath492 . [ lem : revhammquasi ] let @xmath493 be either @xmath17 or @xmath18 . then for an indecomposable @xmath7 , the set @xmath494 contains a unique quasisimple @xmath495 , and a necessary condition for an object @xmath41 to be in @xmath494 is that @xmath496 . [ lem : t1inr(x ) ] let @xmath497 . then @xmath498 if and only if @xmath499 . we now want to assign to each indecomposable in @xmath487 a uniquely defined string in the quiver of @xmath267 . in the main result of this section we will show that the images under the hom - functor are given by these strings . the first step is to encode information about @xmath17-maps and @xmath18-maps in separate strings , which will be joined to one string at a later stage . [ lem : revhamm ] let @xmath493 be either @xmath17 or @xmath18 . for @xmath497 we have the following . * @xmath500 is empty if and only if @xmath501 . * for any @xmath5-subwing triple @xmath235 , at most one of @xmath213 and @xmath174 can be in @xmath494 . * if @xmath502 is non - empty , there is a unique string in the quiver of @xmath267 traversing each of the vertices corresponding to the indecomposables in @xmath502 exactly once ( and no other vertex ) and ending in the vertex corresponding to the summand in @xmath502 of highest quasilength . * a string of the type in part ( iii ) contains no @xmath18-arrow ( or inverse of such ) . * ( i ) : * we need to show that @xmath503 if and only if @xmath504 . if @xmath501 , then there are no non - zero maps from @xmath5 to @xmath7 since @xmath5 is rigid . for the converse , assume that the intersection is empty . using lemma [ lem : t1inr(x ) ] , we get that @xmath156 . moreover , we have that @xmath505 since @xmath506 , the object @xmath7 , and consequently @xmath507 , is rigid . so @xmath508 for some @xmath204 since @xmath5 is maximal rigid , and we can conclude that @xmath509 . * ( ii ) : * we know that @xmath225 and @xmath177 are disjoint . the claim then follows from lemma [ lem : revhammquasi ] . * ( iii ) : * assume @xmath502 is non - empty , and let @xmath510 and @xmath511 be elements in this set with minimal and maximal quasilength , respectively . by lemma [ lem : revhammquasi ] , the unique quasisimple @xmath495 which is in @xmath494 is now in both @xmath512 and @xmath513 . so in particular @xmath512 and @xmath513 have non - empty intersection , and therefore by lemma [ lem : subwingintersection ] we know that @xmath514 . also by lemma [ lem : subwingintersection ] we see that @xmath511 and @xmath510 are uniquely determined . there is some @xmath5-subwing triple @xmath515 , and it can easily be seen that if @xmath495 were in @xmath516 , say , then also @xmath517 would be in @xmath494 , which would violate the minimality condition on @xmath510 . thus @xmath510 is the summand of smallest quasilength such that @xmath495 is in the corresponding wing . see figure [ fig : revhammandsubwings ] . and @xmath511 are the summands of @xmath5 in @xmath494 of lowest and highest quasilength , respectively , then the entire rectangle indicated must be in @xmath494.,width=302 ] also , for summands @xmath518 we have that @xmath519 if and only if @xmath520 , again by lemma [ lem : subwingintersection ] and the maximality of @xmath511 . now the desired string is a string of the type described in lemma [ lem : stringsubwing ] , oriented in the suitable way . * ( iv ) : * by part ( ii ) , a @xmath18-arrow associated with a non - degenerate @xmath5-subwing triple @xmath235 could not be traversed by a string of the type described in part ( iii ) . moreover , the loop is disallowed as well , since then the loop vertex would be traversed twice , contrary to the condition in part ( iii ) . for an indecomposable @xmath497 such that @xmath502 is non - empty , where @xmath493 is either @xmath17 or @xmath18 , we denote the string in lemma [ lem : revhamm ] part ( iii ) by @xmath521 . if the intersection is empty , we define @xmath521 to be the zero string . the two next lemmas tell us that different objects in @xmath522 can be distinguished by their associated strings . [ lem : differentstrings ] let @xmath523 . if @xmath524 and @xmath525 , then @xmath526 . we note first that for any indecomposable @xmath53 , the unique quasisimple @xmath527 in @xmath528 determines the first coordinate of @xmath53 . similarly , the unique quasisimple @xmath529 in @xmath530 determines the sum of the coordinates of @xmath53 modulo the rank @xmath10 . so if the first coordinate of @xmath53 is known , the quasisimple @xmath529 determines the second coordinate _ modulo @xmath10_. let now @xmath7 and @xmath41 be in @xmath522 such that @xmath531 and @xmath525 . we aim to show that @xmath7 and @xmath41 must be equal . by lemma [ lem : revhamm ] part ( i ) , at least one of @xmath531 and @xmath525 is non - zero . assume first that both are non - zero . we claim that @xmath532 and @xmath533 . as in the proof of lemma [ lem : revhamm ] part ( iii ) , we observe that if @xmath510 is the @xmath5-summand of smallest quasilength in @xmath479 , then @xmath534 is the unique quasisimple which is in @xmath512 but not in the wing of any @xmath5-summand of smaller quasilength . since @xmath510 is also the summand of smallest quasilength in @xmath535 , we must have @xmath532 . similarly , we deduce that @xmath533 . thus @xmath7 and @xmath41 have the same first coordinate , and the same second coordinate modulo @xmath10 . but since @xmath7 and @xmath41 are in @xmath487 , this means that unless @xmath7 and @xmath41 are equal , one of them is in @xmath536 and the other is in the non - rigid part . if they are not equal , there is then a contradiction to lemma [ lem : t1inr(x ) ] : if @xmath7 is in @xmath536 and @xmath41 is in the non - rigid part , then by lemma [ lem : t1inr(x ) ] @xmath537 but @xmath538 , which is impossible since we have assumed @xmath539 and @xmath540 . we conclude that if both @xmath541 and @xmath542 , then @xmath526 . assume then that @xmath541 and @xmath543 , and furthermore that @xmath544 . then @xmath7 and @xmath41 have the same first coordinate . moreover , at least one of @xmath7 and @xmath41 must be in @xmath536 , since the only other possible positions for an object @xmath545 such that @xmath546 and @xmath547 are on the coray @xmath548 . suppose ( without loss of generality ) that @xmath7 has smaller quasilength than @xmath41 . so in particular , @xmath7 is in @xmath536 . let @xmath511 be the object in @xmath549 with highest quasilength . since @xmath550 , we have that @xmath551 . therefore , there is some @xmath5-subwing triple @xmath552 , where @xmath511 is necessarily on the left side since @xmath479 contains whole corays , and so by the maximality of @xmath511 , there can be no more summands of @xmath5 on @xmath553 . assume first that this triple is non - degenerate . since @xmath480 does not contain any summands of @xmath5 , there is in particular no @xmath18-map @xmath554 . so @xmath7 is in @xmath555 but not in @xmath556 . moreover , we see that @xmath557 , by the maximality of @xmath511 . so @xmath7 must be on the coray @xmath558 . if the triple is degenerate , then @xmath7 must be on the right edge of @xmath513 , since @xmath559 . in any of these two cases , we get a contradiction : since @xmath41 and @xmath7 have the same first coordinate , and @xmath41 has higher quasilength , @xmath560 must be in @xmath535 . this contradicts the equality of @xmath561 and @xmath562 , and our assumption that @xmath544 must be wrong . see figure [ fig : equalsigmas ] . and @xmath563 , then @xmath7 and @xmath41 have the same first coordinate . if @xmath7 is inside some non - degenerate @xmath5-subwing triple , then @xmath7 must be on the coray @xmath558 , since otherwise @xmath564.,width=340 ] the situation where @xmath565 and @xmath542 can be proved in a similar manner . the following lemma is used to show that if two different objects have exactly one associated @xmath566- or @xmath567-string , then the strings are different . [ lem : tstringisnotdstring ] if @xmath568 , then @xmath562 is non - zero . suppose that @xmath568 , and let @xmath230 be the summand in @xmath569 which has highest quasilength . we claim that @xmath256 . to see this , assume that @xmath570 . then there is some ( degenerate or non - degenerate ) @xmath5-subwing triple @xmath571 or @xmath572 . since @xmath479 contains whole corays , and @xmath573 contains whole rays , the summand @xmath174 must be in one of these two reverse hom - hammocks . but this contradicts our choice of @xmath230 . so @xmath256 . now observe that if @xmath41 is in @xmath487 , and there is a @xmath43-map @xmath574 , then there is also a @xmath17-map @xmath574 , so in particular @xmath575 . we now show that if both strings associated with an indecomposable in @xmath487 are non - zero , then there is a larger string containing both of them . [ lem : dstrings ] let @xmath497 . if both @xmath561 and @xmath576 are non - zero , there is a @xmath18-arrow @xmath577 from the end vertex of @xmath561 to the end vertex of @xmath576 . so in particular , @xmath578 is a well - defined string . we consider four cases , depending on the position of @xmath7 in @xmath487 . the first case is when there is a @xmath18-map @xmath579 . one readily verifies that there is then also a @xmath17-map @xmath579 , so in this case @xmath95 is in both @xmath479 and @xmath480 , and the claim holds with the loop as @xmath577 . the second case is when there is a @xmath17-map @xmath579 , but no @xmath18-map from @xmath95 to @xmath7 . this happens exactly when @xmath7 is on the coray @xmath548 , and in this case there are no @xmath18-maps from any summands of @xmath5 to @xmath7 , so @xmath580 is empty , and there is nothing to prove . the third case is when @xmath7 is located on the ray @xmath581 . then there are no @xmath17-maps from @xmath5 to @xmath7 , so again there is nothing to prove . the only remaining situation is then when @xmath7 is in the wing @xmath536 . as in the proof of lemma [ lem : differentstrings ] , let @xmath511 be the summand in @xmath582 of highest quasilength . since @xmath551 , there is some @xmath5-subwing triple @xmath552 with @xmath511 necessarily on the left , since @xmath479 contains whole corays . assume first that this triple is non - degenerate . then , since @xmath583 and @xmath584 by the maximality property of @xmath511 , we note the following about the position of @xmath7 : * @xmath585 , but not on the right edge of @xmath586 , since then there would be a @xmath17-map @xmath587 * @xmath588 , where @xmath41 is the object on @xmath589 which has an irreducible map to @xmath590 , since then there would be no @xmath17-map @xmath591 * if @xmath592 , then it is on the right edge , since otherwise there would be no @xmath17-map @xmath591 our aim is to show that @xmath590 is in @xmath480 , and moreover that it is the summand of @xmath5 with highest quasilength appearing in @xmath480 . with the above remarks about the position of @xmath7 , we see that the only allowed positions such that there is no @xmath18-map @xmath593 are positions on the coray @xmath558 . but if @xmath7 were on this coray , then @xmath580 would be empty , contrary to our hypothesis : namely , assuming this position for @xmath7 , suppose there was some summand @xmath594 . an equivalent condition to this ( cf . lemma [ lem : homspaces ] ) is that there is a non - zero @xmath17-map @xmath595 , which again is equivalent to the existence of a @xmath17-map @xmath596 . but since @xmath597 , we have that @xmath507 is on the right edge of @xmath586 . so there would be a non - zero @xmath17-map @xmath598 , which is impossible since @xmath560 and @xmath599 are ext - orthogonal . so @xmath600 . let @xmath601 be the @xmath5-summand of highest quasilength which appears in @xmath480 . then , since both @xmath602 and @xmath603 contain the quasisimple @xmath604 from lemma [ lem : revhammquasi ] , lemma [ lem : subwingintersection ] tells us that @xmath605 . but @xmath590 is in @xmath586 as well , so by lemma [ lem : subwingintersection ] again , either @xmath606 or @xmath607 . the former case is not possible , since it would imply that @xmath608 , which is impossible since @xmath609 . so the remaining possibility is that @xmath610 , that is , @xmath590 is the @xmath5-summand in @xmath480 of highest quasilength . by the description of the quiver of @xmath267 in section [ sec : endoring ] , there is a @xmath18-arrow @xmath577 associated with the non - degenerate @xmath5-subwing triple , from the vertex corresponding to @xmath511 to the vertex corresponding to @xmath590 . since @xmath561 ends in the vertex corresponding to @xmath511 , and @xmath611 starts in the vertex corresponding to @xmath590 , the string @xmath612 is well - defined . it remains to consider the case where the @xmath5-subwing triple @xmath552 is degenerate , that is , @xmath613 . in this case , since @xmath584 , the only option is that @xmath7 is on the right edge of @xmath513 . but then @xmath580 is empty : if there was a @xmath18-map @xmath614 for some @xmath5-summand @xmath615 , then @xmath560 and @xmath615 would have an extension , as can be seen from an argument similar to the above . by virtue of the preceding considerations , we can now associate to each indecomposable object @xmath616 a unique indecomposable @xmath267-module @xmath617 where @xmath618 \sigma^{\d}_x & \textrm{if } \sigma^{\t}_x \textrm { is zero } \\[4pt ] ( \sigma^{\d}_x)^{-1 } \beta_x \sigma^{\t}_x & \textrm{if both } \sigma^{\t}_x \textrm { and } \sigma^{\d}_x \textrm { are non - zero } \end{array } \right.\ ] ] we can now describe the action of the hom - functor on objects in @xmath487 . [ thm : actiononf ] let @xmath5 be a maximal rigid object of @xmath19 , and @xmath369 the endomorphism ring . * let @xmath7 be an object in @xmath619 . then the @xmath267-module @xmath620 is isomorphic to the string module @xmath617 . * the functor @xmath621 provides a bijection between @xmath622 and the set of isoclasses of indecomposable @xmath267-modules . * ( 1 ) * by lemma [ lem : revhamm ] part ( i ) , the module is non - zero . let @xmath623 be the idempotent corresponding to the indecomposable projective @xmath624 . let @xmath623 be the idempotent of @xmath267 corresponding to the vertex @xmath204 , which again corresponds to the summand @xmath230 of @xmath5 . then the vector space @xmath625 decomposes @xmath626 where each vector space @xmath627 and @xmath628 is at most 1-dimensional and is spanned by a @xmath17-map @xmath629 and a @xmath18-map @xmath630 respectively , where these maps are zero if no non - zero such maps exist . by the definition of @xmath561 , the vertices for which @xmath631 are exactly the vertices that are traversed by @xmath561 . similarly , the vertices for which @xmath632 are the ones traversed by @xmath576 . in particular , there is an equality of dimension vectors @xmath633 we need to establish that the action of @xmath267 on @xmath625 is the same as the action on @xmath617 . each map which is irreducible in @xmath238 corresponds to an arrow in the quiver of @xmath267 , and the arrow acts by composition with the irreducible map . unless both the start vertex and the end vertex of this arrow are vertices in the support of @xmath625 , then clearly this map ( equivalently , this arrow ) has a zero action on both the modules @xmath620 and @xmath617 . so we must show that for each @xmath17-arrow @xmath634 appearing in @xmath561 acts by an isomorphism @xmath635 , and each @xmath17-arrow appearing in @xmath576 acts by an isomorphism @xmath636 , and finally that if @xmath637 is defined , then the action of this is given by an isomorphism @xmath638 . our first goal is now to show that whenever @xmath639 is a @xmath17-arrow such that @xmath360 itself or @xmath377 appears in @xmath640 , then the action of @xmath360 is given by a pair of linear transformations @xmath641 which are isomorphisms when their domains and codomains are both non - zero . ( and necessarily zero otherwise . ) let @xmath642 be the irreducible @xmath17-map corresponding to @xmath360 . then what we need is that if @xmath643 and @xmath644 are both non - zero , then @xmath645 up to a non - zero scalar , and similarly that if @xmath646 and @xmath647 are both non - zero , then @xmath648 . the first assertion is clearly true by the structure of the tube . the second assertion holds by an application of lemma [ lem : rayfactoring ] part ( i ) , and remark [ rem : rayfactoring ] , where we use that @xmath649 must be a composition of maps which are irreducible in @xmath650 , and follows a ray or coray ( along the edge of a wing ) , and thus all the indecomposables that @xmath649 factors through are also in @xmath480 . next let @xmath7 be such that the @xmath18-arrow @xmath651 is defined , and thus appears in the string @xmath640 . then we know that @xmath627 and @xmath652 are non - zero . the action of @xmath577 is given by composition with a @xmath18-map @xmath653 . we wish to show that ( up to multiplication by a non - zero scalar ) this action is given by a linear transformation @xmath654 in other words , it sends @xmath644 to @xmath647 and annihilates @xmath646 . the composition @xmath655 is clearly zero , as all compositions of two @xmath18-maps are . consider the image @xmath656 of @xmath644 . we need that the @xmath17-map @xmath644 does not factor through any indecomposable to which there is no @xmath18-map from @xmath213 . this holds , since by lemma [ lem : onecomponent ] , @xmath657 forms a rectangle - shaped subgraph of the tube , and the map @xmath644 can not factor through any indecomposable outside this subgraph . we can then conclude from lemma [ lem : rayfactoring ] part ( ii ) and remark [ rem : rayfactoring ] that @xmath658 , which is what we wanted . it remains to show that if there exists an arrow which connects two vertices in the support of @xmath620 , but which does nt appear in @xmath640 , then the action of this arrow is zero on @xmath620 . by lemmas [ lem : stringsubwing ] , [ lem : revhamm ] and [ lem : dstrings ] the only case to consider is when @xmath577 is the loop vertex , and there is a @xmath5-subwing triple @xmath235 such that @xmath659 and @xmath660 or vice versa . since the action of the arrow @xmath661 is given by composition with the @xmath18-map @xmath662 , we only need to study the case where @xmath561 traverses @xmath405 and @xmath576 traverses @xmath8 . so we need to show that @xmath663 is zero . but by examining the hom - hammocks of @xmath213 and @xmath174 , we see that if there is a @xmath17-map @xmath664 and a @xmath18-map @xmath665 , then either @xmath7 is in @xmath224 , which contradicts the fact that @xmath204 must be traversed by @xmath640 , or @xmath644 factors through objects on the coray @xmath666 . in the latter case , the composition must be zero , since there are no @xmath18-maps from @xmath174 to any objects on this coray . * ( 2 ) * counting the number of elements of @xmath487 , we find that it contains @xmath667 objects with quasilength less than @xmath10 and @xmath668 with quasilength @xmath10 or more , that is , a total of @xmath669 elements . since @xmath5 has @xmath189 summands , the cardinality of @xmath670 is @xmath671 which , by theorem [ thm : finitetype ] , is also the number of indecomposables in @xmath672 . by lemmas [ lem : differentstrings ] and [ lem : tstringisnotdstring ] , if @xmath7 and @xmath41 are different objects in @xmath673 , then @xmath674 . it then follows from part ( 1 ) that @xmath675 . so we @xmath621 provides a bijection . we now turn to the indecomposables which are not in @xmath487 . it is easily seen that lemma [ lem : revhamm ] parts ( ii)-(iv ) hold also for indecomposables which are not in @xmath487 . so we can define @xmath561 and @xmath576 in this case as well . the following theorem now completes the description of the action of @xmath621 on objects . let @xmath7 be an indecomposable object in @xmath19 , where @xmath676 . then we have the following . * @xmath503 if and only if @xmath677 , where @xmath678 . * if @xmath7 is not of the type described in @xmath679 , then @xmath680 where @xmath389 is the zero module if @xmath390 is the zero string . * ( 1 ) : * when @xmath676 we know that @xmath503 if and only if @xmath681 . there are no @xmath17-maps @xmath579 if and only if @xmath7 is on the ray @xmath581 , that is @xmath682 for some @xmath683 . moreover , there are no @xmath18-maps @xmath579 if and only if @xmath7 is on the coray @xmath684 . the indecomposables that are in the intersection of @xmath581 and @xmath684 and outside @xmath487 are exactly the ones with coordinates @xmath685 such that @xmath686 . the claim follows . * ( 2 ) : * the proof of theorem [ thm : actiononf ] goes through , with the following exception , which is exactly what is needed . the action of @xmath577 ( which in this case is always the loop , as we see from the argument for ( 1 ) above ) is zero : the @xmath17-map @xmath687 factors through ( at least ) one object on the coray @xmath684 . we know that there are no @xmath18-maps from @xmath95 to any object on this coray . it then follows that the composition @xmath688 , where @xmath689 is the @xmath18-endomorphism of @xmath95 , is a zero map . the result follows . the author is very grateful to aslak bakke buan for many helpful comments on earlier versions of this manuscript . much of the work presented here was carried out during a visit at the school of mathematics , university of leeds . the author would like to thank robert marsh for his hospitality as well as for many discussions on the subject .

the gravity probe b satellite is scheduled to fly in the year 2000 @xcite . it contains a set of gyroscopes intended to test the predictions of general relativity ( gr ) that a gyroscope in a low circular polar orbit , with altitude @xmath3 will precess about 6.6 _ arcsec / year _ in the orbital plane ( geodetic precession ) and about 42 _ milliarcsec / year _ perpendicular to the orbital plane ( lt precession , see @xcite ; @xcite , secs . 4.7 and 7.8 ; @xcite , sec . 9.1 ) . in this paper we review the theoretical derivation of these effects and in particular consider the contributions of the earth s quadrupole and higher multipole fields . we first review the derivation of the metric for a rotating body using the standard lgrt approach ( @xcite ; @xcite ; @xcite , secs . 4.7 and 7.8 ) . the metric is characterized by a newtonian scalar potential ( the gravito - electric field ) and a vector potential ( the gravito - magnetic field ) ( @xcite , sec . we then obtain the same result with a simple transformation argument which clarifies the physical meaning of the metric ( @xcite , sec . specifically it makes clear that if the metric of a point mass contains fundamental parameters such as the eddington parameters @xmath0 and @xmath1 , then to lowest order the metric of a rotating body contains no * new * fundamental parameters @xcite . thus there is no new lense thirring or frame - dragging parameter to be measured by gp - b or any other experiment . we then derive the precession equations for a gyroscope in a general weak field system , that is for any scalar and vector potential fields @xcite . the calculation is valid to first order in the fields and velocities of the source body and the satellite . the gravitational field of the earth is described by the scalar and vector potentials which depend on the shape of the body and the mass and velocity distribution inside it . we treat both of these fields by a multipole expansion and express the precessions in series of spherical harmonics . we do not limit ourselves with the axially symmetric case which has been thoroughly studied by teyssandier ( @xcite ) , ( @xcite ) . in particular we show that for a solid body rotation up to the order @xmath4 both precessions depend only on the tensor of inertia of the earth . the major nonspherical contributions to the gp - b precessions are from the earth s quadrupole moment . the contribution to the geodetic precession which has a magnitude of about 1 part in @xmath5 is detectable by gp - b and quite important for the determination of the parameter @xmath1 which is to be measured to about 1 part in @xmath6 , the most accurate measurement envisioned ( @xcite , sec . 3.5 and in particular table 14.2 ) . this contribution has been calculated independently by wilkins @xcite and barker & oconnel @xcite , and then in the most elegant and general way by breakwell @xcite . the contribution to the lense thirring precession is beyond the expected accuracy of gp - b and close enough to the result by teyssandier ( @xcite ) obtained from a different earth s model . we also estimate the influence of moon and sun to find that only the geodetic contribution of the sun must be included in the gp - b data reduction , as anticipated . we first briefly review this standard derivation of linearized general relativity theory . the metric of a rotating body such as the earth is obtained by introducing a small perturbation of the lorentz metric @xmath7 , that is @xmath8 the perturbation @xmath9 is assumed to be expressed in isotropic space coordinates so that @xmath10 . similarly we suppose the matter producing the metric field is described by the energy momentum tensor of slow - moving and low density matter with negligible pressure , @xmath11 where @xmath12 is the 4velocity and @xmath13 is the matter density . the field equations are : @xmath14 where @xmath15 is the einstein tensor . the calculation of @xmath15 and @xmath16 to the lowest order in the perturbation is straightforward and results in the following equations : @xmath17 @xmath18 here indices are raised and lowered with the lorentz metric , which is consistent to lowest order , and the slash denotes differentiation . the last expression imposes the so - called lorentz condition which can always be achieved by a coordinate transformation ( to lowest order ) and involves no loss of generality ; it is also the analog of a gauge choice in electromagnetism . because we have used isotropic coordinates equation ( [ eq:4 ] ) leads immediately to @xmath19 , and the standard classical correspondence gives @xmath20 . we thereby obtain the wave equations for the scalar and vector potentials : @xmath21 @xmath22 where @xmath23 is the velocity of the source . these equations may be solved in the standard way by means of a retarded green s function . in the case of a time independent system , or a system which changes so slowly that retardation effects may be ignored , the solution is @xmath24 @xmath25 note that these expressions are analogs of the equations of electrostatics and magnetostatics , which is why one may speak about gravito - electric and gravito - magnetic effects in the linearized theory . in summary , we may write the ( lense thirring ) line element as @xmath26 this is valid to first order in field intensity and source velocity , and will serve as a basis for calculating the gyroscope precession . 1 shows the general aspects of the scalar and vector fields of a spinning body ; the vector field generally points in the same direction as the velocity of the surface of the body . it is , in fact , possible to obtain the above result from a different and physically interesting perspective , and moreover introduce parameters convenient for discussing experimental measurements . following eddington , consider the metric of a massive point with a geometric mass @xmath27 at a large distance @xmath28 , so that @xmath29 . we expand the schwarzschild solution in isotropic coordinates for this situation as @xmath30 @xmath31 eddington suggested that this be written in terms of parameters as @xcite , @xcite @xmath32 where @xmath0 and @xmath33 and @xmath1 are equal to @xmath34 for general relativity . the power series ( [ eq:11 ] ) is clearly a rather general form for the metric far from a spherical body . since the parameter @xmath27 which appears in the metric is a constant of integration representing the mass of the central body ( specifically @xmath35 ) , we may absorb the parameter @xmath0 into it , which is equivalent to taking @xmath36 . this is consistent as long as no independent nongravitational determination of the mass of the body is considered . we will nevertheless retain @xmath0 in our calculations as a book - keeping device . indeed , the parameters in ( [ eq:11 ] ) may all be viewed as a tool for tracking which terms in the metric contribute to a gravitational phenomenon such as the perihelion shift of mercury or the deflection of the starlight by the sun . alternatively they may be viewed as numbers which may _ not _ be equal to @xmath34 if a metric theory other than general relativity is actually valid . in either case they provide a convenient way to express the results of experimental tests of gravity as giving value to the parameters . this parameterized approach has been extended to include many other parameters and has been highly developed under the name parameterized post newtonian theory , or ppn @xcite . in this paper we take the viewpoint that general relativity is to be tested and emphasize that we are _ not _ using the more general ppn approach . solar system measurements give @xmath37 and @xmath38 , which are of course entirely consistent with general relativity . we consider now only phenomena in which the quadratic term , @xmath39 , in @xmath40 is unimportant , that is in which we may ignore @xmath33 and assume that the underlying gravitational theory is linear . then for a stationary mass , @xmath41 since this is nearly the lorentz metric , we may generalize it to a moving mass point by simply transforming to a moving system using a transformation that is lorentzian to the first order in velocity : @xmath42 here the subscript @xmath28 labels the system in which the mass is at rest , and which moves at velocity @xmath43 in the @xmath44 direction in the laboratory system . this transformation gives the metric for the moving mass point as @xmath45 @xmath46 this obviously generalizes for motion in any direction to @xmath47 @xmath48 as we assume that our theory is linear to this order ( like general relativity ) , we can superpose the fields of a distribution of such point masses and replace @xmath49 by @xmath50 from ( [ eq:8 ] ) and @xmath51 by @xmath52 from ( [ eq:9 ] ) , resulting in @xmath53 @xmath54 @xmath55 this agrees with the general relativity result ( [ eq:12 ] ) when @xmath56 but now contains appropriate combinations of eddington parameters . we emphasize that this line element has been obtained for any slowly moving mass distribution from the parameterized metric for a stationary mass point by transformation and superposition , and thus no new parameter appears in the expression . therefore a measurement of a phenomenon which depends on the cross term in ( [ eq:14 ] ) provides a value for @xmath57 and not for some new parameter associated with gravito - magnetism . note also that the result is rather strong since it depends on the observationally verified schwarzschild metric ( [ eq:12 ] ) , the well - tested lorenz transformation , and the principle of superposition , valid for any linear theory . an orbiting gyroscope has its spin axis parallel displaced in accord with the metric ( [ eq:14 ] ) . we calculate this motion with minimal assumptions about the potentials , which could be the potentials of a nearly spherical and rather uniform rigid body such as the earth or a potato - shaped body such as in fig . 1 , with some interior mass distribution . we will work always to the first order in the potentials and in the velocities , @xmath23 and @xmath58 , of the source and orbiting gyroscope . the parallel displacement equation for the gyro spin @xmath59 is ( @xcite ; @xcite , secs . 4.7 and 7.8 ; @xcite , sec . 9.1 ) : @xmath60 we suppose that the gyro spin 4-vector is perpendicular to the velocity 4-vector , which is equivalent to assuming that the gyro spin has no zero component in its rest frame . from this assumption the zero component in another frame is easily obtained , and to first order in the satellite velocity it is given by @xmath61 . we calculate the euler lagrange equations in the standard way , and then put them into canonical form to give the christoffel symbols . to lowest order in the potential and velocity the christoffel symbols are @xmath62 @xmath63 the roman indices in these expressions are space indices and run from 1 to 3 , a slash denotes an ordinary derivative , and @xmath64 . note that the gravitational vector potential @xmath65 occurs in a gauge invariant way , that is only its @xmath66 appears . substitution of the christoffel symbols into the spin equation of motion gives @xmath67 we break the drift rate into two parts , the geodetic drift rate due to the scalar potential @xmath68 , and the lense thirring precession rate due to the vector potential @xmath65 . separating also symmetric and antisymmetric parts of the geodetic effect , we arrive at , in a 3dimensional vector notation , @xmath69 and @xmath70 + \gamma\bigl(\vec v\cdot\nabla\phi\bigr)\vec s \biggr\},\ ] ] @xmath71 @xmath72 where @xmath73 and @xmath74 are the instantaneous values of the lense thirring and geodetic precessions , respectively . since @xmath73 is the _ curl _ of the gravitational vector potential the lense thirring precession rate is the analog of the magnetic field in magnetostatics theory . the geodetic effect is of the order of the scalar potential times the orbital velocity of the satellite , while the lense thirring terms are of the order the scalar potential times the velocity of the central body ( the earth ) , which is typically much smaller than the orbital velocity . symmetric geodetic terms are responsible for stretching the spin vector @xmath75 , and the reason for separating them is that their effect almost vanishes when averaged over any reasonable satellite orbit . to see this consider the last term in the symmetric part of the geodetic drift rate . using newton s law in the form @xmath76 we may write @xmath77 where @xmath78 is the change in the velocity squared in total time @xmath79 , and it is assumed that the drift rate is small and @xmath75 does not change significantly during this period of time . if the orbit is periodic , this quantity will be zero , and for a nearly periodic orbit it will be very small . similarly for the remaining symmetric terms in the geodetic drift we may write : @xmath80 @xmath81 again , this is zero for a periodic orbit and very small for a nearly periodic orbit . in summary the average precession rate of the gyro spin is @xmath82 with the values of the geodetic and precessions given in ( [ eq:17 ] ) and ( [ eq:18 ] ) . note that the geodetic and lense thirring effects are approximately perpendicular to each other for an approximately polar orbit around a central body of reasonable shape ; for a circular polar orbit about a spherical body they are perpendicular . 2 shows the orientation of the various vectors for a fairly general situation , a satellite in a roughly planar polar orbit about an oddly shaped body . the geodetic precession vector @xmath74 is roughly perpendicular to the orbit plane and the lense thirring precession vector @xmath73 has the general appearance of a dipole magnetic field . if there are relatively distant masses , such as the moon , in the neighborhood of the central body , their density distribution @xmath83 may be expressed as @xmath84 the effect on the scalar and vector potentials is @xmath85 it then follows that the precessions due to these distant masses are @xmath86 where @xmath87 is the angular momentum of the distant mass . for the specific case of the moon and the gp - b experiment the numerical values will be discussed in sec . we will henceforth study the case of a body such as the earth which is rigidly rotating , that is @xmath88 . motivated by expression ( [ eq:9 ] ) we introduce a new vector potential quantity @xmath89 : @xmath90 the vector @xmath89 is a harmonic function outside the body which we will expand in spherical harmonics ; for a special spherically symmetric case @xmath89 is collinear with @xmath91 . in terms of @xmath89 , the metric ( [ eq:14 ] ) may be written as @xmath92 using ( [ eq:8 ] ) and ( [ eq:9 ] ) it is possible to express the divergence of @xmath89 via the scalar potential , namely , @xmath93 which allows us to rewrite formula ( [ eq:17 ] ) for the lense thirring precession as @xmath94 \vec\omega_{lt}= -(\alpha+\gamma)\omega\,\biggl[\vec\pi_{|z}+\bigl(\phi+r\phi_{|r}\bigr)\biggr]\hat z \label{eq:24}\ ] ] here we have chosen the @xmath95 axis to be along the spin , @xmath96 . we now put the origin of spherical coordinates @xmath97 at the center of mass of the body , introduce spherical harmonics with the notation @xmath98 and expand the potentials @xmath68 and @xmath99 in corresponding series : @xmath100 \label{eq:26}\ ] ] @xmath101 \label{eq:27}\ ] ] [ @xmath102 terms in ( [ eq:26 ] ) and @xmath103 terms in ( [ eq:27 ] ) are missing because the origin is at the center of mass of the body , see expressions ( [ eq:28 ] ) and ( [ eq:29 ] ) below ] . these expansions are in the inertial frame in which the body rotates ; the direction of the coordinate polar axis z is so far arbitrary ; @xmath104 is the characteristic size of the body , which we take to be the equatorial radius for the earth . the potentials are slowly varying functions of time due to the earth rotation , so we write the coefficients as explicit functions of time . these coefficients are related to the mass distribution in the standard way by @xmath105 @xmath106 note that generally the domain of integration here is also time dependent . in particular , we write for convenience @xmath107 . the time dependence in ( [ eq:28 ] ) and ( [ eq:29 ] ) may be easily analyzed for a rigidly rotating body provided that the rotation axis is fixed both in the inertial space and in the body . in that case in the inertial frame with the @xmath95 axis along the spin , @xmath96 , the earth density is @xmath108 where @xmath109 is the time independent density measured in the frame rotating with the earth . substituting this for the density in ( [ eq:28 ] ) and ( [ eq:29 ] ) and transforming to the earth fixed frame we find that the time dependent coefficients may be written in terms of constant moments of density according to @xmath110= \left|\left|\matrix{\cos\omega t&-\sin\omega t\cr\sin\omega t&\cos\omega t \cr}\right|\right| \left[\matrix{a^{c}_{lm}\cr a^{s}_{lm}\cr}\right ] \label{eq:30}\ ] ] @xmath111= \left|\left|\matrix{\cos\omega t&-\sin\omega t\cr\sin\omega t&\cos\omega t \cr}\right|\right| \left[\matrix{p^{ic}_{lm}\cr p^{is}_{lm}\cr}\right ] , \label{eq:31}\ ] ] and the time independent coefficients are explicitly given by @xmath112 @xmath113 if the rotation axis wanders in the body and/or in the inertial space , the relation between the time dependent and time independent coefficients is given by a combination of appropriate rotation matrices which is more complicated than the one matrix in ( [ eq:30 ] ) and ( [ eq:31 ] ) ( see c. f. @xcite ) . generally , a time dependent coefficient with the indices @xmath114 and @xmath27 is a linear combination of the appropriate time independent coefficients with the indices @xmath114 and @xmath115 . introducing now the time dependent coefficients from ( [ eq:32 ] ) and ( [ eq:33 ] ) back into ( [ eq:26 ] ) and ( [ eq:27 ] ) we obtain the time dependent potentials in a convenient form using time independent coefficients only : @xmath116 \label{eq:34}\ ] ] @xmath117 \label{eq:35}\ ] ] this is of course what one should expect intuitively ; in general this form will be most useful for our purposes . the constant coefficients @xmath118 in ( [ eq:34 ] ) are those that are measured very accurately for the earth ( up to @xmath119 their values are found in @xcite ) . however , for the earth of arbitrary shape and composition it is impossible to express @xmath120 through @xmath118 , in other words , their values , and hence the scalar and vector potentials , are independent . nevertheless , a useful relationship between the two sets of coefficients exists ; to describe it , we need a notation for a general moment of the density , @xmath121 in particular , @xmath122 using the definitions ( [ eq:32 ] ) and ( [ eq:33 ] ) and the recurrence relations for legendre functions @xcite , we derive the following equalities relating @xmath120 to @xmath118 : @xmath123 @xmath124 \left[m^{\nu}_{l+1l-1m+1}-(l+m-1)(l+m)m^{\nu}_{l+1l-1m-1}\right ] \bigr\}\ ] ] @xmath125 @xmath124 \left[m^{\mu}_{l+1l-1m+1}-(l+m-1)(l+m)m^{\mu}_{l+1l-1m-1}\right ] \bigr\ } \label{eq:38}\ ] ] @xmath126 in the second line of ( [ eq:38 ] ) , the minus sign is taken and @xmath127 when @xmath128 , the plus sign and @xmath129 when @xmath130 . from ( [ eq:34 ] ) and ( [ eq:35 ] ) using the definition ( [ eq:17 ] ) of @xmath73 a rather compact multipole expansion for the lense thirring precession may be computed : @xmath131\biggl({r\over r}\biggr)^l y^{\nu}_{lm}(\theta,\varphi-\omega t ) \label{eq:39}\ ] ] the corresponding expansion for the geodetic precession is too cumbersome to be useful . if the shape of the central body and the mass distribution inside it are known , then all the pertinent quantities , including multipole expansion coefficients @xmath118 and @xmath120 , may be found by integration , but this is rarely the case . even when @xmath118 are measured , as for the earth , all the @xmath120 , and the lense thirring effect with them , remain entirely undetermined . however , for a body of any shape and composition , the coefficients @xmath132 and @xmath133 can be expressed in terms of the elements @xmath134 of the tensor of inertia @xmath135 determined in a standard way , @xmath136 writing @xmath137 , we find : @xmath138 @xmath139 @xmath140 @xmath141 this is done by comparing the integrals ( [ eq:32 ] ) ( with @xmath142 ) and ( [ eq:33 ] ) ( with @xmath102 ) to @xmath134 using explicit expressions of legendre functions with @xmath143 ; formulas for @xmath144 and @xmath145 are known and used in geodesy for the determination of the earth s moments of inertia @xcite . introducing ( [ eq:40 ] ) into ( [ eq:34 ] ) , ( [ eq:35 ] ) and dropping the terms with @xmath146 for @xmath68 and @xmath147 for @xmath99 , we first obtain the @xmath4 formulas for the potentials : @xmath148,\quad \vec\pi(\vec r)=-{g\over r^3 } \biggl [ { \bf i}\vec r-{1\over2}\,\bigl({\rm tr}\,{\bf i}\bigr)\vec r \biggr ] \label{eq:41}\ ] ] from these we then obtain the approximation for the precessions by differentiation of the expressions ( [ eq:18 ] ) and ( [ eq:17 ] ) ( see also ( [ eq:22 ] ) ) : @xmath149 \bigl(\vec r\times\vec v\bigr)+ { 3\over r^2}\,\bigl({\bf i}\vec r\times\vec v\bigr ) \biggr\ } \label{eq:42}\ ] ] @xmath150\vec\omega+ 3\,{\vec\omega\cdot\vec r\over r^2}\,\biggl[{1\over2}\,\bigl({\rm tr}\,{\bf i}\bigr)\vec r-{\bf i}\vec r\biggr ] \biggr\ } \label{eq:43}\ ] ] [ of course , the same expression for @xmath73 is also obtained from ( [ eq:40 ] ) and ( [ eq:39 ] ) with @xmath142 ] . note that formulas ( [ eq:40 ] ) and ( [ eq:41 ] ) hold in both the body - fixed and inertial frames , and the coefficients and moments of inertia in the latter depend on the time if the rotation is not axisymmetric , @xmath152 , @xmath153 , @xmath154 . the expressions ( [ eq:42 ] ) and ( [ eq:43 ] ) for the precessions are meaningful in the inertial frame where generally @xmath155 , @xmath156 being a rotation matrix which converts the body - fixed radius - vector into the inertial one . in the above case of a simple rotation about an inertially - fixed axis , @xmath157 , @xmath158 the results ( [ eq:41])([eq:43 ] ) are valid under either of the two conditions : \1 ) _ far field _ , @xmath159 ; \2 ) _ high symmetry _ , i.e. , all higher order moments are small these results alter somewhat our usual notion of the geodetic and lense thirring effects : the first one is generally proportional not only to the orbital momentum , but to @xmath160 as well , and the lense - thirring precession generally points not only in the direction of the angular momentum @xmath161 , but has also components parallel to @xmath162 , and @xmath163 . two particular cases of the inertia tensor are of special interest . _ a ) spherical symmetry _ , @xmath164 . in this case , the standard formulas follow immediately from ( [ eq:43 ] ) for @xmath56 : @xmath165 \label{eq:44}\ ] ] note that we have thus shown this to be the exact result for a spherical earth with any _ radial _ density distribution @xmath166 . _ b ) symmetric top , _ @xmath167 , @xmath168 . it turns out that for @xmath157 , i. e. , for the rotation about the material symmetry axis , the previous expression for @xmath169 remains true ; to lowest order in the oblateness , this also proves to be the exact result for a slightly oblate uniform ellipsoid of revolution rotating about its semiminor axis . the corresponding expression for @xmath170 is given and discussed in sec . to go beyond the @xmath4 approximation , one must make some assumptions about the shape of the central body and density distribution inside it , and use the data of gravitational potential measurements , as available . bearing in mind the application to gp - b , that is , the earth , we use the following set of assumptions . _ 1 . gravitational potential . _ we assume it to be measured , i.e. , the gravitational coefficients @xmath171 known . our problem is thus to determine the vector potential in terms of the coefficients @xmath172 using as general an earth model as possible . _ 2 . shape . _ we assume that the earth is a slightly oblate ellipsoid of revolution , so its surface equation to lowest order in the eccentricity @xmath173 is @xmath174 where @xmath104 is the semi - major axis ( equatorial radius ) ; the eccentricity is the ratio of the difference of the semi - axes to the major one . _ 3 . mass distribution . _ we examine two different models : _ a ) _ with @xmath175 and @xmath176 being arbitrary functions of their arguments , we set @xmath177 the first term here describes any depth variation of the _ average _ density , and the only assumption is that the angular variations are depth independent . _ b ) _ for arbitrary functions @xmath175 and @xmath178 , we set @xmath179 contrary to the previous model , here all angular variations of the density are concentrated at the earth s surface . we will call model a the set of assumptions @xmath180 and @xmath181 , and model b the assumptions @xmath180 and @xmath182 ; in both cases the mass distributions are , of course , assumed consistent with the gravitational data of the assumption @xmath34 . as far as the earth is concerned , assumption 2 ) reflects the classical clairaut formula ( see c. f. @xcite ) ( the eccentricity is @xmath183 ) ; also the mass model @xmath182 seems rather realistic for the earth since the estimated thickness of the layer where its mass distribution varies significantly in the angular directions is only about @xmath184 . note that an entirely different earth model is used in geodesy : it assumes that the sum of gravitational and centrifugal potentials is constant at the earth s ellipsoid s surface ( @xcite ) . this allows one to relate @xmath185 to the eccentricity and the earth angular velocity only ( in particular , to obtain the earth s gravitational oblateness @xmath186 with a surprisingly good accuracy ) , but gives zero values to @xmath187 and leaves @xmath120 undetermined . evidently , our two models should give bounds for the corrections to the @xmath188 values of the precessions . moreover , for both of them it is possible to find the corrections explicitly by means of the following four steps : @xmath189 calculate @xmath171 by ( [ eq:32 ] ) via spherical harmonics coefficients @xmath190 of the function @xmath191 for model a , [ or @xmath192 for model b ] ; @xmath193 since @xmath171 are assumed known , solve the resulting equations for @xmath190 , hence having them expressed via @xmath171 ( thus @xmath176 or @xmath178 are uniquely determined at this stage through the gravitational data ) ; @xmath194 using that , calculate the needed moments @xmath195 of the density by ( [ eq:36 ] ) in terms of @xmath171 ; @xmath196 using the found values of @xmath195 , express @xmath172 through @xmath171 according to ( [ eq:38 ] ) . in fact , what is described here is a fit of our density distributions ( [ eq:46 ] ) and ( [ eq:47 ] ) to the known gravitational coefficients @xmath171 which allows us , in these two cases , to express the former through the latter and thus determine uniquely the coefficients @xmath172 , i.e. , the gravito - magnetic part of the field . the implementation of this procedure is rather cumbersome , though basically straightforward , so the details are given in the appendix . the results of the calculations to lowest order in the earth s eccentricity are : @xmath197 @xmath198+o\bigl((\epsilon l)^2\bigr)\ ] ] @xmath199 @xmath200+o\bigl((\epsilon l)^2\bigr ) \label{eq:48}\ ] ] @xmath201 + o\bigl((\epsilon l)^2\bigr),\ ] ] with the minus sign and @xmath127 for @xmath128 , plus sign and @xmath129 for @xmath130 in the second of these formulas , and @xmath202 it is remarkable that the first non - zero oblateness correction is quadratic , and that the results for the two models differ by just a factor @xmath203 in front of a couple of terms . these expressions are needed only for @xmath188 since for @xmath102 formulas ( [ eq:40 ] ) always hold . we now apply the above results to the gp - b satellite which will be circling the earth on a low ( @xmath204 ) polar nearly circular orbit . our aim is to check the gr predictions for the geodetic and lense - thirring drift of a gp - b gyroscope taking into account all the peculiarities of the real earth s gravito - electric and gravito - magnetic field pertinent to the expected experimental error of one part in @xmath205 for the geodetic effect and about one part in @xmath206 for the lense - thirring effect . more precisely , one needs to check whether the spherically symmetric approximation ( [ eq:44 ] ) , which has been used for many years in the theoretical discussion and planning of the experiment , needs any corrections to match the above experimental accuracy . we use the gr values for the eddington parameters @xmath56 in the sequel . we always use the assumption 2 of the previous section that the earth has a shape of a slightly oblate ellipsoid of revolution with the eccentricity @xmath207 . its semi - minor axis is assumed fixed in the inertial space and coincident with the earth rotation axis and the @xmath95 axis of the cartesian coordinates . _ 1 . gravitational potential_. for the required accuracy , it is enough to include only the earth s quadrupole moment into @xmath68 , i. e. , to set @xmath208 this is because all gravitational coefficients other than the earth s oblateness @xmath2 are at least 2 orders of magnitude smaller @xcite . by ( [ eq:49 ] ) and ( [ eq:40 ] ) , the @xmath4 expression ( [ eq:40 ] ) for @xmath68 is valid with @xmath209 which gives the gravitational potential in the usual form : @xmath210= -\frac{gm}{r}\left[1-j_2\frac{r^2}{r^2}p_2(\cos\theta)\right]\ ] ] _ 2 . instantaneous geodetic precession_. just as for the potential , ( [ eq:49 ] ) assures that the @xmath4 formula ( [ eq:42 ] ) is valid for @xmath211 . the explicit form of the geodetic precession is obtained either from ( [ eq:42 ] ) and ( [ eq:50 ] ) or from its general definition ( [ eq:18 ] ) by differentiating the above expression for the potential . the result reads : @xmath212 \right\ } , \label{eq:51}\ ] ] this expression contains the @xmath2 corrections to the classical spherically symmetric expression ( [ eq:44 ] ) . breakwell @xcite obtained it but in a different form , pinning the value of the precession to a point on a given satellite orbit rather than to a given point in space through which the spinning particle passes with a velocity @xmath58 . _ 3 . gravito - magnetic field and the instantaneous lense thirring precession_. we now invoke our assumption about the mass distribution ( models a and b of the previous section ) and substitute the values ( [ eq:49 ] ) of the gravitational coefficients into the expressions ( [ eq:48 ] ) for @xmath120 . in this way it turns out that for _ both _ models a and b the only additional to @xmath102 non - zero coefficients are those with @xmath213 , and their values are @xmath214 @xmath215 so , for both our models the only addition to the field of a symmetric top is the gravito - magnetic field ( [ eq:52 ] ) . in terms of the lense thirring precession that means that the standard expression ( [ eq:44 ] ) has to be appended by the terms ( [ eq:52 ] ) from the general formula ( [ eq:39 ] ) for the precession . the result reads : @xmath216-\ ] ] @xmath217 , \label{eq:53}\ ] ] where ( recall @xmath218 ) @xmath219 @xmath220 the second line of ( [ eq:53 ] ) is the correction to the classical lense thirring expression in the first line which is induced by the gravito - magnetic field ( [ eq:52 ] ) . [ recall from section 7 , b ) , that the dependence of @xmath73 on @xmath2 for a symmetric top rotating about its symmetry axis cancels out - remarkably!. thus the first line of ( [ eq:53 ] ) contains no @xmath2 ] . note that the only difference between the earth s models a and b is a factor of 7/5 in the constants @xmath222 and @xmath223 . note also that allowing for more multipoles in the gravitational potential than given in the equation ( [ eq:49 ] ) we automatically obtain higher order multipoles in the gravito - magnetic field by the expressions ( [ eq:48 ] ) for @xmath120 . however , for gp - b all these higher order corrections prove to be too small to be taken into account . orbit of the gp - b satellite and the instantaneous values of precessions_. the perfect orbit for the gp - b satellite would be a circular polar one with the altitude @xmath224 which is described by @xmath225 . in reality it is slightly distorted by the quadrupole moment of the earth s gravitational field ; with the lowest order in @xmath2 corrections included , the orbit becomes ( @xcite ) , ( @xcite ) : @xmath226 , \label{eq:55}\ ] ] @xmath227 \left [ \hat { \theta}+{1\over2}\,j_2\left({r\over r_0}\right)^2\sin2\theta\,\hat { r } \right],\ ] ] where @xmath228 are the corresponding unit vectors . ( in view of the orbit symmetry , it is enough to consider a half - orbit from one pole to the other , @xmath229 ; otherwise , the second half may be assigned to @xmath230 , with the proper direction of the unit vector @xmath231 ) . introducing the orbital velocity ( [ eq:55 ] ) into the formula ( [ eq:51 ] ) and dropping some terms @xmath232 , we obtain the geodetic precession for the gp - b gyroscope in the following form : @xmath233 \biggl [ 1+{3\over2}\,j_2\biggl({r\over r_0}\biggr)^2\,\biggl(1 - 3\cos^2\theta\biggr ) \biggr]\hat e_{\varphi}\ ] ] using now the orbit radius ( [ eq:55 ] ) and keeping only the corrections linear in @xmath2 , we arrive at the expression we need : @xmath234\hat e_{\varphi } \label{eq:56}\ ] ] to derive a similar formula for @xmath73 from ( [ eq:53 ] ) we , of course , need no orbital velocity but only the orbit radius ( [ eq:55 ] ) ; with accuracy @xmath235 the instantaneous value of the lense thirring precession for gp - b turns out to be @xmath236 @xmath237 ; \label{eq:57}\ ] ] @xmath238\ ] ] for convenience , we have assumed that the orbit is in the plane @xmath239 ; the values of @xmath240 and @xmath241 are given in ( [ eq:54 ] ) , while @xmath242 note that the expressions ( [ eq:56 ] ) and ( [ eq:57 ] ) contain the corrections coming both directly from the non - spherically symmetric gravito - electric and gravito - magnetic field and also through the influence of the former on the orbit . the first contribution can be recognized from the second in the lense thirring precession ( [ eq:57 ] ) by the presence of the factor @xmath243 , however , in the geodetic precession ( [ eq:56 ] ) this difference is concealed . orbit - averaged precessions : the gp - b drift rates_. with the above results , it is now straightforward to carry out the orbit averaging which reduces to a simple integration over @xmath244 . in this way , to lowest order in the earth s oblateness @xmath245 using ( [ eq:56 ] ) we find @xmath246\hat e_{\varphi}={3gmv\over 2r_0 ^ 2}\ , \biggl [ 1 - 1.003\times10^{-3 } \biggr]\hat e_{\varphi } , \label{eq:59}\ ] ] where the values @xmath247 and @xmath248 for the earth s and orbit radius have been taken . the result exactly coincides with the one found by breakwell @xcite in a different way ( he gave the formula for the orbit of any inclination ) . since gp - b is intended to measure the geodetic precession and the eddington parameter @xmath1 to about a part in @xmath6 , a @xmath249 correction is critically important . note that the gp - b gyro spin axis will be initially aligned with the reference direction to the guide star in the orbit plane , so the geodetic effect will cause the in - plane drift of the spin in the north south direction . in a similar fashion , from ( [ eq:57 ] ) we derive @xmath250\hat z \label{eq:60}\ ] ] for our model b , @xmath251 according to ( [ eq:54 ] ) ; therefore , with @xmath252 for the earth , @xmath253\hat z= { gi\omega\over 2r^3_0}\ , \biggl [ 1 - 2.97\times10^{-4 } \biggr]\hat z \label{eq:61}\ ] ] thus the @xmath2 correction for the model b is @xmath254 , which is essentially beyond the expected gp - b accuracy for the lense thirring effect . for the model a the correction is of the opposite sign and almost one order of magnitude smaller , about @xmath255 . for a different model of the mass distribution inside the earth teyssandier @xcite has obtained a slightly larger correction of about @xmath256 , with the same relevance to gp - b . effect of the moon_. according to formula ( [ eq:21 ] ) from section 5 the geodetic effect from a distant mass such as the moon scales with the mass and inversely with the square of distance . since the moon has a small mass and is at a large distance we expect its effect to be small . a rough estimate for the geodetic precession due to the moon is @xmath257 which is too small to be of significance for gp - b experiment . the lense thirring effect scales with the angular momentum and inversely with the cube of the distance . since the velocity of the moon in its orbit is small we expect its effect to be small , and the estimate @xmath258 shows that it is really far beyond the gp - b accuracy . thus the moon is of no consequence for the gp - b experiment . effect of the sun_. although the sun is quite distant its mass is large compared to the earth so it is not apparent how large its effect on the precession will be . we can calculate this using ( [ eq:18 ] ) . in the inertial frame centered on the sun the velocity of the satellite will be the sum of the satellite velocity in the earth orbit plus the orbital velocity of the earth @xmath259 but the satellite velocity averages to zero in the course of one orbit , while the other factors in ( 18 ) change very little , so for long time averages we may neglect the satellite velocity compared to the earth orbital velocity . treating the sun as a point mass and approximating the earth orbit as a circle we may thus write ( [ eq:18 ] ) as @xmath260 where the unit vector @xmath261 is perpendicular to the plane of the ecliptic and @xmath262 is the earth - sun distance . the numerical value of this , with @xmath56 , is @xmath263 this contribution to the gp - b precession is not negligible and must be included in the data analysis . it was first discussed by desitter in 1916 @xcite , @xcite , and further information and references can be found in the book of will @xcite . note an important fact concerning the direction of the above precession vector : it is perpendicular to the plane of the ecliptic , whereas the geodetic precession vector due to the earth lies mainly in the equatorial plane of the earth , so the two are not parallel . indeed the precession due to the sun is roughly in the direction of the lense - thirring precession due to the earth . finally we may estimate the precession due to the spin of the sun by using ( [ eq:44 ] ) . since the spin period of the sun is about 1 month this gives roughly @xmath264 which is far too small to be of relevance to the gp - b experiment . this work was supported by nasa grant nas 8 - 39225 to gravity probe b. we are grateful to the members of gp - b theory group , especially , to c.w.f.everitt , g.m.keiser , r.v.wagoner , and p.w.worden , for many fruitful discussions and enlightening comments . let us show briefly the implementation of the four - step procedure described in section 8 which allows us to obtain expressions ( [ eq:48 ] ) for gravito - magnetic multipole coefficients @xmath120 . we do it for the earth s model a ; calculations for the model b are similar . first , we introduce the model a density distribution ( [ eq:46 ] ) into the definition ( [ eq:36 ] ) of the general moment @xmath265 of the density and calculate this moment for the shape of a slightly oblate ellipsoid of revolution ( assumption 2 of section 8) . working to the first order in the eccentricity @xmath173 , we find @xmath265 in terms of the spherical harmonics coefficients @xmath266 of the function @xmath191 from ( [ eq:46 ] ) , namely : @xmath267+o\bigl((\epsilon k)^2\bigr ) \eqno(a.1)\ ] ] here @xmath268 are known positive rational fractions of @xmath114 and @xmath27 bounded for all pertinent values of those parameters , @xmath269 @xmath270 formula ( a.1 ) is slightly different for the case @xmath271 , but we do not need it . for @xmath272 the left - hand side of the equality ( a.1 ) is given via @xmath273 according to ( [ eq:37 ] ) : @xmath274 @xmath275+o\bigl((\epsilon l)^2\bigr)\ ] ] this is a tri - diagonal system of linear algebraic equations for @xmath276 with small off - diagonal elements . solving it for @xmath276 , we express the latter in terms of @xmath273 : @xmath277 @xmath278+o\bigl((\epsilon l)^2\bigr),\ ] ] where @xmath279 introducing now the expression ( a.2 ) back into the general formula ( a.1 ) , we get _ all _ moments , with _ any k _ , expressed through the gravitational coefficients : @xmath280 @xmath281 \biggr\ } + o\bigl((\epsilon l)^2\bigr),\ ] ] with the quantities @xmath282 simply related to @xmath268 , respectively : @xmath283 we finally set @xmath284 in ( a.3 ) and use the resulting expressions in the relation ( [ eq:38 ] ) , completing thus the last fourth step of the procedure described in section 8 and obtaining the equalities ( [ eq:48 ] ) for @xmath120 ( model a ) .

field theoretical models in external electromagnetic fields are of great interest and have attracted a lot of attention @xcite . they are relevant to the study of many physical systems whose properties depend on the effects produced by external fields , especially when such effects are non - perturbative in nature . a particular example of such a situation is the so - called magnetic catalysis of chiral symmetry breaking @xcite ( see also @xcite for some earlier studies ) which has many potential applications in condensed matter physics @xcite and in studies of the early universe @xcite . it is well known that a sufficiently strong attractive interaction between massless fermions can result in a chiral symmetry breaking condensate and a dynamically generated fermion mass . in the nambu - jona - lasinio ( njl ) model , for example , this occurs when the interaction strength is at a particular large critical value . it is also now known that the presence of an external magnetic field can be a strong catalyst for this effect , leading in some cases to the generation of a dynamical mass for fermions even by very weak attractive interactions . for example , in an external magnetic field , the critical value of the njl coupling is reduced to zero . it was suggested in @xcite that this magnetic catalysis of chiral symmetry breaking is a rather universal phenomenon and that its main features are model independent . the arguments were based on the observation that the effect is primarily due to the dynamics of the fermions from the lowest landau level which are gapless and should therefore dominate the behavior of the system at long wavelengths . in this paper we will develop this idea further . we shall investigate the low - energy dynamics of chirally symmetric models with a short range interaction in an external magnetic field . the idea is to construct an effective action for the low energy degrees of freedom . in strong external magnetic fields , these are the modes of charged particles in the lowest landau level . in the absence of interactions , their effective dynamics is described by a field theory whose dimensionality is two less than the spacetime dimension . we shall take the hypothesis that , at least for strong magnetic fields and short ranged interactions , the dynamics of chiral symmetry breaking can be understood in the lower dimensional effective theory . as an illustration , we shall argue that this assumption is very natural in 2 + 1 dimensions . there , the modes of the lowest landau level are non - propagating , and in the absence of interactions , the ground state is highly degenerate and contains both chirally symmetric and chirally non - symmetric states . interactions , even with infinitesimally small coupling constants , will split this degeneracy , and whether chiral symmetry breaking occurs or not depends on whether the interaction favors the chirally non - symmetric or chirally symmetric ground states . one by - product of our analysis will be the fact that the dynamical generation of parity violating masses seems to be suppressed by magnetic fields . in the more complicated case of ( 3 + 1)-dimensional gauge theory , we shall argue that the infrared dynamics is governed by an effective ( 1 + 1)-dimensional gross - neveu like theory with an infinite number of flavors of fermions . we derive the most general low energy effective action which is compatible with the symmetries of the four dimensional theory . we then study the resulting model using the renormalization group . we will present a plausible argument for our main conclusion : that the presence of chiral symmetry breaking depends crucially on the sign ( but not the magnitude ) of one particular moment of the coupling constants which we denote by @xmath0 . if it is of the appropriate sign , its infrared renormalization group flow is to strong coupling and chiral symmetry is broken . with the other sign it flows to zero coupling and chiral symmetry remains unbroken . consider a system of relativistic ( 2 + 1)-dimensional fermions with the dirac hamiltonian @xmath1 where @xmath2 is the covariant derivative and @xmath3 are the first two pauli matrices . this hamiltonian is obtained in the continuum limit of some tight - binding lattice models relevant to condensed matter physics on a square lattice with half of a quantum of magnetic flux through each plaquette ( see @xcite and references therein ) or on a honeycomb lattice without a magnetic field @xcite.it has also been discussed in the context of the continuum limit of the _ d_-wave state of high @xmath4 superconductors @xcite . @xmath5 commutes with any linear combination of the matrices @xmath6 which generate the lie algebra of a @xmath7 flavor symmetry which we will call chiral symmetry . possible mass terms which can be added to @xmath5 must be matrices which anti - commute with it . the basic matrix which anti - commutes with @xmath5 and which commutes with all of the flavor generators is @xmath8 and a mass term which one could add to @xmath5 which would preserve the flavor symmetry is @xmath9 . generally , in 2 + 1 dimensions , fermion mass terms violate parity symmetry . a parity transformation which is a symmetry of the massless dirac equation @xmath10 and which commutes with the flavor symmetry generators @xmath11 is @xmath12 the mass term @xmath9 is odd under this transformation and it would therefore violate parity symmetry . more generally , all mass terms of the form @xmath13 are odd under the parity defined in eq . ( [ pparity ] ) . moreover , there is no modification of the parity transformation which would make the mass @xmath9 parity symmetric . it is possible to get a parity invariant mass by breaking the flavor symmetry . for this , we must define parity as a combination of the transformation ( [ pparity ] ) and a discrete flavor transformation , for example , @xmath14 this is a symmetry of @xmath5 and it also commutes with the mass term @xmath15 . this mass term breaks the @xmath7 flavor symmetry to @xmath16 . this existence of the unitary matrix @xmath17 which anti - commutes with @xmath5 implies that @xmath5 has a symmetric spectrum : if @xmath18 , then @xmath19 and , with a particular choice of phase , @xmath20 . thus , for each positive energy state , there is a negative energy state and the probability measure @xmath21 is a symmetric function of @xmath22 . the interesting feature of @xmath5 is the appearance of zero energy states in a background magnetic field . consider the equation @xmath23 where the background electromagnetic field is static and has the property @xmath24 . it has the solution @xmath25 v(x),\ ] ] where @xmath26 and @xmath27 the second equation in eqs . ( [ flux ] ) has the solution @xmath28 with asymptotic behavior @xmath29 ( this is strictly correct if the magnetic field goes to zero sufficiently quickly at infinity for the total flux @xmath30 to be finite . if it is not finite , there is an infinite number of zero modes and their density is given by @xmath31 where @xmath32 is the volume . ) asymptotically , @xmath33 in order to have a normalizable zero mode , the spinor @xmath34 should be an eigenvector of @xmath35 with eigenvalue @xmath36 ( @xmath37 if @xmath38 and @xmath39 if @xmath40 ) . it then must be a normalizable solution of the equations @xmath41 solutions of this equation are @xmath42^k$ ] for @xmath43 - 1 $ ] where @xmath44 $ ] is the largest integer less than @xmath45 and the maximum value of @xmath46 is the largest power allowed by normalizability of the wave function . thus there are @xmath47 $ ] normalizable zero modes . this fact depends only on the total magnetic flux and is independent of the profile of the magnetic field . the result , which we have found by explicit computation , is also a result of the atiyah - singer index theorem generalized to open spaces@xcite . some other applications of this idea are discussed in @xcite . there are thus @xmath48 $ ] zero mode wave functions of @xmath5 . all of them obey the equation @xmath49 @xmath47 $ ] of them obey the equation @xmath50 and @xmath47 $ ] of them obey @xmath51 of course , any linear combination of them is also a zero mode and could be chosen to be eigenvalues of any of the flavor symmetry breaking mass operators @xmath52 . all linear combinations would satisfy eq . ( [ eig ] ) . the existence of zero modes implies that the dirac ground state of the second quantized system of fermions is degenerate . in the ground state , all negative energy states of the fermions are full and all positive energy states are empty . for overall charge neutrality , half of the fermion zero modes must be occupied but the occupation of zero modes is otherwise unconstrained . this means that the degeneracy of the ground state is @xmath53 ^ 2 $ ] where @xmath54 $ ] is the number of zero mode wave - functions . in second quantization , the expectation value of any fermion bilinear operator , using the dirac commutator for normal ordering , is @xcite @xmath55 since , in the dirac ground state , all positive energy states are unoccupied and negative energy states are occupied , and since @xmath17 maps positive energy states onto negative energy states , if the matrix @xmath56 commutes with @xmath17 , only the zero modes survive in this summation . first of all , this means that in a magnetic field the expectation value of the parity violating mass operator vanishes , @xmath57 regardless of the occupation of zero modes . this has the implication that external magnetic fields do not enhance the generation of parity violating mass terms . in fact , it indicates that if the external field is strong enough that the zero modes are well isolated , the expectation value of a parity breaking mass operator , @xmath58 should always vanish . on the other hand , the parity invariant , chiral symmetry breaking mass ( with @xmath59 ) has the expectation value @xmath60 it is possible to choose ground states where this expectation value has any value between two bounds : @xmath61 \leq \left\langle\int d^2x : \bar\psi(x)t^3\psi(x):\right\rangle \leq [ \vert\chi\vert].\ ] ] if we added an infinitesimal mass term @xmath62 to the hamiltonian @xmath5 , the system would choose the ground state with minimal @xmath63 , that is the one where is finite , the dirac hamiltonian has a continuum spectrum whose support begins at zero energy . furthermore , there is a threshold density of continuum states whose contribution to bilinear densities such as eq . ( [ chrl ] ) makes up the difference between @xmath47 $ ] and @xmath45 which is proportional to the spectral asymmetry of the hamiltonian and is the correct result . for details , see @xcite . in the present paper , for simplicity we ignore the relatively small error which we make in neglecting the asymmetry of the continuum spectrum . when the magnetic flux is infinite , for example in the case of constant external field , @xmath47 $ ] should be replaced by @xmath64 where @xmath32 is the spatial volume . ] @xmath65\mbox{sgn}(m\chi ) . \label{chrl}\ ] ] the observation that this is so for infinitesimal mass was made in @xcite . however , this does not mean that chiral symmetry is necessarily broken . it only implies that the ground state is degenerate and there are chirally non - symmetric ground states which are degenerate with chirally symmetric ones . if there are additional weak interactions , of course , the correct ground state should be found by resolving the degeneracy using degenerate perturbation theory . in this case , all that is required to break chiral symmetry is an interaction which favors a chirally non - symmetric population of the zero modes . if the interaction favors the breaking of chiral symmetry , symmetry breaking occurs for even an infinitesimal value of its coupling constant . this is the reason why the critical value of the interaction can be at zero coupling in the presence of a magnetic field . in fact , in the absence of external magnetic fields a critical coupling necessary to break chiral symmetry is typically large . in the presence of the magnetic field it is reduced to zero . for example , it is well known that an interaction of the njl model , @xmath66 when added to the hamiltonian will break the chiral symmetry if the coupling constant is attractive and is greater than a particular critical value . this is seen by analyzing solutions of the schwinger - dyson equation for a mass condensate . it was shown in @xcite that , in the presence of a magnetic field , the critical coupling moves from some finite value to zero even an infinitesimal coupling will break the chiral symmetry . that observation is consistent with our finding here . an interesting test of this idea would be to examine the effect on system with the interaction @xmath67 which can break parity by generating a parity violating fermion mass if @xmath68 has the appropriate sign and is of large enough magnitude @xcite . our analysis seems to suggest that a large magnetic field should in fact tend to increase the magnitude of the critical coupling . in 3 + 1 dimensions , the situation is more complicated . the zero modes of the dirac hamiltonian in a background magnetic field still have some dynamics which is non - trivial and there are more possibilities for interactions . to understand the general strategy that we will take , begin with the lagrangian which describes dirac fermions interacting with an external electromagnetic field , @xmath69 where @xmath70 . for concreteness , we choose the chiral representation of the dirac matrices , @xmath71 and we shall use the notation @xmath72 the lagrangian has the form @xmath73 . \label{s_0}\ ] ] this action has a global @xmath74 chiral symmetry . in the following , we must keep in mind that when there are both background electric and magnetic fields , or if the gauge field is dynamical , this symmetry is reduced to the vector @xmath75 by the axial anomaly . for simplicity , in this paper we will assume that this is not the case . it would be straightforward to extend the present work to models with dynamical @xmath75 fields , by introducing more species of fermions so that there exist chiral symmetries which are unaffected by the axial anomaly . for now , we will consider the case of a constant non - dynamical background magnetic field , @xmath76 in this background field , it is convenient to make use of the mixed spacetime - momentum representation as follows : @xmath77,\ ] ] where , by definition , @xmath78 , @xmath79 and @xmath80 the spectrum of @xmath81 is well known . the equation @xmath82 has eigenvalues @xmath83 for all integers @xmath84 . when @xmath85 , these are dispersion relations of ( 1 + 1)-dimensional dirac fermions with masses given by @xmath86 . [ of course , this mass gap is not lorentz invariant from a four - dimensional point of view . however , it is lorentz invariant in 1 + 1 dimensions . the reason for this is that there is a subgroup of the lorentz group which survives in the background magnetic field it is invariant under boosts in the direction of the field lines . here this will mean that the effective theory for landau levels must be invariant under the ( 1 + 1)-dimensional lorentz transformations . ] zero modes of @xmath87 are solutions of the equation @xmath88 and are given by the infinite set of ortho - normal functions @xmath89 with @xmath90 . here , without loss of generality , we assumed that @xmath91 . in the ( 1 + 1)-dimensional theory , the zero modes of @xmath87 correspond to massless fermions . the lagrangian is ( restoring the space dependence ) @xmath92 where @xmath93 and the new fields are defined as the coefficient functions in the expansions of @xmath94 and @xmath95 over the complete set of eigenstates : @xmath96 note that the kinetic term in the action ( [ massless ] ) for the massless modes appears to have a @xmath97 with @xmath98 symmetry . this is just the unitary symmetry which mixes the different wave functions of the degenerate zero modes , and which is actually there for every landau level . this effective symmetry is not preserved by interactions . we will analyze the possibility that interactions that are added to this field theory drive a spontaneous breaking of the @xmath99 chiral symmetry by generating a mass gap for the fermions in eq . ( [ massless ] ) . we expect that this spontaneous symmetry breaking takes place at very long wavelengths . our approach to the problem in the following sections can be summarized as follows : 1 . we consider the theory of four - dimensional fermions in a magnetic field as described above , with some interactions which should be local and respect ( 3 + 1)-dimensional poincar and chiral symmetry but are otherwise unspecified . we then consider the effective field theory which is obtained by integrating out all of the massive modes of the fermions in eq . ( [ massless ] ) and momentum states of the massless modes above an ultraviolet cutoff . we assume that the resulting effective lagrangian is local . to guarantee locality , we generally have to assume that the interactions between fermions in four dimensions are short ranged . furthermore , the ultraviolet cutoff for the effective theory should be less than the mass gap of the lightest massive mode , @xmath100 . we consider all relevant operators which could be added to the effective lagrangian which are consistent with the symmetries of the theory . since they should be relevant in the sense of two dimensional field theory , these are all possible four - fermion operators which are consistent with symmetry . there are an infinite number of such operators . we compute the beta function for the coupling constants of the relevant operators and look for infrared stable fixed points . these fixed points should govern the behavior of the very long - wavelength degrees of freedom of the theory . 6 . if the coupling constants flow to an infrared stable fixed point , then an infrared limit of the massless theory exists and there is no symmetry breaking . if , on the other hand , the coupling constant flow is not in the domain of attraction of any infrared fixed point , so that it flows to strong coupling in the infrared , we postulate that this implies the dynamical generation of a mass gap and spontaneously broken chiral symmetry . we shall find examples of both kinds of behavior . an example of a local four dimensional interaction which preserves the @xmath101 chiral symmetry is the njl interaction @xcite , @xmath102 . \label{lagnjl}\ ] ] a renormalizable version of this interaction would be one which is mediated by a massive scalar mesons . the constraints of ( 1 + 1)-dimensional lorentz invariance and @xmath103 chiral symmetry allow four - fermion coupling constants as in the effective theory , @xmath104 the coupling constants @xmath105 obey further constraints from charge conjugation , parity and time reversal ( cpt ) symmetry and the symmetry of the underlying theory under translations and rotations about the axis of the magnetic field which are summarized in appendix [ appa ] . a general solution of those constraints would yield the most general allowed structure of the low - energy effective action in eq . ( [ eff - gen ] ) . some particular solutions of those constraints are of interest . for example , the constraints allow the maximally symmetric solution with @xmath97 ( @xmath106 ) symmetry , @xmath107 where @xmath108 is real . similarly , there is a solution with @xmath109 ( with @xmath106 ) symmetry , @xmath110 where @xmath111 is real again . in addition a large class of solutions could be found by the reduction of the interactions in the original ( 3 + 1)-dimensional model to the lowest landau level . after such a reduction of the njl interaction ( [ lagnjl ] ) , we arrive at @xmath112 since it descends from a lorentz and chirally invariant interaction in 3 + 1 dimensions , it must necessarily satisfy the constraints of symmetry [ see eqs . ( [ real ] ) , ( [ parity ] ) , ( [ restrict1 ] ) , ( [ restrict2 ] ) and ( [ nnmm ] ) ] . unlike the first two solutions in eqs . ( [ unun ] ) and ( [ uv ] ) , this last one in eq . ( [ lll - int ] ) does not seem to have any extra symmetry in addition to the required @xmath103 flavor symmetry . now , we envisage having obtained the effective action ( [ eff - gen ] ) by integrating out all modes in the higher landau levels , as well as all momentum modes of the fermions in the lowest landau level which are above a certain cutoff . of the many interactions that this procedure would produce , we have kept only the local four - fermion operators . this procedure is legitimate only if the ultraviolet cutoff of this model is lower than the lowest mass gap of the fields which have been eliminated , i.e. @xmath113 . by chiral symmetry and lorentz invariance , the effective action can not contain mass terms for the fermions . furthermore , the only lorentz invariant four - fermion operator is of the form given in eq . ( [ eff - gen ] ) . the renormalization group procedure examines how the coupling constants in eq . ( [ eff - gen ] ) change as we further lower the cutoff to isolate the very long wavelength excitations . this information is encoded in the beta function . the @xmath17 function for the general coupling constant in eq . ( [ eff - gen ] ) is computed in @xmath114 dimensions to two loop order in appendix [ appb ] . the result is @xmath115 \nonumber\\[3 mm ] -\frac{1}{8\pi^2 } \sum_{k_1,k_2 } & & \left [ g\left(\begin{array}{cc } n_1 & k_2 \\ k_1 & n_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+k_1-k_2 & n_2 \\ m_1 & m_2+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } m_2+k_1-k_2 & k_1 \\ k_2 & m_2 \end{array}\right)\right . \nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2 \\ m_1 & m_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1 & n_2-m_1+k_2 \\ k_2 & m_2 \end{array}\right ) g\left(\begin{array}{cc } m_1+k_1-k_2 & n_2 \\ n_2-m_1+k_2 & k_1 \end{array}\right ) \right ] \nonumber\\[3 mm ] + \frac{2}{(4\pi)^2 } g\left(\begin{array}{cc } n_1 & n_2 \\ m_1 & m_2 \end{array}\right ) & & \sum_{k_1,k_2 } \left [ g\left(\begin{array}{cc } n_1 & k_2+n_1 \\ k_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & k_1 \\ k_2+n_1 & n_1 \end{array}\right ) + g\left(\begin{array}{cc } k_1 & k_2+n_2 \\ n_2 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & n_2 \\ k_2+n_2 & k_1 \end{array}\right)\right.\nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2+m_1 \\ m_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & m_1 \\ k_2+m_1 & k_1 \end{array}\right ) + g\left(\begin{array}{cc } m_2 & k_2+m_2 \\ k_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & k_1 \\ k_2+m_2 & m_2 \end{array}\right)\right ] . \nonumber\end{aligned}\ ] ] finally , before we proceed to the next section , we note that we may have to occasionally cut off the summation over modes in the first landau level . we do this by summing modes up to some maximum number , @xmath116 . such a situation could be produced by considering the fermions in an external magnetic field with finite magnetic flux , or with uniform field and finite transverse area . in that case , @xmath117 which can be thought of as the number of flavors " which goes to infinity as @xmath118 where @xmath119 is the area of the two - dimensional perpendicular ( with respect to the direction of the magnetic field ) subspace . in appendix [ appb ] , we have computed the two - loop @xmath17 function ( [ beta2loop ] ) for the most general coupling constants , subject to the restriction in eq . ( [ nnmm ] ) . the next natural step would be locate infrared stable fixed points of this beta function . this problem is far more complicated than we can solve at present . instead , in our analysis below , we shall restrict our attention to the beta function for coupling constants with close to maximal symmetry . for some scalar field theories , where there are many components of a scalar field and renormalizable interactions which couple them , it is known that the only infrared fixed points of the renormalization group flow of relevant couplings are those with maximal symmetry @xcite . to our knowledge , there is no similar theorem for gross - neveu like models . on the other hand , we consider it plausible that similar arguments can be applied : in particular , if no maximally symmetric infrared stable fixed point exists , then there are no infrared stable fixed points at all . maximally symmetric couplings can easily be shown to be contained in other combinations of coupling constants . in appendix [ appc ] we show that the njl coupling contains the maximally symmetric ones . consider the renormalization group flow in the special case when @xmath97 and @xmath109 couplings , as defined in eqs . ( [ unun ] ) and ( [ uv ] ) , are the only non - zero couplings in the effective action . then , from our general result in the previous section , we extract the following expressions for the @xmath17 functions of interest : @xmath120 . \label{beta_v}\end{aligned}\ ] ] note that we have been forced to introduce a cutoff on @xmath117 . in order to get a sensible result when taking the limit @xmath106 , it is convenient to rescale the couplings as follows : @xmath121 and @xmath122 . as is clear from the definition , the @xmath17 functions will also get rescaled accordingly , @xmath123 and @xmath124 . after taking this into account and performing the limit @xmath106 , we arrive at @xmath125 it must be emphasized that we have adopted the rescaling of the most symmetric @xmath108 coupling by @xmath126 rather than @xmath127 . since the rescaling performed above led to the well - defined @xmath17 functions , we conclude that the ( 1 + 1)-dimensional action in eq . ( [ eff - gen ] ) with @xmath108 of order @xmath128 describes a consistent and non - trivial interacting theory in the limit @xmath106 . by @xmath127 is also meaningful . the resulting @xmath17 functions are @xmath129 the corresponding theory is less interesting . indeed , the expressions for the @xmath17 functions in eq . ( [ b_lr ] ) and ( [ b_v ] ) in the limit @xmath130 describe the situation when the @xmath131 coupling does not run at all , while the @xmath132 experiences asymptotic freedom from the @xmath133 side and infrared freedom from the @xmath134 side . we shall see in a moment that this picture corresponds to a special case ( @xmath135 ) of the flow described by eqs . ( [ beta_lr ] ) and ( [ beta_v ] ) . ] when @xmath136 we can solve the renormalization group equations ( [ beta_lr ] ) and ( [ beta_v ] ) explicitly , and the analytical solution reads @xmath137 this flow is presented graphically in fig . [ fig6 ] where the arrows show the flow direction toward ultraviolet . in the upper half - plane @xmath133 , the simple analysis of the flow given by eqs . ( [ sol - lr ] ) and ( [ sol - v ] ) reveals infrared ( with @xmath138 ) and ultraviolet ( with @xmath139 and @xmath138 ) landau poles at @xmath140 respectively . we argue that the strong infrared dynamics ( with @xmath138 ) in this half - plane of couplings is an indication of a mass generation and breaking of the chiral @xmath75 symmetry . indeed , the generation of the fermion mass in the infrared region seems to be the only way one can avoid running into the problem of the physical landau pole . the generation of the fermion mass in the @xmath133 half - plane , in its turn , is consistent with the expectation of the universality of the magnetic catalysis in a wide range of ( 3 + 1)-dimensional models ( such as the njl ) with a short range interaction . indeed , as we established , the low energy dynamics in such models is described by the ( 1 + 1)-dimensional effective action in eq . ( [ eff - gen ] ) with the coupling satisfying the set of restriction in eqs . ( [ restrict1 ] ) , ( [ restrict2 ] ) and ( [ nnmm ] ) . the generic coupling [ say , like that in eq . ( [ lll - int ] ) coming from the interaction of the lowest landau level modes ] which does not have any extra symmetry is still expected to have the @xmath97 and @xmath109 contribution [ see eqs . ( [ glll - lr ] ) and ( [ glll - v ] ) ] . then , if this @xmath109 contribution is positive , @xmath133 , it is going to drive the system to the generation of mass . now let us study the flow in the lower half - plane @xmath141 . as is easy to see , there is an infrared fixed point at @xmath142 and the ultraviolet landau pole ( with @xmath143 while @xmath131 is either fixed or approaches @xmath144 ) at the following values of @xmath145 : @xmath146 since the infrared fixed point @xmath147 corresponds to weakly coupled dynamics , there is apparently no mass generation in this half - plane of the coupling space . this is also in full agreement with our general expectation . indeed , the negative values of @xmath132 correspond to the repulsion rather than attraction in the fermion - antifermion channel which is responsible for the generation of mass , breaking chiral @xmath75 symmetry . our conclusion here is that whether this model breaks chiral symmetry or not is entirely dependent on the sign of the coupling constant @xmath0 . the reader might be puzzled by the fact that a @xmath75 chiral symmetry can be broken in an effectively two dimensional system . we emphasize here that this phenomenon is identical to that in the chiral gross - neveu model . strictly speaking , chiral symmetry can only be broken in the large @xmath117 limit . the finite @xmath117 system should still be chirally symmetric , as the low dimensionality of the system would not allow for spontaneous breaking of a continuous symmetry . indeed , if we consider the case of a large but finite value of @xmath117 , an infrared stable fixed point appears for @xmath148 at @xmath149 which goes to infinity with @xmath106 . this fixed point is at strong coupling , so a conclusion based on perturbation theory is speculative at best , but its appearance is consistent with the expectation that the massless limit of this model is well defined ( albeit strongly coupled ) when @xmath148 and @xmath117 is finite and chiral symmetry breaking need not take place . when @xmath98 this fixed point moves to infinite coupling and chiral symmetry breaking is possible . in this paper we have shown that the effective low - energy dynamics of the @xmath75 chirally symmetric models with a short range interaction in a background magnetic field is described by a ( 1 + 1)-dimensional gross - neveu like model with an infinite number of flavors [ see eq . ( [ eff - gen ] ) ] . here we established that different flavors come out as the representation space of the magnetic translations in the original ( 3 + 1)-dimensional model . the number of flavors is infinite and proportional to the area of the two dimensional space perpendicular to the magnetic field . based exclusively on the arguments of symmetry , we established a set of conditions , given by eqs . ( [ real ] ) , ( [ parity ] ) , ( [ restrict1 ] ) , ( [ restrict2 ] ) and ( [ nnmm ] ) , that the couplings of the effective theory have to satisfy . to show that they allow a non - trivial solution , we presented a few ( out of infinitely many possible ) examples of couplings that satisfy all the constrains . among them , there are , in particular , the highly symmetric @xmath97 and @xmath109 ( with @xmath106 ) couplings . these latter are of special interest because their renormalization group flow is self - contained and allows an analytical solution [ see eqs . ( [ sol - lr ] ) and ( [ sol - v ] ) ] . at the level of the effective theory , we calculated the two - loop @xmath17 function and analyzed the renormalization group flow in the two - dimensional subspace of the @xmath97 and @xmath109 couplings . the general result is argued to indicate the generation of the fermion mass in the @xmath133 half - plane of couplings . in the other half - plane , the infrared dynamics is weakly coupled and there is no mass generation . this mass generation pattern is consistent with the earlier suggested conjecture of the universality of the so - called magnetic catalysis @xcite in , at least , the models from the same universality class as the chiral @xmath75 njl model . we would like to thank v. miransky for many interesting discussions . the work of g.w.s . was supported by the natural sciences and engineering research council of canada and by the niels bohr fund of denmark . the work of i.a.s . and l.c.r.w . was supported by the u.s . department of energy grant no . de - fg02 - 84er40153 . let us clarify the origin of the action as well as the meaning of different flavors in eq . ( [ eff - gen ] ) in terms of the original ( 3 + 1)-dimensional model defined in eqs . ( [ s_0 ] ) and ( [ lagnjl ] ) . we start from the analysis of the space - time symmetries in the model . notice that , due to the presence of the background field , the standard translations in the two - dimensional plane perpendicular to the magnetic field are not symmetries of the original model in eq . ( [ s_0 ] ) . nevertheless , there are other transformations , the so - called magnetic translations , which leave the action invariant . in contrast to the case of the ordinary translations , the two generators of the magnetic translations do not commute . in the symmetric gauge given in eq . ( [ dmsu1 ] ) , the explicit representation of the generators and their commutation relation read @xmath150=i , \label{x&yc } \end{aligned}\ ] ] where we assume that @xmath91 . it is easy to check that the operators @xmath151 and @xmath152 commute with the hamiltonian of our model . instead of these @xmath151 and @xmath152 , it is convenient to introduce the creation " and annihilation " operators @xmath153=1 . \label{a&c}\end{aligned}\ ] ] having introduced these operators in the problem , we realize that the fock space is spanned by the set of states @xmath154 where the quantum number @xmath155 denotes the eigenvalue of @xmath156 operator and the multi - index @xmath157 represents all the other quantum numbers ( say , the landau level number , the fermion spin projection and the chirality ) . in the absence of any vacuum rearrangement ( symmetry breaking ) , the above set of states ( in coordinate representation ) reads @xmath158 where @xmath159 is the magnetic length , @xmath160 , @xmath161 , @xmath162 and @xmath163 is the spinor with a given spin projection @xmath164 and chirality @xmath30 . note that the expression in eq . ( [ states ] ) is well defined even for the case when @xmath165 ( both numbers are positive ) due to the rodrigues formula for the generalized laguerre polynomials , @xmath166 one has to remember also that all the modes in the lowest landau level ( @xmath167 ) have the same projection of the spin , while the modes in the higher landau levels ( @xmath85 ) have both projections . since we are interested in the structure of the effective action , describing the infrared dynamics ( @xmath168 ) , it is sufficient to take into account only those degrees of freedom that originate from the lowest landau level modes ( @xmath167 ) . these modes freely propagate in the ( 1 + 1)-dimensional parallel @xmath169 subspace , and are classified by the chirality and the eigenvalue of @xmath156 [ see eq . ( [ states ] ) ] . in what follows , we denote the effective degrees of freedom accordingly , @xmath170 , where the superscript @xmath171 or @xmath172 denotes the states that result from the ( 3 + 1)-dimensional states of definite chirality . in the parallel subspace @xmath169 , @xmath173 and @xmath174 have the interpretation of the left and right moving along the @xmath175 axis modes , respectively . while restricting the kinetic term in the njl model ( [ lagnjl ] ) to the lowest landau level modes , @xmath176 we check that @xmath177 . after performing the integration over the perpendicular space coordinates in the original njl action , we arrive at the effective model as in eq . ( [ eff - gen ] ) . remarkably , while the njl coupling @xmath178 is dimensionful , the effective coupling in eq . ( [ eff - gen ] ) is dimensionless , @xmath179 . from the derivation above , we see that the fields of different flavors in the effective ( 1 + 1)-dimensional theory ( [ eff - gen ] ) correspond to different eigenstates of @xmath156 operator . this simple observation , as we show below , has far reaching consequences . now , let us establish the allowed structure of coupling constants in eq . ( [ eff - gen ] ) . the most general restriction on the couplings comes from the condition of reality of the action . this requires that @xmath180 similarly , the invariance under the parity ( @xmath181 ) leads to another restriction : @xmath182 these two conditions are too general and so are not of great interest or of great power by themselves . it turns out , however , that there are other , more restrictive conditions . by recalling the origin of flavors in eq . ( [ eff - gen ] ) , we realize that the effective theory should enjoy some kind of flavor symmetry that results from the symmetry under magnetic translations of the original ( 3 + 1)-dimensional theory . this flavor symmetry , as will become clear in a moment , puts further restrictions on the allowed structure of the four - index coupling in eq . ( [ eff - gen ] ) . the infinitesimal transformations of the magnetic translations and the related rotation in the perpendicular plane are given by the following operators : @xmath183 to determine the transformation properties of the fields of different flavors , we again recall that , by construction , these fields are the eigenstates of the @xmath156 operator . then , by doing a simple exercise , we find that the action of the creation and annihilation operators , @xmath184 and @xmath185 , on the properly normalized fields should read @xmath186 making use of these properties , we check that the kinetic term in the effective action is invariant under the transformations in eqs . ( [ trans1 ] ) , ( [ trans2 ] ) and ( [ trans12 ] ) . in fact , if we had started with a more general , non - diagonal kinetic term in the effective action [ which is not forbidden by the chiral @xmath75 symmetry ] , the requirement of invariance under the magnetic translations would lead us back to the diagonal form as in eq . ( [ eff - gen ] ) . the invariance of the four - fermion interaction @xmath187 in eq . ( [ eff - gen ] ) under the set of transformations in eqs . ( [ trans1 ] ) , ( [ trans2 ] ) and ( [ trans12 ] ) leads to @xmath188 to derive the two - loop @xmath17 function of the effective spinor theory in eq . ( [ eff - gen ] ) , we apply the method of ref . @xcite that was used for the @xmath189 scalar theory in @xmath190 dimensions . in dimensional regularization ( with @xmath191 ) , our renormalized lagrangian density reads @xmath192 where the coupling @xmath178 includes the coupling constant renormalization , @xmath193 , and the ellipsis denotes the counterterms . we remind the reader that the four - index coupling @xmath187 is non - zero only for @xmath194 . this means that the four - fermion term in eq . ( [ ren - lag ] ) contains the sum only over three indices ( say , @xmath195 , @xmath196 , and @xmath197 ) , while the fourth one ( @xmath198 ) is superfluous . before proceeding with the actual calculation of the @xmath17 function , we need to specify how to handle the infrared divergences that appear in the calculation of the feynman diagrams @xcite . such divergences usually come from the propagators of the massless fermions . if treated improperly , they could easily obscure the calculation of the relevant diagrams and eventually lead to a wrong result . to avoid the problem , in what follows , we modify the infrared region by changing the fermion propagators as follows : @xmath199 this infrared regularization procedure respects all the symmetries of the model and does not change the ultraviolet region . now , let us calculate the @xmath17 function . first of all , we recall that the relation between the bare and the renormalized couplings , in the dimensional regularization ( @xmath191 ) , reads @xmath200 this is going to be used in the definition of the @xmath17 function . in calculation , we impose the following renormalization conditions : @xmath201 note that the introduction of the effective infrared cutoff @xmath202 in the fermion propagators earlier allows us to use the renormalization point at @xmath203 . the feynman diagrams of the relevant contributions to the two - point functions at two - loop order are given in figs . [ fig1]a and [ fig1]b . by extracting the divergent ( of order @xmath204 ) terms of these two - loop corrections , we arrive at the equations @xmath205 where , by definition , @xmath206[(q - k)^2-m^2](q^2-m^2)}\right|_{p=0 } \nonumber\\ & = & -\frac{1}{(4\pi)^2\varepsilon}+o(1 ) , \label{def - c_r } \\ ( \mbox{fig.~}\ref{fig1}\mbox{b } ) \to \frac{c^{(l ) } } { ( 4\pi)^2\varepsilon}+o(1 ) & \equiv&-\frac{1}{\mu^{2d-4 } } \left.\frac{\partial}{\partial p_{+ } } \int\frac{d^{d}k d^{d}q}{(2\pi)^{2d } } \frac{(k+p)_{+ } ( q - k)_{- } q_{+ } } { [ ( k+p)^2-m^2][(q - k)^2-m^2](q^2-m^2)}\right|_{p=0 } \nonumber\\ & = & -\frac{1}{(4\pi)^2\varepsilon}+o(1 ) . \label{def - c_l } \end{aligned}\ ] ] thus , we see that @xmath207 . in a similar way , the perturbative expansion for the four - point function reads @xmath208 \nonumber\\[3 mm ] + & & \frac{b_0+b_1\varepsilon}{(4\pi\varepsilon)^2 } \sum_{k_1,k_2}\left [ g\left(\begin{array}{cc } n_1 & n_2 \\ k_1 & n_1-n_2+k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_1 & k_1 \\ k_2 & n_1-n_2+k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_2 & k_2 \\ m_1 & m_2 \end{array}\right)\right.\nonumber\\ & & + \left . g\left(\begin{array}{cc } n_1 & k_1 \\ m_1 & n_1+m_1-k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1+m_1-k_1 & k_2 \\ k_1 & n_1+m_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+m_1-k_2 & n_2 \\ k_2 & m_2 \end{array}\right)\right ] \nonumber \\[3 mm ] + & & \frac{c_0}{(4\pi)^2\varepsilon } \sum_{k_1,k_2 } \left [ g\left(\begin{array}{cc } n_1 & k_2 \\ k_1 & n_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+k_1-k_2 & n_2 \\ m_1 & m_2+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } m_2+k_1-k_2 & k_1 \\ k_2 & m_2 \end{array}\right)\right . \nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2 \\ m_1 & m_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1 & n_2-m_1+k_2 \\ k_2 & m_2 \end{array}\right ) g\left(\begin{array}{cc } m_1+k_1-k_2 & n_2 \\ n_2-m_1+k_2 & k_1 \end{array}\right ) \right ] \nonumber\\[3 mm ] + & & \frac{d_0+d_1\varepsilon}{(4\pi\varepsilon)^2 } \sum_{k_1,k_2 } \left [ g\left(\begin{array}{cc } n_1 & k_2 \\ k_1 & n_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_1 & k_1 \\ m_1 & m_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+k_1-k_2 & n_2 \\ k_2 & n_1-n_2+k_1 \end{array}\right)\right . \nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2 \\ m_1 & k_1-k_2+m_1 \end{array}\right ) g\left(\begin{array}{cc } n_1 & n_2 \\ n_2-n_1+k_1 & k_1 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2+m_1 & n_2-n_1+k_1 \\ k_2 & m_2 \end{array}\right)\right . \nonumber\\ & & + \left.g\left(\begin{array}{cc } n_1 & n_1+k_1-k_2 \\ k_1 & k_2 \end{array}\right ) g\left(\begin{array}{cc } m_1+k_2-k_1 & n_2 \\ n_1+k_1-k_2 & m_2 \end{array}\right ) g\left(\begin{array}{cc } k_2 & k_1 \\ m_1 & m_1+k_2-k_1 \end{array}\right)\right . \nonumber\\ & & + \left . g\left(\begin{array}{cc } n_1+m_1-k_2 & n_2 \\ k_1+k_2-m_2 & k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1 & k_2 \\ m_1 & n_1+m_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1 & k_1+k_2-m_2 \\ k_2 & m_2 \end{array}\right)\right ] . \label{gam-4 } \end{aligned}\ ] ] note that the restriction @xmath194 is satisfied for each four - index coupling that appears here . the coefficients @xmath209 , @xmath210 , @xmath211 and @xmath212 are defined by the following expressions : @xmath213 } \nonumber\\ & = & -\frac{1}{2(4\pi)^2\varepsilon}+o(\varepsilon ) , \label{def - c}\end{aligned}\ ] ] and @xmath214}\nonumber\\ & = & -\frac{1}{(4\pi\varepsilon)^2 } \left(\frac{1}{2}+\varepsilon(1+\gamma ) + \varepsilon\ln\frac{m^2}{4\pi\mu^2}\right)+o(1 ) . \label{def - d}\end{aligned}\ ] ] here @xmath215 is the euler constant . from eqs . ( [ def - a ] ) ( [ def - d ] ) , we obtain @xmath216 after expressing the function @xmath178 in terms of @xmath187 , we arrive at @xmath217 -\frac{b_0 - 2a_0 ^ 2+(b_1 - 4a_1a_0)\varepsilon } { ( 4\pi\varepsilon)^2 } \nonumber\\[3 mm ] & & \times \sum_{k_1,k_2 } \left [ g\left(\begin{array}{cc } n_1 & n_2 \\ k_1 & n_1-n_2+k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_1 & k_1 \\ k_2 & n_1-n_2+k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_2 & k_2 \\ m_1 & m_2 \end{array}\right)\right.\nonumber\\ & & + \left . g\left(\begin{array}{cc } n_1 & k_1 \\ m_1 & n_1+m_1-k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1+m_1-k_1 & k_2 \\ k_1 & n_1+m_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+m_1-k_2 & n_2 \\ k_2 & m_2 \end{array}\right)\right ] \nonumber \\[3 mm ] - & & \frac{c_0}{(4\pi)^2\varepsilon } \sum_{k_1,k_2 } \left [ g\left(\begin{array}{cc } n_1 & k_2 \\ k_1 & n_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+k_1-k_2 & n_2 \\ m_1 & m_2+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } m_2+k_1-k_2 & k_1 \\ k_2 & m_2 \end{array}\right)\right . \nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2 \\ m_1 & m_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1 & n_2-m_1+k_2 \\ k_2 & m_2 \end{array}\right ) g\left(\begin{array}{cc } m_1+k_1-k_2 & n_2 \\ n_2-m_1+k_2 & k_1 \end{array}\right ) \right ] \nonumber\\[3 mm ] - & & \frac{d_0+a_0 ^ 2+(d_1 + 2a_0a_1)\varepsilon } { ( 4\pi\varepsilon)^2 } \sum_{k_1,k_2 } \left [ g\left(\begin{array}{cc } n_1 & k_2 \\ k_1 & n_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_1 & k_1 \\ m_1 & m_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+k_1-k_2 & n_2 \\ k_2 & n_1-n_2+k_1 \end{array}\right)\right . \nonumber\\[3 mm ] & & + g\left(\begin{array}{cc } k_1 & k_2 \\ m_1 & k_1-k_2+m_1 \end{array}\right ) g\left(\begin{array}{cc } n_1 & n_2 \\ n_2-n_1+k_1 & k_1 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2+m_1 & n_2-n_1+k_1 \\ k_2 & m_2 \end{array}\right ) \nonumber\\ & & + g\left(\begin{array}{cc } n_1 & n_1+k_1-k_2 \\ k_1 & k_2 \end{array}\right ) g\left(\begin{array}{cc } m_1+k_2-k_1 & n_2 \\ n_1+k_1-k_2 & m_2 \end{array}\right ) g\left(\begin{array}{cc } k_2 & k_1 \\ m_1 & m_1+k_2-k_1 \end{array}\right ) \nonumber\\ & & + \left . g\left(\begin{array}{cc } n_1+m_1-k_2 & n_2 \\ k_1+k_2-m_2 & k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1 & k_2 \\ m_1 & n_1+m_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1 & k_1+k_2-m_2 \\ k_2 & m_2 \end{array}\right)\right].\label{g(g)}\end{aligned}\ ] ] the @xmath17 function is defined as follows @xcite : @xmath218 = 2\varepsilon\frac { g\left(\begin{array}{cc } n_1 & n_2 \\ m_1 & m_2 \end{array}\right ) } { \sqrt{z^{(l)}_{n_1 } z^{(r)}_{m_1 } z^{(r)}_{n_2 } z^{(l)}_{m_2}}}. \label{def - beta}\end{aligned}\ ] ] by making use of this definition , we calculate the two - loop @xmath17 function , @xmath219 \nonumber\\[3 mm ] + \frac{b_1 - 2a_0a_1}{4\pi^2 } \sum_{k_1,k_2 } & & \left [ g\left(\begin{array}{cc } n_1 & n_2 \\ k_1 & n_1-n_2+k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_1 & k_1 \\ k_2 & n_1-n_2+k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_2 & k_2 \\ m_1 & m_2 \end{array}\right)\right.\nonumber\\ & & + \left . g\left(\begin{array}{cc } n_1 & k_1 \\ m_1 & n_1+m_1-k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1+m_1-k_1 & k_2 \\ k_1 & n_1+m_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+m_1-k_2 & n_2 \\ k_2 & m_2 \end{array}\right)\right ] \nonumber \\[3 mm ] + \frac{c_0}{4\pi^2 } \sum_{k_1,k_2 } & & \left [ g\left(\begin{array}{cc } n_1 & k_2 \\ k_1 & n_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+k_1-k_2 & n_2 \\ m_1 & m_2+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } m_2+k_1-k_2 & k_1 \\ k_2 & m_2 \end{array}\right)\right . \nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2 \\ m_1 & m_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1 & n_2-m_1+k_2 \\ k_2 & m_2 \end{array}\right ) g\left(\begin{array}{cc } m_1+k_1-k_2 & n_2 \\ n_2-m_1+k_2 & k_1 \end{array}\right ) \right ] \nonumber\\[3 mm ] + \frac{d_1+a_0a_1}{4\pi^2 } \sum_{k_1,k_2 } & & \left [ g\left(\begin{array}{cc } n_1 & k_2 \\ k_1 & n_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1-n_2+k_1 & k_1 \\ m_1 & m_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+k_1-k_2 & n_2 \\ k_2 & n_1-n_2+k_1 \end{array}\right)\right . \nonumber\\ & & + g\left(\begin{array}{cc } k_1 & k_2 \\ m_1 & k_1-k_2+m_1 \end{array}\right ) g\left(\begin{array}{cc } n_1 & n_2 \\ n_2-n_1+k_1 & k_1 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2+m_1 & n_2-n_1+k_1 \\ k_2 & m_2 \end{array}\right ) \nonumber\\ & & + g\left(\begin{array}{cc } n_1 & n_1+k_1-k_2 \\ k_1 & k_2 \end{array}\right ) g\left(\begin{array}{cc } m_1+k_2-k_1 & n_2 \\ n_1+k_1-k_2 & m_2 \end{array}\right ) g\left(\begin{array}{cc } k_2 & k_1 \\ m_1 & m_1+k_2-k_1 \end{array}\right ) \nonumber\\ & & + \left . g\left(\begin{array}{cc } n_1+m_1-k_2 & n_2 \\ k_1+k_2-m_2 & k_1 \end{array}\right ) g\left(\begin{array}{cc } n_1 & k_2 \\ m_1 & n_1+m_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1 & k_1+k_2-m_2 \\ k_2 & m_2 \end{array}\right)\right]\nonumber\\[3 mm ] + \frac{2}{(4\pi)^2 } g\left(\begin{array}{cc } n_1 & n_2 \\ m_1 & m_2 \end{array}\right ) & & \sum_{k_1,k_2 } \left [ g\left(\begin{array}{cc } n_1 & k_2+n_1 \\ k_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & k_1 \\ k_2+n_1 & n_1 \end{array}\right ) + g\left(\begin{array}{cc } k_1 & k_2+n_2 \\ n_2 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & n_2 \\ k_2+n_2 & k_1 \end{array}\right)\right.\nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2+m_1 \\ m_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & m_1 \\ k_2+m_1 & k_1 \end{array}\right ) + g\left(\begin{array}{cc } m_2 & k_2+m_2 \\ k_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & k_1 \\ k_2+m_2 & m_2 \end{array}\right)\right ] , \label{2-loop}\end{aligned}\ ] ] where we already used the fact that @xmath220 which , by the way , is the necessary condition for the renormalizability of the model . note that the last term in eq . ( [ 2-loop ] ) appears due to the renormalization of the two - point function . after taking into account the values of constants in eqs . ( [ a_i ] ) ( [ d_i ] ) , we arrive at our final result for the @xmath17 function : @xmath221 \nonumber\\[3 mm ] -\frac{1}{8\pi^2 } \sum_{k_1,k_2 } & & \left [ g\left(\begin{array}{cc } n_1 & k_2 \\ k_1 & n_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1+k_1-k_2 & n_2 \\ m_1 & m_2+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } m_2+k_1-k_2 & k_1 \\ k_2 & m_2 \end{array}\right)\right . \nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2 \\ m_1 & m_1+k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } n_1 & n_2-m_1+k_2 \\ k_2 & m_2 \end{array}\right ) g\left(\begin{array}{cc } m_1+k_1-k_2 & n_2 \\ n_2-m_1+k_2 & k_1 \end{array}\right ) \right ] \nonumber\\[3 mm ] + \frac{2}{(4\pi)^2 } g\left(\begin{array}{cc } n_1 & n_2 \\ m_1 & m_2 \end{array}\right ) & & \sum_{k_1,k_2 } \left [ g\left(\begin{array}{cc } n_1 & k_2+n_1 \\ k_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & k_1 \\ k_2+n_1 & n_1 \end{array}\right ) + g\left(\begin{array}{cc } k_1 & k_2+n_2 \\ n_2 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & n_2 \\ k_2+n_2 & k_1 \end{array}\right)\right.\nonumber\\ & & + \left.g\left(\begin{array}{cc } k_1 & k_2+m_1 \\ m_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & m_1 \\ k_2+m_1 & k_1 \end{array}\right ) + g\left(\begin{array}{cc } m_2 & k_2+m_2 \\ k_1 & k_1-k_2 \end{array}\right ) g\left(\begin{array}{cc } k_1-k_2 & k_1 \\ k_2+m_2 & m_2 \end{array}\right)\right ] . \label{beta-2-loop } \end{aligned}\ ] ] while having less symmetry , the coupling in eq . ( [ lll - int ] ) still could contain contributions of highly symmetric solutions as in eqs . ( [ unun ] ) and ( [ uv ] ) . in order to extract such contributions , we introduce the projection operators to the corresponding subspaces in the space of couplings , @xmath222&=&\lim_{n\to\infty } i^{(lr ) } \frac{n\mbox{tr}^{(lr)}\left[\cdots\right ] -\mbox{tr}^{(v)}\left[\cdots\right ] } { n(n^2 - 1 ) } , \label{pr - lr } \\ p^{(v)}\left[\cdots\right]&=&\lim_{n\to\infty } i^{(v ) } \frac{n\mbox{tr}^{(v)}\left[\cdots\right ] -\mbox{tr}^{(lr)}\left[\cdots\right ] } { n(n^2 - 1 ) } , \label{pr - v } \end{aligned}\ ] ] where , by definition , @xmath223 & = & \sum_{n , m } g\left(\begin{array}{cc } n & m \\ m & n \end{array}\right ) , \\ i^{(v)}\left(\begin{array}{cc } n_1 & n_2 \\ m_1 & m_2 \end{array}\right ) & = & \delta_{n_1,n_2}\delta_{m_1,m_2},\\ \mbox{tr}^{(v)}\left[g\left(\begin{array}{cc } n_1 & n_2 \\ m_1 & m_2 \end{array}\right)\right ] & = & \sum_{n , m}g\left(\begin{array}{cc } n & n \\ m & m \end{array}\right).\end{aligned}\ ] ] by applying the projection operators in eqs . ( [ pr - lr ] ) and ( [ pr - v ] ) to the coupling in eq . ( [ lll - int ] ) , we easily extract the symmetric contributions , @xmath224&=&\frac{2g}{n+1 } i^{(lr ) } , \label{glll - lr}\\ p^{(v)}\left[g\left(\begin{array}{cc } n_1 & n_2 \\ m_1 & m_2 \end{array}\right)\right]&=&\frac{2g}{n+1 } i^{(v)}. \label{glll - v}\end{aligned}\ ] ] s. w. macdowell and o. tornkvist , , 3833 ( 1992 ) ; k. enqvist , int . j. mod d * 7 * , 331 ( 1998 ) ; r. fiore , a. tiesi , l. masperi and a. megevand , mod . a * 14 * , 407 ( 1999 ) ; v. skalozub and m. bordag , electroweak phase transition in a strong magnetic field , " hep - ph/9807510 ; k. kajantie _ et al . _ , nucl . phys . * b544 * , 357 ( 1999 ) ; t. inagaki , t. muta and s. d. odintsov , prog . . suppl . * 127 * , 93 ( 1997 ) . i. a. shushpanov and a. v. smilga , b * 402 * , 351 ( 1997 ) ; s. kanemura , h. -t . sato and h. tochimura , nucl . phys . * b517 * , 567 ( 1998 ) ; a. yu . babansky , e. v. gorbar and g. v. shchepanyuk , b * 419 * , 272 ( 1998 ) ; t. itoh and h. kato , , 30 ( 1998 ) . k. farakos and n. e. mavromatos , int . j. mod . b * 12 * , 809 ( 1998 ) ; n. e. mavromatos and a. momen , mod . a * 13 * , 1765 ( 1998 ) ; k. farakos , g. koutsoumbas , n. e. mavromatos and a. momen , hep - ph/9905272 . k. farakos , g. koutsoumbas , and n. e. mavromatos , b * 431 * , 147 ( 1998 ) ; int . j. mod . b * 12 * , 2475 ( 1998 ) ; w. v. liu , parity breaking and phase transition induced by a magnetic field in superconductors , " cond - mat/9808134 .

the acceleration and collimation of powerful bipolar outflows and jets in a variety of astronomical settings are often attributed to the action of magnetic fields ( see , e.g. , @xcite and @xcite for reviews ) . the commonly invoked scenario is that magnetic field lines threading a rotating compact object or its surrounding accretion disk can efficiently tap the rotational energy of the source and accelerate gas to supersonic speeds through centrifugal and/or magnetic pressure - gradient forces . it is argued that the hoop stresses of the twisted field lines can account for the narrowness of many jets and that , in many cases , alternative production mechanisms ( such as thermal driving ) can be excluded on observational grounds . although numerical simulations have provided useful insights into various aspects of hydromagnetic jet production , practical limitations have necessitated complementing this approach with analytic studies . owing to the complexity of the problem , the most general semianalytic solutions obtained so far have been time - independent and self - similar , patterned after the pioneering disk - outflow solutions of @xcite . the advantage of pursuing such solutions is that they are exact and self - consistent , and that they can be systematically classified ( e.g. , * ? ? ? furthermore , these solutions are evidently rich enough to capture most of the relevant physics , as corroborated by numerical calculations ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? almost all of the previous semianalytic work on jet magnetohydrodynamics ( mhd ) was done in the newtonian limit of nonrelativistic bulk and random speeds . relativistic outflows are , however , observed quite commonly in nature with examples including active galactic nuclei ( agns ) , galactic superluminal sources ( often referred to as `` microquasars '' ) , pulsars , and gamma - ray bursts ( grbs) and in many of these cases magnetic fields are again implicated as the main driving and collimation mechanism ( e.g. , * ? ? ? this provides a motivation for generalizing the newtonian self - similar outflow solutions , although it is readily seen that this can not be done in a totally straightforward manner . for one thing , special - relativistic mhd ( unlike the nonrelativistic theory ) involves a characteristic speed ( the speed of light @xmath1 ) , which precludes the incorporation of gravity into the self - similar equations and a simple matching of the outflow solution to a particular ( e.g. , keplerian ) disk rotation law . furthermore , again in contrast with the nonrelativistic formulation , the displacement current and the charge density can not be neglected in the constitutive equations ( which now must also satisfy relativistic covariance ) . despite the aforementioned complications , @xcite and @xcite succeeded in generalizing the `` cold , '' radially self - similar solutions of @xcite to the relativistic regime . their solutions are characterized by the thermal pressure playing a negligible role in the flow acceleration and by the flow being trans - alfvnic : the poloidal speed is less than the poloidal component of the alfvn speed at the base of the flow , and comes to exceed it further up . our aim in this paper is to further generalize these solutions to the `` hot '' relativistic case i.e. , we allow both the bulk _ and _ the random speeds to be relativistic . this is motivated primarily by the desire to apply these solutions to grb outflows , in which thermal driving by an optically thick , hot `` fireball '' composed primarily of radiation and electron - positron pairs could play a role . in fact , most previous models of grb outflows were purely hydrodynamical and included only thermal driving ( see , e.g. , @xcite for a review ) . it has subsequently been realized that energy deposition by annihilating neutrinos , which had been one of the main proposed heating mechanisms , would typically be inefficient ( e.g. , * ? ? ? * ) , and it was , furthermore , argued @xcite that only a small fraction ( @xmath2 ) of the energy deposited in the source could be in thermal form to avoid generating strong photospheric emission in the outflow ( for which there has been no observational evidence ) . for these ( and other ) reasons , it is now believed that magnetic fields play the dominant role in the driving of grb jets ( e.g. , @xcite ) . nevertheless , some thermal energy injection , either by neutrinos or by magnetic energy dissipation , is likely to take place ( e.g. , * ? ? ? * ; * ? ? ? * ) , and , as we show in [ soln_a ] , even if it contributes only a small fraction of the initial energy flux in the flow , it can dominate the early phase of the acceleration . there are indications that thermal energy deposition at the source may also contribute to the initial acceleration of relativistic jets in agns ( e.g. , @xcite ; n. vlahakis & a. knigl , in preparation ) . it thus appears that , to fully understand the nature of such jets , it is necessary to model them in terms of a `` hot '' mhd outflow . the ability of large - scale , ordered magnetic fields to guide , collimate , and accelerate relativistic outflows has been previously discussed by various authors ( including , in the grb context , @xcite , @xcite ; @xcite , @xcite , @xcite , @xcite , and @xcite ) . these discussions , however , did not include exact solutions from which detailed quantitative estimates could be made . in a previous paper , vlahakis & knigl ( 2001 , hereafter vk01 ) presented a semianalytic self - similar solution of the `` hot '' relativistic mhd equations and applied it to the interpretation of grbs . the full formalism underlying this solution is described in the present paper , where we also analyze its dependence on the relevant physical parameters and compare it with other characteristic solutions of trans - alfvnic flows . this discussion is extended in the companion paper ( vlahakis & knigl 2003 , hereafter paper ii ) , where we focus on super - alfvnic flows . the self - similar solutions that we derive correspond to ordered magnetic field configurations in the ideal - mhd limit . although it is quite conceivable that the fields that drive the flow from a differentially rotating star or disk are at least in part small - scale and disordered , the statistical ( temporal and spatial ) averages of such fields could in principle have a similar effect to that of large - scale , ordered field configurations in providing both acceleration @xcite and collimation @xcite . the applicability of ideal mhd to the acceleration of ultrarelativistic flows has been questioned by @xcite ( @xcite ; see also @xcite ) , who proposed instead a force - free electromagnetic formulation . we note in this connection that a force - free behavior can be recovered from the relativistic mhd formulation as a limiting case of negligible particle inertia . furthermore , electromagnetic energy dissipation perhaps the strongest argument against the ideal - mhd modeling framework has been claimed to lead , on its own , to an efficient conversion of poynting flux into a highly relativistic bulk motion ( e.g. , * ? ? ? our ideal - mhd solutions may thus be regarded as a first step toward a more comprehensive theory in which dissipation effects will be taken into account ( and possibly further enhance the effectiveness of the magnetic acceleration process ) . although the formulation presented in this paper is quite general , the main application that we consider is to grb sources . for definiteness , we adopt the `` internal shock '' scenario for the origin of the prompt high - energy emission in grbs ( e.g. , * ? ? ? considerations involving variability time scales ( as interpreted in the context of this picture ) as well as source energetics support the identification of an accretion disk around a newly formed black hole as the source of the grb outflow ( e.g. , * ? ? ? * ) ; accordingly , we concentrate on modeling jets from accretion disks . however , our solutions should be at least qualitatively applicable also to other source configurations , such as a rapidly spinning neutron star ( e.g. , * ? ? ? * ; * ? ? ? * ) or a rotating , magnetically threaded black hole ( e.g. , * ? ? ? * ; * ? ? ? since grb outflows have a limited duration ( the value of which is plausibly related to the disk accretion time ) , a naive application of a steady - state similarity solution is not warranted . in previous , purely hydrodynamical models of grb outflows , this difficulty was circumvented by applying the so - called `` frozen pulse '' approximation @xcite . in this paper we prove ( [ governing ] ) that this approximation can be generalized to the relativistic mhd regime , but we also demonstrate ( [ timerecovery ] ) how any inherent time dependence can be recovered . the plan of the paper is as follows . in [ mhd ] we present the equations of time - dependent relativistic mhd , simplify them using the frozen - pulse approximation , and discuss what effect each of the various forces acting on the plasma has on the flow acceleration and collimation . in [ r - ss_section ] we describe the @xmath3 self - similar model and in [ results ] we give the results of the numerical integration of the model equations . we discuss general implications of this work to grb sources and other relativistic jet sources in [ conclusions ] , where we also summarize our conclusions . the stress - energy tensor of relativistic mhd consists of three parts matter ( subscript m ) , radiation ( subscript r ) and electromagnetic fields ( subscript em ) : @xmath4 ( @xmath5 ) . the matter component is given by @xmath6 , where @xmath7 is the total comoving matter energy density , @xmath8 is the particle pressure , @xmath9 is the fluid four - velocity , and @xmath10 is the metric tensor ( assuming a flat spacetime and cartesian space coordinates @xmath11 ) . here @xmath12 is the baryon rest - mass density , @xmath13 is the internal energy density , with @xmath14 denoting the polytropic index ( @xmath15 or @xmath16 in the limit of an ultrarelativistic or a nonrelativistic temperature , respectively ) , @xmath17 is the three - velocity measured in the frame of the central object , and @xmath18 is the lorentz factor . the radiation component is given by @xmath19 , where @xmath20 and @xmath21 are , respectively , the radiation energy density and pressure in the fluid rest frame . radiation forces are typically most important in regions that are sufficiently optically thick that one can take the local radiation field to be nearly isotropic and set @xmath22 . we are most interested in the ultrarelativistic case @xmath23 , in which the matter and radiation can be treated ( under optically thick conditions ) as a single fluid . thus , we henceforth write @xmath24 , @xmath25 , and @xmath26 . introducing the specific ( per baryon mass ) relativistic enthalpy @xmath27 , where @xmath28 and including the contribution of the electric ( @xmath29 ) and magnetic ( @xmath30 ) fields ( measured in the central - object frame ) , the components of the total stress - energy tensor take the form ( @xmath31 ) @xmath32 with @xmath33 , @xmath34 , and @xmath35 representing the energy density , energy flux , and spatial stress contributions , respectively . the electromagnetic field obeys maxwell s equations @xmath36 where @xmath37 is the four - current ( with @xmath38 representing the charge density ) . under the assumption of ideal mhd , the comoving electric field is zero , which implies @xmath39 the baryon mass conservation equation is @xmath40 , or @xmath41 in the absence of a gravitational field or any other external force , the equations of motion are @xmath42 . the momentum conservation equation is given by the @xmath43 components , @xmath44 the entropy conservation equation ( the first law of thermodynamics ) is obtained by setting @xmath45 , @xmath46 which can be rewritten ( using @xmath47 $ ] ) as @xmath48 one can carry out a partial integration of equations ( [ maxwell])([energy ] ) under the assumptions of axisymmetry [ in cylindrical coordinates ( @xmath49 ) , @xmath50 and of a zero azimuthal electric field ( @xmath51 ) if the flow is time - independent ( e.g. , * ? ? ? * ; * ? ? ? we now show that the equations describing a highly relativistic mhd `` pulse '' that could be identified with a grb event may , in fact , be cast in a steady - state form . we start by noting that , with the above assumptions @xmath52 , the poloidal component of faraday s law implies that the poloidal magnetic field is time independent . if the field is anchored in an accretion disk , then the poloidal field near the disk surface would remain quasi steady at least on the timescale of the local radial inflow ( @xmath53 ) . in the case of a grb outflow associated with the emptying up ( by accretion onto a black hole ) of a disk of finite size , the poloidal field may be expected to change significantly only on the timescale of the burst duration . at the end of this time interval , the information that the field has changed starts to propagate with at most the speed of light ( the actual speed of propagation is the fast magnetosonic speed , which is generally @xmath54 in a material medium ) . as we are concerned with highly relativistic outflows , it is reasonable to expect that the poloidal field associated with the outflowing fluid elements that produce the burst will exhibit negligible explicit time dependence ( @xmath55 ) over the duration of the burst . ( note , however , that the azimuthal component of the magnetic field , which is related to the poynting flux , will be time dependent . ) the solenoidal condition on the magnetic field , @xmath56 , implies that there is a poloidal magnetic flux function @xmath57 , defined by @xmath58 , which satisfies @xmath59 where the subscripts @xmath60 and @xmath61 denote the poloidal and azimuthal components , respectively . furthermore , equation ( [ ohm ] ) together with the condition @xmath62 implies @xmath63 , from which it follows that there are functions @xmath64 and @xmath65 ( whose coordinate dependence we discuss below ) such that @xmath66 denoting the arclength along a poloidal fieldline by @xmath67 , we change variables from ( @xmath68 ) to ( @xmath69 ) , with @xmath70 . for any function @xmath71 , we can define the operator @xmath72 we now rewrite the mhd equations using @xmath73 [ mhd - eqs - s ] equation ( [ solenoidal ] ) becomes @xmath74 whereas equation ( [ ohm ] ) gives @xmath75 faraday s law implies @xmath76 and the continuity equation yields @xmath77 where , using equation ( [ v - b ] ) , @xmath78 . turning now to the momentum conservation equation , we employ equation ( [ operators ] ) and the fact that @xmath79 is time independent @xmath80 to write the current density in the form @xmath81 \ .\ ] ] decomposing the vectors using the local cartesian basis @xmath82 we get @xmath83 - e \hat{n } \\ \nonumber & = & -\left(e+ b_{\phi } \right ) \hat{n } + b_{\phi } \left(\hat{n } \cdot { { \mathbf{\nabla}}\!\!_s}\ell \right ) \hat{b } \,,\end{aligned}\ ] ] and hence @xmath84\ .\ ] ] the charge density , in turn , can be written as @xmath85\ .\ ] ] substituting these expressions into equation ( [ momentum ] ) , we obtain @xmath86 { { \mathbf{\nabla}}\!\!_s}a + \frac{\left({{\mathbf{\nabla}}\!\!_s}\times { \boldsymbol{b } } \right ) \times { \boldsymbol{b } } } { 4 \pi } \ .\end{aligned}\ ] ] finally , the entropy conservation equation transforms into @xmath87 for a highly relativistic poloidal motion , when @xmath88 , one can simplify the equations by noting the following : @xmath89 1 ) due to lorentz contraction , the observed width of the outflow is @xmath90 times smaller than its comoving width , @xmath91 . as @xmath92 and @xmath93 , one gets @xmath94 . thus , all terms in equations ( [ mhd - eqs - s ] ) containing @xmath95 are negligible . @xmath89 2 ) the term @xmath96 in equation ( [ continuity - s ] ) is of the order of @xmath97 and is thus negligible in comparison with the first term on the left - hand side of this equation . the arguments above were originally given in the context of a purely hydrodynamic ( hd ) flow by @xcite . @xmath89 3 ) using @xmath98 , @xmath99 , and equations ( [ v - b ] ) and ( [ e_a ] ) , we infer @xmath100 , which remains @xmath101 throughout the flow in view of the scaling @xmath102 ( see [ validity ] ) . thus , all terms in equation ( [ momentum - s ] ) that contain @xmath103 ( i.e. , the first two terms on the right - hand side ) are a factor @xmath104 smaller than the last term on the right - hand side and can be neglected . the same is true in equation ( [ omega_a ] ) , which implies @xmath105 \sim \gamma \partial ( \omega / \gamma^2 ) \ll \omega$ ] . @xmath89 4 ) the pressure - gradient force in the momentum conservation equation can also be neglected . it is much smaller than the lorentz force in the transfield direction , consistent with the fact that the field is everywhere close to being force - free in that direction ( especially so in the region near the origin , where thermal effects are most important ) . along the field , there is a force that is @xmath106 times larger ( namely , the inertial force component associated with the @xmath107 term ) , as in the purely hd case examined in @xcite . in general , the pressure force is important up to the slow magnetosonic point , where , for highly relativistic temperatures , @xmath108 , and its contribution is negligible in the highly relativistic regime . one can therefore replace the term @xmath109 by @xmath110 or even completely neglect the pressure - gradient force in the momentum equation without introducing a significant error . with the above approximations , all the @xmath111 terms in the continuity , momentum , and entropy equations can be eliminated , and the conservation equations simplify to a steady - state form . although the label @xmath112 remains attached to the @xmath113 operator , it now serves only to identify a given outflowing shell ( or pulse ) . the motion remains effectively time independent and can be described as a frozen pulse whose internal profile is specified through the variable @xmath112 . as we noted in [ introduction ] , the frozen - pulse approximation was first applied to relativistic hd outflows in grb sources by @xcite ( @xcite ; see also @xcite ) . we have now shown that this approximation can be extended to relativistic mhd flows . in the remainder of this paper we pursue this effectively steady - state formulation , but we return in [ timerecovery ] to consider time - dependent effects in grb outflows . the full set of effectively steady - steady equations can be partially integrated to yield several field - line constants : [ integrals - s ] \a ) equations ( [ v - b ] ) and ( [ omega_a ] ) yield the field angular velocity , which equals the matter angular velocity at the footpoint of the fieldline at the midplane of the disk , @xmath114 \b ) the continuity equation ( [ continuity - s ] ) and equation ( [ v - b ] ) imply that the mass - to - magnetic flux ratio has the form @xmath115 \c ) the @xmath116 component of the momentum equation ( [ momentum - s ] ) yields the total ( kinetic + magnetic ) specific angular momentum , @xmath117 \d ) dotting @xmath118 into the momentum equation ( [ momentum - s ] ) gives the total energy - to - mass flux ratio @xmath119 , where @xmath120 \e ) the entropy equation ( [ energy - s ] ) gives the adiabat @xmath121 equation ( [ polytropic ] ) is the usual polytropic relation between density and pressure , but in the current application the polytropic index is only allowed to take the values @xmath122 ( if the temperatures are relativistic , in which case matter and radiation are treated as a single fluid ) and @xmath16 ( if the gas is `` cold , '' in which case radiation forces can be neglected ) . other than 4/3 or 5/3 would imply a nonadiabatic evolution and hence require the incorporation of heating / cooling terms into the entropy and momentum equations . ] two integrals remain to be performed , involving the bernoulli and transfield equations . there are correspondingly two unknown functions , which we choose to be the cylindrical radius of the fieldline in units of the `` light cylinder '' radius , @xmath123 and the `` alfvnic '' mach number ( see * ? ? ? * ) @xmath124 we define the alfvn lever arm by @xmath125 [ and correspondingly @xmath126 and use it to scale the cylindrical radius of the fieldline by introducing @xmath127 to obtain nondimensional variables , we adopt a reference length @xmath128 and a reference magnetic field @xmath129 and define @xmath130 the expressions for the physical quantities in terms of the defined variables and the explicit expressions for the bernoulli and transfield equations are given in appendix [ rmhdeqs ] . except for the @xmath112 label , which serves to identify a given shell ( or pulse ) , these equations are precisely those of steady - state , relativistic mhd . solving these equations requires the specification of seven constraints , of which four are associated with boundary conditions at the source and three are determined by the regularity requirement at the singular points ( the modified - slow , alfvn , and modified - fast points ; see @xcite ) . the momentum equation ( [ momentum ] ) can be written as the sum of the following force densities ( for simplicity we use hereafter the term force ) @xmath131 where @xmath132 the `` gamma '' force @xmath133 further decomposes into two terms : @xmath134 , with @xmath135 the poloidal part of the @xmath136 force is @xmath137 where @xmath138 is the angle between the poloidal magnetic field and the disk ( @xmath139 ) , and the derivative @xmath140 \partial / \partial \theta ) $ ] is taken keeping @xmath141 ( and @xmath112 ) constant . the radius of curvature of a poloidal fieldline is @xmath142 . the projection of equation ( [ momentum - f ] ) along the poloidal flow is @xmath143 the terms on the right - hand side of equation ( [ momentum - v_p ] ) are recognized as @xmath144 , @xmath145 , @xmath146 , @xmath147 , and @xmath148 respectively , where a subscript @xmath149 denotes the component of a vector along the poloidal fieldline . the first term on the left - hand side of equation ( [ momentum - v_p ] ) is @xmath150 , whereas the second term is @xmath151 ( note that @xmath152 ) . the magnetic force component @xmath153 decomposes into the azimuthal magnetic pressure gradient @xmath154 and the magnetic tension @xmath155 . these two parts cancel each other when @xmath156 ; if @xmath157 decreases faster than @xmath158 then the gradient of the azimuthal magnetic pressure exceeds the magnetic tension , resulting in a positive @xmath153 . in the nonrelativistic regime ( @xmath159 , @xmath160 ) , the pressure force @xmath147 dominates up to the slow magnetosonic point , but the bulk of the acceleration is either magnetocentrifugal corresponding to the @xmath146 term , which can be interpreted in the `` bead on a wire '' picture ( e.g. , * ? ? ? * ) and @xmath161 ( with @xmath162 ) near the base of the flow , implying that the azimuthal field satisfies @xmath163 and hence that the @xmath153 term is negligible and that a near - corotation ( @xmath164 ) holds . the small value of @xmath165 in turn implies a large density and hence a measurable thermal pressure , resulting in a nonnegligible pressure force at the base . ] or a consequence of the magnetic pressure - gradient force @xmath153 ( which near the surface of the disk can be interpreted in the `` uncoiling spring '' picture ; e.g. , @xcite ) . in the northern hemisphere ( and parallel to @xmath166 in the southern hemisphere ) , and a corresponding outward - directed magnetic pressure gradient @xmath167 . ] the magnetic force generally becomes important also in flows where the centrifugal acceleration initially dominates : in this case the inertia of the centrifugally accelerated gas amplifies the @xmath168 component , and eventually ( beyond the alfvn point ) @xmath153 becomes the main driving force . this force continues to accelerate the flow beyond the classical fast - magnetosonic point ( which separates the elliptic and hyperbolic regimes of the mhd partial differential equations ) in the central object s frame can exist ( i.e. , eq . [ [ waves_fast ] ] with @xmath169 is satisfied ) . the classical fast - magnetosonic point is equivalently defined by the condition that the poloidal proper speed equals the comoving proper phase speed of a fast - magnetosonic wave whose wavevector is parallel to @xmath170 ( see eq . [ [ theorem ] ] ) . ] and all the way to the modified fast - magnetosonic singular point ( see * ? ? ? the modified ( and _ not _ the classical ) fast - magnetosonic surface has the property that it is a singular surface for the steady mhd equations when one solves simultaneously the bernoulli and transfield equations ( e.g. , * ? ? ? * ) . only when the magnetic field geometry is given ( and one solves only the bernoulli equation but not the transfield one ) does the singular surface correspond to the classical fast - magnetosonic point . the modified fast - magnetosonic surface coincides with the limiting characteristic , the `` event horizon '' for the propagation of fast - magnetosonic waves , since only beyond this surface all the points inside a fast mach cone have larger fast mach numbers than at the origin of the cone . at smaller distances , for a part of a given cone , the converse is true : the opening angle of the fast mach cone becomes progressively larger as one advances inside that part of the cone ; consequently , a small disturbance in the super - fast regime can affect the entire flow . the situation is similar to that of light propagating close to a kerr black hole , where the ergosphere ( which corresponds to the classical fast - magnetosonic surface ) marks the boundary between the elliptic and hyperbolic regimes , and where the singular event horizon ( which corresponds to the modified fast - magnetosonic surface ) is located within the hyperbolic regime : only for a spherically symmetric ( schwarzschild ) black hole are the ergosphere and event horizon equivalent in direct analogy with spherically symmetric flows , in which the classical and modified fast - magnetosonic points coincide ( see * ? ? ? * ) . in the case where the outflow attains a highly relativistic speed , the centrifugal acceleration can not play an important role . this is because the nonnegligible @xmath166 that would be required in this case would constrain the maximum value of the poloidal speed : @xmath171 . therefore , in equation ( [ momentum - v_p ] ) , @xmath172 ( and also @xmath173 ) . the @xmath147 force can be neglected since @xmath174 . the two remaining terms are @xmath145 ( a force with a relativistic origin ) and @xmath175 . the expressions for these terms in equation ( [ momentum - v_p ] ) ( or , equivalently , eq . [ [ int - mu ] ] for the total energy - to - mass flux ratio ) indicate that the bulk lorentz factor can increase in response to the decline in either the enthalpy - to - rest - mass ratio @xmath176 ( the thermal acceleration case ) or the poynting - to - mass flux ratio @xmath177 ( the magnetic acceleration case ) along the flow . when the temperature is relativistic , the initial acceleration is dominated by the temperature force , but after @xmath176 drops to @xmath178 the magnetic force takes over : this is the likely situation in grb outflows ( see vk01 and [ soln_a ] ) . when the outflow speed is only mildly relativistic , the magnetocentrifugal force may be important during the initial acceleration phase , especially if the temperatures are nonrelativistic ; this is the situation envisioned in the `` cold '' relativistic - mhd solutions of @xcite . it is , however , also conceivable that the magnetic pressure - gradient force dominates from the start , as might be the case if the azimuthal field component at the disk surface is large enough ; this is the picture outlined in the presentation of the relativistic solutions derived by @xcite ( @xcite ; see also paper ii ) . we can obtain an expression for @xmath179 as follows . by eliminating @xmath180 between equations ( [ int - l ] ) and ( [ int - mu ] ) , we obtain a relation between @xmath181 and @xmath182 : @xmath183 whose divergence along the flow implies @xmath184 employing the relations for @xmath153 , @xmath146 ( see eq . [ [ momentum - v_p ] ] ) , and equations ( [ a1])([rho_p ] ) , we obtain @xmath185 the first term on the right - hand side of equation ( [ f_c - f_b ] ) can give rise to either acceleration ( when @xmath166 decreases along the flowline ) or deceleration ( when @xmath166 increases , as in the corotation regime at the base of the outflow ) . this term , together with @xmath144 , can lead to a situation in which @xmath186 increases ( resp . , decreases ) and @xmath166 decreases ( resp . , increases ) while the lorentz factor remains roughly constant . the second term on the right - hand side of equation ( [ f_c - f_b ] ) , which is proportional to @xmath175 , demonstrates that the centrifugal force also has a magnetic component and accounts for the poynting - to - kinetic energy transfer that underlies the magnetocentrifugal acceleration process ( see also @xcite for a related discussion ) . the form of this term makes it clear why the centrifugal force exceeds the magnetic force during the initial stage of the acceleration , when the flow is still nonrelativistic ( @xmath187 , @xmath188 , with @xmath189 , @xmath190 ) . the conclusion from the above analysis is that , even though centrifugal and thermal effects could dominate initially , the magnetic force eventually takes over and is responsible for the bulk of the acceleration to high terminal speeds . @xcite described the efficient conversion of poynting - to - kinetic energy fluxes in relativistic mhd outflows in terms of a `` magnetic nozzle '' ( see also @xcite ) . the preceding discussion makes it clear that , in essence , this mechanism represents the ability of a collimated hydromagnetic outflow to continue to undergo magnetic acceleration all the way up to the modified fast - magnetosonic point ( which could lie well beyond the classical fast point ) . when interpreted in these terms , it is evident that this effect is not inherently relativistic this conclusion has , in fact , been verified explicitly in the case of the nonrelativistic self - similar solutions constructed by @xcite . the projection of equation ( [ momentum - f ] ) in the direction perpendicular to the poloidal flow is given by equation ( [ forces - perp ] ) . the two largest terms are the magnetic and electric force components , which almost cancel each other . the magnetic force in the meridional plane has two parts : @xmath191 and @xmath192 . the first term ( which usually dominates ) has components in both the flow ( along @xmath193 ) and the transfield ( along @xmath194 ) directions . the @xmath195 component @xmath196 contributes to the acceleration when @xmath197 , where the subscript @xmath198 denotes the vector component along @xmath199 . ( if a thermally dominated acceleration regime exists near the base of the flow , it is in principle possible to have @xmath200 there , corresponding to an enthalpy - to - poynting energy transfer mediated by the magnetic force ; see paper ii . ) the @xmath199 component @xmath201 acts to collimate or decollimate the flow depending on whether the outflow is , respectively , in the current - carrying ( @xmath202 ) or the return - current ( @xmath203 ) regime ( see fig . [ fig1 ] ) . the second term in the decomposition of the meridional magnetic force is related to the curvature radius @xmath204 , @xmath205 , and is directed along @xmath206 for @xmath207 . the electric field always points in the @xmath199 direction , but the electric force @xmath208 could lie along either @xmath209 or @xmath210 , depending on the sign of the charge density @xmath38 . by employing the curvature radius @xmath204 , one can write @xmath211 . when @xmath212 ( a collimating flow ) , the effect of the curvature term in the expression for the electric force is to oppose collimation , but it is possible for the other term in this expression to dominate , leading to @xmath213 . for highly relativistic motion ( @xmath214 ) , @xmath215 , and the electric force has the sign of @xmath216 . in this case @xmath217 acts to decollimate the flow in the current - carrying regime and to collimate it in the return - current regime ( see fig . [ fig1 ] ) . for comparison , note that , when the motion is nonrelativistic ( @xmath187 ) , @xmath217 is negligible relative to @xmath218 . in this case , a flow in the current - carrying regime is easily collimated ( with @xmath219 balancing the inertial force @xmath220 , resulting in a nonnegligible value of @xmath204 ) . in the return - current regime , collimation ( resp . , decollimation ) is produced if @xmath221 ( resp . , @xmath222 ) ; see @xcite . to obtain semianalytic solutions of the highly nonlinear system of equations ( [ bernoulli ] ) and ( [ transfield ] ) , it is necessary to make additional assumptions : in particular , we look for a way to effect a separation of variables . the most complicated expression is the one for @xmath223 ( eq . [ [ a1 ] ] ) . in view of the importance of the azimuthal field component , which plays a crucial and varied role as part of the magnetic pressure - gradient , magnetic tension , and centrifugal acceleration terms in the momentum equation , the only realistic possibility of deriving exact analytic solutions is to assume that the @xmath224 , @xmath225 , and @xmath226 surfaces coincide , i.e. , that @xmath227 @xcite . we aim to find appropriate forms for the functions of @xmath141 such that the expressions ( [ bernoulli ] ) and ( [ transfield ] ) become ordinary differential equations ( odes ) . from an inspection of the bernoulli equation ( [ bernoulli ] ) we conclude that , in order to separate the variables @xmath228 and @xmath141 and get a _ single _ equation that only has a @xmath228 dependence , it is necessary to assume that the @xmath229 term is a product of a function of @xmath141 times a function of @xmath228 . as @xmath141 is a function of @xmath230 ( see eq . [ [ g ] ] ) , there must exist functions @xmath231 such that @xmath232 ^ 2 = { \cal h}_1 \left(\frac{\varpi } { g}\right ) \ { \cal h}_2 ( g)\,.\ ] ] there always exist the trivial possibilities @xmath233 in spherical coordinates @xmath234 [ @xmath235 when the field is radial ] , and @xmath236 ( @xmath237 for a cylindrical field ) , which are not of interest here . after some algebra one can prove that the only nontrivial case is to have @xmath238 , i.e. , @xmath239 . it thus appears that , to obtain an analytic adiabatic solution , it is necessary to assume @xmath3 self - similarity . the remaining assumptions for constructing an @xmath3 self - similar solution are that the cylindrical distance ( in units of the alfvnic lever arm ) , the poloidal alfvnic mach number , and the relativistic specific enthalpy are also functions of @xmath240 only : @xmath241 , @xmath242 , @xmath243 ( with the result for @xmath176 following from the nonlinearity of the expression for @xmath165 ; see eq . [ [ x - m ] ] ) . following the algorithm described in @xcite , we change variables from ( @xmath244 ) to ( @xmath245 ) ( see eq . [ [ alpha ] ] ) and obtain the forms of the integrals under the assumption of separability in @xmath246 and @xmath240 . the results are given in equation ( [ integrals ] ) . , and ignoring gravity , so it is possible to assume a polytropic equation of state ) . ] among the five unknown functions of @xmath240 [ @xmath247 , @xmath248 , @xmath249 , @xmath250 , and @xmath251 dependence ; see [ timerecovery ] . ] there are three algebraic equations ( eqs . [ [ x - rss ] ] , [ [ m - rss ] ] , and [ [ bernoulli - rss ] ] ) and two first - order odes ( eqs . [ [ g - rss ] ] and [ [ transfield - rss ] ] ) . after solving for these functions , the physical quantities can be recovered using [ quantities - rss ] @xmath252 @xmath253 @xmath254 @xmath255 @xmath256 where @xmath257 is the unit vector along the poloidal fieldline , @xmath258 ( already defined in [ governing ] ) is the unit vector in the transfield direction in the meridional plane ( toward the axis of rotation ) , and @xmath259 is the opening half - angle of the outflow ( the angle between the poloidal fieldline and the axis of rotation ) . the @xmath3 dependence of all the physical quantities can be inferred from the expressions ( [ quantities - rss ] ) on the basis of the known @xmath3 dependence of @xmath246 ( @xmath260 ; see eq . [ [ alpha ] ] ) . this is a general characteristic of @xmath3 self - similar models . the parameters of the model are @xmath261 and @xmath14 ( @xmath15 in this study ) , whereas the `` constants '' @xmath262 , @xmath263 , @xmath264 , @xmath265 , and @xmath266 together with two `` initial '' conditions ( corresponding to the two first - order odes ) are related to seven boundary conditions , as we prove in appendix [ appendixrss ] . the @xmath3 self - similar character of the poloidal field - line shape is shown in figure [ fig2 ] . the model described above is the generalization to a `` hot '' ( @xmath267 ) gas of the only known exact semianalytic solution of the relativistic mhd equations , the `` cold '' @xmath3 self - similar wind solution found independently by @xcite and @xcite . ( the force - free model of @xcite can be regarded as a special case of the latter solutions , corresponding to @xmath268 and @xmath269 . ) self - similarity was first employed by @xcite , who examined purely hd flows , but it has become well - known only after @xcite used it to construct a nonrelativistic mhd disk - wind model . the latter work has subsequently been generalized by many authors ( see @xcite and references therein ) . ] it is obvious from equations ( [ quantities - rss ] ) that the alfvn point , where @xmath270 , is singular . at this point @xmath271 in fact , as @xmath63 , this relation holds not only for the @xmath240-components of ( @xmath272 ) , but for components in any direction in the meridional plane . thus , on the alfvn surface , static alfvn waves with wavevector in any direction in the meridional plane ( in the central object s frame ) can exist ( i.e. , eq . [ [ waves_alfven ] ] with @xmath273 and @xmath274 is satisfied ) . ] ) . ] in order for the solution to pass through the alfvn singular point , the alfvn regularity condition ( [ alfv2 ] ) must be satisfied . the latter determines the slope of the `` alfvnic '' mach number at the alfvn point , @xmath275 , which is related to the alfvnic value of the magnetization function @xmath276 ( see appendix [ appendixrss ] ) . it is seen from equation ( [ alfven - dispersion ] ) that the alfvn point is always located inside the light surface @xmath277 . note , however , that if @xmath278 ( corresponding to the force - free limit , @xmath279 ) , the alfvn and light surfaces almost coincide . it is straightforward to obtain an expression for @xmath280 as a function of @xmath281 using the derivative of the bernoulli equation ( [ bernoulli - rss ] ) . after substituting in the transfield equation ( [ transfield - rss ] ) , the latter takes the form @xmath282 , where the denominator can be written as @xmath283 with @xmath284 being the magnetic field amplitude and with the square of the proper sound speed given by @xmath285 singular points appear wherever @xmath286 ; these are the modified slow and fast - magnetosonic singular points . they correspond to points where static slow / fast - magnetosonic waves with wavevectors along @xmath287 in the central object s frame can exist ( i.e. , eq . [ [ waves_fast ] ] with @xmath288 is satisfied ) . is equal to the comoving proper phase speed of a slow / fast - magnetosonic wave propagating along @xmath287 ( see eq . [ [ theorem ] ] ) . ] the modified singular surfaces , which correspond to the `` limiting characteristics '' of the self - similar flow , have previously been considered in connection with the nonrelativistic solutions ( e.g. , @xcite ) . in order for the solution to pass through a singular point where @xmath286 , @xmath289 must simultaneously hold ( yielding the respective regularity condition ) . when solving the steady , axisymmetric , ideal mhd equations under the assumption that the azimuthal electric field ( as measured in the central object s frame ) vanishes ( @xmath51 ) , seven integrations are required , corresponding to seven unknowns : the gas density and pressure , the three components of the velocity , and two functions related to the magnetic field ( e.g. , @xmath141 and @xmath223 ) . correspondingly , seven boundary conditions determine a unique solution . five of them are the integrals @xmath64 , @xmath65 , @xmath290 , @xmath291 , and @xmath292 , which , as discussed in [ governing ] , are conserved quantities along the meridional fieldlines . the other two correspond to `` initial conditions '' on the functions @xmath293 and @xmath165 , which are obtained by integrating equations ( [ bernoulli ] ) and ( [ transfield ] ) . in a physically viable solution , the flow starts with a sub slow - magnetosonic velocity and must satisfy the causality principle : any disturbance in the asymptotic regime can not influence the solution near the origin through magnetosonic or alfvn waves . since the flow starts with a small velocity , it must cross three singular surfaces : the modified slow - magnetosonic , the alfvn , and the modified fast - magnetosonic . , the latter surface represents the `` event horizon '' for the propagation of the fastest waves . the alfvn surface plays a similar role for the alfvn waves , but the slow - magnetosonic singular surface is not the `` event horizon '' for the propagation of slow - magnetosonic waves ; it is just the limiting characteristic in the sub - slow hyperbolic regime ( see @xcite ) . ] the related three regularity conditions are effectively three boundary conditions that must be satisfied in order for the solution to pass smoothly through the singular points . implementing this procedure is a highly nonlinear task , since the positions of the singular points are not known a priori and must be obtained simultaneously with the solution . all in all , only four boundary conditions remain free and can be specified ( e.g. , on a surface near the origin of the flow ) . in the @xmath3 self - similar model that we examine , in which we end up integrating odes in the variable @xmath240 , it is convenient to choose as the initial surface a cone @xmath294 ( where here and in what follows , a subscript @xmath295 denotes an initial value ) . one can start the integration by specifying seven initial conditions ( i.e. , seven functions of @xmath3 ) on this cone ( where @xmath3 is the arclength along the conical surface ; it should not to be confused with the distance from the central object , which is given by @xmath296^{1/2}$ ] see fig . [ fig2 ] ) . for example , one can specify @xmath297 and @xmath298 such that @xmath299 , @xmath300 , @xmath301 , @xmath302 , @xmath303 , @xmath304 , and @xmath305 ( with the @xmath306 being in general functions of @xmath112 ) . note that , in the framework of this self - similar model , the specified functions of @xmath3 must be consistent with equations ( [ quantities - rss ] ) ; if they are not , the given scalings will not be reproduced on subsequent ( @xmath307 ) cones . by inverting the system of equations ( [ quantities - rss ] ) , one can obtain @xmath308 , @xmath309 , @xmath310 , @xmath311 , @xmath312 , @xmath313 , and @xmath314 ( see eqs . [ [ invert ] ] ) . from eq . [ [ bernoulli - rss ] ] and use it in place of @xmath313 as a fieldline constant . ] three of these parameters are adjusted to satisfy the regularity conditions at the three singular points . we recall in this connection that @xmath14 and @xmath261 are regarded in our formulation as model parameters ( see [ construction ] ) , and we note that @xmath313 , @xmath308 , @xmath309 , and @xmath310 correspond to the fieldline constants @xmath291 , @xmath308 , @xmath64 , and @xmath292 , respectively . next we describe our numerical approach . we have found it most convenient to start the integration from the alfvn point , since this makes it easier to satisfy the alfvn regularity condition . we choose a small angular interval @xmath315 and specify the model parameters @xmath297 together with the following six parameters : @xmath316 , @xmath317 , @xmath308 , @xmath309 , @xmath310 and @xmath314 . ( the latter parameter does not appear in the system of equations [ [ theta - eqs ] ] , but it affects the magnitudes of the electromagnetic field , density , and pressure through eqs . [ [ quantities - rss ] ] . ) the seventh parameter is @xmath318 , which is given from the alfvn regularity condition ( see appendix [ appendixrss ] ) . we start the integration from @xmath319 , setting @xmath320 , @xmath321 , and @xmath322 . using the upper ( lower ) signs , we integrate upstream ( downstream ) from the alfvn point . before the first step , we evaluate the parameter @xmath291 ( which is used along the integration path to yield @xmath138 ) from equation ( [ bernoulli - rss ] ) . whenever the solution hits a singular point , we adjust one of the above six parameters until a smooth crossing is achieved . in our model we have ignored gravity , and consequently we expect the flow to be nonsteady in the sub slow - magnetosonic regime . we therefore do not attempt to obtain steady trans slow - magnetosonic solutions , and thus we do not continue the integration upstream of the slow - magnetosonic point . this does not , however , affect our ability to study magnetic effects , as these only become important downstream of this singular point . the polytropic index @xmath14 controls the thermodynamics of the flow . for adiabatic flow problems such as the one considered in this paper , @xmath23 or @xmath16 depending on whether the temperature is relativistic or not . the exponent @xmath261 controls the current distribution . the poloidal current lines are @xmath324 , where , by equation ( [ b - rss ] ) , @xmath325 thus , for @xmath326 , the current @xmath327 is an increasing function of @xmath246 near the base of a trans - alfvnic flow ( where @xmath328 ) , corresponding to the current - carrying regime ( see the left fieldline in fig . [ fig1 ] ) . the larger the value of @xmath329 , the stronger the magnetic pinching force that collimates the flow , and hence the faster the collimation . the case @xmath330 corresponds to the return - current regime ( represented by the right fieldline in fig . [ fig1 ] ) , whereas @xmath331 corresponds to radial meridional current lines . in this paper ( as well as in paper ii ) we concentrate on the case @xmath326 , which should provide a good representation of the conditions near the axis of highly collimated flows such as grb jets . however , in view of the inherent simplifications of the self - similar formulation , this choice is not unique . for comparison , we present an @xmath330 solution in [ soln_d ] , and we also employ a solution of this type in a forthcoming publication ( n. vlahakis & a. knigl , in preparation ) in which we model relativistic agn outflows . a realistic global solution would encompass both the current - carrying and return - current regimes , as sketched in figure [ fig1 ] . this situation might be mimicked with the help of two , properly joined , self - similar models : one ( with @xmath326 ) that applies near the axis , and the other ( with @xmath330 ) that applies further out ( at larger cylindrical distances ) . the parameter @xmath323 ( the @xmath332 coordinate of the disk in the given system of coordinates ; see fig . 2 ) affects only the boundary conditions on the disk . for example , @xmath323 can be used to mimic a keplerian rotation law . ( recall from [ introduction ] that , in the relativistic case considered here , one can not naturally incorporate a keplerian rotation law as in the nonrelativistic @xmath3 self - similar solutions . ) in our model , @xmath333 ; thus , for @xmath334 , @xmath335 along the conical surface of the disk . however , for @xmath336 , points on the surface of the disk at different cylindrical distances @xmath337 correspond to different values of @xmath338 ( see fig . [ fig2 ] ) , and @xmath65 decreases faster than @xmath339 ( with the rate depending on how fast the function @xmath340 decreases along the surface of the disk ) . the solutions we present in this section are motivated by the grb outflow scenario outlined in [ introduction ] . we approximate the outflow from a disk around a stellar - mass black hole as a pair of `` pancakes '' ( see , e.g. , * ? ? ? * ) that move in opposite directions away from the disk surfaces . the flow originates from the inner part of the disk , which extends from the last stable orbit around the black hole at @xmath341 to an outer radius @xmath342 , which for definiteness we choose to be @xmath343 . for simplicity we set @xmath334 , so ( see eq . [ [ alpha ] ] ) @xmath344 . we take into account the baryonic matter , the electron - positron / photon fluid , and the large - scale electromagnetic field . as noted in [ roles ] , we focus on solutions in the current - carrying regime ( @xmath326 ) . for this choice of @xmath261 , equation ( [ current ] ) implies that the current vanishes smoothly as the axis ( @xmath345 ) is approached . a general property of our solutions is that the flow reaches an asymptotic cylindrical region where the acceleration terminates ( see [ soln_a ] ) . since we seek to maximize the acceleration efficiency , we consider flows that do not collimate too fast ( see [ sectioncollimation ] ) , and hence we focus on the smallest possible values of @xmath261 . we therefore choose @xmath346 . near the origin , the thermal energy associated with the radiation and @xmath347 pairs is nonnegligible ; the optical depth is then large enough to ensure local thermodynamic equilibrium . we therefore assume that the gas ( consisting of baryons with their neutralizing electrons as well as photons and pairs ) evolves adiabatically . the polytropic index is fixed at @xmath23 , corresponding to relativistic temperatures for the matter and to a blackbody distribution for the radiation . using the stephan - boltzmann constant @xmath348 ( where @xmath349 is the temperature ) and equations ( [ p - rss ] ) and ( [ m - rss ] ) , the temperature in units of the electron rest energy is @xmath350 or , equivalently , @xmath351 the matter - to - radiation pressure ratio @xmath352 is constant in the large - temperature limit @xmath353 , where , given that the pair number density greatly exceeds the baryon number density , the pair distribution may be approximated by a maxwellian with zero chemical potential . in this case @xmath354 ( a value very close to the more accurate @xmath355 that characterizes a fermi distribution ) . in the large - temperature limit the pair number density has the @xmath23 polytropic scaling ( @xmath356 ) , but at @xmath357 it decreases exponentially , resulting in @xmath358 . our polytropic model captures both limits but not the intermediate temperatures . which of the two approximations ( @xmath359 or @xmath360 ) is more accurate depends on the initial temperature @xmath361 . for @xmath362 we choose @xmath359 , which yields the correct pressure temperature relation during the initial stage of the flow , when thermal effects are important . this choice introduces an error ( leading to an underestimate of the lorenz factor ) in regions where @xmath363 but the radiation energy is nonnegligible in comparison with the baryon rest energy . however , because of the weak ( a power of 1/4 ) dependence of @xmath364 on @xmath365 in equation ( [ theta ] ) , this error remains small . we note in this connection that we also neglect the pair rest - energy density , since it is much smaller than the matter pressure in the @xmath366 regime where the pair contribution is maximized . the mass - loss rate in the outflow is @xmath367 , or , @xmath368 if @xmath369 is the burst duration , then @xmath370 is the total baryon mass ejected . the total energy is @xmath371 , and initially it resides predominantly in the electromagnetic field ; the initial thermal energy is approximately @xmath372 . , @xmath373 , and ( for the initial enthalpy - to - kinetic - energy ratio ) @xmath374 . ] in vk01 we considered two lower limits on the baryon loading , corresponding , respectively , to the requirements ( 1 ) that the flow be optically thick in the region where the pair number density drops below that of the baryons and ( 2 ) that the flow be matter - dominated when it becomes optically thin . we also obtained an upper limit on @xmath12 from the requirement that the flow be optically thin in the internal grb - shock regime . in the solutions presented in this paper we evaluate the optical depths more accurately . specifically , consider two neighboring fieldlines labeled by @xmath141 and @xmath375 . along a direction @xmath376 that makes an angle @xmath377 to the flow velocity , the optical depth is @xcite @xmath378 where @xmath379 is the electron / positron number density , @xmath380 is the thomson cross section , and @xmath381 is the distance between the fieldlines . the optical depth is minimized when @xmath382 , for which @xmath383 ( corresponding to photons moving perpendicular to the flow direction in the comoving frame ) . starting from a point on the inner fieldline , we chart the photon trajectory by using @xmath382 and @xmath383 until the outer fieldline is reached , and use this information to evaluate the optical depth . we now present the results of the numerical integration for four representative solutions ( labeled @xmath384 , @xmath385 , @xmath1 , and @xmath386 ) , for which the parameters are given in table [ table1 ] . the most important physical quantities are plotted in figure [ fig3 ] , in which each column corresponds to a given solution . the properties of these solutions are described in detail in the following subsections . lccccccccccc solution & @xmath261 & @xmath387 & @xmath309 & @xmath310 & @xmath291 & @xmath314(cgs ) & @xmath388 & @xmath389 & @xmath318 & @xmath276 & @xmath128(@xmath390 cm ) + @xmath391 & @xmath392 & @xmath393 & @xmath394 & @xmath395 & @xmath396 & @xmath397 & @xmath398 & @xmath399 & @xmath400 & @xmath401 & @xmath402 + @xmath385 & @xmath392 & @xmath403 & @xmath394 & @xmath360 & @xmath404 & @xmath405 & @xmath398 & @xmath399 & @xmath406 & @xmath407 & @xmath402 + @xmath1 & @xmath392 & @xmath408 & @xmath409 & @xmath410 & @xmath411 & @xmath412 & @xmath398 & @xmath413 & @xmath414 & @xmath415 & @xmath416 + @xmath386 & @xmath417 & @xmath418 & @xmath419 & @xmath420 & @xmath421 & @xmath422 & @xmath423 & @xmath424 & @xmath425 & @xmath426 & @xmath402 + + + this solution represents our fiducial model of a trans - alfvnic grb outflow in which the poynting flux exceeds the enthalpy flux at the origin ( @xmath427 ) . the flow starts from the disk with a nonrelativistic velocity and in a short distance crosses the slow magnetosonic point ( where @xmath428 ) . the acceleration in this regime is due to the pressure gradient force ( the centrifugal acceleration is negligible for @xmath429 ) . the slow magnetosonic point arises from the interplay between the vertical gravitational and pressure - gradient forces . as we neglect gravity , we start the integration slightly above that slow point ( @xmath430 , see fig . [ fig3]a2 ) . in this initial sub - alfvnic regime , @xmath431 . equation ( [ a2 ] ) for the lorentz factor gives @xmath432 thus , a poynting - dominated flow ( @xmath427 ) is always close to being force - free ( @xmath433 ) , and the initial enthalpy - to - poynting flux ratio satisfies @xmath434 ( @xmath435 in the displayed solution ) . as the flow moves downstream it crosses the alfvn singular point . we solve numerically for the slope @xmath318 of the alfvnic mach number that satisfies the alfvn regularity condition ( see [ alfven - section ] ) . in the super - alfvnic regime there are three possible cases : + 1 . the flow recollimates ( @xmath436 ) and there is a termination point where the entire energy is suddenly transformed into kinetic motion ( corresponding to the solution hitting the modified fast singular surface but not being able to cross it ) ; + 2 . the flow starts to decelerate at some heigh above the disk ; + 3 . the flow becomes asymptotically cylindrical . the last case is the only physically acceptable one and corresponds to a specific value of one of the model parameters . ( for example , for the chosen parameter set @xmath261 , @xmath316 , @xmath318 , @xmath308 , @xmath309 , and @xmath310 , there is a unique value of @xmath291 that corresponds to the cylindrical solution ) . this is the `` magnetic nozzle '' mechanism first described by @xcite and discussed in [ pol_accel ] . when the flow has cylindrical asymptotics , the asymptotic regime @xmath437 is the only possible solution of @xmath438 ( where @xmath439 is given by eq . [ [ d ] ] ) . in this case , the modified fast - magnetosonic singular surface ( the `` event horizon '' for the propagation of fast waves ) corresponds to the asymptotic cylindrical regime ( where all the fast mach cones along each fieldline have the same opening angle ) . as in the `` cold '' solutions derived by @xcite , the critical value of @xmath291 is always close to @xmath440 . figure [ fig3]a1 shows the meridional projections of the innermost and outermost fieldlines on a logarithmic scale . our chosen initial cylindrical distances are @xmath441 cm for the innermost fieldline and @xmath442 cm for the outermost one ; these distances are only slightly larger than those of the footpoints of these fieldlines in the disk . figure [ fig3]a1 also depicts the optical paths of photons that originate at three points on the innermost fieldline ; these were calculated according to the procedure described at the beginning of this section and are plotted as dashed lines labeled 1 , 2 , and 3 . figure [ fig3]a7 shows the corresponding optical depths @xmath443 , @xmath444 , and @xmath445 along these paths as a function of @xmath446 . it is seen that , as one moves downstream ( with the initial point on the innermost fieldline shifting to a larger height @xmath332 above the disk ) , the optical depth gets smaller , and the flow eventually becomes optically thin ( @xmath447 for @xmath448 ) . figure [ fig3]a7 also shows the radial profile of the mass - loss rate . for @xmath448 , @xmath449 , corresponding to a total ejected baryonic mass of @xmath450 for a typical burst duration @xmath451 . the total energy flux along the poloidal flow is @xmath452 ( where @xmath453 is the mass flux ) . the main part of the total energy flux ( at least in the initial phase of the flow ) is the poynting flux ( see eq . [ [ tensor_flux ] ] ) , @xmath454 ( note that this term is positive since @xmath455 . ) the remaining part @xmath456 ( see the energy conservation relation , eq . [ [ int - mu ] ] ) includes the kinetic energy flux ( @xmath457 ) and the enthalpy flux . the total energy loss rate is @xmath458 , and for @xmath459s the total energy injected into the two oppositely directed jets is @xmath460 . figure [ fig3]a2 shows the various energy fluxes in units of @xmath461 ( the mass flux @xmath462 ) as well as the lorentz factor @xmath90 as functions of @xmath337 , the distance from the axis of rotation along the innermost fieldline . we distinguish three different regimes : \1 . + from the initial point ( slightly above the slow point ) up to @xmath463 cm , @xmath464 ( the exact initial values for solution @xmath384 are @xmath465 and @xmath466 . ) the acceleration is primarily thermal : enthalpy is transformed into kinetic energy . in figure [ fig4]a we see that the force @xmath467 ( which measures the increase in @xmath90 ) is equal to the temperature force @xmath145 ( which describes the decrease of @xmath176 ) . the poynting - to - mass flux ratio ( @xmath468 ) is essentially constant , which means that the field is force - free ; it only guides the flow . to a good approximation , @xmath469 and @xmath470 . so long as @xmath471 , equation ( [ theta ] ) gives @xmath472 ; thus , @xmath473 . this is verified in figure [ fig3]a3 , which shows the poloidal magnetic field ( in units of @xmath474 g ) and the baryon rest - mass density ( in units of @xmath475 g @xmath476 ) . it is seen that @xmath477 ( as expected from the polytropic relation @xmath478 ) and @xmath479 ( as expected from the constancy of the mass - to - magnetic flux ratio @xmath480 ) . as the azimuthal velocity is negligible ( for a highly relativistic poloidal motion ) , equation ( [ int - omega ] ) implies @xmath481 , as expected in the force - free regime . in general , @xmath482 , so @xmath483 , consistent with @xmath479 and @xmath484 . + figure [ fig3]a4 shows the angle between the total magnetic field and its azimuthal component ( which , given that @xmath479 and @xmath485 , decreases as @xmath486 ) , the `` causal connection '' opening angle @xmath487 , and the opening half - angle of the outflow @xmath488 ( its initial value is @xmath489 ) . + figure [ fig3]a5 shows the pressures associated with the poloidal , azimuthal , and comoving magnetic field components , as well as the total thermal ( lepton + radiation ) pressure contribution ( @xmath490 dashed lines ; @xmath491 solid lines , respectively ) . in general , @xmath492 . as @xmath340 and @xmath90 are both proportional to @xmath337 , their ratio is a constant , @xmath493 . thus , @xmath494 , as verified by the figure . + figure [ fig3]a6 shows @xmath495 and @xmath496 ( the squares of the `` light cylinder '' radius and the alfvnic mach number , respectively ) as well as the square of the `` classical fast - magnetosonic proper mach number '' @xmath497 , where @xmath498 is the larger solution of the quadratic @xmath499 the point where @xmath277 corresponds to the light surface , which is close to the alfvn surface @xmath500 ( more accurately , the alfvn surface is where @xmath501 ) . + the point @xmath502 is the classical fast - magnetosonic point . it is seen that , up to that point , @xmath503 . this can be understood by noting that , since @xmath504 , the solution of equation ( [ cl_fast ] ) yields @xmath505 . it follows that , so long as @xmath506 ( using @xmath507 ) , @xmath508 . + at the classical fast point ( subscript @xmath509 ) @xmath510 , so equation ( [ a2 ] ) for the lorentz factor gives @xmath511 . if this point is located inside the thermally dominated region ( @xmath512 , as in the depicted solution ) , then , using @xmath513 , one infers @xmath514 . as @xmath515 , it follows that @xmath516 . + from the end of the thermal acceleration zone , where @xmath517 , up to @xmath518 cm , it is seen from figure [ fig3]a2 that the lorentz factor continues to increase as @xmath519 , with @xmath520 a constant @xmath521 . this exponent is in general different for different solutions . we find that for solutions with a given @xmath291 but different @xmath522 , @xmath523 is always less than 1 and decreases with increasing @xmath522 . for larger @xmath522 the magnetic effects are less important , resulting in a weaker collimation of the flow and hence in a larger asymptotic width @xmath524 . since the final lorentz factor is @xmath525 , the exponent @xmath526 should be smaller . centrifugal acceleration can also influence the magnitude of this exponent : for mildly relativistic flows we expect this acceleration to increase the lever arm of the flow , resulting in a lower value of @xmath523 . + the acceleration in this region is due to magnetic effects : poynting energy is transformed into kinetic energy . figure [ fig4]a shows that the force @xmath467 ( which describes the increase in @xmath90 ) is equal to the lorentz force @xmath153 ( which derives from the decrease of @xmath527 ) . the poynting - to - mass flux ratio declines from its initial value @xmath528 as @xmath529 , with @xmath530 , a result verified by figure [ fig3]a2 . however , the deviation from the initial value of @xmath119 is not too strong , so the scalings @xmath479 , @xmath477 , and @xmath531 remain approximately the same as at smaller radii ( see figs . [ fig3]a3 , [ fig3]a4 , and [ fig3]a5 ) . + figure [ fig3]a6 shows that @xmath532 but that @xmath165 increases faster than @xmath340 . as the poynting - to - matter energy flux ratio is @xmath533 / \left[\xi \gamma^2 \rho_0 c^2 v_p\right]= ( \mu - \xi \gamma ) / \xi \gamma = ( x^2 - x_a^2 ) / ( m^2 + 1 - x_a^2 ) \approx x^2 / m^2 $ ] , this is another manifestation of the poynting - to - kinetic flux conversion . + the bernoulli equation simplifies in this region to @xmath534 , which can be used , along with @xmath535 and equation ( [ v - rss ] ) , to obtain the fieldline slope , @xmath536 ( the same result can be found using eq . [ [ bernoulli - rss ] ] . ) so long as @xmath532 , the slope is @xmath537 ( as the critical value for @xmath291 is close to @xmath538 , and @xmath539 ) . thus , the shape of the fieldlines is parabolic , @xmath540 , as figure [ fig3]a1 verifies ( cf . + at the end of the magnetic acceleration region the flow becomes cylindrical : figure [ fig3]a1 shows that the fieldlines tend to a constant @xmath337 , whereas figure [ fig3]a4 indicates that the opening half - angle @xmath541 tends to zero . + figure [ fig4]b shows that , although the electric force almost cancels the magnetic force ( @xmath542 ) , their sum @xmath543 . on the other hand , @xmath544 , or ( using eqs . [ [ fcbot ] ] and [ [ fibot ] ] ) , @xmath545 , which means that the poloidal curvature radius vanishes ( i.e. , the poloidal fieldlines become straight ) a characteristic of cylindrical collimation . ( in the initial acceleration region near the disk it is seen from fig . [ fig4]b that @xmath546 , which implies that the fieldlines curve toward the axis . ) self - similarity does not allow such a transition , other types of self - similar solutions can be constructed in which such a transition takes place and the flow continues to accelerate ( see * ? ? ? * ) . ] + in the cylindrical region , the transfield force - balance equation becomes @xmath547 or , using @xmath548 and equations ( [ quantities - rss ] ) , @xmath549 as the left - hand side of equation ( [ transf_cyl ] ) is small but positive , @xmath550 . using @xmath551 and @xmath552 , we conclude that @xmath553 is invariably @xmath554 ; hence , only a current - carrying jet ( @xmath326 ) can have cylindrical asymptotics . + equation ( [ bernoulli_asympt ] ) with @xmath555 gives the final kinetic - to - poynting flux ratio , @xmath556 , corresponding to @xmath557 . the value of @xmath558 becomes as large as that of @xmath559 ( see fig . [ fig3]a6 ) , so a rough equipartition holds between the kinetic and poynting fluxes : @xmath560 ( see fig . [ fig3]a2 ) . + the implied conversion efficiency of @xmath561 between the total energy ( injected largely in the form of a poynting flux ) and the final kinetic energy of the flow is consistent with the internal shock scenario for grbs . specifically , the asymptotic kinetic energy in each jet is @xmath562 for our fiducial parameters . with a radiative efficiency of @xmath563 ( e.g. , * ? ? ? * ; * ? ? ? * ) , this implies a radiated @xmath90-ray energy of a few times @xmath564 ergs , as inferred observationally @xcite . + figure [ fig3]a1 shows that , as the cylindrical region is approached , the lorentz factor increases more slowly than @xmath565 , resulting in a nonnegligible value of @xmath566 . since @xmath567 , this explains the divergence of the @xmath568 and @xmath569 curves in fig . [ fig3]a5 near the cylindrical region . [ [ transf_cyl ] ] can be written as @xmath570 , which shows that @xmath571 in the cylindrical regime . ) + the condition that the grb emission region be optically thin to the prompt @xmath90-ray photons typically implies a lower limit @xmath572 on the lorentz factor of the outflow ( e.g. , @xcite ) . the asymptotic lorentz factor attained in our fiducial solution ( and , in fact , in the other three solutions presented in this section ) readily satisfies this requirement . as noted at the beginning of [ results ] , we approximate the outflow as a pair of pancakes that move in opposite directions away from the disk . in the context of the internal - shock scenario , each pancake consists of a number @xmath573 of shells . the width of each shell is @xmath574 cm , where the total duration of the burst is @xmath575 s. if the ejection of the shells from the disk surface ( @xmath576 ) starts at time @xmath577 and ends at time @xmath369 , then each shell can be labeled by its ejection time @xmath578 , with @xmath579 . the shell `` @xmath112 '' moves along the poloidal fieldline and at time @xmath580 its position is @xmath581 . the distance between two neighboring shells @xmath112 , @xmath582 is ( using @xmath583 ) @xmath584 at early times the second term on the right - hand side of equation ( [ distance_of_shells ] ) is negligible and the frozen - pulse approximation holds ( the distance between the two shells is constant , @xmath585 ) . however , even if the integrand in the second term is tiny , there will eventually come a time ( call it @xmath586 ) when the integral will cancel the @xmath587 term ( for @xmath588 , corresponding to the trailing shell moving faster than the leading one ) . one can integrate the equation of motion for each shell , @xmath589 , with @xmath590 , to show that two neighboring shells with @xmath591 will not collide for as long as they move in the flow acceleration zone . this is a reflection of the fact that the difference in the initial lorentz factors of the two shells is not large enough to compensate for the longer acceleration time of the leading shell as it crosses this region , which also implies that the frozen - pulse assumption continues to hold throughout the acceleration zone . only after the shells reach the constant - velocity , cylindrical - flow region , will the two shells collide ( @xmath592 ) . using equation ( [ distance_of_shells ] ) , this will happen at a distance @xmath593 beyond the end of the acceleration zone . so long as the frozen - pulse approximation is valid , @xmath594 . as we proved in [ governing ] , it is possible to examine the motion of each shell in this regime using the @xmath3 self - similar model . if we focus on the shell @xmath595 , then the parameter @xmath261 and boundary conditions @xmath596 that we used to specify the solution refer to this particular @xmath595 , and the corresponding solution @xmath597 is valid only for this shell . in general , we may choose a different set @xmath598 for a different shell @xmath112 , but we have to be careful to also satisfy the assumption of a quasi - steady poloidal field , which requires all the shells to experience the same poloidal magnetic field at any given location . this requirement constrains the choice of boundary conditions . no constraint is necessary in the force - free regime , in which the electromagnetic field is effectively decoupled from the matter in that there is no feedback from the matter acceleration to the field . the electromagnetic field only guides the flow and the motion is effectively hd . as we proved in vk01 , one simply has to replace the spherical @xmath3 in the radial - outflow scalings ( e.g. , @xcite ) with the cylindrical radius @xmath337 to get the correct scalings in the magnetically guided case . however , in the superfast regime ( @xmath599 , @xmath600 ) , one finds that the ( appropriately simplified ) bernoulli and transfield force - balance equations ( eqs . [ [ bernoulli - rss ] ] and [ [ transfield - rss ] ] ) become completely @xmath112-independent if one writes @xmath601 , @xmath602 , and @xmath603 . ( @xmath261 must also be @xmath112-independent . ) thus , we are free to specify the functions @xmath604 , @xmath262 , @xmath264 , and @xmath265 when @xmath605 , and from the fact that @xmath141 is @xmath112-independent we then also have @xmath606 . these functions correspond to the initial conditions for each shell , obeying the quasi - steady poloidal magnetic field assumption . using @xmath607 in any quantity @xmath608 , we may find either the time dependence following the motion of a particular shell : @xmath609 with @xmath112 held constant , or the time dependence at a given point in space as different shells pass by : @xmath610 , with @xmath611 held fixed . the `` initial '' value @xmath612 corresponds to the time @xmath613 when the shell @xmath112 is ejected from the surface @xmath614 . thus , we see that the @xmath112-dependence in @xmath615 represents the initial conditions at the ejection surface for each shell . to illustrate how the time dependence in our model can be recovered , we now consider the baryon mass density along the flow . from equation ( [ gamma - rho - rss ] ) , @xmath616^{\frac{4}{f } } \mu(s_0 ) x_a^4(s_0 ) g(s ) \ a^{2-\frac{4}{f } } } { 4 \pi c^2 f^{\frac{4}{f } } \sigma_m(s_0 ) g^2(s_0 ) \left[{\cal m}^2(\theta ) + { \cal x}^2(\theta)\right ] \sigma_m(s ) } \nonumber \\ & & \times \frac{{\cal m}^2(\theta ) g(s)-1+x_a^2(s ) } { { \cal m}^2(\theta ) g(s ) } \frac{{\cal x}^2(\theta ) g(s ) + { \cal m}^2(\theta ) g(s ) } { { \cal x}^2(\theta ) g(s ) + { \cal m}^2(\theta ) g(s)-1}\ .\end{aligned}\ ] ] neglecting the second line ( which is approximately equal to 1 in the non force - free regime ) , @xmath617 figure [ fig5 ] shows @xmath618 as a function of the arclength along the inner fieldline at various times ( as in the hd case illustrated in fig . 1a of @xcite ) . at each instant of time @xmath580 , one can identify ( using @xmath619 ) which shell @xmath112 is at the point @xmath611 . knowing the distribution of @xmath620 as a function of @xmath611 for the reference shell @xmath595 , one can obtain the density for all the other shells ( corresponding to a particular choice of the function @xmath621 , such as the one depicted in fig . [ fig5]b ) . the ideal - mhd theory is not always a good approximation ; in some cases it must be replaced by an exact multi - fluid theory including a relativistic ohm s law ( e.g. , * ? ? ? a necessary condition for its applicability , derived from the need to attain the requisite charge density and current , implies a lower limit on the matter density ( e.g. , * ? ? ? the limits are @xmath622 , @xmath623 ( where @xmath624 the electron charge ) , and we have verified that they are well satisfied in all of our solutions . in particular , the protons and the neutralizing electrons have a sufficiently large number density to screen the electric field parallel to the flow and to provide the necessary charge density and current . our neglect of gravity has turned out to be a valid approximation , since the flow near the disk is thermally dominated and the temperature force density near the origin of our fiducial solution ( at @xmath625 , slightly above the slow - magnetosonic point ) is @xmath626 dyne @xmath476 , which is much larger than the gravitational force density associated with a central object of a few solar masses . in [ timerecovery ] we demonstrated that the frozen - pulse assumption is valid throughout the acceleration region . we have also neglected the terms @xmath627 and @xmath628 in equation ( [ momentum - s ] ) . our results by and large verify these approximations , as we find that only near the asymptotic cylindrical region does @xmath629 deviate from @xmath630 ( see fig . [ fig3]a5 ) . even so , the ratio of the terms in equation ( [ momentum - s ] ) that contain @xmath631 over the last term in that equation is @xmath632 , which ( using eq . [ [ ohm ] ] ) becomes @xmath633 . an alternative argument supporting the conclusion that @xmath634 is based on the requirements that @xmath635 and @xmath636 . the former gives ( using @xmath637 and eq . [ [ e_a ] ] ) @xmath638 , whereas the latter yields ( using eq . [ [ v - b ] ] ) @xmath639 . thus , the ratio @xmath640 is bounded in a tiny interval around @xmath641 , from @xmath642 to @xmath643 . in the limit of low temperatures ( @xmath645 , @xmath646 ) , we reproduce the exact cold relativistic solutions of @xcite , using a different parameter regime appropriate to grb outflows . @xcite employed the same model but examined more massive flows that were not close to being force - free and thus were not accelerated as efficiently . we present a cold solution in the second column of figure [ fig3 ] ; it can be readily compared with the fiducial solution @xmath384 . the cold solution is closer to being force - free , since @xmath647 ( using @xmath648 ) , and hence ( by eq . [ [ ximu ] ] ) @xmath649 . ( for comparison , @xmath650 for solution @xmath384 . ) since the flow does not pass through a slow magnetosonic point in this case , we are able to find the solution all the way down to the disk surface . near the origin the flow corotates with the disk ( @xmath651 ) and is accelerated due to the centrifugal force @xmath146 . the lorentz factor does not increase linearly with @xmath337 ( but rather , to a good approximation , as @xmath652 ) . however , the lorentz force soon takes over and thereafter @xmath102 ( see fig . [ fig3]b2 ) . at the classical fast point @xmath653 , equation ( [ a2 ] ) gives @xmath654 , which ( using @xmath655 ) can be transformed into the familiar form @xmath656 ( e.g. , * ? ? ? unlike the purely monopole case @xcite , this point is located at a finite distance from the origin , and the bulk of the magnetic acceleration occurs further downstream . the flow near the cylindrical regime is similar to that in the fiducial solution , as thermal effects are not important anymore . however , a signature of the cold solution can be seen in figure [ fig3]b6 in the weaker deviation of @xmath165 from @xmath657 . ( the solution of eq . [ [ cl_fast ] ] in this case is @xmath658 , and the deviation of @xmath165 from @xmath657 is due only to @xmath659 becoming larger than @xmath660 , which happens near the cylindrical region . ) this solution describes a slow rotator , in which the poynting flux is smaller than the enthalpy flux ( so @xmath662 ) . the angular velocity of the inner fieldline is @xmath663 rad s@xmath664 ( see table [ table1 ] ) , much smaller than the typical value ( @xmath665 rad s@xmath664 ) near a solar - mass black hole . this solution is characterized by significantly higher @xmath666 ratios ( see fig . [ fig3]c4 ) and much lower values of @xmath340 ( see fig . [ fig3]c6 ) in comparison with solutions @xmath384 and @xmath385 . it is also seen that this solution is not force - free : @xmath667 ( see eq . [ [ ximu ] ] ) . the acceleration in this case is thermal ( @xmath668 ; see fig . [ fig3]c2 ) . nevertheless , since the flow is trans - alfvnic , the poloidal magnetic pressure is much larger than the thermal pressure ( see fig . [ fig3]c5 ) . this can be understood from the fact that , in general , one has at the alfvn point @xmath669 only when the flow is super - alfvnic everywhere ( even near the origin ) can one obtain a hot , radial , hd - like solution with @xmath670 . the magnetic field only guides the flow in this case . the collimation , however , is not very efficient ( see fig . [ fig3]c1 ) . the flowlines `` attempt '' to become conical , but since in the framework of @xmath3 self - similarity the radial distance is given by @xmath671 , the only possible asymptotes as @xmath672 are @xmath673 . for @xmath674 we have cylindrical asymptotics ( as in solutions @xmath384 and @xmath385 ) , and the only other possibility is for all the flowlines to have the same asymptotic opening half - angle , which is the situation depicted in figure [ fig3]c1 . this behavior can be understood if we write the transfield force - balance equation as @xmath675 + \nonumber \\ + \frac{b_p^2}{4 \pi \varpi } \left(\frac{m v_{\phi } } { v_p}\right)^2 \hat{n } \cdot \hat{\varpi } - \hat{n } \cdot { { \mathbf{\nabla}}\!\!_s}\left(p+\frac{b_p^2}{8 \pi}\right)\end{aligned}\ ] ] ( see @xcite ) . by neglecting the centrifugal ( @xmath676 ) and the poloidal magnetic pressure terms and using equation ( [ ohm ] ) ( which implies @xmath677 ) , equation ( [ transfield_v03 ] ) can be rewritten as @xmath678 the first term on the right - hand side is positive and acts to collimate the flow ( resulting in a positive poloidal curvature , @xmath212 ) , whereas the second term is negative and has the opposite effect on @xmath204 . in poynting flux - dominated solutions , the pressure term is negligible and the collimation continues ( albeit at a very low rate ; see paper ii ) until the shape becomes cylindrical . however , when the thermal pressure becomes comparable to the comoving magnetic pressure ( as in the present solution ; see fig . [ fig3]c5 ) , the two terms on the right - hand side of equation ( [ transfield_v03_appr ] ) cancel each other . the curvature @xmath679 thus vanishes before the cylindrical regime is reached , resulting in an asymptotic conical flow . for completeness , we also present a return - current ( @xmath680 ) solution . since , near the disk surface , @xmath681 is a decreasing function of @xmath246 in the @xmath330 case under consideration , the magnetic force acts to _ de_collimate the flow ( see [ sectioncollimation ] ) . this results in a weaker collimation than in the current - carrying ( @xmath326 ) solutions ( see fig . [ fig3]d1 ) . the electric force succeeds in collimating the flow despite the counter effect of the magnetic force ; as discussed in [ sectioncollimation ] , this is associated with the positive charge density ( @xmath682 ) expected in the return - current regime . examining the forces in the transfield direction , we find that @xmath683 ; i.e. , the lorentz force is balanced by the inertial force . ( this can be contrasted with the cylindrical solution @xmath384 , where the balance is provided by the centrifugal force : @xmath684 ; see figure [ fig4]b ) . as in the nonrelativistic , cold mhd flow considered by @xcite , the curvature radius is nonnegligible and the solution terminates at a finite height above the disk . we regard the fiducial solution @xmath384 as providing the most plausible representation of an accelerating grb outflow . neutrino energy deposition and magnetic field dissipation are likely to give rise to a nonnegligible thermal component at the origin , as envisioned in the original fireball scenario . such a component is incorporated into solution @xmath384 but is missing in solution @xmath385 . it is , nevertheless , worth reemphasizing that the dominant energy source even in solution @xmath384 is the poynting flux . we note in this connection that one of the potential problems of traditional fireball models has been the `` baryon contamination '' issue : why does the highly super - eddington luminosity that drives the fireball not lead to a more significant mass loading of the flow and thereby keep @xmath685 comparatively low ? if most of the requisite energy were injected in a nonradiative form , then this problem would be alleviated . the expectation that grb outflows are poynting flux - dominated also renders the slow - rotator ( enthalpy flux - dominated ) solution @xmath1 less relevant to their interpretation than the fiducial solution . grb outflows are inferred to be highly collimated at large distances from the origin : this has motivated us to construct our fiducial solution in the current - carrying ( @xmath686 ) regime , which is applicable near the symmetry axis ( see fig . [ fig1 ] ) . however , as demonstrated in figure [ fig3 ] , the illustrative current - carrying solution @xmath386 also becomes well collimated on a similar length scale , so it may also provide an adequate representation of such outflows . in fact , this regime may be a natural choice for modeling the far - downstream region of the flow since it corresponds to a plausible current - configuration on large scales ; furthermore , as shown in @xcite , it also results in a more efficient acceleration . we have opted to focus on the current - carrying regime in this paper since the solutions in this case , unlike the return - current solutions formally extends to infinity . we incorporate the return - current regime into a grb outflow model in @xcite . our fiducial self - similar `` hot '' solution as well as our `` cold '' relativistic mhd solution ( and the corresponding ones derived by @xcite and @xcite ) have cylindrical asymptotics . in the nonrelativistic theory , it was shown @xcite that it is possible to derive exact solutions of radially self - similar flows that have either one of the following three types of asymptotes : cylindrical , parabolic , or conical . although we expect that it should be possible to obtain different asymptotic behaviors also in radially self - similar relativistic flows , this has not yet been studied in detail . our basic conclusions about the magnetic acceleration of the flow do not , however , depend on the shape of the asymptotic flowlines , although the quantitative constraints on the mass loading ( to assure that the flow is optically thin in the shell - collision region ) would be eased if the flow continued to expand rather than collimated to a cylinder . grb jets are often modeled in terms of conical jets , and this picture has gained support from observations of a panchromatic break in the light curve of several grb afterglows ( from which jet half - opening angles in the range of @xmath687 to @xmath688 have been inferred ; e.g. , @xcite ) . however , a sharp physical boundary may not be required for interpreting the data in the context of a beam - patterned outflow ( e.g. , * ? ? ? * ; * ? ? ? furthermore , it has been argued that a cylindrical jet model could provide a better fit to the light curves of at least some afterglows and might perhaps even be able to account for the panchromatic breaks commonly attributed to a finite jet opening angle ( e.g. , * ? ? ? * ; * ? ? ? even if grb jets are not characterized by a global cylindrical geometry , the asymptotically cylindrical solutions that we derived may be applicable to the `` patchy shell '' scenario for variable grb outflows @xcite . in this picture , the outflowing shells are ejected within a given opening angle but do not fill the entire solid angle into which they are injected . in the context of the internal shock model , this scenario alleviates the energy requirements on bright bursts ; it is also consistent with the apparent lack of a strong correlation between the prompt high - energy and the afterglow fluences , and it may account for the early - time variability in the afterglow emission of a source like grb 021004 @xcite . in the phenomenological model considered by @xcite , successive blobs of angular scale @xmath689 are ejected randomly within an opening angle of @xmath690 , and the number of blobs ejected along a given direction ( given that more than one is required for the production of internal shocks ) is also taken to be a random number ( distributed uniformly in a predetermined range ) . this model was intended to mimic the behavior of causally disconnected regions in large - scale ejected shells . an alternative physical picture is that of a collection of magnetically active regions on the surface of the disk that feed the grb outflow . in this picture , the magnetic field associated with each `` patch '' guides successive `` mini '' shells along roughly congruent paths , thereby enhancing the efficiency of internal - shock production . however , the ejection directions from different sites need not be parallel to each other , so some of the beamed @xmath90-ray emission may miss the observer : this might contribute to the appearance of quiescent times between grb pulses @xcite . the high ( @xmath561 ) poynting - to - kinetic energy conversion efficiency exhibited by all our fast - rotator solutions ( solutions @xmath384 , @xmath385 , and @xmath386 ) has made it possible for our models to accommodate the internal - shock emission scenario without requiring a prohibitively large energy input at the source ( see [ results ] ) . however , currently available data on the radiated @xmath90-ray energy @xmath691 ( e.g. , * ? ? ? * ) and the outflow kinetic energy @xmath692 ( e.g. , * ? ? ? * ) , which indicate that they are both approximately equal ( @xmath693 ) , are inconsistent with this scenario for any reasonable radiative energy efficiency of such shocks ( see @xcite for further discussion of this issue ) . one possible resolution of this apparent inconsistency is that the prompt high - energy emission arises directly from electromagnetic energy dissipation without the intermediate step of poynting - to - bulk - kinetic energy conversion ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? our solutions seem to be consistent with this possibility : they imply that about half of the injected energy is converted into bulk kinetic energy ; if the remaining poynting flux can be efficiently converted into high - energy emission , then this would lead to an approximate equality between @xmath691 and @xmath692 . @xcite used ideal - mhd equations with parameterized magnetic energy dissipation to demonstrate ( in the context of a strictly radial flow , i.e. , without taking into account the transfield force - balance equation that determines the fieldline shape ) that the poynting - to - radiation energy conversion could in principle be as high as 50% and that the dissipation may also result in significant bulk acceleration . it would be interesting to combine the approach taken by these authors ( see also @xcite and @xcite ) with the one utilized in this paper to consider the effects of dissipation in a _ nonradial _ flow . we now briefly compare the results of our exact solutions with some of the previous work on mhd effects in grb outflows . the incorporation of `` disposable '' magnetic energy into the traditional fireball model was discussed by @xcite . the behavior that they infer is consistent with our solutions in the regime where @xmath102 ( but not beyond it ) . @xcite interpreted the outflows in terms of pulsar winds that generate intense electromagnetic waves at the point where the ideal - mhd approximation breaks down . he highlighted the fact that the lorentz factors of the particles accelerated by such waves could be as high as @xmath694 and thus greatly exceed the terminal lorentz factors ( @xmath695 ) attained in the @xcite ideal - mhd monopole solution . the acceleration efficiency of our collimating ( and thus nonmonopolar ) ideal - mhd solutions is much higher yet : they yield @xmath696 . @xcite examined a magnetized conical jet with a pure electron - positron composition . they deduced ( and our solutions have confirmed ) that the pair photon fluid is initially accelerated by thermal pressure , with the magnetic field acting only to guide the flow . they also pointed out that radiation drag effects allow the pairs and photons to remain coupled even beyond the point where the scattering optical depth drops below 1 and that at some point pair annihilation freezes out . @xcite subsequently showed that the surviving pairs carry only a fraction @xmath697 of the initial energy , implying that , in the absence of magnetic acceleration , this scenario is extremely inefficient . we note , however , that if a nonnegligible fraction ( @xmath698 ) of the total injected energy is in magnetic form ( [ with the ratio of magnetic to pairs - plus - photons energy injection rates being @xmath699 , as assumed by @xcite ) , then , based on our `` cold '' solution ( see [ soln_b ] ) , half of the magnetic energy could be eventually transformed into kinetic energy of the surviving pairs , reaching overall efficiencies @xmath700 . as the rest - mass density of the final pair population , which represents only a tiny fraction of the initial rest - mass density of pairs . ] for @xmath701 this yields an efficiency of @xmath702 , which is significantly higher than the values estimated by @xcite . as was , however , discussed by these authors , the resulting outflow would have a huge lorentz factor and might not by itself be able to reproduce the observed properties of real grbs . although the presentation in this paper and its companion is focused on the application to grbs , our formalism is quite general and the solutions that we present should be relevant also to other magnetically driven relativistic outflow sources , notably agns and microquasars . the possible relationship between the different classes of beamed relativistic outflows in astrophysics has already been noted before by various authors . one potential attraction of adopting a common modeling framework for these outflows is that it could shed new light on each of the different types of sources . for example , the concept of internal shocks was originally proposed in the context of agn jets @xcite , and it has recently been adopted also for modeling microquasars @xcite . in the latter class of objects , there is direct observational evidence ( in the form of correlated x - ray , infrared , and radio flux variations and the appearance of superluminal radio knots ) for a causal connection between a rapid accretion episode onto the central black hole and the ejection of superluminal blobs ( e.g. , * ? ? ? this is the same basic scenario as the one commonly adopted ( albeit without direct observational support ) for grb outflows . a similar correlation between the x - ray and radio flux behavior and the appearance of radio - bright superluminal knots was recently discovered also in an agn jet @xcite . although the observed behavior in this case does not support the notion that agns are a simple scaled - up version of microquasars , this might be attributable to a different origin of the underlying disk instability that induces the accretion / ejection episodes ( e.g. , * ? ? ? * ) or to different environmental conditions ( e.g. , * ? ? ? there is now also evidence that the high - energy ( gev - range ) @xmath90-ray emission in the blazar class of agns originates in superluminal radio knots which likely represent discrete blobs or shocks in a relativistic jet at a large distance from the origin ( e.g. , * ? ? ? this is essentially the picture envisioned for grb outflows . the superluminal knots in blazar jets such as 3c 345 ( e.g. , * ? ? ? * ) and 3c 279 ( e.g. , * ? ? ? * ) are inferred to move along distinct curved paths , although in the former case it has been argued @xcite that at least two of the knots follow each other along the same trajectory . this behavior is consistent with the `` patchy shell '' model discussed above in connection with grbs . there are also indications that the outflows continue to accelerate on large scales : for example , the lorentz factor of knot c7 in the quasar 3c 345 was inferred to increase from @xmath703 to @xmath704 as it moved from a distance of @xmath705 from the galactic nucleus to @xmath706 @xcite . a scaled - down version of this behavior was found in the radio galaxy ngc 6251 , where knots in the radio jets were inferred to accelerate from @xmath707 at @xmath708 to @xmath709 at @xmath710 @xcite . such large - scale acceleration is most naturally interpreted in terms of a poynting - dominated jet model of the type discussed in this paper ( n. vlahakis & a. knigl , in preparation ) . in conclusion , we have derived in this paper the general formalism for radially self - similar , relativistic mhd outflows and presented exact solutions ( obtained by solving the euler and transfield equations and imposing regularity conditions at the relevant critical points ) to illustrate their basic properties . we focused on trans - alfvnic flows , deferring a discussion of super - alfvnic solutions to the companion paper . we considered `` hot '' and `` cold '' rapid - rotator flows ( the latter reproducing previous results by @xcite and @xcite ) as well as slow - rotator flows in the current - carrying regime ( which should apply near the rotation axis ) , but also described a `` hot , '' fast - rotator solution in the return - current regime . in all cases , we demonstrated that the poynting flux injected at the source can be transformed with high ( @xmath561 ) efficiency into kinetic energy of relativistic baryons at a large ( but finite ) distance from the origin . we concentrated on applications to grb outflows , although we pointed out that similar solutions may describe jet acceleration in agns and microquasars . in the application to grbs , we presented a particular `` hot , '' fast - rotator outflow as a fiducial solution . in this case the outflow is initially accelerated thermally , with the magnetic field acting only to guide and collimate it , but subsequently magnetic effects become fully dominant and are responsible for the bulk of the acceleration ( which occurs between the classical fast - magnetosonic and modified fast - magnetosonic surfaces ) and for the final collimation to cylindrical asymptotics . we stress , however , that this solution is only illustrative and is not meant to provide an accurate description of a typical grb flow . in this solution is about an order of magnitude higher than the minimum lorentz factors typically inferred in real grbs ( e.g. , * ? ? ? ] as we discuss in paper ii , super - alfvnic outflows can also provide plausible representations of grb outflows ; their basic properties , however , are quite similar to those of trans - alfvnic flows . we have , furthermore , concluded that solutions in the return - current regime may also be relevant to these flows ; we address this possibility in greater detail in a separate publication @xcite . this work was supported in part by nasa grant nag5 - 12635 and by the u.s . department of energy under grant b341495 to the center for astrophysical thermonuclear flashes at the university of chicago . n. v. also acknowledges support from a mccormick fellowship at the enrico fermi institute . we may combine equations ( [ int - omega ] ) , ( [ int - l ] ) , and ( [ int - mu ] ) to obtain @xmath712 , and @xmath223 as functions of @xmath713 @xmath340 , and @xmath293 using @xmath714 @xmath715 @xmath716 knowing the field - line constants ( for each @xmath112 ) @xmath717 we can find the quantities @xmath718 ( which in general are functions of @xmath719 , and @xmath112 ) by solving the following system of equations : @xmath720 the bernoulli equation @xmath721 which is obtained after substituting all quantities in the identity @xmath722 using equations ( [ a1],[a2 ] ) and which in the nonrelativistic limit takes the familiar form ( after taylor expanding in @xmath723 ) @xmath724 and the transfield equation ( obtained from the component of the momentum equation along @xmath725 ) @xmath726 \left(\frac{{{\mathbf{\nabla}}\!\!_s}a}{\varpi } \right)^2 + \nonumber \\ & & + \left [ \frac{2 x_a^2}{\varpi_a^3 g } \left({{\mathbf{\nabla}}\!\!_s}a \right)^2 + \frac{\mu^2 x_a^6 a^2}{\varpi_a^5 \sigma_m^2 m^2 g^3 } \left(\frac{g^2-m^2-x^2}{1-m^2-x^2}\right)^2\right ] \hat{\varpi } \cdot { { \mathbf{\nabla}}\!\!_s}a - \nonumber \\ & & -\frac{m^2}{2 } { { \mathbf{\nabla}}\!\!_s}\left[\left(\frac{{{\mathbf{\nabla}}\!\!_s}a } { \varpi } \right)^2\right ] \cdot { { \mathbf{\nabla}}\!\!_s}a -\frac{\gamma-1}{\gamma } { { \mathbf{\nabla}}\!\!_s}\left [ \frac{\xi ( \xi-1)}{m^2 } \frac{a^2 x_a^4}{\sigma_m^2 \varpi_a^4 } \right ] \cdot { { \mathbf{\nabla}}\!\!_s}a - \nonumber \\ & & - \frac{1}{2 \varpi^2 } { { \mathbf{\nabla}}\!\!_s}\left [ \frac{\mu^2 a^2 x_a^6}{\sigma_m^2 \varpi_a^2 } \left(\frac{1-g^2}{1-m^2-x^2}\right)^2\right ] \cdot { { \mathbf{\nabla}}\!\!_s}a = 0\,,\end{aligned}\ ] ] where the operator @xmath727 is related to the curvature radius @xmath728 . the latter equation can be written as [ forces - perp ] @xmath729 where the subscript @xmath198 denotes the component of a force perpendicular to the poloidal fieldline and pointing toward the axis ( along @xmath730 ) , and where the individual terms are given by @xmath731 \cdot \frac{{{\mathbf{\nabla}}\!\!_s}\alpha}{\mid { { \mathbf{\nabla}}\!\!_s}\alpha \mid}- \frac { m^2 \varpi_0 ^ 3 \mid { { \mathbf{\nabla}}\!\!_s}{\cal a}\mid}{4 \alpha g^2 } \l { \cal a } \right\}\ , \label{fibot } \\ & & { { { \mathbf { f}}}}_{p \bot}=\frac{b_0 ^ 2}{4 \pi \varpi_0 } \frac{\gamma-1}{\gamma } \varpi_0 { { \mathbf{\nabla}}\!\!_s}\left [ \frac{\xi ( \xi-1)}{m^2 } \frac{c^2 \psi_a^2}{b_0 ^ 2}\right ] \cdot \frac{{{\mathbf{\nabla}}\!\!_s}\alpha}{\mid { { \mathbf{\nabla}}\!\!_s}\alpha \mid}\ , \\ & & { { { \mathbf { f}}}}_{e \bot}=-\frac{b_0 ^ 2}{4 \pi \varpi_0 } \frac{x_a^2 \varpi_0 \mid { { \mathbf{\nabla}}\!\!_s}{\cal a } \mid}{4 \alpha } \left[\varpi_0 ^ 2 \l { \cal a}+\frac{2 } { g \alpha^{1/2 } } \varpi_0 \hat{\varpi } \cdot { { \mathbf{\nabla}}\!\!_s}{\cal a } + \left(\varpi_0 { { \mathbf{\nabla}}\!\!_s}{\cal a } \right)^2 \left(\frac{1}{x_a}\frac{dx_a}{{\cal a}}-\frac{1}{2 \alpha}\frac{d \alpha}{d { \cal a}}\right)\right]\ , \\ & & { { { \mathbf { f}}}}_{b \bot}=\frac{b_0 ^ 2}{4 \pi \varpi_0 } \left \ { \frac{1}{2 \alpha g^2 } \varpi_0 { { \mathbf{\nabla}}\!\!_s}\left [ \alpha x_a^2 \mu^2 \frac{c^2 \psi_a^2}{b_0 ^ 2 } \left(\frac{1-g^2}{1-m^2-x^2}\right)^2\right ] \cdot \frac{{{\mathbf{\nabla}}\!\!_s}\alpha}{\mid { { \mathbf{\nabla}}\!\!_s}\alpha \mid } + \frac{\varpi_0 \mid { { \mathbf{\nabla}}\!\!_s}{\cal a } \mid}{4 \alpha g^2 } \varpi_0 ^ 2 \l { \cal a } \right\}\ .\end{aligned}\ ] ] in the @xmath3 self - similar model described in [ r - ss_section ] , the integrals of motion have the following form : [ integrals ] @xmath732\ , , \quad \mbox{where } \quad \sigma_0\ , , \alpha_0=const\ , , \quad \mu\ , , x_a\ , , q = const\ , , \quad a=\frac{b_0 \varpi_0 ^ 2}{f } \left(\alpha^{\frac{f}{2}}- \alpha^{\frac{f}{2}}_0 \right)\ , \ ] ] with the fieldline constants given by @xmath733 note that @xmath309 is the `` magnetization parameter '' in the monopole solution of @xcite . in cases where @xmath734 , the requirement that the magnetic flux function @xmath141 vanishes on the rotation axis @xmath345 implies that @xmath735 when @xmath736 and @xmath737 . we examine solutions with @xmath735 throughout this paper . [ theta - eqs ] equations ( [ x - m ] ) and ( [ bernoulli ] ) take the form @xmath738 @xmath739 @xmath740 and can be solved for @xmath741 , and @xmath138 . using @xmath742 and equation ( [ alpha ] ) , we get @xmath743 or , @xmath744 the transfield equation ( [ transfield ] ) becomes @xmath745 = & & \left(f-1\right ) \frac{x_a^4 \mu^2 x^2}{f^2 \sigma_m^2 } \left ( \frac{1-g^2}{1-m^2-x^2 } \right)^2 -\sin^2 \theta \frac{m^2+f x^2 -f + 1}{\cos^2 \left(\psi+\theta\right)}- \nonumber \\ & & -\frac{x_a^4 \mu^2 x^2}{f^2 \sigma_m^2 m^2 } \left ( \frac{g^2-m^2-x^2}{1-m^2-x^2 } \right)^2 + 2 \frac{\gamma-1}{\gamma } \frac{f-2}{f^2 \sigma_m^2 } \frac{\xi \left(\xi-1\right)x^4}{m^2 } \ .\end{aligned}\ ] ] the magnetization function @xmath746 is defined as the poynting - to - matter energy flux ratio : @xmath747 at the alfvn singular point @xmath748 , @xmath749 , @xmath500 , @xmath501 , @xmath750 , @xmath751 , @xmath752 , and , using lh@xmath753spital s rule , @xmath754_a=\frac{1}{\sigma_a+1}\ , , \ \left[\frac{1-g^2}{1-m^2-x^2}\right]_a = \frac{\sigma_a / x_a^2}{\sigma_a+1 } \ , , \ \left[\frac{g^2-m^2-x^2}{1-m^2-x^2}\right]_a = \frac{x_a^2-(1-x_a^2)\sigma_a}{x_a^2 ( \sigma_a+1 ) } \ .\end{aligned}\ ] ] substituting the above ratios into equation ( [ bernoulli - rss ] ) yields @xmath291 as a function of @xmath317 , @xmath316 , and @xmath276 , @xmath755 ^ 2 } \left[x_a^2 \xi_a^2 + \frac{f^2 \sigma_m^2 ( 1-x_a^2)^2 \sin^2 \theta_a } { x_a^2 \cos^2 ( \theta_a + \psi_a)}\right]\,,\ ] ] whereas equation ( [ transfield - rss ] ) gives the alfvn regularity condition @xmath756 \right\ } \nonumber \\ & & = \left[x_a^2-\sigma_a ( 1-x_a^2 ) \right]^2 - ( f-1)\sigma_a^2 ( 1-x_a^2 ) - 2 \frac{\gamma-1}{\gamma}(f-2)\frac{\xi_a-1}{\xi_a } \left\{x_a^2-\left[x_a^2-\sigma_a ( 1-x_a^2 ) \right]^2 \right\ } \ .\end{aligned}\ ] ] next we prove the following statement : if one chooses @xmath757 ( the model parameters ) and specifies seven boundary conditions on a cone @xmath758 , then it is possible to derive all the other model parameters , and the solution is uniquely determined . _ proof : _ suppose that the quantities @xmath759 , @xmath223 , @xmath760 , @xmath761 , @xmath166 , @xmath12 , and @xmath762 are given as functions of @xmath3 along the cone @xmath758 ; i.e. , @xmath763 , @xmath764 , @xmath765 , @xmath766 , @xmath767 , @xmath768 , and @xmath769 . then one finds [ invert ] @xmath770 one is thus in a position to start the integration , using equations ( [ bernoulli - rss ] ) , ( [ g - rss ] ) , and ( [ transfield - rss ] ) , and find a solution that uniquely corresponds to the seven boundary conditions ( or , equivalently , to the parameters @xmath771 ) and the model parameters @xmath14 and @xmath261 . _ qed_. whether the solution actually hits and passes smoothly through any given singular point depends on the choice of the boundary conditions . from a physical standpoint , the most robust solutions cross all the three singular points ( the modified slow - magnetosonic , the alfvn , and the modified fast - magnetosonic points ) . in this case , the three regularity conditions at the ( a priori unknown ) positions of the singular points impose three relationships among the boundary conditions . suppose that we have obtained a solution of the axisymmetric , relativistic , ideal - mhd equations ( [ maxwell])([energy ] ) . if we consider localized , fast varying , axisymmetric disturbances , then we may assume that the unperturbed solution is uniform and time - independent and neglect all its space and time derivatives . we may then look for perturbations having a fourier dependence @xmath772 = \exp [ i(\omega_{\rm co } t_{\rm co } - { \boldsymbol { k}}_{\rm co}\cdot { \boldsymbol { r}}_{\rm co})]$ ] , where , by using the lorentz transformations between the comoving @xmath773 and observer s frame @xmath774 , we get @xmath775 $ ] . lies in the meridional plane , but that , nonetheless , the comoving wavevector @xmath776 has an azimuthal component ! ] it is more convenient to analyze the disturbances in the ( comoving ) flow frame . define a local cartesian system of coordinates @xmath777 such that @xmath778 , and @xmath779 . after linearizing equations ( [ maxwell])([energy ] ) , we may express all the perturbed quantities in terms of the perturbation @xmath780 . after some manipulation , we obtain @xmath781 where the alfvn speed can be expressed in terms of the corresponding proper speed @xmath782 , which satisfies @xmath783 . * alfvn waves : @xmath787 , or @xmath788 , corresponding to a displacement @xmath789 normal to the \{@xmath790 } plane . transforming the dispersion relation to the observer s frame , we get @xmath791\ .\ ] ] * slow / fast - magnetosonic waves : @xmath792 , or , @xmath793 , corresponding to a displacement @xmath789 in the \{@xmath794 } plane . in the observer s frame , we have @xmath795=0 \ .\end{aligned}\ ] ] an interesting property of the waves , related to the discussion on critical / singular surfaces of steady - state mhd , is the following : if the component of the flow proper velocity along the wavevector equals in magnitude , but is opposite in sign , to the comoving proper phase velocity of the wave , then @xmath796 and the wave is static . ( the converse is also true . ) thus , @xmath797 this statement is easily proved by combining @xmath798 with the lorentz invariant @xmath799 and solving for @xmath800 . + equation ( [ theorem ] ) is the generalization of the property of nonrelativistic static waves @xmath801 . it is consistent with the result that proper speeds are the appropriate generalization of ordinary speeds in relativistic theory ( e.g. , * ? ? ? * ) . piran , t. 2001b , in asp conf . ser . , from x - ray binaries to gamma - ray bursts , ed . e. p. j. van den heuvel , l. kaper , & e. rol . , in press ( astro - ph/0111314 ) piran , t. , shemi , a. , & narayan , r. 1993 , , 263 , 861

electromagnetic interactions played a key role in the history of physics since they are related to the first successful example of unification of two apparently different fields , the electric and the magnetic one , into a single body , the faraday tensor @xmath0 . the latter fulfills the so - called maxwell s equations which , on a flat background , are automatically poincar invariant and they yield that @xmath0 can be described in terms of an auxiliary field , the vector potential @xmath1 . even though @xmath0 stays the basic observable of the theory , @xmath1 has the advantage of being apparently easier to handle since every faraday tensor , solution of maxwell s equations , can be reconstructed from a vector potential which solves both the wave equation and a second one , known as the lorenz gauge . with the advent of quantum field theory , this interplay between @xmath1 and @xmath0 has been even more emphasized since the quantization scheme which is canonically employed still focuses on the vector potential and considers @xmath0 , also known as the field strength , as a derived object , albeit it is the real observable . in view of the aharonov - bohm effect this latter assertion might be considered even erroneous since it is often stated that it is actually @xmath1 the true physical object . yet , as noted for example in @xcite , in all idealized and real experiments of the aharonov - bohm kind the true observable is actually the flux of the magnetic field which is present inside an impenetrable region , typically a cylinder . hence even this quantity can be expressed in terms of the components of the field strength tensor . the role of the vector potential becomes primary as soon as interactions are switched on , but , till we consider only a free maxwell system , @xmath0 should contain all the physical information both at a classical and at a quantum level . it is far from the scope of this paper to discuss the details of this procedure , but suffice to say that , on minkowski background and in absence of sources , the result is pretty much satisfactory . yet the situation starts to complicate itself as soon as it is assumed that the spacetime @xmath2 has a non trivial geometry . although we shall provide more details in the main body of the paper , we can easily explain the source of all potential problems . the field strength tensor is best described as a two - form @xmath3 which satisfies maxwell s equations , which in absence of sources can be expressed as @xmath4 and @xmath5 where @xmath6 is the exterior derivative while @xmath7 is the codifferential . it is important to remark that , while the second equation depends on the metric associated to @xmath2 and hence on the geometry , the first one relies only on the smooth differentiable structure of the background and it is thus a constraint . it is at this stage that the scheme employed on minkowski background encounters the first difficulties since , if we leave @xmath2 arbitrary and thus not a priori diffeomorphic to @xmath8 , we can not apply poincar lemma to conclude the existence of @xmath9 such that @xmath10 . in other words it is not true that it is always possible to reconstruct all field strengths fulfilling maxwell s equations , even starting from an auxiliary object such as the vector potential @xmath1 . the consequences of this result of differential geometry has far reaching consequences , since it tells us that , if one wants to quantize maxwell s equations on a curved background , unless @xmath2 is somehow special , one can not use @xmath1 as the building block , but has to refer uniquely to @xmath0 . an example of a field strength which can not be derived from the vector potential can be found in @xcite . the aim of this paper is indeed to develop a quantization scheme for the field strength on an arbitrary four dimensional globally hyperbolic spacetime within the framework of the algebraic formulation of quantum field theory see @xcite for an earlier investigation in this language . this is certainly not the first investigation in this direction and preliminary works are present in @xcite and , as a special case of a much broader analysis , in @xcite . compared to these earlier publications , we rectify some minor problems mostly in the analysis of the classical dynamical system , but our main contribution will be related to the construction of a field algebra of observables for the field strength . in this endeavour we will prove in particular that the commutator between two generators of the algebra is given by the lichnerowicz propagator @xcite regardless of the chosen spacetime . this allows us also to make a direct connection with an old result of ashtekar and sen @xcite , who identified the existence of a two - parameter family of unitary inequivalent representations of the canonical commutation relations for the field strength on schwarzschild spacetime . in our language this translates in the existence of a non trivial centre for the field algebra whenever the second de rham cohomology group of the manifold , either with real or with complex coefficients , is non trivial . as a last point we can address the question whether the field strength tensor can be described as a locally covariant quantum field theory . first introduced in @xcite , the so - called principle of general local covariance was formulated leading to the realization of a quantum field theory as a covariant functor between the category of globally hyperbolic ( four - dimensional ) lorentzian manifolds with isometric embeddings as morphisms and the category of @xmath11-algebras with injective homomorphisms as morphisms . already shown to hold true for scalars @xcite , spinors @xcite and recently for the proca field and for the vector potential ( though in this case it has been assumed that the first de rham cohomology group of the underlying background is trivial ) @xcite , such paradigm turns out not to hold true in the case of a quantum field theory based on the field strength . although we will be more explicit in the main body of the paper , we stress that the obstruction is related to a potential clash between the presence of a non trivial centre for the field algebra of @xmath0 in a globally hyperbolic spacetime @xmath2 and the existence of isometric embeddings of @xmath2 into backgrounds @xmath12 with trivial second de rham cohomology group . as a potential way out , we show that general local covariance can be restored if the category of admissible spacetimes is suitably reduced , although , as we shall comment later in detail , this has certainly far reaching physical consequences . .2 cm the paper will be organized as follows : in section 1.1 we will recollect the notations and conventions we shall use throughout the paper . in section 2 we will instead discuss maxwell s equations and the associated initial value problem , showing that it is well - defined and that the space of solutions can be constructed also in this case with the help of the causal propagator for a suitable second order hyperbolic differential operator . section 3 will be instead entirely devoted to the issue of constructing the associated field algebra and , in particular , we shall prove that the commutator between two generators can be computed via the lichnerowicz propagator . in section 3.1 we shall show that , whenever certain topological invariants of the background are not trivial , the field algebra possesses a non trivial centre and we fully characterize its elements also providing explicit examples . in section 3.2 we tackle the problem whether the principle of general local covariance holds true for the field strength finding in general a negative answer unless the class of admissible spacetimes is reduced . in section 4 we draw some conclusions . in this paper , each background will always be referred to as a _ `` spacetime '' _ , that is a four dimensional differentiable , second countable , hausdorff manifold @xmath2 with a lorentzian metric @xmath13 whose signature is @xmath14 . we shall also assume that @xmath2 is _ globally hyperbolic _ , hence there exists a closed achronal subset @xmath15 whose domain of dependence coincides with @xmath2 itself . note , that a spacetime is usually assumed to be connected in the literature . yet , in this paper and with respect to the construction of the field algebra in particular , we are also interested in disconnected manifolds with finitely many connected components . see in particular the remark after proposition [ universal algebra properties ] . note , that there is no ambiguity in carrying over established results from the connected case to the disconnected case with finitely many connected components . on account of standard results in differential geometry and of the recent analysis in @xcite and of theorem 1.1 in @xcite , global hyperbolicity entails that there exists an isometry @xmath16 between @xmath2 and a smooth product manifold @xmath17 . thus @xmath18 turns out to be a three - dimensional embedded submanifold and theorem 2.1 in @xcite guarantees , moreover , that @xmath19 splits as @xmath20 where @xmath21 is a temporal function , @xmath22 while @xmath23 induces for fixed values of @xmath24 a smooth riemannian metric on @xmath18 . furthermore global hyperbolicity yields that @xmath2 admits an orientation and thus , henceforth , we assume that a choice has been done and all spacetimes are globally hyperbolic as well as time oriented and oriented . on top of the geometric structure we shall consider @xmath25 and @xmath26 , respectively the set of smooth and of smooth and compactly supported @xmath27-forms on @xmath2 with values in the field @xmath28 , usually chosen either as @xmath29 or @xmath30 . here @xmath31 and @xmath32 , where the parenthesis around the subscript indicates that the statement holds true both with and without the subscript itself . let @xmath28 be the complex numbers ; then , on these spaces , one can define two natural operators , the external derivative @xmath33 and the hodge dual @xmath34 . notice that , while @xmath6 is completely independent from @xmath13 , @xmath35 is instead a function of the underlying metric . furthermore , since @xmath35 is invertible , we can introduce a third operator , known as the codifferential @xmath36 . in the main body of the paper we will be often interested in compactly supported smooth forms which are either closed or coclosed and to avoid to be redundant in the exposition we introduce the following novel notation : @xmath37 to conclude , we mention two further ingredients we shall need in the forthcoming discussion . the first is @xmath38 which is the @xmath27-th de rham cohomology group of @xmath2 see @xcite for the definition and for a recollection of the main properties . it is noteworthy that , since such groups are built only out of the external derivative , they are completely independent from the underlying geometry and from @xmath13 in particular . we can combine together @xmath6 and @xmath7 to define the laplace - de rham operator @xmath39 . the second ingredient is instead @xmath40 which stands for the @xmath27-th smooth singular homology group of the manifold and whose main properties are discussed in @xcite . as stated in the introduction , the main objective of this paper is to shed some light on the classical and on the quantum structure of maxwell s equations on curved backgrounds , emphasizing in particular how the underlying topology affects the qualitative behaviour of the system . to start with , we need to introduce the key objects of our analysis : the curved spacetime analogue of maxwell s equations sees @xmath3 as the dynamical variable and the dynamics is ruled by @xmath41 where @xmath42 is the external current such that @xmath43 . a key property of when defined on a generic spacetime @xmath44 lies in the analysis of the first identity . this is a constraint on the form of @xmath0 which usually leads to state both that there exists @xmath9 such that @xmath10 and that one can consider @xmath1 , the so - called vector potential as the underlying dynamical field . this statement is based on the poincar lemma which , alas , can not be always applied since it fails to hold true whenever @xmath45 is not trivial . in this particular case , it turns out that there exist classical field strengths which can not be derived as the external derivative of a suitable one - form . since , from a physical point of view , it is @xmath0 the observable field of the dynamical system , it is natural to wonder whether a full classical and quantum analysis of could be performed without making use of any auxiliary structure such as the vector potential . in order to grasp the classical behaviour of a dynamical system ruled by , we need to prove that this set of equations admits a well - defined initial value problem on every globally hyperbolic spacetime . despite the apparent obviousness of this question , to the best of our knowledge it turns out that this problem has been only partly treated in details and the discussions available in the literature are either partly incomplete or based upon further restrictive assumptions , such as the compactness of the cauchy surface @xmath18 see @xcite , but also @xcite although they work with the vector potential . on the opposite , since we want to cope with the most general scenario , we need the following statement see also @xcite for a similar analysis : [ cauchy problem field strength tensor ] let @xmath46 be an oriented 4-dimensional globally hyperbolic spacetime with at most finitely many connected components whose smooth spacelike cauchy surface is @xmath18 with smooth embedding @xmath47 . then , for each triple @xmath48 such that @xmath49 , @xmath50 with @xmath51 and @xmath52 , there exists a unique solution @xmath53 of the initial value problem as @xmath54 in place of @xmath55 . furthermore we indicate with @xmath56 the hodge dual induced on the cauchy surface @xmath18 to distinguish it from the one on @xmath2 . ] @xmath57 furthermore , @xmath0 depends linearly and continuously on both the source term @xmath58 , and on the initial data @xmath59 , @xmath60 . each solution also enjoys the following support property : @xmath61 where @xmath62 are the causal future and past of @xmath63 respectively . we start with the connected case . since @xmath64 , one can realize by direct inspection that every solution of also solves @xmath65 . yet , in order to use the latter as a starting point to solve maxwell s equations , we need to prove that we can always select suitable initial data for the wave equation so that a solution of the latter yields also one of . to this avail , let us consider @xmath66 where @xmath54 here is the pull - back induced from @xmath47 on the compactly supported sections of the exterior bundle on @xmath2 . in other words @xmath67 and @xmath68 are maps from @xmath18 into @xmath69 such that @xmath70 on account of @xmath2 being isometric to @xmath17 with line element @xmath71 as outlined in section [ notations ] , here @xmath72 is a coordinate patch of @xmath2 adapted to this last metric . it intersects @xmath18 on a non empty open set and it is endowed with a local chart @xmath73 , whereas @xmath74 is the unit normal vector to @xmath18 . hence , the cauchy problem @xmath75 where both @xmath67 and @xmath68 are chosen as in the previously displayed set of two equations , admits a unique solution @xmath76 which , furthermore , on account of ( * ? ? ? * thm.3.2.11 ) , depends linearly and continuously both on the source term and on the initial data @xmath67 , @xmath68 . at the same time it holds @xmath77 where @xmath78 , which , in turn , entails the sought support property . it remains to be shown that the obtained solution @xmath0 of the cauchy problem for the wave equation solves as well . to achieve this , it suffices to show that @xmath0 also satisfies @xmath79 with vanishing initial data . since @xmath80=[\box,\delta]=0 $ ] , the two equations automatically descend from @xmath65 and thus only the initial data have to be checked . it suffices to show it in @xmath81 . from @xmath82 and @xmath83 it follows directly @xmath84 , whereas , from @xmath85 , it descends @xmath86 ; hence @xmath87 . equivalently @xmath88 yields @xmath89 . notice that @xmath90 is a by - product of both @xmath51 and @xmath91 ; thus @xmath92 . the remaining initial condition @xmath93 arises out of @xmath65 , of the properties of @xmath94 $ ] and of the symmetries of the riemannian curvature tensor see @xcite . hence , on account ( * ? ? ? * cor.3.2.4 ) , this suffices for @xmath4 to hold true on @xmath2 . to conclude , @xmath95 is a result of @xmath4 , @xmath65 and of the conservation of the current @xmath43 . as above this suffices to prove @xmath96 on @xmath2 . to conclude , in order to establish that the solution of , it suffices to suppose that there exists two different solutions , say @xmath0 and @xmath97 . their difference @xmath98 must satisfy @xmath99 and @xmath100 with vanishing initial data on the cauchy surface @xmath18 . this entails that @xmath101 must also satisfy @xmath102 with vanishing inital data on @xmath18 and , according to standard results of the theory of partial differential equations , this holds true only if @xmath103 , hence @xmath104 . if @xmath2 is disconnected with finitely many connected components @xmath105 , @xmath106 , we consider the partition of unity @xmath107 subordinated to @xmath108 such that @xmath109 and @xmath110 for @xmath111 . since @xmath44 is globally hyperbolic , so is @xmath112 , with smooth spacelike cauchy surface @xmath113 ; we obtain , therefore , the initial value problem @xmath114 we solve each of these initial value problems as prescribed in the connected case , therefore obtaining unique solutions @xmath115 . the unique solution @xmath0 is then assembled via the partition of unity , @xmath116 . as a by - product of this last proposition , we can construct the solutions of maxwell s equations on a globally hyperbolic spacetime starting from the wave equation . if we focus on the source free case , that is @xmath117 , we can generate all solutions of @xmath118 with compactly supported initial data as @xmath119 where @xmath120 and where @xmath121 is the causal propagator @xcite . here @xmath122 are the uniquely defined advanced and retarded green operators such that @xmath123 and @xmath124 , for all @xmath120 . notice that these properties of @xmath125 also entail that every compactly supported smooth solution of @xmath118 must vanish identically . an additional noteworthy property of the causal propagator associated to the laplace - de rham operator originates from its structure and from the fact that @xmath126 intertwines between the codifferential operator @xmath7 acting on @xmath27 and on @xmath127-forms , that is @xmath128 . to wit , at a level of solutions of the corresponding wave equation with smooth and compactly supported initial data , this entails that , if @xmath129 with @xmath120 , then @xmath130 . the very same properties hold with respect to the exterior derivative @xmath6 . since not all @xmath131 fulfil also the source free maxwell s equations , one needs to impose some further constraints on the set of initial test functions @xmath132 in order to take into account only the two - forms solving . the following proposition amends this deficiency : a smooth complex @xmath133-form @xmath0 is a solution of with @xmath117 and with compactly supported smooth initial data if and only if there exist @xmath134 and @xmath135 such that @xmath136 . `` @xmath137 '' since @xmath138 and @xmath139 are of compact support and since g commutes with @xmath6 and @xmath7 , it holds that @xmath140 and that @xmath141 . furthermore , on account of the properties of the causal propagator , it is also guaranteed that the initial data of maxwell s equations associated to @xmath142 are smooth , compactly supported and their form fulfils the constraints of . + `` @xmath143 '' since @xmath4 and @xmath5 entail @xmath118 , there must exist @xmath120 such that @xmath119 . furthermore , @xmath144 and @xmath145 entail the existence of @xmath134 and @xmath135 satisfying @xmath146 and @xmath147 . on account of the nilpotency of both @xmath6 and @xmath7 , it holds @xmath148 and @xmath149 which suffices to conclude that @xmath150 , @xmath138 and @xmath139 being compactly supported . the same reasoning entails that the following chain of identities @xmath151 yields @xmath152 , up to an irrelevant sign the sought result . the full control of the classical dynamics of maxwell s equations allows us to address the problem of quantising a field theory with @xmath0 as the main ingredient . as it is customary in the algebraic approach , this is a two - step procedure , the first calling for the identification of a suitable algebra of observables and the second requiring the assignment of a state to represent such an algebra in terms of operators on a suitable hilbert space . in this paper we will focus on the first part of the programme , hence we shall construct the full field algebra and investigate its properties . in the process we will benefit from ideas which first appeared in earlier works @xcite and @xcite ; the sketch of the construction is the following : first we consider a suitable covering of @xmath44 in globally hyperbolic submanifolds @xmath153 , @xmath154 where @xmath155 is a set , which we will specify below . afterwards , we construct the local field algebras @xmath156 of the field strength tensor , whereas the global one is defined as the universal algebra @xmath157 associated to the local algebras @xmath156 . the commutation relations encoded in @xmath157 will turn out to be given by the lichnerowicz s commutator and the algebra itself will have all the properties required to deserve the name _ global _ " field algebra . one could wonder why it is necessary to go through such an involute construction . there are many conceptual reasons but it is noteworthy that the form of the commutator is in principle only known for spacetimes with certain topological restrictions and thus we need to show that a generalization to a more generic spacetime exists . [ universal algebra ] let @xmath155 be a set and @xmath158 a family of unital @xmath35-algebras together with linking unital @xmath35-homomorphisms @xmath159 where the admissible pairs @xmath160 are a suitable subset of @xmath161 . furthermore the compatibility condition @xmath162 holds true whenever @xmath163 and @xmath164 are defined . then , there exists a unique ( up to @xmath35-isomorphism ) unital @xmath35-algebra @xmath165 together with a family of unital @xmath35-homomorphisms @xmath166 such that the following universal property holds true : ( uva ) : : for each unital @xmath35-algebra @xmath60 and for each family @xmath167 of unital @xmath35-homomorphisms such that @xmath168 whenever @xmath164 exists , there exists a unique unital @xmath35-homomorphism @xmath169 which satisfies @xmath170 for all @xmath154 . the pair @xmath171 is called the * universal algebra * of @xmath172 . to begin with , we regard the unital @xmath35-algebras @xmath173 , @xmath154 , as complex vector spaces and construct the associative tensor algebra @xmath174 of their direct sum . with componentwise addition , componentwise multiplication with a scalar , componentwise antilinear involution @xmath35 and multiplication induced by the algebraic tensor product @xmath175 , @xmath174 becomes a unital @xmath35-algebra . next , we consider the two - sided @xmath35-ideal @xmath176 generated by those elements of the form @xmath177 @xmath178 and for all given @xmath164 . @xmath179 denotes the vector in @xmath180 for which every entry is zero except the @xmath58-th one which is precisely @xmath181 . we denote the equivalence class of an element @xmath182 with respect to that quotient by @xmath183 $ ] . now , define @xmath184 and for @xmath185 @xmath186 . \end{aligned}\ ] ] @xmath165 defined in this way is a unital @xmath35-algebra and , per direct inspection , @xmath187 turns out to be a well - defined unital @xmath35-homomorphism for all @xmath154 . the pair @xmath171 satisfies the universal property . to wit , let @xmath60 be any arbitrary unital @xmath35-algebra and @xmath167 a family of unital @xmath35-homomorphisms such that @xmath168 , whenever @xmath164 exists . since any element @xmath183\in a_u$ ] can be written as @xmath188=\underset{n\in\mathbb{n}}{\sum}t_n^{k_1\dots k_n}\underset{l=1}{\overset{n}{\prod}}\underset{j\in i}{\sum}\big[0_\mathbb{c},(a_{k_l}^i)^j_{i\in i } , 0_{(\bigoplus_{i\in i}a_i)^{\otimes 2}},\dots\big],\end{aligned}\ ] ] @xmath189 , by the structure of @xmath174 , a unital @xmath35-homomorphism @xmath169 such that @xmath170 for all @xmath154 is uniquely fixed by @xmath190)=\underset{n\in\mathbb{n}}{\sum}t_n^{k_1\dots k_n}\underset{l=1}{\overset{n}{\prod}}\underset{j\in i}{\sum}\phi_j(a^j_{k_l}).\end{aligned}\ ] ] this shows the existence of the universal algebra . let @xmath191 be another pair consisting of a unital @xmath35-algebra and unital @xmath35-homomorphisms having the universal property ( uva ) . thus we have a unique unital @xmath35-homomorphism @xmath192 fulfilling @xmath193 for all @xmath154 . according to ( uva ) , @xmath194 is the unique unital @xmath35-homomorphism such that @xmath195 for all @xmath154 . @xmath196 and @xmath197 for all @xmath154 . however , since @xmath171 has the universal property , thus the unital @xmath35-homomorphism @xmath198 is unique , and @xmath199 satisfies @xmath200 for all @xmath154 as well , necessarily it holds that @xmath201 . @xmath202 and @xmath203 for all @xmath154 . in the same manner , since @xmath191 has the universal property , the unital @xmath35-homomorphism @xmath204 is unique and @xmath205 satisfies @xmath206 for all @xmath154 , @xmath207 . thereby @xmath165 and @xmath60 are isomorphic via a unital @xmath35-isomorphism . let us recall that every connected component @xmath208 , @xmath209 for an @xmath106 , of an oriented 4-dimensional globally hyperbolic spacetime @xmath46 with at most finitely many connected components will turn into a connected , oriented , 4-dimensional , globally hyperbolic , embedded subspacetime , if endowed with the structures induced by @xmath46 . consequently any of these @xmath208 can be foliated up to isometries as @xmath210 . here @xmath211 being a smooth spacelike cauchy surface endowed with the natural structures inherited from @xmath212 , @xmath213 is a smooth spacelike cauchy surface of @xmath46 . therefore each @xmath214 lies precisely in one connected component @xmath208 and at least on one of such surfaces , which we denote by @xmath215 , and we can always construct an open subset @xmath216 that is either contractible or disconnected with finitely many contractible connected components . the net advantage is that its associated cauchy development @xmath217 is in turn a contractible open subset or a disconnected open subset of @xmath2 with finitely many contractible connected components that will become an oriented , 4-dimensional , globally hyperbolic , embedded subspacetime , if equipped with the structures induced by @xmath46 . since this procedure can be repeated for all points of the manifold , we can always cover @xmath2 with contractible open subsets and disconnected open subsets with finitely many contractible connected components such that these open subsets will become oriented , 4-dimensional , globally hyperbolic , embedded subspacetimes , if endowed with the structures induced by @xmath46 . but not any such cover will do the trick . we need a very specific cover , namely @xmath218 that of all contractible open subsets of @xmath2 and all disconnected open subsets of @xmath2 with finitely many contractible connected components such that @xmath219 becomes an oriented , 4-dimensional , globally hyperbolic , embedded subspacetime for all @xmath154 , if endowed with the structures induced by @xmath46 and such that in addition the image of the inclusion @xmath220 is causally convex . the endpoint are oriented 4-dimensional globally hyperbolic spacetimes with at most finitely many contractible connected components , which we denote as @xmath153 , @xmath154 . @xmath155 is actually a set because this cover @xmath221 is contained in the power set of @xmath2 . for each @xmath154 , we associate to the oriented 4-dimensional globally hyperbolic spacetime @xmath153 with at most finitely many contractible connected components the local field algebra @xmath156 of @xmath0 . notice that , since @xmath219 has at most finitely many contractible connected components contractible , the first equation in entails via poincar lemma that @xmath10 where @xmath222 . also , there is no ambiguity in constructing the field algebra for a disconnected spacetime with finitely many contractible connected components in the same manner as it is done in the connected contractible case . [ fieldal ] we call the field algebra of the field strength tensor on an oriented 4-dimensional globally hyperbolic spacetime @xmath44 with at most finitely many contractible connected components , @xmath223 , the unital @xmath35-algebra generated by the elements @xmath224 with @xmath225 together with the defining relations @xmath226=i\,(\int_m g\delta\omega\wedge*\delta\omega')\,1_{\mathfrak{f}(m)},\enspace\forall\omega , \omega'\in\omega^2_0(m , \mathbb{c}),\\ \end{array}\ ] ] where @xmath227 is the causal propagator associated to the @xmath228-operator and @xmath229 is the identity element of the algebra . the @xmath35-operation is the complex conjugation . we remark that , in the above definition , the first two conditions entail the fulfillment of the equations of motion and the equalities are meant in a distributional sense , _ i.e. _ , @xmath230 and similarly for _ eom 2)_. the form of the commutator descends from earlier analyses , see in particular @xcite . notice also that isotony is automatically implemented , that is , for given @xmath231 and @xmath232 with @xmath233 and @xmath234 , then @xmath235 . in other words there always exists an injective unital @xmath35-homomorphism of algebras @xmath236 , realised by @xmath237 with the help of the inclusion @xmath238 and subject to the additional compatibility condition @xmath239 , whenever we consider three oriented 4-dimensional globally hyperbolic spacetimes with at most finitely many contractible connected components such that @xmath240 . consequently , we obtain a family @xmath241 of unital @xmath35-algebras together with linking unital @xmath35-homomorphisms @xmath242 for @xmath243 that meet the compatibility condition @xmath162 whenever @xmath244 . since the system of unital @xmath35-algebras together with the unital @xmath35-homomorphisms as specified before satisfies the conditions of theorem [ universal algebra ] , its application yields the universal algebra @xmath245 . we define the global field algebra of the field strength tensor of an arbitrary oriented 4-dimensional globally hyperbolic spacetime @xmath44 with at most finitely many connected components to be the unital @xmath35-algebra @xmath157 . its properties are clarified in the following statements : @xmath157 satisfies the local compatibility condition @xmath246 whenever @xmath225 such that @xmath247 . since @xmath44 is globally hyperbolic , its standard topology coincides with the alexandrov one , i.e. that which has the diamonds @xmath248 as its basis . as a result of that , there exists for every point @xmath249 a contractible diamond @xmath250 containing @xmath251 and @xmath252 , in particular @xmath253 . with the structures induced by @xmath44 these diamonds @xmath250 become 4-dimensional oriented globally hyperbolic embedded subspacetimes and therefore belong to our chosen cover @xmath218 . let @xmath254 be a partition of unity subordinated to that open cover @xmath255 . since a partition of unity is locally finite , all appearing sums are actually finite and we can compute @xmath256 which is the sought result . [ universal algebra properties ] in @xmath223 , we can define _ global _ smeared field strength operators @xmath257 for all @xmath225 such that 1 . @xmath223 is generated by the global smeared field strength operators @xmath257 , @xmath120 , 2 . @xmath258 fulfils maxwell s equations in a weak sense , i.e. eom 1 ) @xmath259 for all @xmath260 and eom 2 ) @xmath261 for all @xmath262 , 3 . @xmath263 for all @xmath264 , for all @xmath265 ( linearity ) , 4 . @xmath266 for all @xmath120 ( hermicity ) . furthermore , @xmath223 obeys the principle of locality , i.e. @xmath267=0 $ ] for all @xmath268 that are spacelike separated . choose any partition of unity @xmath269 subordinated to the cover @xmath218 and define @xmath270 first of all , the sum is finite because the partition of unity is locally finite and @xmath132 is of compact support . secondly , this definition does not depend on the chosen partition of unity for let @xmath271 be another partition of unity , then @xmath272 where we applied the local compatibility of @xmath223 , i.e. the foregoing lemma . note , that this implies another local compatibility property for @xmath223 , namely let @xmath225 be completely contained in one @xmath273 for a @xmath185 , i.e. @xmath274 , then @xmath275 linearity , hermicity and maxwell s equations in a weak sense follow from this definition and from their implementation at a level of local field algebras @xmath156 , @xmath154 . to show locality , we start with a slightly simpler statement . let @xmath225 and @xmath276 be spacelike separated and @xmath277 for a @xmath154 and @xmath278 for a @xmath185 such that @xmath219 and @xmath273 are spacelike separated , i.e. @xmath279 . @xmath280 can be regarded as an oriented , 4-dimensional , globally hyperbolic , embedded submanifold of @xmath44 with finitely many contractible connected components . hence , we have an injective unital @xmath35-homomorphism @xmath281 for @xmath282 . we find that @xmath283&=\big[\alpha_{i\sqcup j}\big(\widehat{\textbf{f}}_{i\sqcup j}(\iota^*_{i\sqcup j}\omega)\big ) , \alpha_{i\sqcup j}\big(\widehat{\textbf{f}}_{i\sqcup j}(\iota^*_{i\sqcup j}\omega')\big)\big]\\ & = \alpha_{i\sqcup j}\big[\widehat{\textbf{f}}_{i\sqcup j}(\iota^*_{i\sqcup j}\omega ) , \widehat{\textbf{f}}_{i\sqcup j}(\iota^*_{i\sqcup j}\omega')\big]\\ & = \alpha_{i\sqcup j}\big(i\int_{m_i\sqcup m_j}g\delta\iota_{i\sqcup j}^*\omega\wedge*\delta\iota_{i\sqcup j}^*\omega'\enspace1_{\mathfrak{f}(m_i\sqcup m_j)}\big)\\ & = 0\end{aligned}\ ] ] because @xmath132 and @xmath284 are spacelike separated and , therefore , so are @xmath285 and @xmath286 . now , let @xmath268 be spacelike separated without further restrictions or assumptions . since the topology of @xmath44 coincides with the alexandrov one and since @xmath132 and @xmath284 being spacelike separated implies @xmath287 in particular , we can cover @xmath288 with diamonds @xmath250 , @xmath289 , and @xmath290 with diamonds @xmath291 , @xmath292 in such a way that @xmath293 for all pairs @xmath294 ( smooth manifolds are @xmath295 ) . these diamonds will turn into contractible 4-dimensional oriented globally hyperbolic spacetimes if we endow them with the structures induced by @xmath44 . let @xmath296 be a partition of unity subordinated to @xmath297 and @xmath298 be a partition of unity subordinated to @xmath299 respectively . on account of the previous results , it holds @xmath283&=\big[\widehat{\textbf{f}}_u(\underset{x}{\sum}\psi^x\omega ) , \widehat{\textbf{f}}_u(\underset{y}{\sum}\varphi^y\omega')\big]=\big[\widehat{\textbf{f}}_u(\underset{x}{\sum}\omega^x ) , \widehat{\textbf{f}}_u(\underset{y}{\sum}\omega'^y)\big]\\ & = \underset{x , y}{\sum}\big[\widehat{\textbf{f}}_u(\omega^x ) , \widehat{\textbf{f}}_u(\omega'^y)\big]\\ & = 0,\end{aligned}\ ] ] since @xmath300 are pairwise spacelike separated and @xmath301 and @xmath302 . note that the sums are finite because the partition of unity is locally finite . the proof of @xmath157 obeying the principle of locality makes it clear why we wanted to consider disconnected spacetimes with finitely many connected contractible components in the first place . what remains to be shown is our claim that the commutator in @xmath157 is given by the lichnerowicz commutator : [ propagator ] the commutator between two algebra elements in @xmath157 is provided by the so - called lichnerowicz commutator @xcite @xmath303=i\,(\int_mg\delta\omega\wedge*\delta\omega')\,1_{\mathfrak{f}_u(m)}\enspace\forall\omega,\omega'\in\omega^2_0(m , \mathbb{c}),\end{aligned}\ ] ] where @xmath304 is the identity element of the universal algebra and @xmath227 the causal propagator of the @xmath228-operator . we sketch here the main steps of the proof pointing an interested reader to @xcite for the details of some lengthy albeit straightforward computations . choose a cauchy surface @xmath305 in the future of @xmath306 and @xmath307 and consider the compact set @xmath308 . cover @xmath309 with finitely many contractible open subsets @xmath310 , @xmath311 of @xmath305 whose cauchy developments will be called @xmath312 . without loss of generality , all @xmath312 belong to the chosen cover of @xmath2 . let @xmath313 be a finite refinement of @xmath310 such that @xmath314 such a refinement exists , because , whenever ( [ refinement ] ) is not fulfilled by two sets @xmath313 and @xmath315 , we can replace them with finitely many other sets satisfying such condition and all other constraints of our construction . construct the cauchy developments @xmath316 of each set @xmath313 in the refinement ; automatically it holds that there exists @xmath317 , such that @xmath318 . if @xmath319 then @xmath316 and @xmath320 are spacelike separated . if instead @xmath321 then there exists an @xmath317 such that @xmath322 . since there exist finitely many sets @xmath316 , we can always find a spacelike cauchy surface @xmath323 in the future of @xmath305 and a spacelike cauchy surface @xmath324 in the past such that @xmath325 , @xmath326 , @xmath327 and @xmath328 are all contained in @xmath329 . now , let @xmath330 be chosen in such a way that @xmath331 and @xmath332 is identically @xmath333 in @xmath334 and @xmath335 in @xmath336 . consider @xmath337 and @xmath338 . on account of the construction and of the properties of @xmath339 , @xmath340 , @xmath341 and @xmath342 are compactly supported and their supports lie in @xmath343 . choose a partition of unity @xmath344 belonging to @xmath345 . on account of the properties eom 1 ) and eom 2 ) , @xmath346 and @xmath347 . hence @xmath348=\sum_{k , k'}\big[\widehat{\bf{f}}(\psi_k\tilde{\omega } ) , \widehat{\bf{f}}(\psi_{k'}\tilde{\omega}')\big]=\sum_{k\sim k'}\big[\widehat{\bf{f}}(\psi_k\tilde{\omega } ) , \widehat{\bf{f}}(\psi_{k'}\tilde{\omega}')\big],\ ] ] where @xmath349 means that we consider only the pairs @xmath350 such that @xmath351 and where in the last equality we used that spacelike separated observables do commute . hence @xmath352 = \sum_{k\sim k'}i\,\left(\int_mg\delta(\psi_k\tilde{\omega})\wedge*\delta(\psi_{k'}\tilde{\omega}')\right)\,1_{\mathfrak{f}_u(m)}=\\ = \sum_{k , k'}i\,\left(\int_mg\delta(\psi_k\tilde{\omega})\wedge*\delta(\psi_{k'}\tilde{\omega}')\right)\,1_{\mathfrak{f}_u(m ) } = i\,\left(\int_mg\delta\tilde{\omega}\wedge*\delta\tilde{\omega}'\right)\,1_{\mathfrak{f}_u(m)}\end{aligned}\ ] ] @xmath353 where , in the second equality we consider all possible values for @xmath354 and @xmath355 since the additional ones contribute @xmath335 to the integral . notice that in the various identities we used the fact that all test functions are compactly supported , that @xmath227 commutes with both @xmath6 and @xmath7 and that all sums are over a finite set of indices . we have shown that @xmath157 enjoys all the properties wanted by a genuine algebra of observables ; hence we can start investigating its additional features . to start with , [ time slice axiom ] the universal algebra @xmath157 satisfies the time slice axiom , that is , if @xmath18 is a cauchy surface of @xmath46 and @xmath356 a globally hyperbolic subset of @xmath2 containing @xmath18 , it holds that @xmath357 . let @xmath358 be an open neighbourhood of @xmath18 . it is sufficient to show that for every @xmath120 there exists a @xmath276 with @xmath359 such that @xmath360 . since @xmath358 is an open neighbourhood of @xmath18 and @xmath361 is compact , there exist cauchy surfaces @xmath362 and @xmath363 respectively in the future and in the past of @xmath18 such that @xmath364 and @xmath365 . let @xmath366 lie in @xmath367 and let us fix them in such a way that @xmath368 and that @xmath332 vanishes in the past of @xmath363 , whereas it is equal to @xmath333 in the future of @xmath362 . then , if we define @xmath369 it holds that @xmath359 is compact due to the properties of @xmath370 and @xmath371 . furthermore , from the conditions eom 1 ) and eom 2 ) on @xmath157 , it follows that @xmath360 . the aim of this subsection is to investigate a distinguishing aspect of the universal algebra , namely the appearance of new features which have no counterpart in the local algebras , above dubbed as @xmath372 . from a technical point of view , this translates in the existence of a non trivial centre in @xmath157 , that is there exists a non trivial subalgebra whose elements are commuting with all those of the universal algebra . yet we want to stress that this happens only if the topology of the underlying background is rather peculiar , namely if @xmath373 . if , on the contrary , the second de rham cohomology group is trivial , then the equation @xmath4 in entails the existence of a global one - form @xmath1 such that @xmath10 . in this case the field algebra of the field strength tensor could be globally defined as the differential of that of the vector potential and no non - trivial centre would appear . therefore , we will henceforth assume that @xmath373 and , with the next lemma , we show how to characterize the elements of the centre of @xmath157 . [ centre ] an algebra element @xmath374 lies in the centre of @xmath157 if and only if @xmath375 with @xmath376 and @xmath377 . @xmath374 is in the centre of @xmath157 if and only if @xmath378=\int_m\delta g\omega\wedge*\delta\omega'=\int_md\delta g\omega\wedge*\omega'=0 $ ] for all @xmath276 . since @xmath69 comes endowed with the non - degenerate scalar product @xmath379 , then the commutator between @xmath374 and @xmath380 vanishes if and only if @xmath381 . in turn , this last equality holds if and only if @xmath382 and @xmath383 , @xmath384 . we can exploit the properties of the green s functions to conclude that the following chains of identities hold @xmath385 and equivalently @xmath386 . furthermore , it holds true that @xmath387 . notice that the proposition guarantees that the centre is trivial if and only if @xmath388 since , in this case , the closedness of @xmath138 and the coclosedness of @xmath139 would guarantee the existence of @xmath389 and of @xmath390 such that @xmath391 . under this assumption , on account of @xmath392 and of @xmath393 for @xmath157 , the field strength operator vanishes . in order to better understand this feature , it is worth to construct explicitly non trivial elements of the centre whenever @xmath394 , the latter bound being assumed only for the sake of simplicity . notice that in the forthcoming analysis we will work with real forms , thus dropping the reference to @xmath30 ; this does not clash with the previous results and it is assumed still only for the sake of simplicity . out of the non - degenerateness of the scalar product on @xmath45 , @xmath2 being four dimensional , and out of poincar duality , ( * ? ? ? 1 ) , the following chain of isomorphisms holds true : @xmath395 where the subscript @xmath396 here stands for compact support . notice that , in the first and in the third isomorphism , the hypothesis of @xmath45 being finite dimensional plays a key role . to wit both @xmath45 and @xmath397 are finite - dimensional vector spaces and hence isomorphic to their dual . consequently every element @xmath398 of @xmath399 can be represented as @xmath400\longmapsto \lambda([f])=\int_m f\wedge\eta.\end{aligned}\ ] ] notice that the symbol @xmath401 $ ] to indicate an equivalence class in @xmath45 has been chosen for a notational reason which will be manifest in the forthcoming discussion . furthermore , on the right hand side , @xmath0 is an arbitrary representative of @xmath401 $ ] as well as @xmath402 is an arbitrary representative of a unique equivalence class @xmath403\in h^2_c(m)$ ] . by direct inspection , one can realize that the integral does not depend on the various choices . since every @xmath404\in h^\infty_2(m)$ ] defines a linear map @xmath405 , there exists a unique @xmath406\in h^2_c(m)$ ] such that @xmath407\in h^2(m ) \end{aligned}\ ] ] where all formulas are independent from the choice of a representative in the various equivalence classes . we can interpret the above remarks as follows : on account of the hypothesis @xmath373 , there exists at least an equivalence class of non - exact field strength tensor @xmath401 $ ] . as a result of that , there exists @xmath404\in h^\infty_2(m)$ ] and @xmath406\in h^2_c(m)$ ] fulfilling regardless of the chosen representative @xmath408 hence we have constructed a classical field strength @xmath0 whose associated algebra element @xmath409 is a non - trivial element of the centre which can be interpreted as the magnetic flux through the 2-cycle @xmath410 . the very same discussion holds true also for @xmath411 in place of @xmath411 because of @xmath412 for all @xmath401\in h^2(m)$ ] . from a physical point of view @xmath413 can be interpreted as the electric flux through @xmath410 . we would like to draw the attention to the fact that these non - trivial elements of the algebra give rise to superselection sectors as discussed in @xcite . as the very last point of our investigation on the algebra of observables for the free maxwell field , we address the question whether it defines a local covariant quantum field theory as per definition 2.1 in @xcite . in this section we shall use both the terminology and the nomenclature of this last cited paper ; we refer to it for an extensive analysis and here we recollect instead just the definition of the main ingredients we need : * @xmath414 : the category whose objects are @xmath44 , that is four dimensional oriented and time oriented globally hyperbolic spacetimes with at most finitely many connected components , endowed with a smooth metric of signature @xmath14 . a morphism between two objects @xmath44 and @xmath415 is a smooth embedding @xmath416 such that @xmath417 is causally convex of a globally hyperbolic spacetime is called _ causally convex _ if @xmath418 all causal curves connecting @xmath251 to @xmath419 lie entirely inside @xmath356 . ] , preserves orientation and time orientation and @xmath420 on @xmath2 . * @xmath421 : the subcategory of @xmath414 whose objects are those @xmath422 and @xmath423 . a morphism between two objects @xmath44 and @xmath415 is a smooth embedding @xmath416 such that @xmath417 is causally convex , preserves orientation and time orientation and @xmath420 on @xmath2 . notice that , since @xmath417 is diffeomorphic to @xmath2 , its cohomology groups are isomorphic to those of @xmath2 ( * ? ? ? * @xmath424 : the category whose objects are unital @xmath35-algebras whereas morphisms are injective unit - preserving @xmath35-homomorphisms . since the composition map between morphisms and the existence of an identity map are straightforwardly defined in every case we shall consider , we will omit them . we shall start proving a weaker form of general local covariance , where the class of spacetimes we consider is not the most general one . we wish to postpone the explanation for this choice to after the proof of the following proposition since we feel that reading it will make our point clearer than an abstract a priori argument . there exists a covariant function @xmath425 which assigns to every object @xmath44 in @xmath421 the @xmath35-algebra @xmath157 with the induced action on the morphisms . in diagrammatic form : @xmath426 here @xmath427 is the unit - preserving @xmath35-homomorphism defined by its action on the generators as @xmath428 where @xmath429 is the pull - back of @xmath132 via @xmath430 . furthermore , such local covariant quantum field theory is causal and it fulfils the time slice axiom . as discussed at the beginning of the session , we can associate to each @xmath422 the universal algebra along the lines of the previous section . hence , if we consider any morphism @xmath431 between two objects @xmath44 and @xmath415 , we can consider @xmath432 as a globally hyperbolic spacetime on its own . since @xmath431 is an isometry , it means that any covering of @xmath2 via globally hyperbolic contractible subsets @xmath219 , @xmath311 induces a cover of @xmath417 via @xmath433 . it is easy to realize that @xmath434 where @xmath427 acts on each generator @xmath224 , @xmath435 yielding @xmath436 . notice that , since @xmath6 is independent from the metric and @xmath7 is constructed out of @xmath6 and of the hodge dual @xmath35 , they both commute with isometries . hence @xmath437 and @xmath438 . this also suffices to claim that , if we call @xmath439 the causal propagator of the @xmath228-operator on @xmath417 , it holds that @xmath440 . hence we can consider the commutator between two generators to prove @xmath441=i\int_{\mu(m)}\!\!g_\mu(\mu_*\omega)\wedge * d\delta(\mu_*\omega')=\\ i\int_{\mu(m)}\!\!\mu_*(g\omega)\wedge \mu_*(*d\delta\omega')=\int_{\mu(m)}\!\!\mu_*(g\omega\wedge * d\delta\omega')=\\ = i\int_m g\omega\wedge * d\delta\omega'=\left[\widehat{\bf f}(\omega),\widehat{\bf f}(\omega')\right].\end{gathered}\ ] ] since complex conjugation is not affected by isometric embeddings , we have proven that @xmath442 actually defines a unit preserving @xmath35-homomorphism between @xmath372 and @xmath443 . we can now without loss of generality assume that the collection of @xmath433 is part of a covering of @xmath12 with globally hyperbolic contractible spacetimes . on account of the structural properties of the universal algebra and of the absence of a centre in both @xmath157 and @xmath444 this entails that @xmath427 is indeed an injective @xmath35-homomorphism . furthermore on account of the commutator being defined out of the causal propagator , the theory is causal and the time - slice axiom is fulfilled as already proven in lemma [ time slice axiom ] . we need to answer why one is forced to restrict the attention to backgrounds with trivial second de rham cohomology group . as one can realize from the above proof , if we would have considered @xmath414 , one would have to consider the homomorphism induced by the embedding @xmath431 from @xmath2 into @xmath12 . since @xmath2 is diffeomorphic to @xmath417 , it is known that these two spacetimes have isomorphic cohomology groups , but we have to go one step further and see @xmath417 as an open subset of @xmath12 . here is the source of potential problems since , even if @xmath373 , there is no reason why @xmath445 should be isomorphic to @xmath45 ; actually it can also be trivial . we provide an explicit example : let us consider the ultrastatic globally hyperbolic spacetime @xmath446 endowed with the line element @xmath447 where @xmath448 is the canonical metric of the unit @xmath133-sphere . by knneth formula , 5 ) , @xmath449 which is non trivial since @xmath450 . let us now consider as @xmath12 , the ultrastatic spacetime @xmath451 whose metric coincides in a local chart to @xmath452 . it is manifest that @xmath2 is isometrically embedded in @xmath12 , but still knneth formula entails that @xmath453 . since @xmath29 is contractible , only @xmath454 contributes and therefore the second cohomology group of @xmath451 is trivial . let us now consider in the framework outlined above @xmath455 , then @xmath224 lies in the centre of @xmath157 thanks to proposition [ centre ] . under the isometric embedding @xmath456 , one obtains @xmath457 . yet , since @xmath442 commutes with @xmath7 , @xmath429 is coclosed and since @xmath445 is trivial , there exists @xmath458 such that @xmath459 . this entails that @xmath460 on account of maxwell s equation . barring a minor generalization , this entails that every element of the centre of @xmath157 is mapped into ( the equivalence class of ) @xmath335 in @xmath444 . this is tantamount to claim that @xmath427 can not be an injective @xmath35-homomorphism , injectivity failing to be achieved . in this paper we have developed a full - fledged quantization scheme for the field strength tensor obeying maxwell s equations . since we wanted to keep the discussion as general as possible we have neither used the vector potential as an auxiliary tool nor we have assumed the compactness of the cauchy surface of the underlying globally hyperbolic spacetime @xmath2 . this forced us to use two - forms @xmath0 obeying as the building block of the theory ; we have shown in particular that it still possible to construct a field algebra whose generators obey the commutation relations provided by the lichnerowicz propagator . yet we have also proven that the overall procedure does not fit in the scheme of general local covariance as developed in @xcite since there exist spacetimes @xmath2 with @xmath461 . in this case the universal algebra @xmath157 possesses a non trivial centre whose elements have been fully characterized in proposition [ centre ] . nonetheless it is possible to conceive that @xmath2 is isometrically embedded in a second globally hyperbolic spacetime @xmath12 which has a trivial second de rham cohomology group and thus the associated field algebra has a trivial centre . this translates in the failure of the homomorphism from @xmath157 into @xmath444 from being injective and thus the embedding translates in a loss of a qualitative feature of the field algebra of @xmath2 when seen from @xmath12 , such as the presence of superselection sectors as first discussed in @xcite . as we have proven in the previous section , a potential way out is to restrict the class of spacetimes we consider and general local covariance is restored as soon as we assume to work only with backgrounds with vanishing second de rham cohomology group . yet it is fair to admit that the situation is rather puzzling : on the one hand the proposed solution would discard spacetimes , such as schwarzschild , which are certainly of physical relevance , while on the other hand the requirement that @xmath388 vanishes entails that all field strength tensors would descend from a vector potential . this feature is certainly desirable as soon as we want to move from a free field theory to an interacting one such as quantum electrodynamics where the spinor fields are known to interact via @xmath9 rather than via the faraday tensor . yet we feel it is still early to claim we have a total loss : as a matter of fact , if we focus on any equivalence class @xmath401\in h^2(m)$ ] , we are considering all elements of the form @xmath462 where @xmath9 while @xmath3 . in other words each non trivial cohomology class is composed of two parts . the first , is responsible for qualitative features such as global topological charges or , from the quantum perspective , for the identification of a specific superselection section and , hence , it is strictly tied to the specific chosen spacetime . the second is instead tied to a @xmath333-form , a sort of vector potential , and it is well - suited both to discuss interactions and to apply the principle of general local covariance . although we are aware that this is simply a remark which does not necessarily solve all problems we have at hand , we still feel it is a starting point for further investigations which is worth to consider in detail . the work of c.d . is supported partly by the university of pavia and by the indam - gnfm project `` stati quantistici di hadamard e radiazione di hawking da buchi neri rotanti '' . we gratefully acknowledge the kind hospitality of the ii . institute fr theoretische physik of the university of hamburg during the realization of part of this work . we are also indebted for enlightening discussions and comments with klaus fredenhagen , thomas - paul hack , valter moretti , nicola pinamonti , jan schlemmer as well as with the whole lqp research group in hamburg . we are grateful to andrey saveliev for drawing the picture present in the main body of the text . the content and result of this paper are partly inspired to those present in the diploma thesis of b.l . the details of the construction of the universal algebra were worked out in york and we would like to thank chris fewster for numerous discussions , in particular for his suggestion to consider disconnected spacetimes as well in order to realise that the universal algebra obeys the principle of locality .

we have extended the search of @xcite for periodicities in the intensities of x - ray sources using light curves produced by the _ rossi x - ray timing explorer ( rxte ) _ all - sky monitor ( asm ) . @xcite analyzed data accumulated during the first 8.5 years of the mission and detected 41 periodicities and 5 possible periodicities or quasiperiodicities . two of the periodicities detected in their search likely represent neutron star spin and @xmath1 represent the precession of tilted accretion disks or other phenomena that produce superorbital periods . the 33 other periodicities are believed to represent evidence of the orbital motion of a close - binary star system . most of the orbital periods were associated with high - mass x - ray binaries ( hmxbs ) where the normal type star is either a supergiant or a be star . eight of the periodicities were associated with the orbital periods of low - mass x - ray binaries ( lmxbs ) , and one was associated with a cataclysmic variable ( cv ) . the search of @xcite was one of the most comprehensive searches for periodicities in the range of hours to years in the x - ray intensities of the bright galactic x - ray sources . however , relatively few new periods were found in the course of the search in spite of the conventional wisdom that nearly all of the bright galactic sources are accreting binaries with periods in the range of sensitivity of the search . the orbital periods of many galactic sources are still not determined , and so searches for evidence of these presently unknown orbital periods remain of interest . the present search improves on the sensitivity achieved in the earlier search through the analysis of light curves obtained over a longer course of time , i.e. nearly 14 years as opposed to 8.5 years , and through two major changes in the analysis techniques . we have also analyzed subintervals of the mission - long data sets . a preliminary report of the results of this work was presented in @xcite . the asm consists of three scanning shadow cameras ( sscs ) mounted on a rotating drive assembly @xcite . each ssc acts as a coded - aperture camera and comprises a position - sensitive proportional counter ( pspc ) that detects x - ray photons and provides one - dimension position information via the charge - division technique . the entrance window of each pspc has a total geometric area of about 68 @xmath2 . a mask that is perforated with a pseudorandom pattern of open slits is held 30 cm above the pspc entrance window . the mask open fraction is close to 50% . the field of view of each ssc is @xmath3 ( fwhm ) . the asm observations have almost entirely been made as series of 90-s exposures known as dwells . in each dwell the aspect of the assembly of sscs is held fixed . for each exposure , a coded - aperture analysis of the data from each ssc yields intensity estimates of the sources that are listed in a catalog ( hereafter the asm catalog ) and that fall within that ssc s field of view . the intensity estimates are derived for each of four spectral bands which nominally correspond to photon energy ranges of 1.5 - 3 , 3 - 5 , 5 - 12 , and 1.5 - 12 kev and typically achieve a sensitivity of a few ssc counts s@xmath4 ( the crab nebula produces intensities of 27 , 23 , 25 , and 75 ssc counts s@xmath4 in the 4 bands , respectively ) . any given source has been viewed in a large number , e.g. , 30,000 to 100,000 , of the 90-s exposures over the @xmath0 years of operation of the asm . each exposure generally yields intensity estimates for the four spectral bands . the present search has been carried out , except where otherwise noted , using the asm data acquired from the beginning of operations in early 1996 through 2010 january 28 ( mjd 55224 ; modified julian date = julian date @xmath5 2,400,000.5 ) . the coded - aperture analysis is applied to all asm observations . the results are sorted source - by - source , filtered , and used to construct light curve files for each of 571 potential sources for each of the four energy bands . the intensity measurements contained in these files are the input data for the present periodicity search . the instrument properties have changed over the 14 years of operation in orbit because of at least three effects . first , the gain of ssc 1 ( of sscs 1 - 3 ) increased at about 10% per year for the first 13 years or so because of a leak in the counter entrance window ( the rate of increase appears to have moderated in year 14 ) . this has resulted , first , in progressive reductions in the effective sensitivity of ssc 1 in the 5 - 12 kev band and , recently , to some extent in the 3 - 5 kev band because in the last few years the pulse heights of higher energy events are often saturated and thus unusable . the second of the effects is present because the one - dimensional position - sensing capability of the proportional counters is based on the charge - division technique and the resistivity of carbon - coated quartz - fiber anodes . the process of event detection in the counters results in the redistribution or removal of the carbon . the resistivity of each anode thus evolves over time generally becoming larger and less uniform , and so the functions that relate charge ratio to position have changed over the course of the mission . third , in the last few years there is often a high background of events that have pulse heights that suggest that they are due to the detection of al k@xmath6 photons . we do not understand in detail why this is happening but we speculate that it may be due to leakage of ultraviolet radiation through the al - coated kapton thermal shields that are immediately above the coded mask plates . uv radiation would produce photoelectrons at the counter body surfaces that happen to be , by design of the instrument , at potentials of approximately @xmath7 kv . the photoelectrons would be accelerated toward and would hit , e.g. , the aluminum collimators that are near the ground potential . some fraction of the al k@xmath6 photons that would be produced by this process would then be detected . both the gain changes and charge ratio to position calibration changes are accounted for in detail in the detector model used in our coded - aperture analysis . however , the calibrations are imperfect and the derived source intensities are therefore somewhat affected by systematic errors on both short and long time scales . the coded aperture analysis is also designed to eliminate the effects of backgrounds that are smooth and slowly - varying with regard to position in the detector ; this is largely but not entirely effective . searches for periodicities where large numbers of measurements are involved have generally proceeded by averaging the measured intensities in equally - spaced time bins , fourier transforming the resulting time series , producing a power density spectrum by taking the absolute square of the complex amplitude for each frequency , and searching for peaks in the power spectrum . our search largely follows this outline but we have implemented a few significant modifications as described below . in some searches for periodicities , a lomb - scargle periodogram ( @xcite ; see also @xcite ) is used in place of a fourier transform , but , when large numbers of measurements and time bins are being analyzed , the differences between the two are not important except at extremely low frequencies . for the present analyses , the conventional discrete fourier transform is adequate so we do not make use of the lomb - scargle variation . we employ two unusual strategies with the goal of improving the search sensitivity . the first strategy involves the use of weights such as the reciprocals of the variances in fourier and other types of analysis since the individual asm measurements have a wide range of associated uncertainties . since in the present instance the amplitudes in the fourier transforms may be regarded as statistical averages , they may be better determined by the use of appropriate weights . this is discussed briefly in @xcite and at greater length in @xcite . these two reports also discuss the use of weights which involve the intrinsic source variability , but we do not use such weights in the work reported herein . the second strategy stems from the fact that the observations of the source are obtained with a low duty cycle , i.e. , the window function is sparse but nevertheless highly structured . the properties of the window function , in combination with the presence of slow variations of the source intensity , act to hinder the detection of variations on short time scales . the window function power density spectrum ( pds ) has substantial power at high frequencies , e.g. , 1 cycle d@xmath4 and 1 cycle per spacecraft orbit ( @xmath8 minute period ) . since the data may be regarded as the product of a ( hypothetical ) continuous set of source intensity measurements with the window function , a fourier transform of the data is equivalent to the convolution of a transform of a continuous set of intensity measurements with the window function transform . the high frequency structure in the window function transform acts to spread power at low frequencies in the source intensity to high frequencies in the calculated transform ( or , equivalently , the pds ) . this effectively raises the noise level at high frequencies . thus , in most of our analyses , a smoothed version of the light curve is subtracted from the binned data prior to the fourier analysis . the details of the analysis of each asm dwell - by - dwell light curve are as follows . the time of each intensity measurement was adjusted to approximately remove the effects of the motion of the earth around the solar system barycenter and the adjusted times were used to assign the measurements to equally - spaced time bins of duration 5 minutes ( 0.0034722 d ) or more . weights are used in averaging those measurements that fall into a given time bin . the @xmath9 measurements that fall into the time bin which is labelled by the index @xmath10 are denoted with index @xmath11 that runs from 1 to @xmath9 . we denote the @xmath11th measured source intensity in a time bin by @xmath12 and its associated 1-sigma uncertainty by @xmath13 . the averaging in this stage uses weights @xmath14 where @xmath15 is a selected weighting index ( see values in table [ tbl : mthds ] ) . the weighted - average intensity in bin @xmath10 is then @xmath16 for convenience we denote the denominator of this expression by @xmath17 so that @xmath18 where it should be understood that @xmath19 for those time bins that contain no measurements . the @xmath17 s are used to weight the time bins in the subsequent analysis . lccc unwt & none & 2.0 & 0.0 + wt & none & 2.0 & 2.0 + ungs & gaussian & 2.0 & 0.0 + wtgs & gaussian & 2.0 & 2.0 + wtgs - v4 & gaussian & 2.5 & 1.5 + wtgs - v8 & gaussian & 2.0 & 3.0 + wtbx & box & 2.0 & 2.0 as stated above , a smoothed version of the light curve is , in many of our analyses , subtracted from the binned data . the smoothed version is given by a time bin by time bin ratio in which the numerator is computed by convolving a kernel function with a weighted version of the binned light curve . the denominator is computed by convolving the same kernel function with a weighted version of the window function . the value of the @xmath10th bin of the smoothed data is given by @xmath20 where @xmath21 represents the kernel function , @xmath22 and @xmath23 are the above defined functions of the time bin index @xmath10 , and `` @xmath24 '' denotes convolution . the kernel function @xmath21 may be either a gaussian or a box function . in the case of the gaussian kernel , the full width at half maximum response is taken to be equal to the smoothing time parameter . the gaussian is calculated out to @xmath25 standard deviations from its center . in the case of the box function , the box width is taken to be twice the smoothing time parameter . the convolution is accomplished using fourier transforms . we repeated the analyses using seven different values of the smoothing time parameter ( see table [ tbl : tmscls ] ) to optimize the sensitivity in each of several different frequency ranges . in each case the smoothed light curve was subtracted from the unsmoothed light curve , weights were applied , and the results were fourier transformed . in symbols , this part of the analysis produces a filtered light curve described by : @xmath26 where the weights @xmath27 are given by @xmath28^{k_b / k_a}\ ] ] and @xmath29 is a second weighting index . weight index values are given in table [ tbl : mthds ] . we also did the analysis without subtracting a smoothed version of the light curve . the unweighted average of the set of @xmath30 s which correspond to bins with nonzero exposure is computed and then subtracted from each of the @xmath30 values for those bins . the resultant array is extended with zeroes by a factor of four and fourier transformed . the extension of the array yields oversampling in the frequency domain . the oversampled transform is converted into a power density spectrum which is normalized to have an average value of unity . the normalized power spectra are compressed by saving the average and maximum values for sets of contiguous frequency bins ; the number of contiguous bins depends on the frequency range and smoothing time parameter ( see table [ tbl : tmscls ] ) . we found that for the analyses where short smoothing time scales were used , i.e. , time scales of 0.3 , 0.9 , and 3.0 days , the compressed power spectra have a non - white appearance . the effect is strongest for the shortest smoothing time scale , i.e. , 0.3 days . it is not clear how to apply a threshold for detection of a periodicity in a nonwhite spectrum . therefore we added a procedure ( to the code that performs the compression described above ) to compute a `` background '' average power around each compressed frequency bin , and to then obtain whitened maximum values . each whitened maximum value is then the ratio of a maximum value in an unwhitened compressed spectrum to the corresponding background average . the resulting whitened spectra generally appear to have little systematic structure ( see fig . [ fig : whiten ] ) . table [ tbl : mthds ] lists , at the highest level , the variations of the analysis method that we have used . we found that the results from the methods where both weighting and filtering were applied ( `` wtgs '' , `` wtgs - v4 '' , `` wtgs - v8 '' , and `` wtbx '' ) were almost always nearly equal or superior to the results obtained with the other methods ( `` unwt '' , `` wt '' , `` ungs '' ) . a typical , i.e. , not the extreme best case , of the differences is illustrated in figure [ fig : pdscalc ] . therefore we did not search for significant peaks in the power spectra obtained using the latter methods . the `` wtgs '' , `` wtgs - v4 '' , `` wtgs - v8 '' , and `` wtbx '' methods were all applied to the 1.5 - 12 kev band light curves , but the results generally produced power spectra that were quite similar in terms of both detailed appearance and sensitivity . therefore only the `` wtgs '' method was applied to the other three energy bands . the results from the four energy bands are , in some cases , quite different from each other . table [ tbl : tmscls ] lists smoothing time scales and the corresponding bin durations , parameters used for compression and whitening , and parameters used in searching for significant peaks . ccrrrcrr none & 2.0 & 1 & 300 & 3 & 0.0007 & 1.0 & 17.7 + 500.0 & 2.0 & 1 & 300 & 3 & 0.004 & 1.0 & 17.7 + 100.0 & 1.0 & 1 & 300 & 3 & 0.022 & 1.0 & 18.4 + 30.0 & 0.15 & 10 & 60 & 1 & 0.1 & 4.0 & 20.3 + 10.0 & 0.05 & 20 & 30 & 1 & 0.2 & 10.0 & 21.4 + 3.0 & 0.005 & 150 & 12 & 1 & 0.66 & 100.0 & 23.7 + 0.9 & 0.0034722222 & 300 & 6 & 1 & 2.2 & 144.0 & 24.1 + 0.3 & 0.0034722222 & 300 & 6 & 1 & 6.6 & 144.0 & 24.1 the powers in a spectrum of gaussian white noise are expected to be well characterized by an exponential probability distribution . we make the assumption that the powers in the whitened spectra computed as described above similarly follow an exponential distribution . in the present case it is not clear how to precisely compute the number of independent bins that we examined because 1 ) some of the power spectra for a given source involve the same data and differ only on account of slight variations in the analysis procedure , and 2 ) there is a significant degree of dependence among the powers in an oversampled power spectrum . although it is not fully justified on an a priori basis , we simply estimate the number of independent bins @xmath31 by the product of the number of frequency bins that would have applied if the fourier spectrum had not been oversampled , the number of energy bands analyzed ( 4 ; this neglects the non - independent bins among the four slightly different analyses of the 1.5 - 12 kev band ) , and the number of sources ( 571 ) . the neglect of the oversampling factor and the multiple similar analyses of the sum band does not appear to have led to significant underestimates of the detection thresholds , and , thus , this procedure appears to be justified . we have chosen a detection threshold @xmath32 for previously unknown periodicities to limit the expected number of false detections @xmath33 produced by statistical fluctuations ( based on the assumption that the powers are exponentially distributed ) in the power spectra calculated using a given smoothing time parameter . thus we have @xmath34 so that @xmath35 the values of @xmath32 are given in table [ tbl : tmscls ] for @xmath36 . as can be seen in figs . [ fig : whiten ] and [ fig : pdscalc ] , the window functions of the light curves often contain strong peaks at or close to particular frequencies . even though the analysis procedure tends to minimize the appearance of these frequencies in the whitened power spectra , there are occurrences of peaks at these frequencies . in the analyses done with filtering based on smoothing time scales of 10.0 days and under , we ignore all peaks in the frequency ranges @xmath37 , @xmath38 , @xmath39 , @xmath40 , @xmath41 , and @xmath42 d@xmath4 . for the other analyses , we ignore all peaks in the frequency ranges @xmath43 , @xmath44 , @xmath45 , @xmath46 , @xmath47 , @xmath48 , @xmath49 , and @xmath50 d@xmath4 . a substantial fraction of the 571 potential sources are extragalactic targets such as active galactic nuclei , extended galactic targets such as supernova remnants , or x - ray binaries that have been in quiescence since the beginning of the rxte mission and for which we have no expectation of finding detectable periodicities . the power spectra of these sources act effectively as controls that show the actual statistics of the power spectra and that bring attention to periodicities that must be the product of statistical fluctuations or artifacts . indeed , the search procedure described here was tailored in part by examination of the results on the `` controls '' . furthermore , since it would be difficult to understand the reality of peaks that exceed the detection threshold in the power spectra of these sources , the effective number of sources must be less than 571 . we have not adjusted the significance thresholds downward to account for this effect . since we searched large numbers of bins , the thresholds for the detection of previously unreported periodicities given in table [ tbl : tmscls ] when a periodicity has previously been reported , it may be confirmed even if the power at or very close to the previously reported frequency is below the threshold given in the table . uncompressed power density spectra are useful for precise estimation of the frequencies of the detected periodicities . since the uncompressed spectra were not saved in the original analysis for smoothing time scales less than or equal to 30.0 days , the analyses that produced the most significant detection of each periodicity were repeated , if necessary , and the resulting uncompressed spectra were saved . for each periodicity a small frequency range around the major peak was fit with a gaussian to obtain estimates of the central frequency , its uncertainty , and the peak width . the uncertainty in the frequency may be estimated based on the result of horne and baliunas ( 1986 ) that @xmath51 where @xmath52 is the length of the time interval covered by the light curve and @xmath53 is the peak power in a normalized pds . we also use the simple estimate @xmath54 which , in most cases , should be a conservative upper limit to the uncertainty ( but , if the periodicity persisted for only a fraction of the time covered by the light curve , the uncertainty could be larger than this value ) . in the estimates of frequencies and periods that we present below , the uncertainty derived using eq . ( 8) is given first and the uncertainty derived using eq . ( 9 ) is given second in square brackets . in the present search we find clear evidence of the periodicities in gro j1008@xmath557 and ss 433 that have been previously detected in the asm data but had not been detected in the analysis reported by @xcite . @xcite reported the detection of the 9.56-day period of 4u 2206 + 543 , but did not find evidence of variation with a period of @xmath55 days . such variation was first reported by @xcite . since we now find clear evidence of variation at the @xmath55-d period , we report its detection herein . in the previous section , we noted that periodicities at previously identified frequencies can be found with better sensitivity than periodicities at undetermined frequencies . therefore we examined the power spectra for peaks at the frequencies corresponding to the periods listed in the catalogs of @xcite and @xcite and thereby identified evidence of the presence of a number of periodicities that were not found by @xcite . table [ tbl : detect ] lists the periodicities that emerge from the present search that are not listed in tables 1 - 3 of @xcite . only one of these was , when first detected in the course of this search , previously unreported on the basis of optical or other x - ray observations , namely the 18.545-day periodicity in igr j18483@xmath50311 . the photon energy band , filter time scale , and method that yielded the most significant detection are given for each detected periodicity . the estimated frequencies and associated uncertainties are listed in table [ tbl : detect ] with the exception of the very low frequencies of the sources gro j1008@xmath557 and 2s 1845@xmath5024 . the frequencies or , equivalently , periods of these latter sources are better obtained from epoch - folding analyses and even more precisely from pulse timing studies . we give in table [ tbl : ampl ] average source intensities , peak powers seen in the whitened and renormalized power spectra , and estimated average amplitudes of the detected periodic signals . the source names , photon energy bands , and periods are taken from table [ tbl : detect ] . the estimated average amplitudes of the detected periodic signals were obtained by superposing sine waves onto the measured intensities in the appropriate light curves and redoing the complete analysis using the appropriate method , time scale , and energy band . the sine wave frequencies were chosen to yield peaks in the power spectrum close to , but not on top of , the frequencies of the detected signals . the sine wave amplitudes were chosen to give peak powers close to or above , within a factor of two , the peak powers of the detected signals . the results were then interpolated to estimate the amplitudes that would give rise to the peak powers of the detected signals . the uncertainty in the amplitude for a given source was estimated by the rms dispersion in the amplitudes obtained when sine waves were injected at four to ten different frequencies . in the last column of table [ tbl : ampl ] we give the ratio of the inferred average signal amplitude ( a ) to the weighted average source intensity ( i ) . this is an indicator of the fractional variation of the source intensity at the orbital frequency . it does not include the effects of harmonics which are detected in some of the power density spectra and should , therefore , be used only as a general indicator of the degree of modulation . llcccccc 2s0921@xmath5630 & x0921@xmath5630 & 1.5 - 12 & 30.0 & wtbx & 0.110994 ( 13 ; 100 ) & 216.23 ( 3 ; 19 ) h & 52738.13 ( 13 ) + 4u1254@xmath569 & x1254@xmath5690 & 1.5 - 12 & 3.0 & wtgs - v4 & 6.101690 ( 15 ; 100 ) & 3.933337 ( 10 ; 64 ) h & 52733.0546 ( 31 ) + 4u1323@xmath562 & x1323@xmath5619 & 1.5 - 12 & 3.0 & wtgs & 8.157930 ( 15 ; 100 ) & 2.941923 ( 5 ; 36 ) h & 52733.2395 ( 18 ) + 4u1636@xmath5536 & x1636@xmath5536 & 1.5 - 12 & 0.9 & wtgs & 6.327232 ( 14 ; 100 ) & 3.793128 ( 8 ; 60 ) h & 52733.0343 ( 18 ) + 4u1728@xmath516 & gx9 + 9 & 1.5 - 12 & 0.9 & wtgs & 5.723207 ( 7 ; 100 ) & 4.193453 ( 5 ; 73 ) h & 52733.152 ( 1 ) + 4u1746@xmath537 & x1746@xmath5370 & 1.5 - 12 & 3.0 & wtgs - v8 & 4.648202 ( 15 ; 100 ) & 5.163287 ( 18 ; 111 ) h & 52733.156 ( 3 ) + grs1758@xmath5258 & grs1758@xmath5258 & 3 - 5 & 500.0 & wtgs & 0.052706 ( 19;100 ) & 18.973 ( 7 ; 36 ) d & 52748.27 ( 23 ) + 4u1820@xmath530 & x1820@xmath5303 & 1.5 - 12 & 0.9 & wtgs & 126.129380 ( 13 ; 100 ) & 0.19028081 ( 2 ; 15 ) h & 52733.00590 ( 7 ) + a0535 + 26 & x0535 + 26 & 1.5 - 12 & 500.0 & wtgs & 0.00906 & 110.4 & 52852.59 ( 27 ) + lmc x-1 & lmcx1 & 1.5 - 12 & 100.0 & wtbx & 0.255821 ( 14 ; 100 ) & 3.90898 ( 21 ; 153 ) d & 52733.952 ( 45 ) + groj1008@xmath557 & groj1008@xmath557 & 3 - 5 & 500.0 & wtgs & & & + 2s1417@xmath562 & x1417@xmath5624 & 5 - 12 & 100.0 & wtgs & 0.023611 ( 45 ; 100 ) & 42.35 ( 8 ; 18 ) d & 52770.04 ( 91 ) + igrj16320@xmath54751 & igrj16320@xmath54751 & 5 - 12 & 100.0 & wtgs & 0.111234 ( 11 ; 100 ) & 8.9901 ( 9 ; 81 ) d & 52742.79 ( 12 ) + igrj16418@xmath54532 & igrj16418@xmath54532 & 5 - 12 & 30.0 & wtgs & 0.267461 ( 20 ; 100 ) & 3.73886 ( 28 ; 140 ) d & 52735.84 ( 7 ) + exo1722@xmath5363 & exo1722@xmath5363 & 5 - 12 & 30.0 & wtgs & 0.102655 ( 19 ; 100 ) & 9.7414 ( 18 ; 95 ) d & 52738.25 ( 19 ) + igrj18027@xmath52016 & igrj18027@xmath52017 & 5 - 12 & 30.0 & wtgs & 0.218881 ( 16;100 ) & 4.5687 ( 3;21 ) & 52736.36 ( 42 ) + 2s1845@xmath5024 & x1845@xmath5024 & 5 - 12 & 500.0 & wtgs & & & + igrj18483@xmath50311 & igrj18483@xmath50311 & 5 - 12 & 100.0 & wtgs & 0.053923 ( 9 ; 100 ) & 18.545 ( 3 ; 34 ) d & 52754.37 ( 10 ) + 3a1909 + 048 & ss433 & 5 - 12 & 100.0 & wtgs & 0.076452 ( 20 ; 100 ) & 13.080 ( 3 ; 17 ) d & 52745.19 ( 18 ) + 4u2206 + 543 & x2206 + 543 & 1.5 - 12 & 100.0 & wtbx & 0.104620 ( 13 ; 100 ) & 9.5584 ( 12 ; 91 ) d & 52740.97 ( 6 ) + lccccccc 2s0921@xmath5630 & 1.5 - 12 & 0.239(4 ) & 0.084(6 ) & 216.2 h & 23 & 0.052 & 0.22 + 4u1254@xmath569 & 1.5 - 12 & 2.455(4 ) & 2.291(6 ) & 3.933 h & 22 & 0.052 & 0.021 + 4u1323@xmath562 & 1.5 - 12 & 0.575(4 ) & 0.449(6 ) & 2.942 h & 27 & 0.055 & 0.096 + 4u1636@xmath5536 & 1.5 - 12 & 9.069(6 ) & 9.339(9 ) & 3.793 h & 22 & 0.087 & 0.0096 + 4u1728@xmath516 ( gx9 + 9 ) & 1.5 - 12 & 18.879(8 ) & 19.01(1 ) & 4.193 h & 123 & 0.283 & 0.015 + 4u1746@xmath537 & 1.5 - 12 & 2.275(6 ) & 2.06(1 ) & 5.163 h & 30 & 0.094 & 0.041 + grs1758@xmath5258 & 3 - 5 & 0.485(5 ) & 0.391(7 ) & 18.97 d & 13 & 0.054 & 0.11 + 4u1820@xmath530 & 1.5 - 12 & 20.347(8 ) & 20.63(1 ) & 0.1903 h & 30 & 0.12 & 0.0059 + a0535 + 26 & 1.5 - 12 & 0.275(4 ) & 0.550(6 ) & 110.4 & 11 & 0.21 & 0.78 + lmc x-1 & 1.5 - 12 & 1.524(4 ) & 1.369(5 ) & 3.909 d & 29 & 0.064 & 0.042 + groj1008@xmath557 & 3 - 5 & 0.036(2 ) & 0.006(3 ) & & 10 & & + 2s1417@xmath562 & 5 - 12 & 0.096(3 ) & 0.035(6 ) & 42.4 d & 12 & 0.035 & 0.36 + igrj16320@xmath54751 & 5 - 12 & 0.224(4 ) & 0.119(7 ) & 8.99 d & 34 & 0.063 & 0.28 + igrj16418@xmath54532 & 5 - 12 & 0.165(5 ) & 0.060(9 ) & 3.739 d & 17 & 0.047 & 0.28 + exo1722@xmath5363 & 5 - 12 & 0.132(4 ) & 0.029(7 ) & 9.741 d & 19 & 0.051 & 0.39 + igrj18027@xmath52016 & 5 - 12 & 0.189(7 ) & 0.03(1 ) & 4.569 d & 16 & 0.069 & 0.37 + 2s1845@xmath5024 & 5 - 12 & 0.069(3 ) & -0.065(7 ) & & 5 & + igrj18483@xmath50311 & 5 - 12 & 0.115(3 ) & -0.021(7 ) & 18.54 d & 60 & 0.099 & 0.86 + 3a1909 + 048 ( ss433 ) & 5 - 12 & 0.256(3 ) & 0.189(6 ) & 13.08 d & 20 & 0.037 & 0.14 + 4u2206 + 543 & 1.5 - 12 & 0.376(3 ) & 0.252(5 ) & 9.558 d & 17 & 0.08 & 0.21 + the asm light curves of the sources for which we report detections of periodicities in table [ tbl : detect ] were folded at periods close to , i.e. , within a small fraction of the quoted 1@xmath56 errors , those listed in that table . weights were used in averaging the intensities in each phase bin . a smoothed version of the light curve was subtracted from each light curve before it was folded . the time scales used for the smoothing are listed in table [ tbl : detect ] . after the folding was accomplished , the overall average intensity was then added to the folded light curve to show the proper mean ( the mean was removed when the smoothed version was subtracted from the light curve ) . the folded light curves are shown in figures [ fig : fold1 ] and [ fig : fold2 ] . the uncertainties shown in the plots of the folded light curves were determined from propagation of the 1@xmath56 uncertainties associated with each measurement in a light curve assuming the measurements are statistically independent . this neglects the effects of random source variability in at least two respects . first , the effects of source variability are not included in the formal uncertainties of the individual measurements . second , when variability is present it almost always involves correlations from measurement to measurement . the epochs of phase zero used in plotting the folded light curves are listed in table [ tbl : detect ] . these epochs were determined by fitting each folded 1.5 - 12 kev band light curve with a combination of constant , sine , and cosine functions of phase . the epoch is then obtained from the phase of minimum or maximum intensity of the fitted function . actual minima or maxima may deviate somewhat from these epochs ; the differences may be seen in the folded light curves . in this section we briefly discuss our results for each of the periodicities of the lmxbs listed in table [ tbl : detect ] . the results in table [ tbl : detect ] show that , with only one exception , the most significant detections of these systems were made in the overall 1.5 - 12 kev energy band . * 2s0921@xmath5630 * is a weak x - ray source in a low - mass system with a relatively long orbital period . photometry of the bright optical counterpart by @xcite and @xcite suggested that the observed variability could be periodic , but did not yield correct estimates of the actual period . x - ray observations with the _ exosat _ satellite showed intensity changes , interpreted as partial eclipses , that were key to obtaining the first relatively precise period estimate @xcite . recent optical observations have also yielded relatively precise estimates of the orbital period ; @xcite obtained @xmath57 d and @xcite obtained @xmath58 d. the orbital period is detected in the present study ; this is shown in the power spectrum plotted in fig . [ fig : pds0921 ] . we derive an estimate of the orbital period of @xmath59 $ ] d ( see table [ tbl : detect ] ) that is consistent with these recently reported values . @xcite and @xcite report further work on the binary parameters including the component masses . the folded asm light curves show that the periodic modulation has an amplitude of @xmath6030% relative to the average source strength . however , the average source intensity could be affected by systematic errors in the baseline level of order 0.1 ssc counts s@xmath4 ( 1.5 - 12 kev band ) . nonetheless , the modulation fraction must be rather large , and the periodicity may be easily detectable by other instruments . 630 . the `` wtbx '' method and a 30.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds0921],height=336 ] * 4u1254@xmath569 * is a so - called dipping lmxb . its @xmath61 hr orbital period was revealed in both x - ray and optical observations @xcite . @xcite report the results of recent observations of this source with the _ xmm - newton _ and _ integral _ satellites . they also discuss the detection of the 3.9-hr periodicity in asm data , present a folded light curve , and use the asm data to obtain an estimate of the orbital period of @xmath62 d ( @xmath63 h ) . we also detect the orbital period in the asm light curves ( see fig . [ fig : pds1254 ] ) and derive a value for the period of @xmath64 $ ] h ( see table [ tbl : detect ] ) . this value is close to that of @xcite but our estimate of the uncertainty is larger . we do not believe that the asm data alone warrant a more precise estimate of the orbital period than we have given . the `` wtgs - v4 '' method and a 3.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1254],height=336 ] * 4u1323@xmath562 * is also a dipping lmxb . @xcite report on _ exosat _ x - ray observations wherein the orbital period of the binary system was manifest in terms of intensity dips that recurred with a period of @xmath65 h. later , @xcite derived a period of @xmath66 h from x - ray observations with _ bepposax_. even though the source is rather weak , the orbital period is easily detected in the asm light curves ( see fig . [ fig : pds1323 ] ) . we derive a value for the period of @xmath67 $ ] h ( see table [ tbl : detect ] ) . it is interesting to compare this value with the conclusion of @xcite that consideration of the times of dips in two observations separated by one year imply that the period must be more precisely given by @xmath68 h. the ambiguity in this formula reflects an ambiguity in the number of orbital cycles in the one year interval between the observations . the present determination of the orbital period implies @xmath69 and that @xmath70 h where the uncertainty is dominated by the uncertainty of the correction for @xmath69 . given that the difference between this period and the asm - derived value is only @xmath71 h and considering the uncertainties in the two measurements , we conclude that the two values are mutually consistent . recent studies of this source may be found in @xcite and @xcite . the `` wtgs '' method and a 3.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1323],height=336 ] * 4u1636@xmath5536 * is a persistent lmxb that often displays type i x - ray bursts . the orbital period was found through optical photometry @xcite . it is also evident through radial velocity measurements made from emission line profiles obtained in optical spectroscopic observations ( e.g. , * ? ? ? * ; * ? ? ? estimates of the orbital period have been successively improved by @xcite since the discovery by @xcite . to our knowledge , the best optical ephemeris is that of @xcite , who estimate that the period is @xmath72 h. the asm detection ( see figure [ fig : pds1636 ] ) is secure , given knowledge of the optical period , but would not be sufficiently strong to have been detected in our blind search . it is , nonetheless , the first detection of the orbital period in x - rays . the period estimated from the asm results is @xmath73 $ ] h ( table [ tbl : detect ] ) . this value is consistent with that of @xcite though it is somewhat less precise . the `` wtgs '' method and a 0.9 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1636],height=336 ] * 4u1728@xmath516 ( gx9 + 9 ) * is a rather bright persistent lmxb . the orbital periodicity at the period of @xmath74 hours was found in _ heao1 _ a-1 x - ray data by @xcite . soon afterwards , @xcite found variations at essentially the same period in the brightness of the optical counterpart . the asm detection and an account of the variability in the asm data of the modulation amplitude is described in full by @xcite . the periodicity is manifest with even greater significance in the longer data set used in our most recent analyses ; compare fig . [ fig : pdsgx9p9 ] with fig . 2 of @xcite . in table [ tbl : detect ] we also give an updated estimate of the orbital period that is only slightly different from that given by @xcite . , height=336 ] * 4u1746@xmath537 * is an lmxb in the globular cluster ngc 6441 . x - ray observations carried out with _ exosat _ , _ ginga _ , and _ rxte_/pca showed the presence of dips that occur roughly every 5 hours and quickly led to the interpretation that their recurrence period is the orbital period @xcite . the current best estimate of the orbital period , @xmath75 h , was given by @xcite . this periodicity is seen with high significance in the asm data ( see fig . [ fig : pds1746 ] ) . we obtain the period estimate @xmath76 $ ] h that is consistent with the results of @xcite but is substantially more precise . the `` wtgs - v8 '' method and a 3.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1746],height=336 ] * grs1758@xmath5258 * most likely comprises a black hole accreting from a k - type giant star ( * ? ? ? * ; * ? ? ? * and references therein ) . @xcite found a periodicity with @xmath77 days in the data obtained in extensive _ rxte _ pca observations done in the years 1997 through 2000 and interpreted the period as the orbital period . @xcite performed near - ir observations and identified an early k - type star as the likely companion if the orbital period is as long as 18.45 days . we find a peak in the asm power spectrum shown in figure [ fig : pds1758 ] that corresponds to a period of @xmath78 $ ] days and that we therefore identify with the orbital period . the `` wtgs '' method and a 500.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the uncompressed pds . [ fig : pds1758],height=336 ] * 4u1820@xmath530 * is a very bright lmxb in the globular cluster ngc 6624 . @xcite analyzed data from x - ray observations with _ exosat _ and discovered the 11 minute orbital period . the periodicity is only evident as a 3% or less peak - to - peak modulation in the overall x - ray intensity . @xcite used measurements made by a number of x - ray satellites of the times of maxima of the 11-m variation to obtain an ephemeris wherein both linear and quadratic terms are significant . to our knowledge , this is the most precise ephemeris in the literature at this time . we detect the periodicity in the asm light curve with high significance ; see fig . [ fig : pds1820 ] and table [ tbl : detect ] . the period we obtain , @xmath79 $ ] s , is an average over the duration of the asm data set and effectively applies near the mid - time of this interval , i.e. , near mjd 52620 . we can compare our period measurement with the period of @xmath80 s predicted for this epoch by the @xcite ephemeris . if we assume that the smaller of the two uncertainties in the asm period that are given in table [ tbl : detect ] is applicable , then the asm period is about 4 standard deviations below the prediction of the @xcite ephemeris . this suggests that the coefficient of the quadratic term in that ephemeris should be more negative . however , we have not undertaken a proper joint analysis of all of the timing results and so this conclusion must be regarded as tentative . the `` wtgs '' method and a 0.9 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1820],height=336 ] the orbital periods of a number of hmxbs are listed in table [ tbl : detect ] . in contrast with the results on lmxbs , table [ tbl : detect ] shows that the majority of the most significant detections of these systems were made in the 5 - 12 kev energy band . this is in accordance with expectations based on the observation that many of the accreting compact objects in hmxbs are x - ray pulsars that tend to have relatively hard x - ray spectra . * a0535 + 26 * is a be / x - ray pulsar that can be rather bright in x - rays during transient outbursts . the early observations are reviewed by @xcite . pulse timing analyses of observations obtained with the batse instrument were used to determine that the orbital period is @xmath81 days ( * ? ? ? * and references therein ) . a portion of an asm power density spectrum is shown in figure [ fig : pdsj1008 ] . this figure demonstrates that the periodicity is apparent , but is not detected with sufficient significance to allow us to obtain a useful improvement in the precision of the orbital period . seven separate peaks in the intensity of this source are apparent in the latter part of the asm light curve . their average separation is close to 114 days and is probably not consistent with the period determined through pulse timing . the outbursts of some other be star / neutron star binaries do not occur at precisely the orbital periods determined by pulse timing , so there is no reason in this case to doubt the accuracy of the pulse timing value . * lmc x-1 * is a luminous hmxb in the large magellanic cloud which highly likely comprises a stellar - mass black hole and its normal ob - type companion . the properties of the system are described in detail by @xcite . @xcite also present the asm results and how they relate to other observations including other period determinations . @xcite adopt the value of @xmath82 days as their best estimate of the period based on optical photometry and spectroscopy . an updated asm power spectrum is shown in figure [ fig : pdslmcx1 ] and revised estimates , based solely on the asm power spectrum , of the orbital frequency and period are given in table [ tbl : detect ] . we note that the asm detection provides the only reported evidence to date of the orbital period in x - rays and that our x - ray - based period estimate of @xmath83 $ ] days is fully consistent with the optically - based value of @xcite . , height=336 ] * groj1008@xmath557 * is a transient x - ray pulsar in a be / x - ray binary discovered with the batse experiment on the _ compton gamma - ray observatory _ ( see references in * ? ? ? * ; * ? ? ? @xcite used @xmath84 years of asm data to obtain an early estimate of the outburst period of @xmath85 d. @xcite reported a relatively precise determination of the outburst period , @xmath86 d , that was based on the time intervals between widely spaced outbursts . we believe this is still the best current estimate of the outburst cycle time . the asm power spectrum is shown in the top panel of figure [ fig : pdsj1008 ] . though the peaks due to the outburst periodicity are clear , they are not sufficiently well - defined to yield a superior period estimate . @xcite have estimated the orbital period through timing of the 93-s pulsations seen in the batse data . they obtained the result @xmath87 d and noted that is it is in good agreement with the result of @xcite . the `` ringing '' appearance derives from a light curve where a small number of strong outbursts punctuate the otherwise low source strength . ( middle ) power density spectrum made from the 3 - 5 kev light curve of groj1008@xmath557 . ( bottom ) power density spectrum made from the 5 - 12 kev light curve of 2s1845@xmath5024 . the previously determined orbital frequencies are 0.009066(25 ) d@xmath4 ( a0535@xmath88262 ) , 0.004018(8 ) d@xmath4 ( groj1008@xmath557 ) , and 0.0041292(2 ) d@xmath4 ( 2s1845@xmath5024 ) . [ fig : pdsj1008],height=336 ] * 2s1417@xmath562 * is also a transient pulsar in a be / x - ray binary system . a brief description of the early history and of a pulse - timing analysis using batse data may be found in @xcite . the batse timing analysis yielded an estimate of the orbital period of @xmath89 d as well as the projected semimajor axis , eccentricity , epoch of periastron passage , and other orbital elements . a slight revision to the orbital period and time of periastron passage have been given by @xcite . evidence of outbursts recurring at intervals that are close in duration to the orbital period ( or multiples thereof ) is seen in the asm power spectrum ( see fig . [ fig : pds1417 ] ) . frequency and period estimates obtained from the spectrum shown in fig . [ fig : pds1417 ] are listed in table [ tbl : detect ] but are not as precise as the pulse - timing values . 62 . the `` wtgs '' method and a 100.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . the orbital frequency determined through pulse timing is 0.02374 ( 2 ) d@xmath4 ( see text ) . [ fig : pds1417],height=336 ] * igrj16320@xmath54751 * is an x - ray pulsar with the rather long pulse period of @xmath90 s @xcite . @xcite found the @xmath91 day orbital period by analyzing a _ swift _ burst alert telescope 14 - 200 kev light curve , and determined the epoch of maximum flux to be mjd @xmath92 . @xcite used observations with _ integral _ to measure the period , obtaining @xmath93 d. the orbital period is seen at a high level of significance in the asm power spectrum shown in figure [ fig : pds1632 ] . we use the asm power spectrum to estimate that the period is @xmath94 $ ] d. 4751 . the `` wtgs '' method and a 100.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1632],height=336 ] * igrj16418@xmath54532 * is an x - ray source in the galactic plane that was discovered using _ integral _ @xcite . @xcite found and measured the orbital period using both _ swift_/bat and asm data and obtained the values @xmath95 days and @xmath96 days , respectively . our current asm power spectrum ( fig . [ fig : pds1641 ] ) shows a peak that is moderately significant given prior knowledge of the period . the frequency of this peak corresponds to the period @xmath97 $ ] days . this period is essentially identical to that determined earlier from the asm data by @xcite . the `` wtgs '' method and a 30.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1641],height=336 ] * exo1722@xmath5363 * is an x - ray pulsar in a high - mass binary system seen edge - on ( * ? ? ? * and references therein ) . eclipses were found in x - ray observations made in scans of the galactic plane with the _ rxte _ pca and their temporal spacing was used to get the first real measurement of the orbital period , i.e. , @xmath98 days @xcite . @xcite have estimated the orbital parameters through a pulse timing analysis and obtain a value for the period of @xmath99 days . one might expect a strong signal from an eclipsing system but the source intensity is rather modest , about 4 mcrab ( 1.5 - 12 kev ) on average in the asm observations . thus only a modest peak is evident close to the frequency in the asm power spectrum ( see figure [ fig : pds1722 ] ) . we estimate the period to be @xmath100 $ ] days . this is consistent with but somewhat less precise than the pulse - timing - based estimate of @xcite . @xcite have recently reported estimates of the masses of the two components of the binary system and confirm that it is indeed an hmxb . the `` wtgs '' method and a 30.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1722],height=336 ] * saxj1802.7@xmath52017 * = * igrj18027@xmath52016 * is only @xmath101 away from the bright x - ray source gx9 + 1 . it was discovered using _ bepposax _ and , despite its proximity to a much brighter source , was found to be an x - ray pulsar in a high - mass x - ray binary system @xcite . changes in the apparent pulse period led @xcite to infer that the pulsar is in an orbit with a period of @xmath102 days . @xcite realized that the _ integral _ source igrj18027@xmath52016 is associated with the sax source and refined the orbital period to be @xmath103 days . @xcite used _ swift_-bat , _ integral_-isgri , and _ rxte_-asm data to estimate that the orbital period is @xmath104 days . this periodicity is apparent in our asm power spectra ( see fig . [ fig : pds18027 ] ) . we use the present results to estimate @xmath105 $ ] days . the `` wtgs '' method and a 30.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds18027],height=336 ] * 2s1845@xmath5024 * = * gs1843@xmath502 * is an accreting pulsar that is evident during widely spaced outbursts ( * ? ? ? * and references therein ) . a fairly precise value for the orbital period of @xmath106 days was determined by @xcite by a pulse - timing analysis of batse data . a small number , @xmath107 , of outbursts are evident in the mission - long asm light curve and the outburst periodicity is weakly evident in the asm power spectrum ( see the lower panel of figure [ fig : pdsj1008 ] ) . however , the asm results do not support a determination of an orbital period that can approach the precision of the pulse - timing - based value determined by @xcite . hence , we do not present an asm - based estimate of the period . * igrj18483@xmath50311 * was discovered in _ integral_/ibis observations of the galactic center region @xcite . the orbital period was independently found around the same time by @xcite in the asm light curve and by @xcite in the _ integral _ light curve . the reported values for the period were @xmath108 days and @xmath109 days , respectively . @xcite have recently analyzed _ swift_/bat and asm data and estimate the period to be @xmath110 days . the period is detected with high significance in the present study ( see figure [ fig : pds1848 ] ) . we use that power spectrum to estimate that the orbital period is @xmath111 $ ] days . the `` wtgs '' method and a 100.0 day smoothing time scale were employed in producing the pds . ( bottom ) a small region of the corresponding uncompressed pds . [ fig : pds1848],height=336 ] * 3a1909 + 048 * is the x - ray counterpart to the well - known object * ss433*. this object most likely consists of a stellar - mass black hole that is accreting at super - eddington rates from a strong wind emitted by a high - mass normal - type star , and produces jets wherein the outflow velocity is approximately @xmath112 . an early review may be found in @xcite and a more recent one in @xcite . the precession of the jets with a period of @xmath113 days is not only evident in the moving optical lines , but also in other phenomena including the apparent x - ray intensity ( see * ? ? ? * ; * ? ? ? a number of reports discuss the evidence from optical spectroscopy for doppler shifts of the compact and normal components and their interpretation in terms of orbital parameters and component masses ( see , e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? @xcite reported a precise measurement of the orbital period of the binary , i.e. , @xmath114 days , on the basis of the times of eclipse - like dips in the optical flux . the x - ray intensity also shows eclipse - like dips ( e.g. , * ? ? ? * ; * ? ? ? * ) but these have not yet been exploited for orbital period estimation . @xcite reported the first evidence of the orbital period in the asm x - ray intensity measurements . very recently , @xcite obtained a new time of minimum optical brightness and used it , together with the earlier observations , to obtain a revised value of the period , @xmath115 d. in the present analysis , we find unambiguous evidence of both the precession and the orbital periods in the power spectrum ( see figure [ fig : pds433 ] ) . our estimate of the orbital period is @xmath116 $ ] days . this value is consistent with , but somewhat less precise than , those of @xcite and @xcite . if we fold the asm light curve with the period and epoch of @xcite , we find that the time of x - ray minimum is coincident with the time of optical minimum . the shape of the folded asm light curve is similar to the eclipse - like shapes seen by _ integral _ ( e.g. , * ? ? ? * ) and suggests that the eclipse - like decrease in the flux is the result of periodically occurring partial occultations of the x - ray emitting regions , broadly construed to include regions where the x - ray flux may be scattered ( see fig . [ fig : fold2 ] ) . , height=336 ] * 4u2206 + 543 * is a high - mass x - ray binary which manifests a distinct variation at a period of @xmath117 days in the asm light curve @xcite . this source is included in this report even though the @xmath118-day period was reported by @xcite because the underlying period may actually be twice as long , or @xmath119 days ( @xcite ; see also @xcite ) . in this regard , @xcite found a signal at a period of @xmath120 days in _ swift_/bat data and a signal near the same period in asm data . @xcite found a peak near a period @xmath121 days in the asm data . our power spectrum of asm data also shows peaks at frequencies corresponding to both periods ( see figure [ fig : pds2206 ] ) . the more significant peak corresponds to a period of @xmath122 $ ] days . assuming that the subharmonic corresponds to a period twice this value , we obtain @xmath123 $ ] days . radial velocity measurements from optical or near ir spectroscopy are needed to unambiguously determine the true orbital period . , height=336 ] lcccccc crab nebula & 1.5 - 12 & 75.50(1 ) & & 3.0 & 0.12 & 0.0016 + scox-1 & 1.5 - 12 & 867.81(6 ) & 0.789 d & 3.0 & 7.2 & 0.0082 + scox-1 & 1.5 - 12 & 867.81(6 ) & 0.789 d & 10.0 & 8.5 & 0.0099 + gx340@xmath880 & 1.5 - 12 & 28.657(8 ) & & 3.0 & 0.36 & 0.013 + groj1655@xmath540 & 1.5 - 12 & 5.311(7 ) & 2.62 d & 30.0 & 0.30 & 0.056 + gx339@xmath54 & 1.5 - 12 & 4.337(6 ) & 1.75 d & 10.0 & 0.083 & 0.019 + 4u1735@xmath5444 & 1.5 - 12 & 12.739(7 ) & 4.65 h & 3.0 & 0.11 & 0.0089 + gx5@xmath51 & 1.5 - 12 & 69.30(1 ) & & 3.0 & 0.75 & 0.011 + v4641 sgr & 1.5 - 12 & 0.458(9 ) & 2.80 d & 30.0 & 0.13 & 0.28 + 4u1957 + 115 & 1.5 - 12 & 2.366(4 ) & 9.33 h & 3.0 & 0.040 & 0.017 + the estimates of the amplitudes of the periodic signals that are reported in table [ tbl : ampl ] fall , for most of the cases , just below the detection thresholds of the present blind search . they thereby illustrate the typical sensitivity of the present search . we have also estimated the amplitudes that approximately correspond to our detection thresholds in a set of additional sources chosen either to help cover a wide range of intensities or because they are of interest for other reasons . to accomplish this we have computationally superimposed sine waves onto the asm light curves and repeated the search analyses on the light curves for those particular sources in order to determine the signal amplitudes that would have resulted in peaks reaching renormalized whitened powers of @xmath124 in the power density spectra . these estimates , which may be taken as upper limits since we would have recognized peaks near these frequencies that reached powers as small as 15 , are listed in table [ tbl : uplim ] . this table also lists previously reported periods for some of the sources as well as mission - average intensities . column 7 of the table gives the effective upper limits on the amplitudes as fractions of the average source intensities . @xcite reported modest significance detections of periodicities in asm power density spectra for gro j1655@xmath540 , gx339@xmath54 , v4641 sgr , and 4u1957 + 115 at or very close to the periods previously reported in the literature . in the analyses using the entire asm light curves that we have done recently we do not confirm these possible detections . results of sensitivity analyses for these sources are included in table [ tbl : uplim ] . table [ tbl : uplim ] also includes entries for two bright sources , sco x-1 and 4u1735@xmath5444 , with well - established periodicities that we do not detect as well as entries for three sources , i.e. , the crab nebula , gx340 + 0 , and gx5@xmath51 , that do not have known orbital periodicities . table [ tbl : uplim ] gives upper limits , calculated as described above , on any modulation near the reported periods . we note that for sco x-1 the noise is dominated by the intrinsic source variations . in rare cases , such as that described by @xcite , specialized techniques may be available to partially mitigate the effects of the intrinsic variation and improve the sensitivity of a periodicity search . we have not tried to apply any such techniques in the present general search . the upper limits on periodicities in the power spectra for the crab , gx340 + 0 , and gx5@xmath51 are included in this table to illustrate the typical sensitivity of the present search in the cases of bright sources . we also did not detect the alias of the 290-s pulse period in x1145@xmath5619 that was reported in @xcite . this is not so unexpected since pulse periods tend to change more rapidly than orbital periods and could change to a degree that makes them significantly noncoherent and therefore hard to detect in a 14-yr data set . in short , tables [ tbl : ampl ] and [ tbl : uplim ] show that the search sensitivity depends on a number of factors including source strength , degree of intrinsic source variability , and quality of the asm light curve ( which , in turn depends on the source location relative to nearby bright sources , etc . ) . the present search for periodicities in the asm light curves has proved to be more sensitive than the earlier search done by @xcite . the weighting and filtering aspects of our search algorithms have proved to be major factors in the sensitivity improvements . another major factor is the use of light curves that cover @xmath605 more years . the result is the detection of the orbital periods in 8 lmxbs and in @xmath6012 hmxbs that were not detected by @xcite . our detection of the 18.55-d period of igrj18483@xmath50311 represents the co - discovery with @xcite of the orbital period of this system . our detections of the orbital periods of 4u1636@xmath5536 and lmc x-1 are the first for these systems , to our knowledge , that have been accomplished on the basis of x - ray observations . it is interesting to compare the present list of detections together with those of @xcite with tables 1.1 and 1.3 of @xcite that list the well - established orbital periods of lmxbs and hmxbs as of 1994 or 1995 . table 1.1 of @xcite lists 32 lmxbs with known orbital periods . of these , seven have been in quiescence through the entire rxte mission and one ( cal 87 ) is a supersoft source that emits at energies below the asm band ; these systems would not be expected to yield periodicity detections in asm data . the periodicities of 17 of the remaining 24 systems have now been detected in the asm light curves if we include sco x-1 per the results of @xcite even though its orbital period was not detected in the present study . seven systems that are detected by the asm by virtue of their intensities do not yield period detections in the present search . the asm periodicity searches have yielded clear detections of the orbital periods of 13 of the 16 systems listed in table 1.1 of @xcite that had , at that time , shown evidence of the orbital period in x - rays ( see column 5 therein ) . two of the remaining three systems have been in quiescence throughout the rxte mission . the last remaining system is cyg x-2 for which we find no evidence of the 9.8-d period even though it is rather bright in x - rays . table 1.3 of @xcite lists 24 hmxbs with known orbital periods . of these , only one ( x0535@xmath5668 ) has been in quiescence during the entire rxte mission . the orbital periodicities of 16 of the remaining 23 systems have been clearly detected in power density spectra of the asm light curves ; we do not count x0535 + 262 as a clear detection in the power spectrum . the periods of four of the systems ( x0115 + 634 , x0535 + 262 , gx304@xmath51 , and x1145@xmath5619 ) are apparent in the asm light curves as limited sequences of short transient outbursts that are separated by the orbital period or small multiples thereof . the hmxbs x1553 - 542 and x0331 + 530 are evident in the asm light curves during short transient outbursts , but the asm yields no evidence of their orbital periods . we also find no evidence of the 1.7-d period of lmc x-3 even though it detectable over most of the 14-year time of the rxte mission . not unexpectedly , we find that the amplitudes of the variations at the orbital period , when expressed as fractions of the average intensity , are typically low for lmxbs and high for hmxbs ( see tables [ tbl : ampl ] and [ tbl : uplim ] ) . this is consistent with the relatively low rate of detections of orbital modulation in lmxbs and the high rate in hmxbs . there are some exceptions such as the high fractional modulation of the weak lmxb 2s0921@xmath5630 and the low degree of modulation of the hmxbs lmc x-1 and lmc x-3 . our results for 2s0921@xmath5630 suggest that there could be other weak lmxbs with fractional amplitudes that are not particularly low and that would not be difficult to detect with an instrument somewhat more sensitive than the asm . indeed , it is likely that the maxi instrument on the international space station @xcite will produce light curves useful for extending the search for periodicities in x - ray light curves and will be able to find many additional periodicities and orbital periods . we gratefully acknowledge the efforts of the rxte / asm science teams at mit and nasa / gsfc , and the rxte mission support groups at gsfc . we thank alex camacho for computing the upper limits presented in table [ tbl : uplim ] . we acknowledge the support of nasa through contract nas 5 - 30612 ( mit ) .

compact stellar mass objects emit gravitational waves in the good sensitivity band of the planned laser interferometer space antenna ( lisa ) during their last year of inspiral into a supermassive black hole . the gravitational waveforms of emris , extreme mass ratio insiprals , are therefore of interest , and when observed may shed light on the spacetime of the central black hole , including testing the kerr hypothesis @xcite , in addition to being a sensitive tool to measure the central black hole s parameters . this possibility makes such sources of gravitational waves extremely interesting , despite the relatively low number of sources expected during the lifetime of the lisa mission . specifically , tens to hundreds of such sources are expected to redshifts of @xmath1@xmath2 @xcite . in addition , the solution of the two body problem in general relativity remains challenging , and emris orbits and waveforms correspond to its solution in the extreme mass ratio limit . construction of theoretical templates is important both for detection of emris gravitational waves and for accurate parameter estimation . numerical waveforms can be constructed using the frequency domain ( fd ) or the time domain ( td ) approaches . the former approach has been developed to very high accuracy , and is considered robust and accurate . on the other hand , advances to the td approach have been hindered first by the success of the fd approach @xcite , and by the crudity of the initial attempts to evolve numerically the fields coupled to a point - like source with the teukolsky equation @xcite . the breakthrough in the accuracy of td solutions of the inhomogeneous teukolsky equation , i.e. , the 2 + 1d solution of the teukolsky equation coupled to a point mass , was recently achieved in @xcite . for the first time , it was shown that td calculations can be as accurate as fd calculations . the td method of @xcite was improved with the introduction of the discrete delta " model of the source @xcite and an appropriate low pass filter that makes the discrete delta useful also for eccentric or inclined orbits @xcite . specifically , correlation integrals of gravitational waveforms done for the same system in the fd and td approaches show that the two agree to a high level @xcite . one may therefore argue that the two methods are comparable in the results they are capable of producing . we therefore contend that the viewpoint that the td solution of the inhomogeneous teukolsky equations is far from being competitive from the fd solution can no longer be supported . however , we believe that one should not seek competition of the two approaches , but rather how they complement each other , as either method has non - overlapping strengths . in order to achieve this goal one needs to compare the computational efficiency of the two approaches . the question of the efficiency and the computation time with which the results are obtained remains an open question though . the common wisdom is that the fd computation is more effective computationally than its td counterpart : fd approaches are particularly convenient when the system and the emitted gravitational waves exhibits a discrete set of frequencies . indeed , as shown by schmidt @xcite and by drasco and hughes @xcite , all bound kerr orbits have a simple , discrete spectrum of orbital frequencies . however , generic orbits and in particular high - eccentricity orbits , although in principle amenable to a fourier decomposition and a fd construction of the waveforms , require the summation of many terms in the fourier series . this problem limits the accuracy and increases the computation time in fd calculations . while this statement , or similar ones , appear in the literature @xcite , it has not been quantified . the motivation of this paper is to study the relative computational efficiency of fd and td codes for the solution of the inhomogeneous teukolsky equation . specifically , we study the question of how much computational time is needed to find the fluxes of energy and angular momentum to infinity from an eccentric and equatorial orbit around a fast spinning kerr black hole . we restrict the analysis here to equatorial orbits . analysis of equatorial orbits is non - trivial , and teaches much of the method and properties also of non - equatorial orbits . analysis of non equatorial orbits increases considerably the volume of the parameter space , and we leave its study to the future . we do , nevertheless , comment on the analysis of generic orbits . the estimation of the total needed computation time for td codes is relatively simple : after the desired accuracy level is set , one needs to estimate the number of azimuthal @xmath0 modes ( associated with the @xmath3 orbital angular frequency ) required for the total sum over @xmath0 modes to achieve the desired accuracy , and then compute each @xmath0 mode at the same accuracy level . the number of needed @xmath0 modes is not hard to estimate , because the partial fluxes approach a geometric progression in @xmath0 for large @xmath0 values , with the asymptotic factor between two successive partial fluxes depending on the eccentricity of the orbit . after the needed number of @xmath0 modes is determined , one can use the td code of @xcite ( possibly with the improvements included in @xcite ) to calculate the partial fluxes and sum over them . using the asymptotic geometric progression structure one may also estimate the error associated with the summation over the partial fluxes , and verify that the desired accuracy level is indeed achieved . this is done in section [ m_modes ] . the td calculation of the individual partial fluxes can be done more efficiently than a straightforward calculation of the fields to very late times and great distances , that are required for an accurate estimate of the fluxes at infinity . specifically , one can make use of the peeling properties of the weyl scalars at great distances , and fit the fields at finite distances to that behavior . one may then use the fitted behavior to extract the behavior at infinity . this is done in appendix [ appendixa ] . we find this method to be cleaner and better motivated physically than the fit done in @xcite , that uses a general two parameter fit function of a form which is not strongly motivated physically . the estimation of the total needed computation time for fd codes starts similarly to the td analysis : specifically , one needs to first determine the number of @xmath0 modes needed for the determined accuracy level . this can be done in a similar way to how it is done in the td . the calculation of each @xmath0 mode in the fd approach requires the summation over a number of @xmath4 modes , that correspond to the radial and angular ( about the equatorial plane ) orbital angular frequencies and the radiative multipole @xmath5 . because we focus attention on equatorial plane orbits , the @xmath6 modes trivialize , and in practice only the @xmath7 modes are of importance . we find in section [ k_mode ] that the number of @xmath7 modes that one needs to sum over to achieve the desired accuracy level increases with the eccentricity and with the corresponding value of @xmath0 in an intricate manner . next , in section [ comp_time ] we estimate the computation time of each @xmath7 mode , and find it is a function of @xmath7 . the behavior of the computation time of the @xmath7 mode is found to have a rather intricate dependence of the value of @xmath7 . we use this behavior to approximate the computation time over a sum of @xmath7 modes up to some @xmath8 , and argue it is approximately quadratic in @xmath8 for high @xmath8 values . then , in section [ compare ] we find the total computation time for all the fd modes on the same machine on which we perform the td computation , and compare the two . we find that the fd code is more efficient at low @xmath0 values ( for a twofold reason : because those require few @xmath7 modes , and for low @xmath0 values each @xmath7 mode takes less time to compute ) , but the computation time increases rapidly with the value of @xmath0 . we find the growth rate of the computation time with the mode number @xmath0 to increase as a power of @xmath0 , with the value of the exponent increasing with the eccentricity . finally , we estimate that for generic orbits the two methods become comparable already for moderately high values of the eccentricity , in the range of @xmath9@xmath10 . higher eccentricity orbits are more efficiently calculated using the td approach . even when the calculation of the sum of all @xmath0 modes is done more efficiently using one method over the other , one may still compute the low @xmath0 modes using the fd approach and the high @xmath0 modes using the td approach . such a hybrid method may prove to be the most efficient computationally . one should also consider the fact that the total needed number of @xmath7 modes is an empirically found number . when the full parameter space is mapped , one may tabulate the numbers of required modes as functions of the orbital parameters . until this is done , extra computation time is needed to find the number of needed modes , including the computation of some modes that are found _ a posteriori _ to be unneeded . no similar problem occurs in the td approach . to obtain the sum over all @xmath0 modes we need to obtain results for high @xmath0numbers , and find a way to determine ( i ) how many @xmath0 modes we need to sum over to get the total flux to a certain pre - determined accuracy , and ( ii ) estimate the error in neglecting all the higher @xmath0 modes . in the case of circular and equatorial orbits , finn & thorne @xcite show that ( see @xcite for more details and for definitions ) @xmath11 ^ 2}\,\eta^2{\tilde \omega}^{2 + 2m/3}\,{\dot{\cal e}}_{\infty\ , m}\ ] ] which has the nice property that @xmath12 as @xmath13 . ( notably , this property depends on the factor of @xmath14 . ) we may therefore approximate the sum over infinitely many modes by taking the series to be a geometric progression . for eccentric orbits we no longer have the finn thorne formula , but we can do numerical experiments to test whether the same results holds also for such orbits . we first show in table [ table5 ] the average fluxes obtain for a number of eccentric orbits ( with a fixed value of the semilatus rectum @xmath15 ) as a function of the mode @xmath0 . the data presented is table [ table5 ] is intentionally coarse , and is not intended to be more accurate than at the 510% level . below , we also present similar data with higher accuracy , where the latter are needed . we plot in fig . [ fig1 ] @xmath16 as a function of @xmath0 for several eccentric orbits . we see that for all values of @xmath17 , the drop off rate is exponential in @xmath0 for large values of @xmath0 . notably , the highest flux is in the fundamental mode , @xmath18 , and is highest for @xmath19 . this result is indeed expected : the radiated power in gravitational waves averaged over one period for a point mass @xmath20 in a keplerian orbit around a schwarzschild black hole of mass @xmath21 is given by the quadrupole formula to be @xcite @xmath22 recalling that the semimajor axis @xmath23 is related to the semilatus rectum @xmath15 by @xmath24 , we find that for fixed @xmath15 , @xmath25 which has a maximum at @xmath26 @xcite . as the partial flux in any other mode is suppressed by over an order of magnitude compared with the fundamental mode ( @xmath18 ) , this result for the total flux is carried over to the fundamental mode . notably , the larger @xmath0 , we find numerically that the larger the eccentricity for which most flux is obtained . 0.7 & 0.8 & 0.9 1 & @xmath27 & @xmath28 & @xmath29 & @xmath30 & @xmath31 2 & @xmath32 & @xmath33 & @xmath34 & @xmath35 & @xmath36 3 & @xmath37 & @xmath38 & @xmath39 & @xmath40 & @xmath41 4 & @xmath42 & @xmath43 & @xmath44 & @xmath45 & @xmath46 5 & @xmath47 & @xmath48 & @xmath49 & @xmath50 & @xmath51 6 & @xmath52 & @xmath53 & @xmath54 & @xmath55 & @xmath56 7 & @xmath57 & @xmath58 & @xmath59 & @xmath60 & @xmath61 8 & @xmath62 & @xmath63 & @xmath64 & @xmath65 & @xmath66 9 & & @xmath67 & @xmath68 & @xmath69 & @xmath70 10 & & @xmath71 & @xmath72 & @xmath73 & @xmath74 11 & & @xmath75 & @xmath76 & @xmath77 & @xmath78 12 & & @xmath79 & @xmath80 & @xmath81 & @xmath82 13 & & @xmath83 & @xmath84 & @xmath85 & @xmath86 [ table5 ] epsf mode to the partial - sum ( up to @xmath0 ) flux compared with the preceding mode , as a function of @xmath0 , for various values of the eccentricity @xmath87 . same parameters as in fig . [ fig1].,title="fig:",width=415 ] mode to the partial - sum ( up to @xmath0 ) flux compared with the preceding mode , as a function of @xmath0 , for various values of the eccentricity @xmath87 . same parameters as in fig . [ fig1].,width=415 ] we next consider the ratio of the relative contribution of the @xmath0 mode to the partial - sum ( up to @xmath0 ) flux compared with the preceding mode , as a function of @xmath0 . an asymptotic drop off of the flux corresponds to this ratio approaching a constant value as @xmath13 . in fig . [ fig2 ] we plot this ratio as a function of @xmath0 for various values of the eccentricity . lastly , we extrapolate the curves in fig . [ fig2 ] to find the asymptotic value for the ratio , and plot it in fig . [ fig3 ] as a function of the eccentricity @xmath17 . the best fit curve is given by @xmath88 with a correlation coefficient @xmath89 . epsf as a function of the eccentricity @xmath87 . the individual error bars are shown , in addition to the best fit curve ( solid ) and @xmath90 confidence curves ( dotted).,title="fig:",width=415 ] as @xmath16 behaves asymptotically like a geometric progression , we can approximately sum over all modes ( provided sufficiently many modes are calculated , so that the sequence of partial fluxes already converges approximately to the asymptotic behavior ) . specifically , calculating the sequence @xmath91 , we can calculate the partial sums @xmath92 , and the remainder is approximated by @xmath93 , where @xmath94 is the asymptotic ratio evaluated above . one can then approximate the total flux by @xmath95 , and use @xmath96 as a measure for the error . we can therefore answer questions such as how many modes @xmath0 do we need to sum over to obtain a desired accuracy level ? " ( assuming each mode individually has the desired accuracy level ) . we show this in fig . [ fig4 ] , that shows contour curves at a fixed accuracy level on the number of @xmath0-modes eccentricity plane . epsf -modes necessary for the sum over @xmath0-modes to have a certain accuracy . we plot contour curves at a fixed level of accuracy , showing the number of @xmath0-modes needed as a function of the eccentricity @xmath87 . the contours are at accuracy levels of @xmath97 , and @xmath98.,title="fig:",width=415 ] based on fig . [ fig4 ] , to determine the energy flux to 10% accuracy requires 4 modes at all values of the eccentricity @xmath17 . we may therefore find the approximate total waveform by summing over the 4 modes of greatest energy flux . for @xmath99 , @xmath100 , and @xmath101 , we show in figs . [ fig5re ] and [ fig5im ] the waveform obtained when summing over these 4 modes , specifically @xmath102 . indeed , the phase information is captured well by the sum of these 4 modes , and is no longer changing significantly by addition of more modes . epsf for @xmath99 , @xmath100 , and @xmath101 for @xmath103 ( dotted ) , @xmath104 ( dash - dotted ) , @xmath105 ( dashed ) and @xmath106 ( solid ) , as a function of @xmath107 . , title="fig:",width=415 ] for @xmath99 , @xmath100 , and @xmath101 for @xmath103 ( dotted ) , @xmath104 ( dash - dotted ) , @xmath105 ( dashed ) and @xmath106 ( solid ) , as a function of @xmath107 . , width=415 ] epsf . shown are the phases corresponding to @xmath103 ( @xmath108 ) , @xmath104 ( @xmath109 ) , @xmath105 ( @xmath110 ) and @xmath106 ( @xmath111 ) , as a function of @xmath107 . , title="fig:",width=415 ] the behavior of the @xmath0-modes that corresponds to the angular frequency @xmath112 is common to both time - domain and frequency - domain approaches . in the td case , however , the study of the behavior of the @xmath0-modes summarizes all the modes that need to be considered to get the full waveform and the total fluxes radiated . in the fd case , however , one needs to consider also the modes that arise from the three different frequencies of the problem : the @xmath7-modes , corresponding to the radial angular frequency @xmath113 , and the @xmath114-modes , corresponding to the inclination angle oscillations with angular frequency @xmath115 . ( we note that various authors exchange the notation of the latter two frequencies , @xmath116 . ) here , we concentrate on the equatorial plane , and defer discussion on motion outside the equatorial plane to a sequel . the fd code that we use is based on the code presented in @xcite , where more details on the code can be found in addition to descriptions of the tests done to check the code and the sources for errors . the fd numerical method is based on that of @xcite and solves the sasaki nakamura equation @xcite using burlisch stoer integration . the accuracy level of the code is at the level of @xmath98@xmath117 . values were noticed . these inaccuracies occur for larger values of @xmath7 than those used here , and therefore do not affect our results . ] first indication to the increase in the number of necessary @xmath7-modes with increasing eccentricity was shown in @xcite . it should be noticed , however , that the results presented in @xcite for @xmath118 may be more indication of the failure of the solution of the radial teukolsky equation in the frequency domain and the need for the sasaki nakamura formulation thereof , than a true behavior of the @xmath7-modes . figure [ fig6 ] shows the flux in each @xmath7 mode for @xmath18 , after all @xmath5-modes were summed over , for @xmath99 and @xmath100 for different values of the eccentricity . similar figures were obtained also for other values of @xmath0 . figures [ fig6 ] and [ fig6a ] suggest the following types of behavior : first , the value of @xmath7 for which the flux is maximal shifts to higher values with the increase in eccentricity of the orbit . second , with increasing eccentricity , the flux curve broadens ( mostly for positive values of @xmath7 , so that the curve becomes less and less symmetric in @xmath7 for high @xmath17 ) . we also find ( fig . [ fig6b ] ) that the peak of the flux curves become @xmath0-modal , i.e. , at high values of @xmath17 the flux curve breaks into a number of peaks equal to the mode @xmath0 . similar behavior was reported first in @xcite . the peaks background exhibits interesting oscillations , that appear to be independent ( or , at the most , weakly dependent ) of @xmath119 . the broadening of the flux curves suggests that more @xmath7-modes are required in order to obtained the total flux . in what follows we quantify this statement and make it precise . epsf for @xmath120.,title="fig:",width=415 ] for @xmath120.,width=415 ] epsf , for @xmath18 ( dotted ) , @xmath120 ( dashed ) , and @xmath121 ( solid ) , for @xmath122 . notice that the first two curves repeat data already presented in figs . [ fig6 ] and [ fig6a].,title="fig:",width=415 ] the number of @xmath7-modes required at a fixed value of the eccentricity @xmath17 to obtain the total flux to a desired accuracy increases logarithmically with the accuracy . this behavior is described in fig . [ fig7 ] that plots the error involved in the inclusion of the @xmath123 modes ( of the greatest partial fluxes ) as a function of @xmath123 for various values of the eccentricity @xmath17 . in all cases this behavior is exponential . notice , however , that the slope of the fitted curve becomes steeper with increasing eccentricity . this behavior suggests that as the eccentricity increases , the rate at which the number of @xmath7-modes that are required in order to obtain the same level of accuracy increases . this rate also increases with the mode number @xmath0 ( fig . [ fig8 ] ) . notice , from fig . [ fig6b ] , that also the number of _ negative _ @xmath7-modes increases with the mode number @xmath0 ( for fixed eccentricity ) , although not as fast as the number of positive @xmath7-modes . in fig . [ fig9 ] we show the number @xmath123 of @xmath7-modes required for a given accuracy as a function of the eccentricity @xmath17 , for different values of the mode number @xmath0 . the increase in @xmath123 is exponential in the eccentricity @xmath17 . for a given accuracy level , the exponential increase with @xmath17 becomes steeper as the mode number @xmath0 increases , as is shown in fig . [ fig10 ] . epsf in summing over a number @xmath123 of @xmath7-modes , after summation over all @xmath5 , for @xmath18 , for @xmath99 and @xmath100 : @xmath101 ( @xmath111 ) , @xmath124 ( @xmath125 ) , @xmath126 ( @xmath108 ) , @xmath127 ( @xmath109 ) , @xmath128 ( @xmath129 ) , and @xmath122 ( @xmath130 ) . for each value of the eccentricity , the solid line describes a fitted line as follows , corresponding to increasing eccentricity : @xmath131 ( @xmath132 ) , @xmath133 ( @xmath134 ) , @xmath135 ( @xmath136 ) , @xmath137 ( @xmath138 ) , @xmath139 ( @xmath140 ) , @xmath141 ( @xmath142 ) . for each fitted curve we include the square of the correlation coefficient @xmath143,title="fig:",width=415 ] epsf in summing over a number @xmath123 of @xmath7-modes , after summation over all @xmath5 , for @xmath99 , @xmath100 , and @xmath126 : @xmath18 ( @xmath111 ) , @xmath120 ( @xmath109 ) , and @xmath121 ( @xmath125 ) . for each value of the eccentricity , the solid line describes a fitted line as follows , corresponding to increasing value of @xmath0 : @xmath135 ( @xmath136 ) , @xmath144 ( @xmath145 ) , @xmath146 ( @xmath147 ) . for each fitted curve we include the square of the correlation coefficient @xmath143,title="fig:",width=415 ] epsf of @xmath7-modes required for an accuracy of @xmath117 as a function of the eccentricity @xmath17 . shown are 3 values of @xmath0 , and their corresponding contours , based on a fit to an exponential : @xmath18 ( @xmath111 ) , @xmath120 ( @xmath125 ) , and @xmath121 ( @xmath108 ) . , title="fig:",width=415 ] of @xmath7-modes required for an accuracy of @xmath117 as a function of the eccentricity @xmath17 . shown are 3 values of @xmath0 , and their corresponding contours , based on a fit to an exponential : @xmath18 ( @xmath111 ) , @xmath120 ( @xmath125 ) , and @xmath121 ( @xmath108 ) . , width=415 ] the computation time of a single @xmath7 mode is a function of @xmath7 . to make a prediction of the @xmath7 dependence of the computation time for a @xmath7 mode we first write the radial teukolsky equation @xmath148 where the potential @xmath149 is a complex valued function of @xmath150 and the angular frequency @xmath151 @xcite . the source term @xmath152 is a certain known function of the stress energy of the particle and its trajectory , and of the background geometry . as we are interested here only in the qualitative features of the solutions for eq . ( [ radial_teukolsky ] ) , the details of the source are unimportant for us here . we next focus attention on the far field limit of eq . ( [ radial_teukolsky ] ) . to gain a qualitative understanding of the structure of the solution , we next consider only the asymptotic solution as @xmath153 . in that limit , @xmath154 ( corresponding to an outgoing solution ) , where @xmath155 is the usual kerr spacetime tortoise coordinate defined by @xmath156 . the solution is an oscillatory solution in @xmath150 with angular frequency @xmath151 . the typical length scale over which the solution is oscillating is therefore @xmath157 : the greater @xmath151 , the shorter the distance over which the solution oscillates . the method we use to solve the frequency - domain equation is by a burlisch stoer algorithm @xcite , that successively divides a single step into many substeps , until a ( polynomial or rational ) interpolation is accurate enough . if the original step happens to be too large ( i.e. , after a certain pre - determined substep is calculated and the required accuracy is still not obtained ) , the original step will be bisected , and the process proceeds . therefore , the greater @xmath151 and the shorter the oscillation length scale , the more divisions of the initial step are required , and the greater the number of substeps used to find the solution . as the angular frequency @xmath158 , it is clear that the larger @xmath7 ( for a fixed value of @xmath0 ) , the shorter the distance scale @xmath159 . we therefore expect the computation time of a single @xmath7 mode to increase with @xmath7 . in fact , the burlisch stoer algorithm suggests that as @xmath160 increases , the more substeps are needed , until for some value of @xmath7 dividing the step into substeps is no longer sufficient , and the step is bisected . then , division into substeps is again sufficient , until another @xmath7 value is obtained for which the step is bisected . each time the step is bisected , the total number of substeps computed jumps discontinuously , so that we expect the @xmath7 mode computation time to increase as a staircase function . in fig . [ fig11 ] we show the computation time of an individual @xmath7 mode as a function of @xmath7 , and also the number of burlisch stoer iterations done . they both behave as expected . the preceding discussion suggests that a similar behavior of the computation time is found also as a function of @xmath0 . this is indeed seen in fig . [ fig12 ] . epsf for the @xmath0 modes . here , @xmath161 and @xmath162 . , title="fig:",width=415 ] for the @xmath0 modes . here , @xmath161 and @xmath162 . , width=415 ] as can be seen from figs . [ fig11 ] and [ fig12 ] the discontinuous jumps in the mode computation times occur at different values of @xmath163 . indeed , our discussion about explains this behavior : jumps occur when the angular frequencies @xmath160 arrive at some threshold values , corresponding to threshold wavelengths . as these threshold values should be about the same , we expect jumps to occur for @xmath164 satisfying @xmath165 . notice that @xmath166 ( @xmath167 ) is defined here for vanishing @xmath0 ( @xmath7 ) . for the data presented here , for @xmath127 we find @xmath168 , while @xmath169 ; for @xmath128 we find @xmath170 , while @xmath171 ; and for @xmath122 we find @xmath172 , while @xmath173 , so that the difference between @xmath174 and @xmath175 is at order 10% . we argue that these results support our interpretation of the discontinuous jumps in the computation time . our discussion suggests that for a fixed value of @xmath7 ( @xmath0 ) and varying @xmath0 ( @xmath7 ) , there are values of the latter for which the computational time drops . these are the values that approximately satisfy @xmath176 . ( recall that @xmath177 can be either positive or negative . ) that is , while fixing either @xmath7 or @xmath0 and varying the other , we expect to find a drop in the computation time ( or , equivalently , in the number of burlisch stoer iterations . ) indeed , we find this behavior , as is shown in fig . [ fig13 ] . notice , that for @xmath161 the drop in computational time is found for @xmath178 , and as @xmath7 increases , so does the value of @xmath0 for which the drop is found . we also comment that this drop is broadened with increasing values of @xmath7 . epsf for fixed values of @xmath7 , for the same orbital parameters as above , for @xmath101 . , title="fig:",width=415 ] the dependence of the computational time on the mode number @xmath7 is therefore rather intricate . it can be approximated as follows : for each value of the eccentricity @xmath17 we take in practice the computational time of a single mode @xmath7 to be a linear function of @xmath7 , that fits the computational data . this approximation underestimates the computation time for some @xmath7 modes , specifically those immediately following a discontinuous jump in the computation time , and overestimates the computation time for @xmath7 modes just before a jump . however , as we are mostly interested in the total computation time of the sum over all modes , this approximation may be quite reasonable for the sum over all modes . in this approximation the sum over @xmath7 modes up to some @xmath8 is a quadratic function of @xmath8 . notice , that we can use fig . [ fig7 ] and [ fig8 ] to find how many @xmath7 modes we need to sum over to obtain the required accuracy level . epsf modes from @xmath161 to @xmath179 as a function of the eccentricity @xmath17 . the computation time is presented in units of the maximal computation time in the chosen range of parameters . , title="fig:",width=415 ] finally , we put together all the previous results , to compare the actual total calculation time of td and fd codes . we choose an orbit that gives no special preference for one approach over the other , i.e. , a moderate value for the eccentricity . in practice , we take @xmath128 and @xmath180 . for the fd calculation we first set the desired level of accuracy , which in practice we choose here to be @xmath181 , and then use fig . [ fig4 ] ( or the data in table [ table5 ] ) to determine the number of @xmath0 modes one needs to sum over to guarantee the desired accuracy level . the needed number of @xmath0 modes is found in practice to be 7 , for @xmath182 . for each individual @xmath0 mode we then determine the number of @xmath7 modes necessary to obtain the desired accuracy level . this determination is done with data as in figs . [ fig9 ] and [ fig10 ] . the values of the @xmath7 modes we used in practice are as follows : for @xmath18 we used @xmath183 , for @xmath120 we used @xmath184 , for @xmath121 we used @xmath185 , for @xmath186 we used @xmath187 , for @xmath188 we used @xmath189 , for @xmath190 we used @xmath191 , for @xmath192 we used @xmath193 , and for @xmath194 we used @xmath195 . in all cases we computed @xmath196 . the data are summarized in table [ table6 ] and in fig . [ fig15 ] . the fd calculations were optimized for most efficient calculation for the desired accuracy level . improvements to the computation time can be achieved , but only at the level of the code , e.g. , making the burlisch stoer algorithm more efficient . we believe that while such improvements can be made , their effect would be moderate . in this sense , the fd curve in fig . [ fig15 ] represents a lower bound on the fd calculation time . increasing the eccentricity of the orbit @xmath17 would result in the fd curve of fig . [ fig15 ] moving up in the figure ( increase in the calculation time of each @xmath0 mode , because of increasing number of required number of @xmath7 modes and increase in the computation time of individual @xmath7 modes ) and also having a faster growth rate . modes & fd time ( sec ) & td time ( sec ) 2 & @xmath183 & 13 & 51 & 954 3 & @xmath184 & 17 & 99 & 945 4 & @xmath185 & 20 & 116 & 952 5 & @xmath187 & 24 & 179 & 942 6 & @xmath189 & 25 & 218 & 953 7 & @xmath191 & 28 & 254 & 953 8 & @xmath193 & 30 & 303 & 952 9 & @xmath195 & 35 & 391 & 942 [ table6 ] for the td calculations we used the 2 + 1d code of @xcite , with radial resolution of @xmath197 ( and temporal resolution of @xmath198 ) , and angular resolution of @xmath199 . we placed the inner and outer boundaries at @xmath200 and at @xmath201 , respectively . we approximate the scattering problem on the boundaries by assuming sommerfeld boundary conditions , i.e. , no incoming radiation from outside the boundaries . one may of course make this approximation better by pushing the outer boundary outwards ( and the inner boundary inwards ) . reflections from the boundaries then do not contaminate the center of the computational domain ( at @xmath202 ) until @xmath203 , and we integrate in time until then . in practice , we wait at each evaluation point until the initial spurious waves ( that result from imprecise initial data that correspond to a particle suddenly appearing at @xmath204 ) , and then average the flux over a full period of the orbit . the angular period is @xmath205 , where @xmath150 is the semimajor axis . for our choice of parameters , @xmath206 . the radial period is found to be @xmath207 , and it is the latter ( radial ) period we use for our analysis . we extract the ( radial period averaged ) flux at a number of extraction distances , in practice at @xmath208 , and then fit the ( finite extraction distance ) fluxes to an inverse - square function as in eq . ( [ ansatz ] ) . this allows us to achieve fluxes that agree with the fd fluxes to @xmath181 , while not having to integrate to very late times ( and commensurately increase also the spatial computational domain ) . notably , we can obtain higher accuracy than that reported in @xcite while integrating to a shorter time because we make use of the finite time corrections to the energy flux . we discuss this method in detail in the appendix . the td computational time is presented in table [ table6 ] . it can be made more efficient using code improvements , and data analysis improvements . first , we used gaussian source modeling as in @xcite , and the discrete @xmath209 approach may improve efficiency considerably . in fact , ref . @xcite has argued for a full order of magnitude reduction in computation time . while this statement may be an overestimate of the numerical capabilities especially for generic ( i.e. , eccentric or inclined ) orbits it is certainly possible to improve the efficiency of the td code by changing the method of calculation of the source term . ( the discrete delta approach has also other advantages over the gaussian model . see @xcite for more detail . ) an important property of the td code is that it is very naturally parallelizable . certainly the fd computation is also parallelizable , most naturally by having each @xmath7 mode computed on a different processor . in such a case , the total fd computation time is controlled by the @xmath7 mode that takes the longest to compute , i.e. , the largest @xmath7 ( see fig . [ fig11 ] ) . another improvement of efficiency of the fd code may come from improvements of the burlisch stoer engine , specifically the solution for @xmath210 instead of @xmath150 . it is currently unknown how the computational efficiency and numerical accuracy would be affected by such a change . one factor that limits a possible reduction of the td computation time is the need to average over an orbital ( radial ) period to find the average flux per period . even for the strong field orbit considered here , with @xmath100 and @xmath128 , the radial period is @xmath211 , which puts a lower bound on how short the computational integration can be . one could perhaps shorten the computation time if only half an orbital period is computed , but because the phase of the orbit at the end of the spurious wave epoch is arbitrary , on the average at least three quarters of an orbit need to be computed . however , we were interested here in comparison of the fluxes to infinity . if one is interested in the wave function itself , shorter computation times can be used , thus saving much of the td computation time . no equivalent save in time can be achieved with the fd computation . reduction of the td computation time would result in the dashed line in fig . [ fig15 ] moving lower , thus making the computational time of td and fd comparable already for lower @xmath0 values . most importantly , when higher eccentricity orbits are considered , the td computation time remains unchanged , whereas the fd computation time increases considerably . more specifically , not only is the solid curve in fig . [ fig15 ] shifted up , it also becomes steeper ( i.e. , its rate of growth increases ) , so that the td and fd computation times become comparable for lower @xmath0 values . in addition , the td calculation does not require any extra computation to find the waveform , as it is already available . in fact , an extra computation was needed to find the fluxes . on the other hand , the fd calculation requires an extra computation to produce the waveform , which we have not done here . therefore , the greater efficiency of fd over td for the modes shown here is much less impressive when the computation time for waveforms is of interest . we therefore suggest that at high eccentricity orbits the computational efficiency of the td and the fd approaches is comparable for high @xmath0 values . we believe that the question of the total ( sum over @xmath0 ) computation time is not necessarily a very important one , and suggest that for actual computations the low @xmath0 modes are calculated using the fd approach , and the high @xmath0 modes are calculated using the td approach . the determination of which @xmath0 values are low and which are high depends of course on the parameters of the system , and the accuracy level of the computation . using a single method , however , could still possibly make td more efficient than fd for very high values of the eccentricity . epsf and @xmath128 as a function of the mode number @xmath0 . the circles ( @xmath111 ) denote fd computation time , and the squares ( @xmath109 ) show td computation times . both sets of runs were done on the same computer , in this case a single processor of a dual 2.7ghz powerpc g5 . the red curve is the fitted curve @xmath212 , that fits the numerical data with a squared correlation coefficient of @xmath213 . for both cases the accuracy level is set at @xmath181 . , title="fig:",width=415 ] in the analysis of this paper we have focused on equatorial orbits . such orbit simplify the problem not just computationally : the number of dimensions in the fd parameter space is significantly lower than that for generic orbits . while the analysis of equatorial orbits presented here is non - trivial , the added complications of handling generic orbits warrant separate treatment . when generic orbits are considered , finding the number of @xmath5 modes required to achieve a set accuracy level with the fd approach becomes a non - trivial problem , as is the question of finding the computational time of each @xmath5 mode as a function of the mode number . notably , the td calculations remain unchanged , as is the associated computation time . we therefore expect the fd computation time to increase significantly with the transition from equatorial to inclined and then generic orbits , an increase that has no counterpart in the td case . to do such an analysis , one needs to check the behavior of all combinations of the modes @xmath214 , find the minimal number of modes required for a set accuracy level , and then compute those modes . the td calculation requires that the same number of @xmath0 modes are computed , and this computation time depends only on the duration of the computation , the size of the computational domain , and of course the grid resolution . modest improvements to the td computation time may be achieved by improving the efficiency of the source calculations ( e.g. , the discrete delta approach ) , but they are not expected to be as dramatic as previously suggested @xcite . we expect the threshold eccentricity above which the td approach is more computationally efficient than the fd approach for some of the needed @xmath0 modes to be moderate high , specifically in the range of @xmath9@xmath10 . the authors are indebted to kostas glampedakis for discussions and for making his fd code available to us . lmb was supported in part by nasa / gsfc grant no . ncc5580 , nasa / ssc grant no . nnx07al52a , and by a minigrant from the office of the vice president for research at uahuntsville . gk is grateful for research support from the university of massachusetts ( dartmouth ) college of engineering , the hpc consortium and sony corporation of america ( scea ) . djl was supported by a research experiences for undergraduates in science and engineering fellowship , sponsored by the alabama space grant consortium under contract nng05ge80h and by the uahuntsville president s office . the peeling theorem describes how the various components of the gravitational field of a bound gravitational system behave as the observer moves away from the source ( see , e.g. , @xcite ) . specifically , in the wave zone of a radiating system , the @xmath215 component of the weyl curvature which describes in the wave zone the outgoing transverse radiative degrees of freedom typically drops off as @xmath216 , where @xmath150 is the distance from the radiating system . the peeling theorem also gives the drop off rates for the remaining weyl scalars , and also for spin coefficients and other newman penrose quantities . in addition to the leading order decay rate , the peeling theorem has been extended also to include correction terms in @xmath217 . specifically , newman and unti showed that @xcite @xmath218 where @xmath219 is a spin coefficient and @xmath220 is an integration constant " from the radial equations . a superscript @xmath221 " means the coefficient of the slowest decaying term in an expansion in @xmath217 . as the flux of energy in gravitational waves at infinity @xmath222 , @xmath223 is directly related to the flux at infinity . notably , the leading order corrections in @xmath217 to @xmath215 are proportional to the _ asymptotic _ value of the weyl scalar @xmath224 and its gradient . in the teukolsky formalism , that conveniently describes black hole perturbations , the weyl scalar @xmath225 ( transverse frame " ) , so that the leading order correction to @xmath215 is at the next order : @xmath226 . therefore , @xmath227 ^ 2\sim ( \psi_4 ^ 0)^2\,[1+o(r^{-2})]\ , .\ ] ] we therefore are motivated to introduce the ansatz @xmath228 \label{ansatz}\ ] ] where @xmath229 , and @xmath230 is the boyer lindquist radius of the orbit . notice that @xmath231 , so that our ansatz can be written as @xmath232 $ ] . note that the expansion parameter @xmath233 , as the field is evaluated far from the radiating system , specifically outside the light cylinder . " this ansatz is also suggested by fig . 1 of ref . @xcite and by table vi of ref . @xcite , where a best fit was done to the ansatz @xmath234 $ ] , and the free parameter @xmath15 was found numerically to be rather close to 2 ( with deviations of @xmath235 ) for a large range of orbits . we test our ansatz by fitting the outgoing flux of energy for a number of circular and equatorial orbits around a kerr black hole as detected at a sequence of distances from the radiating system to the ansatz ( [ ansatz ] ) . in table [ table1 ] we show the outgoing fluxes from particles in circular orbits around a kerr black hole , taken at a sequence of distances . table [ table2 ] shows the squared correlation coefficient for the various cases we have checked . in all cases the high value of the correlation coefficient corroborates our ansatz . in table [ table3 ] we show the values of the free parameter @xmath236 for the same cases . lastly , in table [ table4 ] we compare the flux to infinity based on the ansatz ( [ ansatz ] ) to the flux obtained from a frequency - domain calculation . notably , the associated errors appear to be larger for @xmath237 than for @xmath238 . we attribute these higher errors to the lower values of the flux for high values of @xmath0 . specifically , to obtain the flux with the time domain one needs to first subtract the flux " due to the spurious radiation associated with the initial time of the simulation @xcite . as the fluxes become smaller , this subtraction becomes less accurate . @xmath239 & @xmath240 & @xmath241 & @xmath242 & @xmath243 & @xmath244 & @xmath245 1 & 4.0 & 0.99 & @xmath246 & @xmath247 & @xmath248 & @xmath249 & @xmath250 & @xmath251 2 & 4.0 & 0.99 & @xmath252 & @xmath253 & @xmath254 & @xmath255 & @xmath256 & @xmath257 3 & 4.0 & 0.99 & @xmath258 & @xmath259 & @xmath260 & @xmath261 & @xmath262 & @xmath263 4 & 4.0 & 0.99 & @xmath264 & @xmath265 & @xmath266 & @xmath267 & @xmath268 & @xmath269 5 & 4.0 & 0.99 & @xmath270 & @xmath271 & @xmath272 & @xmath273 & @xmath274 & @xmath275 1 & 10 & 0.90 & @xmath276 & @xmath277 & @xmath278 & @xmath279 & @xmath280 & @xmath281 2 & 10 & 0.90 & @xmath282 & @xmath283 & @xmath284 & @xmath285 & @xmath286 & @xmath287 3 & 10 & 0.90 & @xmath288 & @xmath289 & @xmath290 & @xmath291 & @xmath292 & @xmath293 4 & 10 & 0.90 & @xmath294 & @xmath295 & @xmath296 & @xmath297 & @xmath298 & @xmath299 5 & 10 & 0.90 & @xmath300 & @xmath301 & @xmath302 & @xmath303 & @xmath304 & @xmath305 1 & 10 & 0.99 & @xmath306 & @xmath307 & @xmath308 & @xmath309 & @xmath310 & @xmath311 2 & 10 & 0.99 & @xmath312 & @xmath313 & @xmath314 & @xmath315 & @xmath316 & @xmath317 3 & 10 & 0.99 & @xmath318 & @xmath319 & @xmath320 & @xmath321 & @xmath322 & @xmath323 4 & 10 & 0.99 & @xmath324 & @xmath325 & @xmath326 & @xmath327 & @xmath328 & @xmath329 5 & 10 & 0.99 & @xmath330 & @xmath331 & @xmath332 & @xmath333 & @xmath334 & @xmath335 1 & 12 & 0.0 & @xmath336 & @xmath337 & @xmath338 & @xmath339 & @xmath340 & @xmath341 2 & 12 & 0.0 & @xmath342 & @xmath343 & @xmath344 & @xmath345 & @xmath346 & @xmath347 3 & 12 & 0.0 & @xmath348 & @xmath349 & @xmath350 & @xmath351 & @xmath352 & @xmath353 4 & 12 & 0.0 & @xmath354 & @xmath355 & @xmath356 & @xmath357 & @xmath358 & @xmath359 5 & 12 & 0.0 & @xmath360 & @xmath361 & @xmath362 & @xmath363 & @xmath364 & @xmath365 0.999899201 & 0.9956862 & 0.99507507 ii & 0.99999139 & 0.99996189 & 0.99996031 & 0.999946517 & 0.99994224 iii & 0.99999644 & 0.99996171 & 0.99995783 & 0.998752311 & 0.99997298 iv & 0.99978769 & 0.99994700 & 0.99997873 & 0.99999386 & 0.99997505 0.86647(5 ) & 1.9655(3 ) & 1.9622(4 ) ii & 2.3812(2 ) & 0.6156(4 ) & 0.6915(4 ) & 0.7751(5 ) & 0.7716(5 ) iii & 2.4132(1 ) & 0.6224(4 ) & 0.6976(4 ) & 0.799(2 ) & 0.7734(5 ) iv & 2.250(2 ) & 0.5857(7 ) & 0.6287(5 ) & 0.6950(3 ) & 0.6917(5 ) @xmath239 & @xmath366 & @xmath367 & @xmath368 1 & 4.0 & 0.99 & @xmath369 & @xmath370 & -0.0187 2 & 4.0 & 0.99 & @xmath371 & @xmath372 & -0.0047 3 & 4.0 & 0.99 & @xmath373 & @xmath374 & 0.0023 4 & 4.0 & 0.99 & @xmath375 & @xmath376 & -0.0155 5 & 4.0 & 0.99 & @xmath377 & @xmath378 & -0.0051 1 & 10 & 0.90 & @xmath379 & @xmath380 & -0.0154 2 & 10 & 0.90 & @xmath381 & @xmath382 & 0.0018 3 & 10 & 0.90 & @xmath383 & @xmath384 & 0.0010 4 & 10 & 0.90 & @xmath385 & @xmath386 & -0.0125 5 & 10 & 0.90 & @xmath387 & @xmath388 & -0.0123 1 & 10 & 0.99 & @xmath389 & @xmath390 & -0.0162 2 & 10 & 0.99 & @xmath391 & @xmath392 & -0.0002 3 & 10 & 0.99 & @xmath393 & @xmath394 & 0.0007 4 & 10 & 0.99 & @xmath395 & @xmath396 & -0.0023 5 & 10 & 0.99 & @xmath397 & @xmath398 & -0.0042 1 & 12 & 0.0 & @xmath399 & @xmath400 & -0.0073 2 & 12 & 0.0 & @xmath401 & @xmath402 & 0.0033 3 & 12 & 0.0 & @xmath403 & @xmath404 & 0.0010 4 & 12 & 0.0 & @xmath405 & @xmath406 & -0.0114 5 & 12 & 0.0 & @xmath407 & @xmath408 & -0.0107 j. r. gair , c. li , and i. mandel , phys . d*77 * , 024035 ( 2008 ) and references cited therein . j. r. gair , l. barack , t. creighton , c. cutler , s. l. larson , e. s. phinney , and m. vallisneri , class . quantum grav . * 21 * , s1595 ( 2004 ) . k. glampedakis and d. kennefick , phys . d * 66 * , 044002 ( 2002 ) . s. a. hughes , phys . d * 61 * , 084004 ( 2000 ) . r. lopez aleman , g. khanna , and j. pullin , class . quantum grav . * 20 * , 3259 ( 2003 ) . g. khanna , phys . d * 69 * , 024016 ( 2004 ) . l. m. burko and g. khanna , europhysics letters * 78 * , 60005 ( 2007 ) [ arxiv : gr - qc/0609002 ] . a. sundararajan , g. khanna , and s. a. hughes , phys . rev . d ( in press ) [ arxiv : gr - qc/0703028 ] . a. sundararajan , g. khanna , s. a. hughes , and s. drasco , arxiv:0803.0317 . w. schmidt , class . quantum grav . * 19 * , 2743 ( 2002 ) . s. drasco and s.a . hughes , phys . rev . d * 69 * , 044015 ( 2004 ) . l. s. finn and k. s. thorne , phys . d * 62 * , 124021 ( 2000 ) . p. c. peters and j. mathews , phys . rev . * 131 * , 435 ( 1963 ) . l. s. finn , private communication . hughes , phys . d * 61 * , 084004 ( 2000 ) ; erratum - ibid . d * 63 * 049902 ( 2001 ) . c. cutler , l. s. finn , e. poisson , and g. j. sussman , phys . d * 47 * , 1511 ( 1993 ) . m. sasaki and t. nakamura , prog . * 67 * , 1788 ( 1982 ) . w.h . press , s.a . teukolsky , w.t . vetterling , and b.p . flannery , _ numerical recipes : the art of scientific computing _ , third ed . ( cambridge university press , cambridge , 2007 ) . h. stephani , _ relativity : an introduction to special and general relativity _ , 3@xmath409 ed . , ( cambridge university press , cambridge , 2004 ) . j. stewart , _ advanced general relativity _ ( cambridge university press , cambridge , 1990 ) . e. t. newman and t. unti , j. math . phys . * 3 * , 891 ( 1962 ) .

reliable determination of total mass of spiral galaxies is a crucial step in the search for non - baryonic dark matter . in particular , laboratory experiments aimed at detection of dark matter particles , could succeed only if there is sufficient amount of dark matter inside the disk of our galaxy in the sun s vicinity . present models of mass distribution in spiral galaxies assume that in addition to stellar disk , bulge , luminous halo , there is massive halo of non - baryonic dark matter , most often assumed spherical . it dominates the total galaxy mass and also governs the dynamics of outer parts of the stellar disk . this last feature makes it possible to test in some cases , if spherically symmetric rather than flattened dark matter distribution is actually present . recently , we have analyzed mass distribution in several spiral galaxies @xcite . we find that for some galaxies it can not be spherical at larger radii . since we expect in this case a flattened mass distribution to better approximate the gravitational potential at larger radii than the spherical one , we apply the global disk model . we then find that mass distribution of luminous matter accounts for rotation curves of the examined galaxies . there is also increasing number of studies whose authors , using more involved galaxy models than so far , conclude , contrary to earlier findings , that luminous matter accounts for rotation in the internal galactic regions @xcite . there is also possibility that rotation of galaxies in the outermost regions can be driven by magnetic fields , not by dark matter @xcite . these results need to be seen properly in the context of non - baryonic dark matter searches . currently , there is strong observational evidence of non - baryonic dark matter in clusters of galaxies , in particular , from gravitational lensing by clusters and from colliding clusters , as e.g. the bullet cluster @xcite . also , the cosmological model provides support to the non - baryonic dark matter hypothesis , albeit at a more theoretical level . on the other hand , direct searches give null results , except for dama / libra @xcite , however , these results are disputable and have not been confirmed by other searchers . the consistent picture might emerge if one assumed that clustering scale of non - baryonic dark matter is that of galaxy clusters . at smaller scales , non - baryonic dark matter could be dilute . in particular , its contribution to spiral galaxy masses could be much less than anticipated in spherical halo models . it can not also be excluded that spiral galaxies may differ from each other in abundance of dark matter some of them could be reach in dark matter , the other might be devoid of it . before presenting results for some spiral galaxies obtained with our model based on a global disk approximation , we briefly signalize some important issues concerning rotation curve modeling . the methods for ascertaining mass distribution in spiral galaxies commonly use very simplified , one would even say naive , one- or many - component galaxy models , that assume symmetries of mass distribution ( axial , spherical ) , properties of rotation fields ( concentric circular ) , and composition characteristics ( constant @xmath0 ratios , _ etc _ ) , which at most approximately agree with observations or sometimes even contradict them . our model of spiral galaxies is also simplified , however , it possess some features which , according to us , gives it the advantage over other models e.g. it does not assume in advance any m / l ratio or its constancy , but is obtained at output as a resulting m / l ratio profile , which seems more realistic . in the past , the prevailing opinion was that one needs dark matter inside galaxies to account for their rotation and to stabilize their allegedly cold disks , whereas there is increasing number of papers modeling the same galaxies without dark matter . this is achieved by considering sufficiently complex and thus also more `` flexible '' models , for example , those assuming wide , not thin , stellar disks , ect . there is also too much freedom in choosing galaxy models , and there is no general agreement which models are closer to reality . but , why to assume circular orbits , axial or spherical symmetry , constant mass - to - light ratios when this is not the case for real galaxies ? as is pointed out by @xcite , most of spirals are not differentially rotating axisymmetric disks with only tiny disturbances due to spiral arms . there are large scale deviations from the symmetry , often with very large amplitudes or even with warped hi layers @xcite . the predictions of the simple galaxy models might further be modified if one preferred to reconstruct whole doppler images rather than rotation curves , the more the latter are only some interpretations of doppler images obtained not without additional assumptions . going even further , what would happen if , instead of using the simplified models , one used a sort of templates for inferring matter distribution in real galaxies based on their doppler images and the observed qualitative characteristics like the number of arms , the presence of a bar _ etc _ ? such templates , relating mass distribution and its qualitative shape features to simulated doppler images could be catalogued by elaborating a large number of galaxy - like virialized stationary systems from n - body simulations . would nt this approach be the right step toward realistic modeling of mass distribution in galaxies ? to what extent could this approach change our understanding of spiral galaxies and current opinions about their composition ? the question how much dark matter is present in galaxy s interiors remains still open @xcite , @xcite @xcite point out that the better are the data used for constraining the properties of galaxy halos and the better is the quality of simulations , there is more difficulties with dark matter models on galaxy scales . also @xcite says that the case for cdm , having < < a number of stubborn problems > > , is weak judged on the data from galactic scales only . this is especially evident from the still growing number of examples of morphologically various galaxies whose rotation curves can be accounted for without dark matter hypothesis by using non - canonical , more realistic approach to mass modeling . for example , even though still assuming the simplifying unrealistic axi - symmetry and that the radial mass profile should follow that of luminosity with constant @xmath0 ratio , @xcite were able , under the strict maximum ( thick ) disk hypothesis , that is , not including a dark halo , to reproduce with i - band @xmath0 ratio of @xmath1 and consistently with stellar population synthesis models the overall structure of the optical rotation curves for most of galaxies in their sample of 74 spiral galaxies with various surface brightness profiles and rotation curve morphologies . they link the poorest fits with non - axisymmetric features , like bars or strong spiral arms , that influence the brightness profiles , the determination of inclination , major - axis position angle , _ etc_. this situation , however , might be cured by considering more realistic models . that the interiors of spiral galaxies can not be dominated by dark matter is also suggested by dynamical arguments . as noted by @xcite , if the average projected surface mass density of dark matter halo was greater than that of the optical disk , the common instabilities would be suppressed @xcite together with lopsided modes @xcite appearing in many disk galaxies . in addition , as they point out , there are models that reproduce features of rotation curves within the optical disks on different scales and that the models rarely require dark matter halos @xcite . these interesting observations of absence of dark matter inside galaxies are best expressed by @xcite . they speculate , having succeeded in fitting some galaxies without dark matter , whether occam s razor argument would not be applicable . why to introduce something that is not needed ? the authors also form less authoritative statements about dark matter presence in galaxies than others . what they merely say is that dark matter presence is only statistically significant . the presence of a dark matter component that dominates the outer mass distribution in spiral galaxies , seems evident from their optical rotation curves which do not fall off sufficiently fast close to their optical edge @xcite and from their extended hi rotation curves @xcite . however , in the context of what has been said above , one can also conjecture that the halo would disappear if the galaxy models were more realistic allowing for changing mass - to - light ratios , assuming non - axisymmetric mass features ( such as spiral arms and bars ) , if the assumed extinction models was altered_etc_. such models should be also taking into account that the velocity field in galaxies is much more complicated than circular and that the published rotation curves are obtained by reducing a two - dimensional doppler image of a galaxy into a single one - dimensional curve ( such an image by itself carries already a reduced , partial information about rotation ) , and that this reduction is performed not without additional assumptions concerning the interpretation , averaging methods , _ etc _ , of the doppler image ( for example , an averaging method which produces rotation curve by calculating pairs @xmath2 $ ] scanning a doppler image , where the averaging at a given @xmath3 is carried out over a window of size @xmath4 and weighted by , say , a function of local intensity of the image within the window , is presumably good for circular orbits of test bodies in a spherically symmetric gravitational potential , but not necessarily for a self - gravitating system which may be flattened ) . as concerns dark matter in outer galaxy regions , there is also possibility , never taken into account in modeling of flattened galaxies , that rotation of slightly ionized gas at large radii may be driven also by magnetic fields whose strength may be comparable with that of gravitation of whole galaxy at these radii @xcite , helping to lower ( or even eliminate ) the amount of dark matter needed in the outskirts of galaxies where rotation curves are usually flat @xcite . it is true that realistic modeling of galaxy rotation is very challenging and might be very difficult to achieve , but , at the same time it is not clear whether the naive models used so far , lead to trustworthy results has the opposite ever been rigorously proved ? in modeling of galactic optical discs , it is assumed they are highly flattened , axisymmetric and cold . there is a common opinion linked with stability argument of @xcite , that for such disks to be stable within the optical radius , one necessarily needs the galaxies to be immersed within massive dark matter haloes . this conclusion comes about as follows . as observed in numerical simulations , when random stellar motions are small compared to the streaming rotational motion in flattened galaxies , without such a halo a bar would form in one rotation period , and only afterward axial symmetry would be regained . during this stage , approaching stationary state , the disk would heat up and could not remain cold ( this would not agree with the assumption that disks are cold ) . therefore , cold disks must be immersed in ( usually spherical ) halos of sufficiently large mass . there are various criteria known for deciding whether a virialized rotating flattened system could be stable . @xcite conclude that a spherical halo with halo - to - disk mass ratio of @xmath5 to @xmath6 and initial value of total rotational kinetic energy to total gravitational energy ratio of @xmath7 renders the disk stable ( of course , the halo mass beyond the disk s outer radius has no effect and should not be taken into account ) . their conclusions are supported by many computer simulations and also by linear stability analysis carried out by toomre for mestel disk @xcite . note , however , that for nonlinear differential equations ( and hydrodynamical description of galaxies leads to such equations ) it is not a priori obvious that linear approximation appropriately describes solutions , the more that the gravitational potential also evolves along with matter , thus also the instability modes and the associated spectrum of eigenvalues . @xcite criticized the criterion , as well as various other criteria , to be of little use . nevertheless , he points out , they all lead to similar demands on mass distribution . as shown in @xcite , based on @xcite models , one comes to the conclusion that , for stability , total spherical mass interior to the half - mass radius of the disc should be more than twice the disc mass in the same region . @xcite and @xcite find similarly , using their numerical models , that for stability of differentially rotating disks the ratio of spheroidal mass to that of the disk interior to the outer radius of the disk , should be approximately @xmath8 , while for rigid disks the mass could be smaller @xcite . this spheroidal - to - disk mass criterion is much more handy to use than the ostriker - peebles like criteria in modeling galaxies , since mass functions of different galaxy components are much simpler to find . it should be noted , however , that the assumption of cold disks is based on the belief that the local observation of small dispersion of velocities in solar neighborhood in our own galaxy can be extrapolated onto whole galaxy and even onto all other flattened galaxies . even if the assumption was true , there is an argument due to @xcite that the halo is not so efficient a stabilizer as a bulge alone could be . a massive halo would change rotation field of the bar in a way not measured in real galaxies @xcite . as for modest halo , it would be difficult to decide about its presence since its effect would be small @xcite . it is also known that random motions of stars of the disc near the center have stabilizing property as well as a bulge or a thick disk . neglecting this effect could result in overestimating the bulge+halo mass required for global stability . @xcite surmises that most galaxies might have only moderate fractions of mass in the bulge+halo . as concerns the internal galaxy regions comprising the optical disk , the amount of dark matter is very poorly known @xcite . on the other hand , contrary to the common opinion that only spheroidal halos can help to stabilize cold galactic disks , bulges @xcite or thick disks @xcite are now considered to be more efficient stabilizers @xcite . what s more , the presence of dark matter in large amounts , dominating the disk mass , would simply suppress the instabilities required to form the observed structures such as bars and spiral arms @xcite . because spherical halo mass outside the optical radius has no effect on disk stability , the facts mentioned above , reverse the disk stability argument by @xcite , frequently used as a support for the need of a massive dark matter halo within the optical radius @xcite . here , it is worth to add that @xcite construct a model disk galaxy with an almost flat rotation curve which is stable without presence of dark matter . in modeling of mass distribution in galaxies , provided one has indeed a model that properly describes a given galaxy , one in principle should check whether the found mass distribution is stable . if disks were indeed cold , then for stability the spherical mass component should be sufficiently large . it seems however , that for estimating of the mass distribution in galaxies , the stability issue is not so important for astronomers , the more that the galaxy models used in this respect are very simple , one could say even naive . majority of them assume idealized substructures such as disks and bulges with simple parametric profiles , and assume symmetries that are not observed ( spherical and axial ) . for example , the spherical mass to disk mass ratio in model used by @xcite(things ) is much less than 2 , _ e.g.,_for galaxies ngc 4736 , ngc 3031 . despite the fact that a dark matter halo was included , the ratio would suggest that the found mass distribution is instable as it does not meet the stability criteria mentioned previously , that is , the disk - to - halo mass is too large . there is a big but , however . all mass models for flattened galaxies are merely some idealizations that do not take into account different processes that make real galaxies stable . the role of the models is rather to roughly estimate the gravitational potential from rotation curves assuming circular orbits and axial symmetry , and so on , and then to derive the corresponding mass distribution which one believes to some degree approximate the actual amount of matter in real , and otherwise stable galaxies . for example , maximal disk models give `` effective '' surface density which should be treated rather as a substitution for different galaxy components both those flattened and spheroidal , described altogether by a single function of radius ; only after a suitable de - projection one would infer about various galaxy components , thus it would be at least exaggeration to apply to such models the stability criteria which says that axisymmetric cool disks are instable . of course , in real galaxies , masses are neither distributed in infinitely thin disks nor in ideally spherically symmetric structures , orbits are not circular , rotation field has a very complicated structure , and mass distribution does not have axial symmetry axial symmetry is clearly broken by bars and spiral structure . thus it would be at least inappropriate to qualify or disqualify different models on the basis of stability arguments even if they do not contain a halo of dark matter , since the models , by construction , do not describe real galaxies and ignore many subtleties . the models are only slightly more accurate than would be a dimensional analysis , which states that the amount of electromagnetic energy radiated by a galaxy should be given by @xmath9 , where @xmath10 is some characteristic velocity derived from doppler image , @xmath11 is its radial size estimated based on the spatial range of the luminous stuff , @xmath12 is some mass - to - light ratio derived on population analysis ground , and @xmath13 is some unknown dimensionless factor related to the particular geometry of mass distribution and of rotation field . all parameters apart from @xmath13 can be estimated from measurements . the models only better approximate the parameter @xmath13 by giving its concrete model , more or less accurate . for example , the estimation should be better when for an elliptical galaxy one uses a spherical rather than a disk - like model , or , for a spiral galaxy , a disk model rather than spherical . the approximation should improve when one distinguishes spherical and disk - like subcomponents , further , one can add a width to the disk , and so on . but it would be hard to treat these models seriously when dynamics and stability of a galaxy is concerned ( especially when maximal disk model is used which treats on equal footing flat and spheroidal internal mass components ) . without realistic modeling such difficult problems as structural stability , can not be answered . by realistic modeling we understand finding , for a given galaxy , such a spatial rotation field and mass distribution for an n - body system a computer model of this galaxy which would be able to reconstruct at least the observed doppler image of the galaxy . reconstruction of a published rotation curve only , is not sufficient , because different groups obtain for the same galaxy different rotation curves depending on the details of their procedures applied for processing doppler images . such images are also often nonsymmetric , unlike simple galaxy models , and lot of information about galaxy kinematics they convey is missing by the fact that the images are by construction a sort of projection of a three dimensional rotational field onto the observation plane . @xmath0 ratio is important for studying the formation and evolution of galaxies , thus already from this standpoint it is rather strange to assume it is constant from the start , while it would be more desirable to determine it as a function of radius . unfortunately , the theoretical @xmath0 for an individual galaxy is not well known even when assumed constant . its derived value depends strongly on the assumed mass function for the stellar population , on the star formation history , on metallicity and depends also on extinction . it is also affected by the assumed internal extinction law , by the errors in luminosity measurements , and by the distance dimming @xcite . this ratio is generally assumed constant separately for the bulge and for the disk . accordingly , one tacitly assumes that the contribution of bulge and that of disk to the rotation curve should be controlled only by the light distribution via a single number found by best fits as if brightness profiles were to trace the overall features of rotation curves . however , the assumption of constant @xmath0 ratio may not be correct , since it requires that extinction and population gradients across the luminous disk should be small @xcite . constant @xmath0 ratio seems also improbable since it would imply that composition of a galaxy is homogenous . overestimating the importance of constant @xmath0 may change considerably and in uncontrolled way the contribution of bulge and that of disk to the overall rotation , the more that optical rotation curves are not featureless , while luminosity is usually fitted by simple laws ( exponential , _ etc _ ) . in particular , this may lead to errors in reconstruction of the relation between mass density and rotation of the disk component , since rotation of a disk is a nonlocal functional of the density profile , sensitive to the detailed local structure of the profile . this in turn may lead to wrong estimates of mass profiles of other mass components . to give only an example , many spirals simply do not have exponential disks or any other given by a simple analytical law , and assuming otherwise , results in large discrepancies in disk scale lengths published by different authors for the same galaxies @xcite . closely related to the non - uniqueness of the estimates of mass profiles is the so called disk - halo degeneracy the uncertainty in the relative contributions of different galaxy components to the overall rotation ; for unambiguous disk - halo - bulge decomposition one requires additional constraints on mass distribution , in particular , such as the mass - to - light ratios or the assumed mass profiles and shapes of dark haloes . in finding such constraints may also be helpful measurements of gravitational lensing in spiral galaxies @xcite . usual rotation - curve fitting methods have too many free parameters and are not unique , require additional assumptions , sometimes quite arbitrary . for example , @xcite present disk - halo decompositions with several different assumptions for stellar @xmath0 ratio for the same galaxy , or assume that only the minimum disk is present and that the observed rotation curve is due entirely to dark matter , or assume that stellar @xmath0 ratio is zero ( ! ) and take into account contribution of the atomic gas ( h i and he ) . in contrast , our approach does not have the arbitrariness problem , since we predict , not assume mass to light ratios . as pointed out by @xcite , there is only limited evidence for constant @xmath0 . what s more , the results of @xcite for m33 suggest that this assumption is not valid at all , @xmath0 ratio of the galaxy s disk increases approximately 5 times over 6 scale lengths of the inner disk . their findings also suggest that there is very little dark matter in m33 s central regions . as we have seen above , the case for dark matter inside galaxies is very weak and the amount of ascertained mass of dark matter is strongly model - dependent . it should be stressed that the models are constructed in the same framework of newtonian gravitation . there is increasing number of papers where more and more complex models are considered compared with previous simple models of galaxies . they account for the observed rotation of internal regions of flattened galaxies without dark matter , and some of them , based on maximal disk hypothesis , account for global rotation also without dark matter . the use of models with extended disk finds additional support from the fact that dark haloes may be very flattened rather than spherical . for flattened systems , however , gravitation gets more complicated and new effects difficult to tackle with , such as influence of external masses on internal orbits , must be taken into account and which would be absent if external masses were distributed spherically symmetric . also stellar mass - to - light ratio need not be constant . assuming otherwise , galaxy mass models become very stiff leading , for example , to disk - halo conspiracy . as we saw , this also leads to discrepancies between predictions . as concerns the presence of dark matter in spiral galaxies , we therefore prefer a more cautious approach than just to assume the presence of the ( unobserved ) dark matter halo from the beginning . we propose to admit dark matter , only if models that composed of baryonic matter distributed in a flattened disk ( of stars and gas ) and of more spherical bulge and stellar halo , fail to properly account for the dynamics of the disk measured through rotation curves . such an approach , we believe , could significantly reduce the undesirable model dependence of the amount of cdm in spiral galaxies @xcite . most models of galaxy formation predict that @xmath0 should be a declining function of distance , e.g. the inner regions of a galaxy form first and contain older , higher @xmath0 ratio star populations @xcite . as follows from the previous section it is not correct to assume that the ratio is constant throughout galaxies . this is the reason why we prefer to determine mass distribution in disk based on measurements other than brightness profiles first , and only then determine mass - to - light ratios as a result of comparison of the obtained mass profile with the measured luminosity profile . in several galaxies we have studied so far using our approach we observe that this ratio indeed decreases while density approaches that of hydrogen and helium known from measurements . this also suggest that dark matter is not needed for some galaxies even in the outermost regions . conceptually simple , direct method for determining the gross mass distribution in galaxies , is rotation curve inversion . maximal disk model is an example . provided rotation curve is known globally , the corresponding mass distribution can be found via single integral functional . the method is independent of any conventional assumptions concerning decomposition of a galaxy into various idealized subsystems with their unknown ( usually correlated with each other ) parameters determinable from best fits . unfortunately , direct rotation curve inversion is not unambiguous if mass distribution does not have special symmetries , e.g. spherical symmetry . various ambiguities pertinent to gravitation of flattened systems , easily hidden by the use of a particular model for the dark mass @xcite , may influence the actual importance of different galaxy components . this fact seems not much appreciated . in addition , there is also ambiguity in mass - to - light ratios within optical disks , since extended rotation curves alone give no indication of where the luminous disk could end @xcite . consequently , as noted in @xcite , there are mass models which fit rotation curves within the optical disk both with and without significant dark component @xcite . what s more , rotation of a galaxy considered dark matter dominated in one model can be devoid of dark matter in another model @xcite . for a flattened mass distribution , local density is a nonlocal functional of the rotation field . and vice versa , velocity on a given orbit is affected both by gravitation of masses external and internal to that orbit . for that example , it would be impossible to reconstruct dynamical mass function from rotation of luminous matter moving in a potential well of unseen external very flattened halo of dark matter . such dark haloes considerably flattened toward the stellar plane , resembling a disk more closely than a sphere @xcite , with axis ratios between 0.1 and 0.3 where suggested by @xcite and @xcite . @xcite analyzed errors resulting from extrapolation of the rotation curve beyond the last measured point and gave a criterion for their estimation in global disk model . these errors are the fundamental obstacle in using disk models . the reconstructed mass density is viable only out to a distance of @xmath14 of the radius of the last measurement point . the less serious is sensitiveness of this method to noise criticized in @xcite . however , as @xcite illustrates , this leads only to @xmath15 uncertainty . therefore one needs a method to minimize such errors . once one accepts that very flattened haloes of unseen matter can exist then one may lose any predictability . for example , since dark haloes are not seen , one may assume anything about them , say , that outside the last measurement point there is present a flattened dark matter ring of a given arbitrarily large mass and given shape , encircling the galaxy . to balance its influence on the luminous matter , one needs correspondingly large amount of mass inside galaxy . this in turn would made impossible even to determine the mass - to - light ratio of the luminous matter since the ratio would depend on the unknown mass of the flattened dark matter halo . there is a simple criterion for deciding , whether the mass distribution at larger radii in some galaxy is flattened rather than spherical . in the presence of spherical dark matter halo , the non - spherical component of mass distribution should be negligible . then , only the radial component of the gravitational force is important . in this case keplerian mass function @xmath16 defined for rotation curve @xmath17 must be a nondecreasing function of radius @xmath18}\geqslant0.\ ] ] in deriving this inequality , we assumed circular orbits of matter in the vicinity of the galactic plane , which is customary in modeling of rotation curves . for spherical systems , the keplerian mass function is identical with the true mass function . thus , if sphericity condition ( [ eq : sphcond ] ) is not satisfied at larger radii , the true mass distribution can not be spherical , and spherical cdm halo is thereby excluded ( as far as the usual assumptions in galaxy modeling are employed ) , or other forms of dark matter are required forming flattened structures . we can use the global disk model especially for galaxies with rotation curves breaking the sphericity condition at larger radii . note , that the distribution of internal masses is not much important for the overall gravitation at larger radii ( contribution from higher gravitational multipoles of internal masses decreases quickly with radius ) , therefore , the disk model can be used also for the inner parts , in spite of the fact that the center may be rather spheroidal since , after de - projection of a part of the disk density onto a spheroidal component , this does not change the central mass significantly this is an approximation , a standard procedure in the maximal disk model . since rotation curves of disks satisfying the sphericity condition are also possible @xcite , one can try to apply the disk model also to spiral galaxies of which rotation curves do not violate this condition . the disk model assumes that whole matter in a galaxy is distributed in the vicinity of the galactic equatorial plane . the model also assumes that orbits of stars are circular and that dispersion of velocities is negligible . the other simplifying assumption is that whole system is axisymmetric . the relation between surface mass density @xmath19 and velocity is then given by @xmath20}{\mathrm{d}\chi}\right)\nonumber , \end{aligned}\ ] ] where @xmath21 and @xmath22 are elliptic functions of the first and second kind @xcite . this integral is understood in the principal value sense ( @xmath23 ) , since both summands are divergent , thus the integral must be skilfully calculated . in practice , however , it suffices to assume that matter is located close to the galactic plane only in the galaxy outskirts where the external orbits are stabilized by gravitational potential of the galactic interior ( this contribution to gravitational potential is almost spherical at large radii ) , while in the vicinity of galactic interior the density can be treated as an effective density understood rather as a column density of masses projected onto the galactic plane a substitution for otherwise spherical and stable bulges _ etc_in internal regions . interpreted in this way , this approximated cold disc model of an otherwise stable spiral galaxy also avoids , or at least weakens , the stability counterarguments often raised against global disk models ( see also the discussion in section [ sec : stabarg ] ) . it is seen form equation ( [ eq:6 ] ) that for a flattened mass distribution of which the disk model is a limiting approximation , rotational velocity of matter on a given orbit is dependent both of masses distributed interior and exterior to that orbit . the inverse relation is a functional of whole rotation curve , and it reads @xcite @xmath24.\end{aligned}\ ] ] in the literature of this subject , for some mysterious reason , there is always used an ( equivalent ) expression containing derivatives of velocities which are hardly measurable with satisfactory accuracy , e.g. @xcite . the crucial point now is that a rotation curve is never known globally , thus the mass distribution even in the region where rotation is known , can not be determined from rotation only . this is the manifestation of `` non - locality '' of gravity of flattened systems one can not determine their mass functions from the observed rotation and , therefore , additional constraints on this mass distribution is required . this is particularly important in the regions where measurements of rotation end , while the analogous error is negligible in the internal galaxy regions . this property is fatal for reconstruction of mass distribution inside a galaxy when outside the luminous part of this galaxy there is present an external ring - like halo composed of dark matter ; if it were spherical there would be no problem at all . having realized the unpleasant properties of flattened galaxies we devised in @xcite a method which may help to minimize this uncertainty or remove it . one does not have to assume anything beyond the last measured point , but simply use the available data of mass distribution as a lower bound for the amount of matter in regions where rotation for some reasons could not be determined , but the amount of gas is still measurable . our goal is to examine how good is our model in reproducing the mass distribution in a sample of spiral galaxies . the thin disk model s role is to find such a global @xmath19 , that the corresponding velocity of rotation @xmath25 defined by integral ( [ eq:6 ] ) , would best overlap with the observed rotation curve @xmath17 . this task , however , is not unique , since for every @xmath26 , there is an infinite set of functions @xmath19 that give rise to the same @xmath25 for @xmath27 , even though the @xmath19 s may significantly differ from one another for @xmath28 and have different total masses . to minimize this non - uniqueness , we look for mass distribution in flattened galaxies by iterations . a particular realization of such iteration proposed in @xcite assumes that mass distribution of gas in a given galaxy is known for radii greater than the range of the rotation curve . we bring to the reader s attention the fact , that mass distribution found in @xcite for galaxy ngc4736 using this method , is consistent with all measurements , and this was possible without dark matter . this result is interesting , since another model of this galaxy @xcite predicts almost 70% abundance of dark matter and does not well explain its rotation at large radii . it is therefore natural to examine whether this galaxy is exceptional or , maybe , if there are other spiral galaxies with lower abundance of dark matter . below , we present results for several galaxies obtained with the use of global disk model . we applied the iteration method analogous to that introduced by @xcite . we are aware this is not a representative sample of galaxies . however , we think , that even when rotation of only a single galaxy could be explained without dark matter , this would be interesting as such and worthy to be presented . initially , we planned to examine only galaxies with rotation curves breaking the condition of spherical mass distribution , because such galaxies are presumably flattened and , therefore , natural candidates to which global disk model is applicable . however , in order to see whether the model produces viable results for other galaxies , we included also galaxies ngc 6946 and ugc 6446 for which we also found good and complete data . as concerns the error analysis , our method by construction exactly reproduces rotation curves , thus also fits rotation measurements within error bars . this is the reason why we show rotation curves without error bars . results obtained in the global disk model for galaxy ngc 4736 at @xmath29 . * a ) * rotation curve : things measurements _ [ solid circles ] _ , the model _ [ solid line ] _ , and comparison with high resolution rotation curve taken from @xcite . * b ) * model global surface mass density _ [ solid line ] _ , surface mass density of hi+he _ [ open circles ] _ ; and surface brightness : b - band _ [ solid squares ] _ , v - band _ [ solid triangles ] _ , i - band _ _ [ solid circles]__@xcite ; * c ) * mass - to - light ratio profile without inclusion of hi+he . ] here , we once again consider galaxy ngc 4736 , which earlier was studied in @xcite . this time , however , we use another , newer rotation curve published by @xcite . in figure [ fig:4736 ] , this new curve is compared with that obtained by @xcite we used last time . both the curves are shown assuming distance @xmath30 different from the previous @xmath31 . although both curves do not overlap we come to comparable conclusions concerning the amount of non - baryonic dark matter . ngc 7436 still remains a galaxy in which the halo is not necessarily required . surface density obtained in the global disk model smoothly approaches that of gas in the outskirts of the galaxy , we obtain also low @xmath0 ratios ( @xmath32 for i - band , @xmath33 for v - band , and @xmath34 for b - band ) . in addition , the local @xmath0 ratio decreases with radius for the largest radii . therefore , despite the new rotation curve does not break sphericity condition so strongly unlike that by @xcite , we still consider this galaxy as containing no dark matter or that abundance of dark matter is at most very low . _ upper : _ comparison of the ngc 7793 rotation curves obtained in things @xcite , @xcite and @xcite . symbols stand for measurements and lines for model fits . _ middle : _ surface mass densities obtained in global disk model compared with surface mass densities of neutral hydrogen and helium and with surface brightness in the b - band . _ bottom : _ mass - to - light ratios derived based on mass densities obtained for different rotation curves . ] we take rotation curves published by a ) @xcite , b ) @xcite and c ) three rotation curves published by @xcite . the latter three curves were merged such to obtain a single rotation curve . in effect we have three distinct rotation curves for the same galaxy . as seen in figure [ fig:7793 ] the curves are similar which shows rotation measurements for this galaxy are trustworthy . the corresponding masses are also close to each other , @xmath35 . surface mass density from the global disk model smoothly converges to that of gas at the outskirts of the galaxy for all three cases . we obtain also low values for @xmath0 ratios ( @xmath36 in b - band ) . we stress that the attempt of @xcite to account for the rotation with spherical cdm halo , gives very poor results . by giving up the halo , we obtain surface mass density in agreement with the amount of gas , and low @xmath0 ratios . therefore , we think it is reasonable to assume that the galaxy does not have any dark matter halo . _ upper : _ comparison of the ngc 1365 rotation curves obtained in @xcite and @xcite . symbols stand for measurements and lines for model fits . _ middle : _ surface mass densities obtained in global disk model compared with a 2nd order exponential decay fit and with surface mass densities of neutral hydrogen and helium and surface brightness in the b - band . _ bottom : _ mass - to - light ratios derived based on mass densities obtained different rotation curves . ] now we consider two rotation curves @xcite and @xcite the latter suitably smoothed out in a way to conform with data points and error bars . we have noticed that surface mass density corresponding to the rotation curve has approximately exponential falloff ( see figure [ fig:1365 ] ) . therefore , for a while one may assume that surface mass density is a superposition of two exponential profiles , in order to check whether such mass distribution accounts for the observed rotation . we find , this is indeed the case , however the mass density becomes larger than that of gas ( although quickly falling off ) . on the other hand , the surface density obtained with our method converges smoothly to that of gas already in a place where galaxy is still bright in b - band . this can be interpreted as that apart from hi and he there must be present other luminous stuff . in general , this is very interesting , that for this galaxy already the doubly - exponential disk alone suffices to account for the rotation curve of @xcite . irrespectively of which rotation curve is considered , the obtained mass - to - light ratio is low ( @xmath37 to @xmath38 in b - band depending on which rotation curve and method is used ) . this ratio falls - off with increasing radius in every case . as so , galaxy ngc 1365 can be modeled without non - baryonic dark matter . _ upper : _ comparison of the ngc 6946 rotation curves obtained in @xcite , @xcite and things @xcite . symbols stand for measurements and lines for model fits . _ middle : _ surface mass densities obtained in global disk model compared with surface mass densities of neutral hydrogen and helium and surface brightness in the b - band . _ bottom : _ mass - to - light ratios derived based on mass densities obtained different rotation curves . ] we take three rotation curves published by @xcite , by @xcite and @xcite ( the latter were smoothed out ) . as is seen , the published rotation data differ significantly from each other , particularly rotation curve of @xcite exceeds much from the other . by using global disk model we obtained surface densities which smoothly converges to that of gas . mass - to - light ratio is low based on rotation by sanders ( @xmath39 in b - band ) , but it becomes twice greater for rotation curve of things ( @xmath40 ) @xmath0 ratio for things grows slightly with radius , for the other two curves it is approximately constant . summing up , based on available rotation data it is difficult to uniquely decide about presence of non - baryonic dark matter in this galaxy . the result depends of which rotation curve is used . using our modeling method we can explain all these rotation curves . it is worth of pointing out that the obtained surface densities smoothly converge to that of gas . however , based on things rotation curve , we obtain @xmath0 ratio much greater than for ngc 4736 , ngc 1365 , ngc 7793 , similarly , for the other two rotation curves . _ upper : _ comparison of the ngc 6446 rotation curve obtained in @xcite . symbols stand for measurements and lines for model fits . _ middle : _ surface mass density obtained in global disk model compared with that of obtained in spherical mass model , with that of neutral hydrogen and helium and with surface brightness in the b - band . _ bottom : _ mass - to - light ratio derived based on mass density obtained different rotation curves . ] the @xmath0 profile of galaxy ugc 6446 grows with radius attaining values much higher than for the other examined galaxies , even after subtraction of hi and he . such a singular behavior of the @xmath0 profile is expected at the edge of luminous disk , where the luminosity per unit of surface tends to zero , @xmath41 , while surface mass density remains finite . then , the local @xmath0 ratio becomes singular , @xmath42 . after subtraction of mass contribution of gas , the increasing @xmath0 ratio indicates , that the stars at the outskirts of luminous disk , contain possible admixture of died stars , `` black '' dwarfs , neutron stars , or stars of too low a mass for nuclear burning ( so called brown dwarfs ) . it is important to note , that the mass density of such a died star component needs only to be residual , such that total mass density is infinitesimally bigger than the mass density of shining stars , @xmath43 for the @xmath0 ratio significantly increases at the edge : @xmath44 when both @xmath45 and @xmath41 . this very small mass density , @xmath46 , of died stars does not show up anywhere except at the very edge where it is strongly enhanced by small @xmath47 . contribution of died stars to the mass of galaxy is thus negligible . the following thought experiment provides a strong argument against the existence of a spherical halo around galaxy ugc 6446 . suppose for a while that this galaxy is dominated by such a halo . then , the mass distribution could be considered spherical with good accuracy , thus the corresponding mass function could be calculated from @xmath48 ( and this can be done , since the rotation curve of this galaxy satisfies the sphericity condition ) . then , the corresponding column mass density @xmath49 can be found by projecting the volume density onto the galactic plane by using the formula @xmath50 } } , & \nonumber\end{aligned}\ ] ] where @xmath26 is the distance of the outermost point on the rotation curve . note , that our formula is very advantageous from the practical point of view , contrary to the ( equivalent ) handbook formula , as it does not require derivatives of velocities , which are measured with very poor accuracy . we compare the column density @xmath49 with total surface mass density obtained from the disk model and with its component surface mass density of gas ( hi+he ) . the result is shown in figure [ fig:6446 ] . we see that close to the `` edge '' , @xmath49 is lower than the amount of gas ! we stress that the two surface mass densities seen in figure [ fig:6446 ] correspond to two extreme idealizations in the first the galaxy is approximated by a thin disc , in the other , by a spherically symmetric mass distribution . the realistic mass distribution must be somewhere in between , but , as is seen , closer to that in disk model . in view of these observations , we classify ugc 6446 as a galaxy without massive spheroidal dark halo . we considered 5 galaxies , among which one was reexamined with a new rotation curve . our aim was to answer the question if the galaxies could be satisfactorily modeled without non - baryonic dark matter . in this regard we check whether the following conditions are met : the found mass distribution should account for the measured rotation curve , it should approximately converge toward that observed for gas ( hi + he ) in the outer parts of galaxies ( as this is the lower bound for the amount of luminous matter ) , the mass - to - light ratio should be low and its profile should be a declining function of radius close to the galaxy `` edge '' . we find that these conditions are satisfied by our results for galaxies ngc 4736 , ngc 7793 and ngc 1365 for all rotation curves . it is reasonable to state that these are galaxies with flattened , disk - like mass distributions , for which the disk model approximation works quite well . the case of galaxy ngc 6946 is not so clear . there are significant discrepancies between published rotation curves , and consequently , in the obtained @xmath0 ratios . galaxy ugc 6446 satisfies some of our criteria without non - baryonic dark matter , however its @xmath0 ratio is highest and , in addition , it grows dramatically toward galaxy `` edge '' . in another test we assumed that the galaxy is dominated by a spherical dark matter halo and , using a spherical model , obtained the resulting bulk mass density . this density , accounting for the galaxy rotation , is in this case lower than that of gas , which can be considered as a strong argument for absence of spherical non - baryonic dark halo . this argument does not exclude the presence of a non - spherical halo . however , the mass needed in the galaxy outskirts to account for the rotation in the disk model is practically that of gas . it would be therefore worthwhile to consider other than cdm causes of the @xmath0 ratio growth . we have also signalized the controversial issue , to what extent the very simplified galaxy models like the ones commonly used by astronomers or like that used in the current paper , can be treated as models of realistic galaxies . in particular , whether stability arguments can be applied to favourize or reject a particular galaxy model , since the role of these models , which serve only as an idealized approximation incorporating many assumptions , is to estimate different global characteristics of galaxies like total mass , the amount of gas , global mass - to - light ratio , dark - to - luminous mass ratio _ etc _ , not to reconstruct exact spatial mass distribution , the more that many effects like the influence of global magnetic fields in a galaxy on the rotation of gas in the galaxy outskirts , are normally not taken into account , and these fields may be important for accounting for flat rotation curves @xcite . global disk model is a natural approximation for flattened galaxies such as spiral ones , as far as it is not assumed in advance , sometimes by force , that each spiral galaxy must necessarily be immersed in an invisible halo of dark matter . moreover , we are able to find such mass distributions which perfectly account for rotation curves , while in models assuming massive halo these rotation curves often can not be satisfactorily reconstructed , eg . in the case of ngc 7793 in @xcite , or the dark halo found is too weak to account for stability criteria ( ngc 4736 in @xcite ) ( whereas , historically , the role of such haloes was not only to explain flat rotation , but also to stabilize disks ) . although our model is very simple , similarly as other galaxy models customarily used by astronomers , it is better by that it does not assume constant mass - to - light ratios but it predicts the ratio as a resultant mass - to - light profile . our results pertain only to the particular galaxies examined by us so far . from the fact that there is no need of non - baryonic dark matter in galaxies ngc 4736 or ngc 7793 , one can not conclude that one does not need it in other spiral galaxies . to claim otherwise , further studies are needed . nonetheless , we find interesting the existence of spiral galaxies without non - baryonic dark matter or galaxies with small abundance of such matter . @r@|@c@|@c@|@c name & ngc 7793 & ngc 1365 & ngc 6946 + incl . angle [ deg ] & @xmath51 @xmath52 & @xmath53 @xmath54 & @xmath55 @xmath54 + distance [ mpc ] & @xmath56 @xmath52 & @xmath57 @xmath58 & @xmath59 @xmath60 + morphological type & sad @xmath52 & sbb @xmath54 & sabc @xmath54 + @xmath61 @xmath62}$ ] & @xmath63 @xmath64 & @xmath65 @xmath66 & @xmath67 @xmath68 + @xmath69 @xmath70}$ ] & @xmath71 @xmath52 & @xmath72 @xmath66 & @xmath73 @xmath60 + @xmath74 @xmath70}$ ] & - & - & @xmath75 @xmath76 + [ cols= " > " , ] + we acknowledge the usage of the hyperleda database ( http://leda.univ-lyon1.fr ) athanas & sellwood ( 1987 ) iaus , 117 , 300a athanassoula , e. , bosma , a. , & papaioannou , s. ( 1987 ) a&a , 179 , 23 bahcall , j. n. , & casertano ( 1985 ) apj , 293 , l7 battaner , e. ; garrido , j. l. ; membrado , m. ; florido , e ( 1992 ) nature , 360 , 6405 , p. 652 bernabei r. , belli p. , montecchia f. at al . , 2007 , nucl . b , proc . suppl . , 166 , 87 binney , j. , & tremaine , s. ( 1987 ) , galactic dynamics ( princeton univ . press ) bosma , a. ( 1978 ) , ph.d . thesis , univ . groningen bosma , a. ( 1999 ) , 1999aspc .. 182 .. 339b boomsma r. , t. a. oosterloo , f. fraternali , j. m. van der hulst and r. sancisi ( 2008 ) a&a , 490 , 2 , 555 - 570 bratek . , jaocha j. , kutschera m. , 2008 , mnras , 391 , 1373 buchhorn , m. ( 1992 ) , ph.d . thesis , australian natl . carignan c. , 1985 , apjs , 58 , 107 carignan c. , puche d. , 1990 , aj , 100 , 394 ciardullo , durrell , at al ( 2004 ) apj , 614 , 167c clowe d. , brada m. , gonzalez a.h . , markevitch m. , randall s.w . , jones c. , zaritsky d. , 2006 , apj , 648 , l109 de blok , mcgaugh , rubin ( 2001 ) aj , 122 , 2396d de blok , w. j. g. and walter , f. and brinks , e. and trachternach , c. and oh , s .- h . and kennicutt , r. c. , 2008 , aj , 136 , 2648 - 2719 dicaire , i. ; carignan , c. ; amram , p. ; marcelin , m. ; hlavacek - larrondo , j. ; de denus - baillargeon , m .- m . ; daigle , o. ; hernandez , o. ( 2008 ) aj , 135 , 6 , 2038 - 2047 . evans ( 2001 ) 2001idm .. conf ... 85e fall , s. m. ; efstathiou , g. ( 1980 ) mnras.193 .. 189f hohl f. , ( 1976 ) aj ..... 81 ... 30h herrmann and ciardullo ( 2005 ) aipc , 804 , 341h hoekstra h. , van albada t.s . , sancisi r. , 2001 , mnras , 323 , 453 hyperleda database ( http://leda.univ-lyon1.fr ) jaocha j. , bratek . , kutschera m. , kolonko m. , 2007 , acta phys . pol . b , 38 , 3859 jaocha j. , bratek . , kutschera m. , 2008 , apj , 679 , 373 jorsater s. , van moorsel g.a . , 1995 , aj , 110 , 2037 kalnajs , a. j. ( 1983 ) in iau symp . 100 , internal kinematics and dynamics of disk galaxies , ed . e. athanassoula ( dordrecht : reidel ) , 87 kalnajs ( 1987 ) iaus , 117 , 289k kent , s. m. 1986 , aj , 91 , 1301 kent s.m . , 1987 , aj , 93 , 816 kutschera , m. ; jaocha , j. ( 2004 ) acta physica polonica b , 35 , 10 , p.2493 lake , g. , & feinswog , l. ( 1989 ) aj , 98 , 166 maller , ariyeh h. ; simard , luc ; guhathakurta , puragra ; hjorth , jens ; jaunsen , andreas o. ; flores , ricardo a. ; primack , joel r. ( 2000 ) apj ... 533 .. 194 m mulder , p. s. , and van driel , w. ( 1993 ) a&a , 272 , 63 munoz - tunon , c. , prieto , m. , beckman , j. , and cepa , j. , 1989 , ap&ss 156 , 3001 olling , r. p. ( 1996 ) , aj , 112 , 481 ostriker , peebles ( 1973 ) apj , 186 , 467o palunas & williams ( 2000 ) aj , 120 , 2884p rix , h .- w . , & zaritsky d. ( 1995 ) apj , 447 , 82 rubin , v. c. , thonnard , n. , & ford , w. k. j. ( 1978 ) apj , 225 , l107 sackett , p. d. , rix , h .- w . , jarvis , b. j. & freeman , k. c. ( 1994 ) , apj , 436 , 629 sackett ( 1996 ) iaus , 173 , 165s sackett ( 1997 ) apj , 483 , 103s snchez - salcedo , f. j. ; reyes - ruiz , m. 2004 apj ... 607 .. 247s sanders r. h. ( 1996 ) apj , 473 , 117 - 129 sellwood ( 1980 ) a&a .. 89 .. 296s sellwood ( 1983 ) iaus .. 100 .. 197s sellwood , j. ( 1998 ) , in asp conf . 182 , galaxy dynamics , ed . d. r. merritt , m. valluri , & j. a. sellwood ( san francisco : asp ) , 339 sellwood , kosowsky ( 2000 ) ? ? arxiv : astro - ph/0009074v1 , gas & galaxy evolution asp conference series , vol . * * volume * * , 2000 j. e. hibbard , m. p. rupen and j. h. van gorkom , eds . sellwood , j. a. , & evans , n. w. ( 2001 ) , apj data sofue y. , 1997 , pasj , 49 , 17 sug - whan k. , muk - suk c. , 1984 , jkas , 17 , 23 toomre , a. ( 1964 ) apj , 139 , 1217 t tully r.b . , verheijen m.a.w . , pierce m.j . , huang jia - sheng , wainscoat r.j . , 1996 , aj , 112 , 2471 van albada , g. d. , & sancisi , r. ( 1986 ) phil . london a , 320 , 447 verheijen m.a.w . , sancisi r. , 2001 , a&a , 370 , 765 williams m. j. , m. bureau & michele cappellari ( 2009 ) mnras.tmp.1591w , young j.s . , scoville n.z . , 1991 , ara&a , 29 , 581 znmar snchez r. , sellwood j. a. , weiner b. j. , williams t. b ( 2008 ) apj , 674 , 2 , 797 - 813

in recent years , there has been an increasing interest to link the discussion about complex phenomena in natural science , such as physics or biology , in particular to those in life science , such as sociology , economy or regional planning . one of the challenges is to reveal cross - links between the dynamic models used in the specific fields , in order to find out pieces for a common theory of self - organization and evolution of complexity . with respect to economic and urban problems , this perspective seems rather new . in his essay about self - organization in economics , krugman ( 1996b ) states : `` so far this movement has largely passed economics by . '' however , recent years have also seen numerous attempts for developing a new , evolutionary view of economics ( anderson _ et al . _ , 1988 , wei bin , 1991 , arthur _ et al . _ , 1997 , silverberg , 1997 , schweitzer and silverberg , 1998 ) in a world of fast growing population and intercontinental trade relations , the _ local _ emergence of new economic centers on one hand , and the _ global _ competition of existing economic centers on the other hand has a major impact on the socio - economic and political stability of our future world . this problem also confronts the natural sciences on their way to a new understanding of complex phenomena . this paper aims to contribute to this discussion by adapting a model of interactive structure formation , which has already proved its versatility in a variety of applications , to the problem of economic agglomeration . spatio - temporal pattern formation in urban and economic systems has been investigated for quite a long time @xcite . there is also a well known german tradition in location theory , associated with names like von thnen , weber , christaller ( 1933 ) and lsch ( 1940 ) . recent approaches ( sparsely recognized especially by american economists ) tackle the problem of settlement formation , location theory and city size distribution within a stochastic dynamical theory , based on master equations and economically motivated utility functions ( dendrinos and haag , 1984 , weidlich and munz , 1990 a , b , haag _ et al . _ , 1992 , 1994 , weidlich , 1991 , 1997 ) . different from these investigations , the current paper focuses on an agent based dynamics of economic concentration . agent models , which have originally been developed in the artificial life community ( maes , 1990 , steels , 1995 ) recently turned out to be a suitable tool for describing economic interaction ( anderson _ et al . _ , 1988 , marimon _ et al . _ , 1990 , holland and miller , 1991 , lane , 1992 , arthur , 1993 , kirman , 1993 , epstein and axtell , 1996 , arthur _ et al . _ , 1997 ) . the rational agent model , one of the standard paradigms of neoclassical economic theroy @xcite , is based , among others , on the assumption of the agent s complete knowledge of all possible actions and their outcomes or a known probability distribution over outcomes , and the common knowledge assumption , i.e. that the agent knows that all other agents know exactly what he knows and are equally rational . in this particular form , the rational agent is just one example of a _ complex agent _ , which can be regarded as an autonomous entity with either knowledge based or behavior based rules @xcite , performing complex actions , such as bdi ( belief - desire - interactions ) ( mller _ et al . _ , 1997 ) . the complex agent i.e. is capable of specialization , learning , genetic evolution , etc . however , the freedom to define rules and interactions for the agents , could very soon turn out to be a pitfall , because of the combinatoric explosion of the state space : for 1000 agents with 10 rules , the state space contains @xmath0 possibilities , hence , almost every desirable result could be produced from such a simulation model . the alternative to the complex agent , promoted in this paper , could be the _ minimalistic agent _ , which acts on the possible simplest set of rules , without deliberative actions . instead of specialization , the minimalistic agent model is based on a large number of `` identical '' agents , and the focus is mainly on cooperative interaction instead of autonomous action . a version of the minimalistic agent model , which allows to apply the advanced methods developed in statistical physics and stochastic theory , is based on active brownian particles ( steuernagel _ et al . _ , 1994 , schimansky - geier _ et al . _ , 1995 , 1997 , schweitzer , 1997 , et al . _ , 1998 ) . generally , active brownian particles are brownian particles with internal degrees of freedom . as a specific action , the active brownian particles ( or active walkers , within a discrete approximation ) are able to generate a self - consistent field which in turn influences their further movement and behavior . this non - linear feedback between the particles and the field generated by themselves results in an _ interactive structure formation _ process on the macroscopic level . hence , these models have been used to simulate a broad variety of pattern formations in complex systems , ranging from physical to biological and social systems ( lam and pochy , 1993 , schweitzer and schimansky - geier , 1994 , ben - jacob _ et al . _ , 1995 , lam , 1995 , schweitzer _ et al . _ , 1997 , et al . _ , 1997 , schweitzer and steinbrink , 1997 ) . in the following , the active brownian particles are regarded as economic agents with two different internal states and a very simple behavior : employed agents , which are immobile , generate a wage field ( as the result of their work ) , while unemployed agents migrate guided by local gradients in the wage field , i.e. they try to move toward those regions in their vicinity with a high productivity and a high value of the wage field . further unemployed agents can be employed , dependent on the local economic situation , but employed agents also can be unemployed . in sect . 2 we present our stochastic approach to the problem of employment and migration . in sect . 3 , the general model is used to derive a special dynamics recently applied by krugman ( 1992 , 1996b ) for economic aggregation . in sect . 4 , we present the economic assumptions involved in our model , while sect . 5 describes the results of computer simulations , which show economic aggregation on two time scales . 6 gives some conclusions and an outlook of possible extension of the model presented . let us consider a two - dimensional system with @xmath1 economic agents , which are represented by active brownian particles . these particles are characterized by two variables : their current location , given by the space coordinate @xmath2 , and an internal state , @xmath3 , which could be either one or zero : @xmath4 . active particles with the internal state @xmath5 are considered employed agents , @xmath6 , active particles with @xmath7 are considered unemployed agents , @xmath8 . with a certain rate , @xmath9 ( hiring rate ) , an unemployed agent becomes employed , while with a rate @xmath10 ( firing rate ) an employed agent becomes unemployed , which can be expressed by the symbolic reaction : @xmath11 employed agents are considered immobile , while unemployed agents are able to migrate . the movement of a migrant may depend both on erratic circumstances and on deterministic forces which attract him to a certain place . within a stochastic approach , this movement can be described by the following overdamped langevin equation : @xmath12 @xmath13 describes the local value of a deterministic force , which influences the motion of the agent . we note here , that the agent is _ not _ subject to long - range forces , which may attract him over large distances , but only to a _ local _ force . that means the migrant does _ not _ count on global information which guide his movement , but responds only to local information , which will be specified later . the second term in eq . ( [ langev - red ] ) describes random influences on the movement of the individuals modelled by a stochastic force @xmath14 , which is assumed to be gaussian white noise : @xmath15 @xmath16 is a measure of the strength of the stochastic force . as eq . ( [ langev - red ] ) indicates the unemployed agent will move in a very predictable way if the guiding force is large and the stochastic influence is small , and he will act rather randomly in the opposite case . the current state of the agent community can be described by the canonical @xmath17-particle distribution function @xmath18 , which gives the probability to find the @xmath17 agents with the internal states @xmath19 in the vicinity of @xmath20 at time @xmath21 . the change of this probability can be described by a multi - variate master equation @xcite , which considers both changes in the internal states and migration of the agents . here , we restrict the discussion to the spatio - temporal densities @xmath22 of unemployed and employed agents , which can be formally derived from the @xmath17-particle distribution function : @xmath23 @xmath24 is the kronecker delta function for discrete variables , which is @xmath25 only for @xmath26 and @xmath27 otherwise , while @xmath28 is dirac s delta function used for continuous variables , and @xmath29 is the system size ( surface area ) . with @xmath4 , we obtain the spatio - temporal density of the employed agents , @xmath30 , and of the unemployed agents , @xmath31 . for simplicity , we omit the index @xmath32 , by defining : @xmath33 we only assume that the total number of agents is constant , while the density of employed and unemployed agents can change in space and time : @xmath34 the density of the employed agents can be changed only by local `` hiring '' and `` firing '' processes , which with respect to eq . ( [ react ] ) , can be described by the reaction equation : @xmath35 the density of the unemployed agents can be changed by two processes , ( i ) migration and ( ii ) hiring of unemployed and firing of employed agents . with respect to eq . ( [ langev - red ] ) , which describes the movement , we can derive the following fokker - planck equation : @xmath36 the first term of the r.h.s . ( [ fpe - t ] ) describes the change of the local density due to the force @xmath37 , the second term describes the migration of the unemployed agents in terms of a diffusion process with @xmath38 beeing the diffusion coefficient , the third and the fourth term describe local changes of the density due to `` hiring '' or `` firing '' of agents . by now , we have a complete dynamic model which describes local changes of employment and unemployment , as well as the migration of unemployed agents . however , so far some of the important features of this dynamic model are not specified , namely ( i ) the deterministic influences on a single migrant , expressed by @xmath37 , ( ii ) the hiring and the firing rates , @xmath39 , which locally change the employment density . these variables depend of course on local economic conditions , hence we will need additional economic assumptions . krugman ( 1992 ) ( see also krugman ( 1996b ) for the results ) discusses a dynamic spatial model , where `` workers are assumed to move toward locations that offer them higher real wages '' . krugman makes no attempt `` to model the moving decisions explicitely '' , but he assumes the following `` law of motion of the economy '' : @xmath40 here , @xmath41 is the `` share of the manufactoring labor force in location @xmath42 '' and @xmath43 is the real wage at location @xmath42 . @xmath44 is the `` average real wage '' , which is defined as : @xmath45 here , the sum goes over all different regions , where @xmath42 serves as a space coordinate , and it is assumed that `` at any point in time there will be location - by - location full employment '' . the assumed law of motion of the economy , eq . ( [ krugman ] ) , then means that `` workers move away from locations with below - average real wages and towards sites with above - average real wages '' . in the present form , the assumed law of motion of the economy involves some shortages : 1 . a constant total number of employed workers is assumed , a change of the total number or unemploment is not discussed . employed workers are considered to move . in fact , if a worker wants to move to a place which offers him higher wages , he first has to be a free , that means an unemployed worker , then moves and then has to be reemployed at the new location , again . this process is completely neglected ( or else , it is assumed that it occurs with infinite velocity ) . workers at location @xmath42 always exactly know about the average wage in the system . it is not explained where they get the information about the average wage from . workers move immediately if their wage is below the average , regardless of the migration distance to the places with higher wages , and with no doubt about their reemployment . in the following , krugman s law of motion of the economy shall be derived from the general dynamic model presented in sect . 2 . , in order to elucidate the implicit assumptions which lead to eq . ( [ krugman ] ) . the derivation is based on four approximations : * ( i ) * in his paper , krugman ( 1992 ) does not discuss unemployment . but , with respect to the dynamic model presented in sect . 2 , the unemployed agents exist in an overwhelming large number . thus , the first assumption to derive krugmans law of motion is , the local change of the density of unemployed agents due to hiring and firing can be simply neglected . * ( ii ) * further , it is assumed that the spatial distribution of unemployed agents is in a _ quasistationary _ state . this does not mean that the distribution does not change , but that the distribution relaxes fast into a quasistationary equilibrium , compared to the distribution of the employed agents . with the assumptions ( i ) and ( ii ) , eq . ( [ fpe - t ] ) for the density of the unemployed agents reduces to : @xmath46 integration of eq . ( [ fpe - st ] ) leads to the known canonical distribution : @xmath47}{{\left\langle \exp \left[\int f(r , t)/d_{n}\ ; dr \right ] \right\rangle } } \\ { \left\langle \exp \left[\int f(r , t)/d_{n}\ ; dr \right ] \right\rangle } & = & \frac{1}{a } \int_{a } \exp \left [ \int f(r',t)/d_{n } \ ; dr'\right ] \ ; dr \nonumber\end{aligned}\ ] ] where the expression @xmath48 describes the mean value , and @xmath49 is the mean density of unemployed agents . * ( iii ) * for a derivation of krugmans equation , we now have to specify @xmath37 in eq . ( [ stat ] ) . here , it is assumed that a single unemployed agent which migrates due to eq . ( [ langev - red ] ) , responds to the _ total local income _ in a specific way : @xmath50}\ ] ] eq . ( [ respo1 ] ) means that the migrant is guided by _ local gradients _ in the _ total income _ , with @xmath51 being the local income and @xmath52 the local density of employed workers . we note here again , that the migrant does not count on information about the highest global income , he only `` knows '' about his vicinity . with assumption eq . ( [ respo1 ] ) , we can rewrite eq . ( [ stat ] ) using the discrete notation , which is preferred by krugman : @xmath53 the corresponding equation ( [ react2 ] ) for @xmath54 reads in the discrete notation : @xmath55 since the krugman equation deals with shares instead of densities , we have to divide eq . ( [ react - disc ] ) by @xmath56 , which leads to : @xmath57 * ( iv ) * krugman assumes that the total number of employed workers is constant , so we use this assumption to replace the relation between the hiring and the firing rate in eq . ( [ change - l ] ) . ( [ react - disc ] ) yields : @xmath58 @xmath59 in eq . ( [ change - l ] ) is now replaced by the quasistationary value , eq . ( [ stat - k ] ) . inserting further eq . ( [ sum ] ) into eq . ( [ change - l ] ) and using the definitions of the average wage @xmath44 , eq . ( [ mean - w ] ) , and @xmath60 , we finally arrive at : @xmath61 eq . ( [ krugman-2 ] ) is identical with krugmans law of motion of the economy , eq . ( [ krugman ] ) , if the prefactor @xmath62 in eq . ( [ krugman ] ) is identified as : @xmath63 . hence , @xmath62 is a slowly varying parameter , which depends both on the firing rate , which determines the time scale , and on the average wage , @xmath44 , which may change in the course of time . it is an interesting question whether the assumptions ( i)-(iv ) which lead to krugmans law of motion in the economy , have some practical evidence in an economic context . from a more theoretical perspective , it is noteworthy that krugmans equation , ( [ krugman ] ) , ( [ krugman-2 ] ) has an obvious analogy to a selection equation of the eigen - fisher type , @xmath64 . here , @xmath65 is the fitness of species @xmath66 and @xmath67 is the mean fitness representing the global selection pressure . it can be proved @xcite that this equation describes a competition process which finally leads to a stable state with only _ one _ surviving species . the competition process may occur on a very long time scale , but asymptotically a stable coexistence of many different species is impossible . in the economic context discussed by krugman ( 1992 ) , ( [ krugman ] ) implies that finally _ all _ workers are located in _ one _ region @xmath42 , where the local real wage @xmath43 is equal to the mean real wage , @xmath44 . however , in his paper , krugman ( 1992 ) ( see also krugman ( 1996b ) ) discusses computer simulations with 12 locations ( on a torus ) which show the stable coexistence of two ( sometimes three ) centers , roughly evenly spaced across the torus , with exactly the same number of workers ( cf . fig . 7 in krugman ( 1992 ) and fig . 2.2 . in krugman ( 1996b ) ) . the paper does not provide information about the time scale of the simulations , and stochastic influences ( apart from a random initial configuration ) are not considered . krugmans model of spatial aggregation includes of course some more complex economic assumptions , such as consideration of transportation costs , distinction between agricultural and manifactured goods , price index etc . we are not going to discuss whether the stable coexistence of two centers in krugmans computer simulations might result from those specific economic assumptions in his model or from the lack of fluctuations ( which could have revealed the instability of the ( deterministic ) stationary state ) . due to krugman ( 1992 ) , already a slight variation in the parameters always led to a single center . in the following , we will focus on the more interesting question , whether the general dynamic model of migration and employment introduced in sect . 2 , is able to produce a stable coexistence of different economic centers under the presence of fluctuations . this may allow us to overcome some of the shortages involved in krugman equation . the derivation of krugmans equation was based on the assumtion ( ii ) that the time scale for hiring and firing of workers is more determining than the time scale of migration . if we explicitely consider unemployment in our model , this assumption implies that there are always enough unemployed workers which can be hired on demand . a growing economy , however , might be determined just by the opposite limiting case : it is important to attract workers / consumers to a certain area before the output of production can be increased . hence , the time scale for the dynamics is determined by migration processes . that means the spatial distribution of the employed agents can be assumed in a _ quasistationary equilibrium _ compared to the spatial distribution of the unemployed agents . @xmath68 hence , the local density of employed agents can be expressed as a function of the local density of the unemployed agents available , which itself changes due to migration on a slower time scale . we further assume that the unemployed agent who is able to migrate , eq . ( [ langev - red ] ) , responds to local gradients in the real wages ( per capita ) , instead of gradients in the total income , as assumed for krugmans equation . the migrant tries to move towards places with a higher wage , but again he only counts on information in the vicinity . hence , the guiding force @xmath13 is determined as follows : @xmath69 for a further discussion , we need some assumptions about the local distribution of the real wages , @xmath70 . it is reasonable to assume that the local wages may be a functional of the local density of employed agents , @xmath71 . some specific assumptions about this dependence will be discussed in the next section . with eq . ( [ change - stat ] ) , we can then rewrite the spatial derivative for the wages as follows : @xmath72 where @xmath73 denotes the functional derivative . using the equations ( [ change - stat ] ) , ( [ respo2 ] ) , ( [ wage - dl ] ) , the fokker - planck equation for the change of the density of the unemployed agents , eq . ( [ fpe - t ] ) can now be rewritten as follows : @xmath74 the r.h.s . of eq . ( [ fpe - eff ] ) now has the form of a usual diffusion equation , with @xmath75 being an _ effective diffusion coefficient _ : @xmath76 here , @xmath77 is the `` normal '' diffusion coefficient of the unemployed agents , eq . ( [ fpe - t ] ) . the additional terms reflect that the unbiased diffusion is changed because of the response of the migrants to local differences in the wage distribution . as we see from eq . ( [ d - eff ] ) , there are two contradicting forces determining the effective diffusion coefficient : the normal diffusion , which keeps the unemployed agents moving , and the response to the wage gradient . if the local wage decreases with the number of employed agents , then , and the effective diffusion increases . that means unemployed agents migrate away from regions where employment may result in an effective decrease of the marginal output . however , if , then their wage effectively _ increases _ in regions with a larger number of employed agents , and they are _ attracted _ to these regions . as we see in eq . ( [ d - eff ] ) , for a certain _ positive feedback _ between the local wage and the employment density , the effective diffusion coefficient can be _ locally negative _ and unemployed agents do not leave these areas once they are there . economically speaking , these agents stay there , because they may profit from the local increase in the employment density . we want to emphasize that this interesting dynamic behavior has been derived _ without _ any explicit economic assumptions . but for a further discussion of the model , the three remaining functions which are unspecified by now : ( i ) , ( ii ) @xmath9 , ( iii ) @xmath10 , have to be specified , and that is of course where economics comes into play . in order to determine the economic functions , we refer to a _ perfectly competitive industry _ ( where `` perfect '' means complete or total ) . in this standard model @xcite , the economic system is composed of many firms , each small relative to the size of the industry . these firms represent seperate economies sharing common pools of labor and capital . new competitors can freely enter / exit the market , hence the number of production centers is not limited or fixed . further , it is assumed that every firm produces the same ( one ) product and every firm uses only _ one _ variable input . then , the _ maximum profit condition _ tells us that `` firms will add inputs as long as the marginal revenue product of that input exceeds its market price . '' in the case of labor as variable input , the price of labor is the wage , and a profit maximizing firm will hire workers as long a the marginal revenue product exceeds the wage of the added worker . the marginal revenue product @xmath78 is the additional revenue a firm earns by employing one additional unit of input , where @xmath79 is the price of output and @xmath80 is the marginal product . in a perfectly competitive industry , however , no single firm has any control over prices . instead , the price results from the interaction of many suppliers and many demanders . in a perfectly competitive industry , every firm sells its output at the market equilibrium price , which is simply normalized to one , hence the marginal revenue product is determined by @xmath80 . the marginal product , @xmath80 , can be derived from a _ production function _ - , which describes the relationship between inputs and outputs ( e.i . the technology of production ) . usually , the production function may include the effect of different inputs , such as capital , public goods or natural resources . here , we concentrate only on one variable input , labor . thus , @xmath81 is assumed a cobb - douglas production function @xmath82 with a common ( across regions ) exponent @xmath83 , @xmath52 being the local density of employees . the exponent @xmath83 describes how a firms output depends on the scale of operations . increasing returns to scale are characterized by a lower average cost with an increasing scale of production , hence @xmath84 in this case . on the other hand , if an increasing scale of production leads to higher average costs ( or lower average output ) , we have the situation of decreasing returns to scale , with @xmath85 . in the following , we will restrict the discussion to @xmath85 , common to all regions . the prefactor @xmath86 represents economic details of the level of productivity . we assume that these influences can be described by two terms : . @xmath87 should summarize those output dependences on capital , resources etc . , which are not explicitely discussed here . the new second term , @xmath88 , considers _ cooperative effects _ which result from interactions among the workers . since all cooperative effects are _ non - linear _ effects , @xmath88 should be a non - linear functional of @xmath52 : @xmath89 . hence , the production function depends explicitely only on @xmath52 , now : @xmath90 \;l^{\beta}(r , t)\ ] ] the marginal product ( @xmath80 ) is the additional output by adding one more unit of input , i.e. @xmath91 , if labor is input . the wage of a potential worker will be the marginal product of labor : @xmath92 if a firm faces a market wage rate of @xmath93 ( which could be bound e.g. by minimum wage laws ) , then , in accordance with the maximum profit condition , a firm will hire workers as long as : @xmath94 to complete our setup , we need to discuss how the prefactor @xmath88 depends on the density of employees , @xmath52 . here , we assume that the cooperative effects will have an effect only in the intermediate range of @xmath95 . for small production centers , the synergetic effect resulting from the mutual stimulation among the workers is too low . on the other hand , for very large production centers , the advantages of the cooperative effects might be compensated by the disadvantages of the massing of agents . thus we will assure , that @xmath96 both in the limit @xmath97 and @xmath98 . these assumtions are concluded in the ansatz : @xmath99 where the utility function @xmath100 describes the mutual stimulation among the workers in powers of @xmath95 : @xmath101 the series will be truncated after the second order . the constants @xmath102 characterize the effect of cooperation , with @xmath103 . especially , the case @xmath104 considers saturation effects in the cooperation , e.i . the advantages of cooperation will be compensated by disadvantages of crowding . this idea implies that there is an optimal size for taking advantages of the cooperative effect , which is determined by the ratio of @xmath105 and @xmath106 . if one believes that cooperative effects are an always increasing function in @xmath95 , then simply @xmath107 can be assumed if cooperative effects are neglected , i.e. @xmath108 , then we should obtain the `` normal '' production function , @xmath109 . the constant @xmath110 is determined by a relation between @xmath111 and @xmath112 which results from the eqs . ( [ prod1 ] ) , ( [ a - u ] ) . within our approach , the constant @xmath111 is not specified . if we , without restrictions of the general case , choose : @xmath113 then the production function , @xmath114 , eq . ( [ prod1 ] ) , with respect to cooperative effects can be expressed as follows : @xmath115\;l^{\beta}\ ] ] fig . 1 presents the production function , eq . ( [ prod2 ] ) dependent on the density of employees , which can be compared with the `` normal '' production function , @xmath109 . [ fig1 ] clearly , we see an increase in the total output due to cooperation effects among the workers . if we assume @xmath116 , this increase has a remarkable effect only in an intermediate range in @xmath95 . both for @xmath97 and @xmath98 the cooperative effect vanishes . once the production function is determined , we also have determined the local wage @xmath117 , eq . ( [ wage ] ) , as a function of the density of employees : @xmath118\;\beta \,l^{\beta-1 } \nonumber \\ & + & \frac { \bar{a}}{2 } \;\exp\left(a_{1}\,l+a_{2}\,l^{2}\right ) \,(a_{1}+2a_{2}l)\;l^{\beta } \end{aligned}\ ] ] hence , also the derivative @xmath119 , used for the effective diffusion coefficient , @xmath75 , eq . ( [ d - eff ] ) is determined . 2 shows both functions dependent on the density of employees . [ fig2 ] the left part of fig . 2 indicates that , within a certain range of @xmath95 , the derivative @xmath119 could be indeed positive , due to cooperative effects . with respect to the discussion in sect . , this means that the effective diffusion coefficient , eq . ( [ d - eff ] ) , can be possibly negative , i.e. unemployed workers will stay in these regions . the transition rate @xmath9 for hiring unemployed workers is implecitely already given by the conclusion , that firms hire workers as long as the marginal revenue product exceeds the wage of the worker . in accordance with eq . ( [ wage - star ] ) we define : @xmath120 here , @xmath121 determines the time scale of the transitions . @xmath9 , which is now a function of the local economic situation , is significant larger than the level of random transitions , represented by @xmath121 , only if the maximum profit condition allows hiring . otherwise , the hiring rate tends to zero . the firing rate @xmath10 could be simply determined opposite to @xmath9 . then , from the perspective of the employee , firing is caused by the local situation of the economy , i.e. by _ external reasons _ : workers are fired if @xmath122 . a more refined description , however , should consider also _ internal reasons _ : a worker can not only _ loose _ his job , he can also _ quit _ his job himself for better opportunities , e.g. because he wants to move to a place where he earns a higher wage . it is reasonable to assume that the internal reasons depend again on spatial gradients in the wage . due to ( [ langev - red ] ) the unemployed agent migrates while guided by gradients in the wage . the employed agent at the same location may have the same information . noteworthy again , this is only a local information about differences in the wage . if the local gradients in the wage are small , the internal reasons to quit the job vanish , and the firing depends entirely on the ( external ) economic situation . however , if these differences are large , the employee might quit his job for better chances . we note that the latter process was already considered in krugmans law of motion in the economy , sect . 3 . different from the assumptions involved in krugmans ( [ krugman ] ) , here the process : employment @xmath123 unemployment @xmath123 migration @xmath123 reemployment is _ explicitely _ modeled . it does not occur with infinite velocity , and there is no guarantee for reemployment . hence , we define the `` firing rate '' which describes the transition from an employed to an unemployed agent , as follows : @xmath124 + \ ; q\ , \frac{{\displaystyle}\partial \omega(r)}{{\displaystyle}\partial r } \right\}}\ ] ] the additional parameter @xmath125 can be used to weight the influence of spatial gradients on the employee . eventually , in this section we have determined the variables @xmath37 , @xmath9 and @xmath10 via a production function @xmath126 , which represents certain economic assumptions . now , we can turn back to the dynamic model described in sect . 2 , which now can be solved by means of computer simulations . before presenting the results of the computer simulations , the simulation technique should be shortly discussed . the computer program has to deal with three different processes , which have to be discretized in time for the simulation : ( i ) the movement of active brownian particles with @xmath127 due to the overdamped langevin eq . ( [ langev - red ] ) , ( ii ) the transition of the particles due to the rates , @xmath9 , eq . ( [ kplus ] ) , @xmath128 , eq . ( [ kminus ] ) , and ( iii ) the generation of the field @xmath70 . considering eq . ( [ langev - red ] ) , the new @xmath129-position of a particle @xmath66 with @xmath127 on the two dimensional surface at time @xmath130 is given by : @xmath131 the equation for the @xmath132-position reads accordingly . @xmath133 is the _ non - constant _ time step , which is calculated below . @xmath38 is the diffusion coefficient , and _ grnd _ is a gaussian random number with mean equals zero and standard deviation equals unity . in oder to calculate the spatial gradient of the wage field @xmath134 , eq . ( [ wage2 ] ) , we have to consider its dependence on @xmath52 , which is a local density . the density of employed agents is calculated assuming that the surface is divided into boxes with the spatial ( discrete ) indices @xmath135 and unit length @xmath136 . then , the local density is given by the number of agents with @xmath137 inside a box of size @xmath138 : @xmath139 we note that @xmath140 , with @xmath141 being the spatial move of the migrating agent into @xmath129-direction during time step @xmath133 . that means that the migration process is really simulated as a motion of the agents on a two dimensional plane , rather than a hopping process between boxes . using the box coordinates @xmath135 , the production function and the wage field can be rewritten as follows : @xmath142 the spatial gradient is then defined as : @xmath143 where the indices @xmath144 , @xmath145 , @xmath146 , @xmath147 refer to the left , right , lower and upper boxes adjacent to box @xmath148 . we further note that for the simulations _ periodic boundary conditions _ are used , therefore the neighboring box is always specified . using the discretized versions , eq . ( [ disc - l ] ) , ( [ disc - y - w ] ) , ( [ disc - grad ] ) , the transition rates can be reformulated as @xmath149 , @xmath150 accordingly . they determine the _ average number _ of transitions of a particle in the internal state @xmath3 , located in box @xmath148 , during the next time step , @xmath133 . in a stochastic simulation however , the actual number of reactions is a _ stochastic _ variable , and we have to assure that the stochastic number of reactions ( i ) does not exceed the actual number of particles available during @xmath133 , and ( ii ) is equal to the average number of reactions in the limit @xmath151 . this problem can be solved by using the stochastic simulation technique for reactions , which defines the appropriate time step , @xmath133 , as a random variable . let us assume that we have exactly @xmath152 particles with the internal state @xmath127 in the system at time @xmath153 . then the probability @xmath154 equals one , and the probability for any other number @xmath155 is zero . with this intial condition , the master equation to change @xmath152 reads : @xmath156 where @xmath17 is the total number of particles , and @xmath157 , @xmath158 are the transition rates for each particle . the solution of this equation yields : @xmath159 here , @xmath160 is the mean life time of the state @xmath152 . for @xmath161 , the probability @xmath162 to find still @xmath152 is almost one , but for @xmath163 this probability goes to zero . the time when the change of @xmath152 occurs , is most likely about the mean life time , @xmath164 . in a stochastic process , however , this time varies , hence , the _ real life time _ @xmath165 is a randomly distributed variable . since we know that @xmath162 has values between @xmath166 $ ] , we find from eq . ( [ m - solv ] ) : @xmath167\big\}\ ] ] ( zero actually has to be excluded . ) . that means , after the real life time @xmath168 , one of the possible processes which change @xmath152 occurs . each of these processes has the probability : @xmath169 thus , with a second random number _ rnd_[0,@xmath170 it will be determined which of these possible processes occurs . it will be the process @xmath125 which satisfies the condition : @xmath171 for the transition probabilities to change @xmath152 , we find in particular : @xmath172 & = & \sum_{i=1}^{n_{0 } } k^{-}_{i } t_{m } \nonumber \\ prob[n_{1},t\to n_{1}-1,t+\tau ] & = & \sum_{i=1}^{n_{1 } } k^{+}_{i } t_{m}\end{aligned}\ ] ] obviously , the sum over these probabilities is one , which means , during the time intervall @xmath168 one of these processes occurs with certainity . further , using the definition of @xmath168 , we see that @xmath173 yields asymptotically , i.e. after a large number of simulation steps , @xmath174 . hence , determining the time step as @xmath175 , eq . ( [ tau ] ) , ensures that only one transition occurs during one time step and the number of transitions does not get out of control . further , in the asympotitc limit , the actual ( stochastic ) number of transitions is equal the average number . the ( numerical ) disadvantage might be that the time step @xmath176 is not constant , so it has to be recalculated after each cycle before moving the particles with respect to eq . ( [ disc - x ] ) . this may slow down the speed of the simulations considerably . let us conclude the procedure to simulate the movement and the transition of the particles : 1 . calculate the density of particles @xmath177 , eq . ( [ disc - l ] ) 2 . calculate the production function @xmath178 and the wage field @xmath179 , eq . ( [ disc - y - w ] ) 3 . calculate the sum over all possible transitions , which determines the mean life time @xmath160 , eq . ( [ m - solv ] ) , of the current system state 4 . calculate from the mean life time the _ actual _ time step @xmath175 , eq . ( [ tau ] ) by drawing a random number 5 . move all particles with @xmath127 according to the langevin equation ( [ disc - x ] ) using the time step @xmath168 6 . calculate which one of the particles undergoes a transition by drawing a second random number , eq . ( [ prob - r ] ) 7 . update the system time : @xmath180 and continue with 1 . with the determination of the three functions @xmath37 , @xmath9 and @xmath10 with respect to some economic assumptions , we have completed our dynamic model , described in sects . 2 and 4 . in this section , we want to discuss some features of the dynamics by means of a computer with @xmath181 active brownian particles . initially , every particle is randomly assigned an internal parameter ( either @xmath27 or @xmath25 ) , and a position on the surface , which has been divided into @xmath182 boxes of unit length @xmath183 . the diffusion coefficient , which describes the mobility of the migrants , is set to @xmath184 ( in arbitrary units ) , so , a simple brownian particle would approximately need a time of @xmath185 for a mean spatial displacement of @xmath25 ( which is the spatial extension of a box ) . the minimum wage @xmath93 is set to @xmath186 . further , @xmath187 and @xmath188 , for the remaining parameters see fig . 1 . in the following we will restrict the discussion to the spatio - temporal evolution of the densities of employed and unemployed agents . other quantities of interest , such as the spatio - temporal wage distribution , the production function and the local values of the `` hiring '' and `` firing '' rates have been of course calculated for the simulation , but will be discussed in a subsequent paper , which also investigates different parameter sets . 3 shows snapshots of the evolution of the spatial density of employed agents , while fig . 4 shows the corresponding spatial density of unemployed agents . 5 presents the evolution of the total number of employed and unemployed agents , in terms of the total share @xmath189 with respect to eq . ( [ total ] ) . [ fig3 ] [ fig4 ] [ fig5 ] the simulation shows that the evolution of the densities occurs on two stages : * ( i ) * during the first stage ( for @xmath190 ) we have a significant higher share of employed agents ( up to 70 percent of the agents community ) . they are broadly distributed in numerous small economic centers , which have an employment density of about 5 . this can be understood using figs . 1 , 2 . the increase of the output due to cooperative effects allows the fast establishment of numerous small firms , which count on the mutual stimulations among their workers and in the beginning offer higher wages ( compared to the model without cooperative effects ) . this growth strategy , however , is not sufficient for output on larger scales , because the marginal product drastically decreases ( before it may increase again for an above critical employment density ) . as a result the wages may fall ( possibly below the minimum wage rate , @xmath93 ) , which prevents further growth . so , the first stage is characterized by the coexistence of numerous small economic centers . * ( ii ) * during the second and much longer stage ( for @xmath191 ) , some of these small centers overcome this economic bottleneck , which allows them to grow further . in the simulations , this crossover occured between @xmath192 and @xmath193 . in the model considered here , we have only one variable input , labor , so the crossover is mainly due to stochastic fluctuations in @xmath52 . after the bottleneck , the marginal output increases again , and also the wage increases with the density of employment ( cf . figs . 1 , 2 ) . this in turn effects the migration of the unemployed agents , which , due to eq . ( [ langev - red ] ) , is determined by two forces : the local attraction to areas with higher wages , and the stochastic influences which keep them moving . with @xmath194 , the attraction may exceed the stochastic forces , so the unemployed agents are bound to these regions , once they got there . this , in turn is important for the further growth of these economic centers , which needs agents to hire . as a result , we observe on the spatial level the concentration of employed _ and _ unemployed agents in the _ same _ regions . for their further growth , the economic centers which overcome the bottleneck , _ locally _ attract the labor force at the expense of the former small economic centers . as a result of the competition process , these small centers , which previously employed about 70 percent of the agents , dissapear . in the new ( larger ) centers however , only 60 percent of the agents can be employed ( cf . 5 ) , so the competition and the resulting large - scale production effectively results also in an _ increase of unemployment_. but the employment rate of 60 percent seems to be a stable value for the given parameter setup , since it is kept also in the long run , with certain fluctuations . the concentration process described above , leads to the existence of different _ extended major economic regions _ , shown in figs . 3 , 4 , which each consist of some subregions ( in terms of boxes ) . in the long run ( up to t=100.000 ) we find some remarkable features of the dynamics : * a _ stable coexistence _ of the major economic regions , which keep a certain distance from each other . this _ critical distance _ - which is a _ self - organized phenomenon _ - prevents the regions from detracting each other . in fact , finally there is no force between these regions which would pull off the employed or unemployed agents and guide them over long distances to other regions . thus , the coexistence is really a stable one . * a _ quasi - stationary non - equilibrium _ within the major economic centers . as we see , even in the long run , the local densities of employed and unemployed agents do not reach a fixed stationary value . within the major regions , there are still exchange processes between the participating boxes , hiring and firing , attraction and migration . hence , the total share of employed agents continues to fluctuate around the mean value of 60 percent ( cf . fig . 5 ) . in this paper , we have proposed a simple dynamic model to descibe migration and economic aggregation within the framework of active brownian particles . we consider two types of particles ( or agents ) ; employed agents , which are immobile , but ( as the result of their work ) generate a wage field , and the unemployed agents which migrate by responding to local gradients in the wage field . further a transition between employed and unemployed agents ( hiring and firing ) is considered . the economic assumptions to be used in the model , are concluded in only three functions : ( a ) the derivative @xmath195 , which describes how the local wage depends on the employment density , ( b ) @xmath9 , ( c ) @xmath10 , which are the hiring and firing rates . these functions can be determined using a production function , @xmath114 , which describes the output of production dependent on the variable input of labor . our ansatz for @xmath114 specifically counts on the influence of cooperative interactions between the agents on a certain scale of production . as the result of our dynamic model , we find the establishment of distinct economic centers out of a random initial distribution . the evolution of these centers occurs in two different stages . during the first stage , small economic centers are formed based on the positive feedback of mutual stimulation / cooperation among the agents . during the second and much longer stage , some of these small centers grow at the expense of others , which leads to the interesting situation that in our model economic growth and decline occur _ at the same time _ , but at different locations , which results in specific _ spatial - temporal _ patterns . the competition finally leads to the concentration of the labor force in different extended economic regions . this crossover to large - scale production is accompanied by an increase in the unemployment rate . although the extended economic regions are in an internal , quasistationary non - equilibrium , we observe the stable coexistence between these regions . this is an important result , which agrees e.g. with the central place theory of economics ( christaller , 1933 , lsch , 1940 ) . different from an attempt by krugman ( 1992 ) , who also focused on this problem , we find that ( i ) the coexistence is stable , even at the presence of fluctuations , ( ii ) the centers not necessarily have to have the same number of employed agents , to coexist , ( iii ) the dynamics does not simply converges into an equilibrium state , but the centers exist in an quasistationary non - equilibrium state and still follow a stochastic eigendynamics . one question , which will be tackled in a forthcoming paper , is about the influence of the different parameters . we have seen in the simulations , that cooperative effects are able to initiate an increase in the economic output , which leads to the establishment of many small firms . but the final situation does not include any of those small firms and seems to be independent of this intermediate state . this is mainly due to the fact , that we considered only _ one _ product . a more complex production function with _ outputs , however , should also result in the coexistence of small ( innovative ) firms and large scale production . another important issue is to understand the relation between the response of the agents to local gradients in the wage field , on one hand , and stochastic influences , on the other hand . most likely , there exist critical parameters which may describe a `` phase transition '' within the agents behavior , which eventually results in different types of spatial - temporal coexistence . finally , one may argue that the simple model proposed here , is not the simplest one to describe the economic aggregation . hence , the minimalistic agents discussed in sect . 1 , could be even more minimalistic . one example is the response of the employed agents to gradients in the wage field , expressed in the parameter @xmath125 which enters the firing rate , @xmath10 . @xmath196 simply means that internal reasons are neglected , and the agent is fired only due to external reasons . this might be also a reasonable assumption , however , it includes a pitfall . in a growing economy , eventually almost all unemployed agents might be hired at some locations . but , as long as @xmath197 , employed agents are almost never fired , thus there is a shortage of free agents for further growth . eventually the dynamics sticks in a dead - lock situation , because agents can not move even if there are locations in their neighborhood which may offer them a higher wage . therefore , it is reasonable to choose @xmath198 , but it is important to understand the role of @xmath125 , for instance , for the final patterns observed . so , we conclude that our model may serve as a toy model for simulating the influence of different social and economic assumptions on the spatial - temporal patterns in migration and economic aggregation . this work has been supported by the deutsche forschungsgemeinschaft ( bonn , germany ) . i would like to thank j. lobo for discussions on an early draft of this paper , and w. ebeling , l. schimansky - geier and g. silverberg for comments on the final version . anderson , p.w . , arrow , k.j . , pines , d. ( eds . ) ( 1988 ) . _ the economy as an evolving complex system _ , reading , ma : addison wesley . haag , g. ; munz , m. ; pumain , p. ; sanders , l. ; saint - julien , th . interurban migration and the dynamics of a system of cities : 1 . the stochastic framework with an application to the french urban system , _ environment and planning a _ * 24 * , 181 - 198 . schimansky - geier , l. ; schweitzer , f. ; mieth , m. ( 1997 ) . interactive structure formation with brownian particles , _ in : _ f. schweitzer ( ed . ) : _ self - organization of complex structures : from individual to collective dynamics _ , london : gordon and breach , pp . 101 - 118 . schweitzer , f. ; steinbrink , j. ( 1997 ) . urban cluster growth : analysis and computer simulation of urban aggregations , in : f. schweitzer ( ed . ) , _ self - organization of complex structures : from individual to collective dynamics _ , london : gordon and breach , pp . 501 - 518 . silverberg , g. ( 1997 ) . is there evolution after economics ? , _ in : _ f. schweitzer ( ed . ) : _ self - organization of complex structures : from individual to collective dynamics _ , london : gordon and breach , pp . 415 - 425 . weidlich , w. ( 1997 ) . from fast to slow processes in the evolution of urban and regional settlement structures , _ in : _ f. schweitzer ( ed . ) : _ self - organization of complex structures : from individual to collective dynamics _ , london : gordon and breach , pp . 475 - 488 .

it is by now well - known that increasing the number of antennas in a wireless system can significantly increase capacity . since the seminal papers of foschini and gans @xcite and telatar @xcite , multiple - input multiple - output ( mimo ) channels have been thoroughly investigated in theory ( see @xcite for a survey ) and implemented in practice @xcite . this capacity gain usually comes at the expense of more complex encoders and decoders and a great deal of work has gone into designing low - complexity mimo architectures . in this paper , we describe a new low - complexity architecture that can attain significantly higher rates than existing solutions of similar complexity . we focus on the case where each of the @xmath0 transmit antennas encodes an independent data stream ( see figure [ f : mimo ] ) . that is , there is no coding across the transmit antennas : each data stream @xmath1 is encoded separately to form a codeword @xmath2 of length @xmath3 . channel state information is only available to the receiver . from the receiver s perspective , the original data streams are coupled in time through encoding and in space ( i.e. , across antennas ) through the mimo channel . the joint maximum likelihood ( ml ) receiver simultaneously performs joint decoding across time and receive antennas . clearly , this is optimal in terms of both rate and probability of error . however , the computational complexity of jointly processing the data streams is high , and it is difficult to implement this type of receiver in wireless systems when the number of streams is large . instead , linear receivers such as the decorrelator and minimum - mean - squared error ( mmse ) receiver are often used as low - complexity alternatives @xcite . traditional linear receivers first separate the coupling in space by performing a linear projection at the front - end of the receiver . in order to illustrate this concept and motivate the proposed new approach , we consider the @xmath4 mimo channel characterized by the following matrix : @xmath5 the simplest choice of a linear receiver front - end _ inverts _ the channel matrix . this receiver is usually referred to as the _ decorrelator _ in the literature . that is , the receiver first applies the matrix @xmath6 to the received signal . overall , this converts the original channel into a new equivalent channel characterized by the identity matrix and colored gaussian noise , i.e. , to two scalar channels with correlated noise . the linear receiver then proceeds by separately decoding the output of each of these two channels . the well - known drawback of this approach is that the _ noise _ vector is also multiplied by the linear receiver front - end matrix given in , which alters the variances of its components . in our simple example , if we assume that the original channel had independent additive white gaussian noise of unit variance , the equivalent noises after the linear receiver front - end have variances of @xmath7 and @xmath8 respectively . that is , while the receiver front - end has nulled out cross - interference , it has also significantly increased the noise levels . the integer - forcing linear architecture advocated here is based on the recent insight that if on all transmit antennas , the same linear or lattice code is used , then it is possible to not only decode codewords themselves , but also _ integer _ linear combinations of codewords directly @xcite . let us denote the codeword transmitted on the first antenna by @xmath9 and the codeword transmitted on the second antenna by @xmath10 then , for the simple example matrix from , the receiver can decode the integer linear combination @xmath11 from the first receive antenna and the combination @xmath12 from the second receive antenna . from these ( following @xcite ) , it is possible to recover linear equations of the data streams over an appropriate finite field , @xmath13 and @xmath14 . these equations can in turn be digitally solved for the original data streams . the key point in this example is that the noise variances remain unchanged , which increases the effective @xmath15 per data stream . note that for more general channel matrices beyond the simple example here , it will also be advantageous to first apply an appropriate linear receiver front - end , albeit following principles very different form merely inverting the channel matrix , as we explain in more detail in the sequel . in this paper , we first consider the standard mimo channel and develop a new _ integer - forcing _ linear receiver architecture that provides multiplexing and diversity gains over traditional linear architectures . our approach relies on the compute - and - forward framework , which allows linear equations of transmitted messages to be efficiently and reliably decoded over a fading channel @xcite . we develop a multiple antenna version of compute - and - forward which employs the antennas at the receiver to rotate the channel matrix towards an effective channel matrix with integer entries . separate decoders can then recover integer combinations of the transmitted messages , which are finally digitally solved for the original messages . we show that this is much more efficient than using the receive antennas to separate the transmitted codewords and directly decoding each individual codeword . our analysis uses nested lattice codes originally developed to approach the capacity of point - to - point awgn and dirty - paper channels @xcite and for which practical implementations were presented in @xcite and subsequent works . next , we generalize the mimo channel model to include interference @xcite and show that the integer - forcing receiver architecture is an attractive approach to the problem of oblivious interference mitigation . by selecting equation coefficients in a direction that depends on both the interference space and the channel matrix , the proposed architecture reduces the impact of interference and attains a non - trivial gain over traditional linear receivers . furthermore , we show that the integer - forcing receiver achieves the same generalized degrees of freedom as the joint decoder . our proof uses techniques from diophantine approximations , which have also recently been used for interference alignment over fixed channels and the characterization of the degrees of freedom for compute - and - forward @xcite . in the remainder of the paper , we start with a formal problem statement in section [ sec : problem ] , and then overview the basic existing mimo receiver architectures and their achievable rates in section [ sec : existingarchitecture ] . in section [ sec : architecture ] , we present the integer - forcing receiver architecture and a basic performance analysis . we show that the rate difference between the proposed receiver and traditional linear receivers can be arbitrarily large in section [ sec : fixedchannels ] . we study the outage performance of the integer - forcing linear receiver under a slow fading channel model in section [ sec : outage ] . we show that in the case where each antenna encodes an independent data stream , our architecture achieves the same diversity - multiplexing tradeoff as that of the optimal joint decoder . in section [ sec : interference ] , we consider the mimo channel with interference and show that the integer - forcing receiver can be used to effectively mitigate interference . we characterize the generalized degrees - of - freedom for the integer - forcing receiver and find that it is the same as for the joint decoder . throughout the paper , we will use boldface lowercase to refer to vectors , @xmath16 , and boldface uppercase to refer to matrices , @xmath17 . let @xmath18 denote the transpose of a matrix @xmath19 and @xmath20 denote the determinant . also , let @xmath21 denote the inverse of @xmath19 and @xmath22 denote the pseudoinverse . the notation @xmath23 will refer to the @xmath24-norm of the vector @xmath25 while @xmath26 will refer to the @xmath27-norm . finally , we will use @xmath28 and @xmath29 to refer to the maximum and minimum singular values of the matrix @xmath19 . the baseband representation of a mimo channel usually takes values over the complex field . for notational convenience , we will work with the real - valued decomposition of these complex matrices . recall that any equation of the form @xmath30 over the complex field can be represented by its real - valued representation , @xmath31 we will henceforth refer to the @xmath32 real - valued decomposition of the channel matrix as @xmath33 . we will use @xmath34 independent encoders and @xmath34 independent decoders for the resulting real - valued transmit and receive antennas .. ] [ def : messages ] each of the @xmath34 transmit antennas has a length @xmath35 _ data stream _ ( or message ) @xmath1 drawn independently and uniformly from @xmath36 . [ def : encoders ] each data stream @xmath1 is mapped onto a length @xmath3 channel input @xmath37 by an _ encoder _ , @xmath38 an equal _ power allocation _ is assumed across transmit antennas @xmath39 while we formally impose a separate power constraint on each antenna , we note that the performance at high snr ( in terms of the diversity - multiplexing tradeoff ) remains unchanged if this is replaced by a sum power constraint over all antennas instead . [ def : rate ] each of the @xmath34 encoders transmits at the same @xmath40 the _ total rate _ of the mimo system is just the number of transmit antennas times the rate , @xmath41 . since the transmitters do not have knowledge of the channel matrix , we focus on the case where the @xmath34 data streams are transmitted at equal rates . we will compare the integer - forcing receiver against successive cancellation v - blast schemes with asymmetric rates in section [ subsec : rateallocation ] . [ def : channel ] let @xmath42 be the matrix of transmitted vectors , @xmath43 \ .\end{aligned}\ ] ] the _ mimo channel _ takes @xmath44 as an input , multiplies it by the _ channel matrix _ @xmath45 and adds noise @xmath46 whose entries are i.i.d . gaussian with zero mean and unit variance . the signal @xmath47 observed across the @xmath48 receive antennas over @xmath3 channel uses can be written as@xmath49 we assume that the channel realization @xmath33 is known to the receiver but unknown to the transmitter and remains constant throughout the transmission block of length @xmath3 . [ def : decoder ] at the receiver , a _ decoder _ makes an estimate of the messages,@xmath50 [ def : achrate ] we say that sum rate @xmath51 is _ achievable _ if for any @xmath52 and @xmath3 large enough , there exist encoders and a decoder such that reliable decoding is possible @xmath53 so long as the total rate does not exceed @xmath51 , @xmath54 many approaches to mimo decoding have been studied in the literature . we provide a brief summary of some of the major receiver architectures and the associated achievable rates , including the joint ml receiver , the decorrelator , linear mmse estimator and the mmse - sic estimator . clearly , the best performance is attainable by joint ml decoding across all @xmath55 receive antennas . this situation is illustrated in figure [ f : mimo ] . let @xmath56 denote the submatrix of @xmath33 formed by taking the columns with indices in @xmath57 . if we use a joint ml decoder that searches for the most likely set of transmitted messages vectors @xmath58 , then the following rate is achievable ( using gaussian codebooks at the transmitter ) : @xmath59 where @xmath60 is the identity matrix . is achievable . ] note that this is also the capacity of the channel subject to equal rate constraints per transmit antenna . the worst - case complexity of this approach is exponential in the product of the blocklength @xmath3 and the number of antennas @xmath55 . one approach to reduce the complexity of the joint ml decoder is to employ a sphere decoder . rather than naively checking all possible codewords , the sphere decoder only examines codewords that lie within a ball around the received vector . if the radius of the ball is suitably chosen , this search is guaranteed to return the ml candidate vector . we refer interested readers to @xcite for more details on sphere decoding algorithms as well as to @xcite for a recent hardware implementation . ( -5,-18)(90,42 ) ( -15,32 ) ( -5,0)@xmath61 ( 1,0)(7,0 ) ( 7,-5)(21,5 ) ( 14,0)@xmath62 ( 28,3.5)@xmath9 ( 21,0)(35,0 ) ( -1,10)@xmath63 ( -15,-15.5 ) ( -5,0)@xmath64 ( 1,0)(7,0 ) ( 7,-5)(21,5 ) ( 14,0)@xmath65 ( 28,3.5)@xmath66 ( 21,0)(35,0 ) ( 20 , -22.5 ) ( 0,0)(30,60 ) ( 15 , 30)@xmath33 ( 52,32 ) ( -2,0)(5,0 ) ( 8,0 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 0,8)(0,3 ) ( 0,12)@xmath67 ( 3,0)(18,0 ) ( 19,4)@xmath68 ( 60,10)@xmath63 ( 52,-16 ) ( -2,0)(5,0 ) ( 8,0 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 0,8)(0,3 ) ( 0,12)@xmath69 ( 3,0)(18,0 ) ( 19,4)@xmath70 ( 78,-22.5 ) ( 0,0)(20,60 ) ( 10,30)@xmath71 ( 86,8 ) ( 12,0)(20,0 ) ( 24,15)@xmath72 ( 24,-15)@xmath73 ( 24,2)@xmath63 rather than processing all the observed signals from the antennas jointly , one simple approach is to separate out the transmitted data streams using a linear projection and then decode each data stream individually , as shown in figure [ f : trad ] . given the observed matrix @xmath74 from @xmath75 , the receiver forms the projection @xmath76 where @xmath77 . each row @xmath78 of @xmath79 is treated as a noisy version of @xmath80 . in traditional linear receivers , the goal of the projection matrix @xmath81 is to separate the incoming data streams . for the decorrelator architecture , we choose the projection to be the pseudoinverse of the channel matrix @xmath82 . in the case where @xmath83 , the resulting channel is interference free . if @xmath33 is orthogonal , then the decorrelator architecture can match the performance of a joint ml decoder . as the condition number of @xmath33 increases , the performance gap between the decorrelator and the joint decoder increases due to noise amplification ( see the example in section [ ex:2 ] ) . the performance of the decorrelator can be improved at low @xmath15 using the mmse architecture which sets @xmath84 . let @xmath85 be the @xmath86 row vector of @xmath81 and @xmath87 the @xmath86 column vector of @xmath33 . the following rate is achievable for the @xmath86 data stream using a decorrelator architecture with gaussian codebooks : @xmath88 since we focus on the case where each data stream is encoded at the same rate , the achievable sum rate is dictated by the worst stream , @xmath89 the complexity of a linear receiver architecture is dictated primarily by the choice of decoding algorithm for the individual data streams . in the worst case ( when ml decoding is used for each data stream ) , the complexity is exponential in the blocklength of the data stream . in practice , one can employ low - density parity - check ( ldpc ) codes to approach rates close to the capacity with linear complexity @xcite . the performance of this class of linear receivers can be improved using successive interference cancellation ( sic ) @xcite . after a codeword is decoded , it may be subtracted from the observed vector prior to decoding the next codeword , which increases the effective signal - to - noise ratio . let @xmath90 denote the set of all permutations of @xmath91 . for a fixed decoding order @xmath92 , let @xmath93 denote the indices of the data streams that have not yet been decoded . let @xmath94 denote the @xmath95 column vector of @xmath33 and let @xmath96 be the submatrix consisting of the columns with indices @xmath97 , i.e. , @xmath98 $ ] . the following rate is achievable in the @xmath95 , stream using successive interference cancellation : @xmath99 where @xmath100 is the projection vector to decode the @xmath95 stream after canceling the interference from the @xmath101 streams . for a fixed decoding order @xmath102 , the achievable sum - rate is given by @xmath103 the above scheme is referred to as v - blast i ( see @xcite for more details ) . an improvement can be attained by selecting the decoding order , and thus the permutation . in the case of v - blast ii , the sum rate is improved by choosing the decoding order that maximizes rate of the worst stream , @xmath104 hence , v - blast i performs worse than v - blast ii for all channel parameters . we postpone the discussion of v - blast iii to section [ sec : outage ] where we introduce the outage formulation . using ml decoding for each individual data stream , the complexity of the mmse - sic architecture is again exponential in blocklength . however , unlike the decorrelator and linear mmse receiver , not all @xmath0 streams can be decoded in parallel and delay is incurred as later streams have to wait for earlier streams to finish decoding . ( -17,-33)(145,52 ) ( -17,28 ) ( 0,0)@xmath61 ( 4,0)(10,0 ) ( 10,-5)(23,5 ) ( 16.5,0)@xmath62 ( 29,3.5)@xmath9 ( 23,0)(37,0 ) ( 0.5 , 12)@xmath63 ( -17,-12 ) ( 0,0)@xmath64 ( 6,0)(10,0 ) ( 10,-5)(23,5 ) ( 16.5,0)@xmath65 ( 30,3.5)@xmath66 ( 23,0)(37,0 ) ( 20 , -22.5 ) ( 0,0)(25,60 ) ( 12.5 , 30)@xmath33 ( 47,32 ) ( -2,0)(5,0 ) ( 8,0 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 0,8)(0,3 ) ( 0,12)@xmath67 ( 3,0)(18,0 ) ( 19,4)@xmath68 ( 68,-22.5 ) ( 5,0)(30,60 ) ( 17.5 , 30)@xmath105 ( 87,28 ) ( 11,0)(25,0 ) ( 18,5)@xmath106 ( 25,5)(40,-5 ) ( 32.5,0)@xmath107 ( 40,0)(48,0 ) ( 54,.5)@xmath108 ( 55 , 12)@xmath63 ( 47,-16 ) ( -2,0)(5,0 ) ( 8,0 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 0,8)(0,3 ) ( 0,12)@xmath69 ( 3,0)(18,0 ) ( 19,4)@xmath109 ( 119 , 12)@xmath63 ( 87,-12 ) ( 11,0)(25,0 ) ( 18.5,5)@xmath110 ( 25,5)(40,-5 ) ( 33,0)@xmath111 ( 40,0)(48,0 ) ( 54.5,.5)@xmath112 another class of linear architectures comes under the name of _ lattice - reduction _ detectors . it has been shown that lattice reduction can be used to improve the performance of the decorrelator when the channel matrix is near singular @xcite and can achieve the receive diversity @xcite . lattice - reduction detectors are symbol - level linear receivers that impose a linear constellation constraint , _ e.g. , _ a qam constellation , on the transmitters . the output of the mimo channel @xmath113 is passed through a linear filter @xmath81 to get the resulting output : @xmath114 where @xmath115 is the effective channel matrix . in lattice reduction , the effective channel matrix is restricted to be unimodular : both its entries and the entries of its inverse must be integers . let @xmath116 be the row vectors of matrix @xmath19 . the lattice - reduction detector produces estimates of the symbols of @xmath117 from @xmath118 . there are two key differences between the proposed integer - forcing receiver and the lattice - reduction receiver . first , the integer - forcing receiver operates on the codeword level rather than on the symbol level . second , the effective channel matrix @xmath19 of the integer - forcing receiver is not restricted to be unimodular : it can be any full - rank integer matrix . we compare lattice reduction to the integer receiver in example 3 by restricting the effective matrix to be unimodular under the integer - forcing architecture . we show that this restriction can result in an arbitrarily large performance gap . two other works have developed lattice architectures for joint decoding that can achieve the optimal diversity - multiplexing tradeoff @xcite . linear receivers such as the decorrelator and the mmse receiver directly decode the data streams after the projection step . in other words , they use the linear projection matrix @xmath81 to invert the channel matrix at the cost of amplifying the noise . although low in complexity , these approaches are far from optimal when the channel matrix is ill - conditioned . in the integer - forcing architecture , each encoder uses the same linear code and the receiver exploits the code - level linearity to recover equations of the transmitted messages . instead of inverting the channel , the scheme uses @xmath81 to force the effective channel to a full - rank integer matrix @xmath19 . as in the case of traditional linear receivers , each element of the effective output is then sent to a separate decoder . however , since each encoder uses the same linear code , each decoder can recover an integer linear combination of the codewords . the integer - forcing receiver is free to choose the set of equation coefficients @xmath19 to be any full - rank integer matrix . the resulting integer combinations of codewords can be mapped back to a set of full - rank messages over a finite field . is prime to ensure invertibility . however , this restriction may be removed as shown in @xcite . ] finally , the individual messages vectors are recovered from the set of full - rank equations of message vectors . the details of the architecture are provided in the sequel and an illustration is given in figure [ f : latticedemod ] . ( -17,-33)(170,52 ) ( -17,28 ) ( 0,0)@xmath61 ( 4,0)(10,0 ) ( 10,-5)(23,5 ) ( 16.5,0)@xmath62 ( 29,3.5)@xmath9 ( 23,0)(37,0 ) ( 0.5 , 12)@xmath63 ( -17,-12 ) ( 0,0)@xmath64 ( 6,0)(10,0 ) ( 10,-5)(23,5 ) ( 16.5,0)@xmath65 ( 30,3.5)@xmath66 ( 23,0)(37,0 ) ( 20 , -22.5 ) ( 0,0)(25,60 ) ( 12.5 , 30)@xmath33 ( 47,32 ) ( -2,0)(5,0 ) ( 8,0 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 0,8)(0,3 ) ( 0,12)@xmath67 ( 3,0)(18,0 ) ( 19,4)@xmath68 ( 68,-22.5 ) ( 5,0)(30,60 ) ( 17.5 , 30)@xmath119 ( 87,28 ) ( 11,0)(25,0 ) ( 18,5)@xmath106 ( 25,5)(40,-5 ) ( 32.5,0)@xmath107 ( 40,0)(54,0 ) ( 47,4.5)@xmath120 ( 55 , 12)@xmath63 ( 47,-16 ) ( -2,0)(5,0 ) ( 8,0 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 0,8)(0,3 ) ( 0,12)@xmath69 ( 3,0)(18,0 ) ( 19,4)@xmath109 ( 119 , 12)@xmath63 ( 87,-12 ) ( 11,0)(25,0 ) ( 18.5,5)@xmath110 ( 25,5)(40,-5 ) ( 33,0)@xmath111 ( 40,0)(54,0 ) ( 47,4.5)@xmath121 ( 137,-22.5 ) ( 4,0)(24,60 ) ( 14,30)@xmath21 ( 161,10)(167,10 ) ( 170 , 24)@xmath122 ( 170 , 12)@xmath63 ( 170 , -6)@xmath123 prior to decoding , our receiver projects the channel output using the @xmath32 matrix @xmath81 to get the effective channel@xmath124 each preprocessed output @xmath125 is then passed into a separate decoder @xmath126 . decoder @xmath127 attempts to recover a linear equation of the message vectors @xmath128\hspace{-0.1 in } \mod{q}\end{aligned}\ ] ] for some @xmath129 . let @xmath130 denote the vector of desired coefficients for decoder @xmath127 , @xmath131^t$ ] . we choose @xmath132 to be linearly independent . and desired coefficient vectors @xmath133 that are real - valued decompositions of a complex matrix or vector . ] decoder @xmath127 outputs an estimate @xmath134 for the equation @xmath135 . we will design our scheme such that , for any @xmath52 and @xmath3 large enough , the desired linear equations are recovered with probability of error satisfying @xmath136 let @xmath137^t$ ] denote the matrix of message vectors , @xmath138^t$ ] denote the matrix of linear equations of message vectors and @xmath139^t$ ] denote the integer matrix of equation coefficients . since @xmath19 is full - rank , the original message vectors can be recovered from the set of linear equations by a simple inverse operation : @xmath140 in the following subsections , we will provide details on the achievable rate , the choice of the coefficients of the integer matrix @xmath19 , and the complexity of our architecture . we use the compute - and - forward framework developed in @xcite to derive the achievable rate of the integer - forcing linear receiver . let @xmath141 be the @xmath86 row vector of @xmath33 . in the case where @xmath142 , the channel output to the @xmath86 decoder is given by @xmath143 and the rate at which the set of equations @xmath144 can be reliably recovered is given in the following theorem . define @xmath145 . [ thm : rate1 ] for any @xmath52 and @xmath3 large enough , there exist fixed encoders and decoders , @xmath146 , such that all decoders can recover their equations with total probability of error at most @xmath147 so long as @xmath148 for the selected equation coefficients @xmath149 . note that the decoders in theorem [ thm : rate1 ] are free to choose any equation coefficients that satisfy ( [ eqn : comprate ] ) . the encoders are completely oblivious to the choice of coefficients . it is instructive to examine the noise term @xmath150 the leading @xmath151 corresponds to the additive noise , which has unit variance in our model . the more interesting term @xmath152 corresponds to the non - integer " penalty since the channel coefficients @xmath87 are not exactly matched to the coefficients @xmath133 of the linear equation . as illustrated in figure [ f : latticedemod ] , we first multiply the channel output matrix @xmath113 by a judiciously chosen matrix @xmath153 that is , the effective channel output observed by the @xmath86 decoder can be expressed as @xmath154 where @xmath155 is the effective channel to the @xmath86 decoder and @xmath156 is the effective noise with variance @xmath157 . the achievable rate of the integer - forcing linear receiver is given in the next theorem . [ thm : rate ] consider the mimo channel with channel matrix @xmath45 . under the integer - forcing architecture , the following rate is achievable : @xmath158for any full - rank integer matrix @xmath159 and any matrix @xmath77 . applying theorem [ thm : rate1 ] with effective channel channel @xmath155 and effective noise variance @xmath157 , it follows that the receiver can reliably recover the set of linear equations @xmath144 where @xmath128\hspace{-0.1 in } \mod{q } \ . \end{aligned}\ ] ] the message vectors @xmath160 can be solved in turn by inverting the linear equations , @xmath161 . theorem [ thm : rate ] provides an achievable rate for the integer - forcing architecture for any preprocessing matrix @xmath81 and any full - rank integer matrix @xmath162 the remaining task is to select these matrices in such a way as to _ maximize _ the rate expression given in theorem [ thm : rate ] . this turns out to be a non - trivial task . we consider it in two steps . in particular , we first observe that for a fixed integer matrix @xmath163 it is straightforward to characterize the optimal preprocessing matrix @xmath153 then , in the next subsection , we discuss the problem of selecting the integer matrix @xmath162 we consider the case when @xmath83 and note that given a fixed full - rank integer matrix @xmath19 , a simple choice for preprocessing matrix is @xmath164 where @xmath165 is the pseudoinverse of @xmath33 . we call this scheme `` exact '' integer - forcing since the effective channel matrix after preprocessing is simply the full - rank integer matrix @xmath19 . we also note that choosing @xmath166 and @xmath167 corresponds to the decorrelator . more generally , the performance of exact integer - forcing is summarized in the following corollary . [ cor : ratesub ] consider the case where @xmath83 . the achievable rate from theorem [ thm : rate ] can be written equivalently as @xmath168 for any full - rank integer matrix @xmath19 by setting @xmath169 . we call the expression in the denominator of the `` effective noise variance . '' the achievable rate in ( [ eq : exactrate ] ) is determined by the largest effective noise variance , @xmath170 hence , the goal is to choose linearly independent equations @xmath171 to minimize the expression @xmath172 in . the integer - forcing receiver provides the freedom to choose any full - rank integer matrix @xmath19 . in the remainder of this section , we characterize the optimal linear projection matrix @xmath81 for a fixed coefficient matrix @xmath19 and provide an equivalent rate expression for theorem [ thm : rate ] . we will then use this expression in the section [ sec : equations ] to provide insight into how to select the optimal coefficient matrix @xmath19 . [ cor : bopt ] the optimal linear projection matrix for a fixed coefficient matrix @xmath19 is given by @xmath173 the linear mmse estimator , given by @xmath174 , is a special case of the integer - forcing receiver with @xmath175 and @xmath167 . @xmath176 . hence , under a fixed channel matrix , exact integer - forcing is optimal as @xmath177 . let @xmath178^t$ ] . we solve for each @xmath179 separately to maximize the achievable rate in theorem [ thm : rate ] , @xmath180 define this quantity to be the function @xmath181 and rewrite as follows : @xmath182 taking the derivative of @xmath183 with respect to @xmath179 , we have that @xmath184 setting @xmath185 and solving for @xmath179 , we have that @xmath186 corollary [ cor : bopt ] follows since @xmath187^t$ ] . using the optimal linear projection matrix from corollary [ cor : bopt ] , we derive an alternative expression for the achievable rate in theorem [ thm : rate ] . [ thm : alterrate ] the achievable rate from theorem [ thm : rate ] under the optimal projection matrix @xmath188 from ( [ eq : bopt ] ) can be expressed as @xmath189 where @xmath190 is the matrix composed of the eigenvectors of @xmath191 and @xmath192 is a diagonal matrix with elements@xmath193 and @xmath194 is the @xmath195 singular value of @xmath33 . let @xmath183 be defined as in ( [ eq : effecnoise ] ) and define @xmath196 to be the singular value decomposition ( svd ) of @xmath33 with @xmath197 , and @xmath190 . note that in this svd representation , @xmath198 and @xmath199 for all @xmath200 . evaluating @xmath183 for the @xmath201 row @xmath179 of @xmath188 yields @xmath202 combining and , it follows that @xmath203 since @xmath204 is an orthonormal matrix , @xmath205 and can be rewritten as follows @xmath206 since @xmath207 is an orthonormal matrix , @xmath208 and can be rewritten as follows @xmath209 putting everything together , we have that @xmath210 in the previous section , we explored choices of the preprocessing matrix @xmath81 and characterized the optimal @xmath81 for a fixed full - rank integer matrix @xmath19 . now , we discuss how to select equation coefficients @xmath211 to maximize the achievable rate in theorem [ thm : rate ] or , equivalently , theorem [ thm : alterrate ] . in the integer - forcing linear receiver , we are free to recover any full - rank set of linear equations with integer coefficients . however , due to the integer constraint on @xmath163 it does not appear to be possible to give a closed - form solution for the best possible full - rank matrix @xmath162 an initially tempting choice for @xmath19 might be @xmath212 as we noted previously , for this choice of @xmath163 selecting @xmath213 reduces to the decorrelator while selecting @xmath214 yields the linear mmse estimator . however , as we show , for most channel matrices , fixing @xmath167 is suboptimal . from theorem [ thm : alterrate ] , the achievable rate under the fixed channel matrix @xmath215^t$ ] is given by @xmath216 in general , for a fixed @xmath15 and channel matrix , finding the best coefficient matrix @xmath19 appears to be a combinatorial problem , requiring an explicit search over all possible full - rank integer matrices . the following lemma shows how the search space can be somewhat reduced . [ lm : asearch ] to optimize the achievable rate in theorem [ thm : alterrate ] ( or , equivalently , in theorem [ thm : rate ] ) , it is sufficient to check the space of all integer vectors @xmath133 with norm satisfying @xmath217 where @xmath218 is the maximum singular value of @xmath33 . this lemma thus shows that an exhaustive search only needs to check roughly @xmath219 possibilities . from , the achievable rate of the integer - forcing receiver is zero for all @xmath133 satisfying @xmath220 the left - hand side is lower bounded by @xmath221 hence , if @xmath222 , then @xmath223 . to conclude this subsection , we will now explicitly show how and why the choice @xmath167 is indeed suboptimal in general . in this context , it is instructive to use lemma [ lm : asearch ] to restate as @xmath224 equation suggests that we should choose coefficient vectors @xmath132 to be short and in the direction of the maximum right eigenvector of @xmath33 . to make this concrete , let us study a particular @xmath225 real mimo channel for which the matrix @xmath33 has singular values @xmath226 and @xmath227 with corresponding right singular vectors @xmath228 and @xmath229 respectively , as illustrated in figure [ fig : equations ] . here , decoder 1 recovers a linear combination of the transmitted message vectors with integer coefficients @xmath230^t$ ] and decoder 2 recovers a linear combination with integer coefficients @xmath231^t$ ] . using the exact integer - forcing rate from corollary [ cor : ratesub ] , the following rate is achievable @xmath232 where @xmath233 can be interpreted as the effective noise variance for the @xmath86 decoder , @xmath234 since @xmath235 , suggests that @xmath236 should be chosen in the direction of @xmath237 subject to linearly independent constraints to reduce the noise amplification by @xmath238 . in the case of the decorrelator ( or mmse receiver ) , the equation coefficients are fixed to be @xmath239^t$ ] and @xmath240^t$ ] . as a result , the noise variance in at least one of the streams will be heavily amplified by @xmath238 and the rate will be limited by the minimum singular value of the channel matrix . with integer - forcing , we are free to choose _ any linearly independent _ @xmath241 since we only require that our coefficients matrix @xmath19 be invertible . by choosing @xmath236 in the direction @xmath242 , we are protected against large noise amplification in the case of near - singular channel matrices . ( 0,35)(255,120 ) ( 0,0 ) ( 15,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 15,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 30,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 45,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 60,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 75,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 90,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 105,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 60,45)(72.5,55 ) ( 60,45)(20,90 ) ( 84,62)@xmath243 ( 28,98)@xmath244 ( 60,45)(75,45 ) ( 60,45)(60,60 ) ( 60,65)@xmath245 ( 82,45)@xmath246 ( 115,0 ) ( 15,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 15,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 30,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 45,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 60,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 75,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 90,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 105,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 120,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 60,45)(72.5,55 ) ( 60,45)(20,90 ) ( 80,45)@xmath247 ( 22.5,98)@xmath248 ( 60,45)(105,75 ) ( 60,45)(75,60 ) ( 112,75)@xmath246 ( 75,65)@xmath245 our architecture has the same implementation complexity as that of a traditional linear receiver with the addition of the matrix search for @xmath19 . the ideal joint ml receiver aggregates the time and space dimensions and finds the ml estimate across both . as a result , its complexity is exponential in the product of the blocklength and the number of data streams . our architecture decouples the time and space dimensions by allowing for single - stream decoding . first , we search for the best integer matrix @xmath19 , which has an exponential complexity in the number of data streams in the worst case . for slow fading channels , this search is only needed once per data frame . afterwards , our receiver recovers @xmath0 linearly independent equations of codewords according to @xmath19 and then solves these for the original codewords . this step is polynomial in the number of data streams and exponential in the blocklength for an ml decoder . in practice , the decoding step can be considerably accelerated through the use of ldpc codes and the integer matrix search can be sped up via a sphere decoder . in this section , we compare the performance of the integer - forcing linear receiver against existing architectures through a series of examples . in example 1 , we compare the performance of different architectures for an ill - conditioned channel matrix and demonstrate that the choice of equation coefficients for the integer - forcing receiver changes with @xmath15 . in example 2 , we compare the performance of the integer - forcing receiver with the decorrelator and show that the decorrelator can perform arbitrarily worse . in examples 3 , we illustrate that the gap between the integer - forcing receiver and lattice reduction can become unbounded . finally , we show that the gap between the integer - forcing receiver and the joint decoder can be arbitrarily large in example 4 . consider the @xmath225 real mimo channel with channel matrix @xmath249 real - valued mimo channel with fixed channel matrix from ( [ e : fixedmatrixex1]).,width=326 ] figure [ f : fixedchannel ] shows the performance of the different architectures . ( recall that we assume equal - rate data streams on both transmit antennas , as in definition [ def : rate ] . ) the achievable rates for traditional linear receiver are given by and that of the joint receiver is given by . the decorrelator and the mmse receiver aim to separate the data streams and cancel the interference from other streams . however , this is difficult since the columns of the channel matrix are far from orthogonal . the integer - forcing architecture attempts to exploit the interference by decoding two linearly independent equations in the direction of the maximum eigenvector @xmath250^t$ ] . for example , at @xmath251 , we choose equation coefficients @xmath252^t$ ] and @xmath253^t,$ ] while for @xmath254 , we choose equation coefficients @xmath255^t$ ] and @xmath256^t.$ ] thus , for different values of @xmath15 , the optimal equation coefficients generally change . consider the @xmath225 real mimo channel with channel matrix : @xmath257 where we assume @xmath258 , @xmath259 is an integer and @xmath260 . we first note that @xmath261 using with @xmath262 , the achievable rate of the decorrelator is @xmath263 we compare the achievable rate of the decorrelator with the exact integer - forcing rate from corollary [ cor : ratesub ] . the equation coefficients selected by the decoders are @xmath264^t\\ \mathbf{a}_2 & = \left[\frac{1}{\sqrt{\epsilon}}~~\frac{1}{\sqrt{\epsilon } } + 1\right]^t \ . \end{aligned}\ ] ] using corollary [ cor : ratesub ] , the achievable rate of exact integer - forcing with equations coefficients @xmath265^t$ ] is @xmath266 where the inequality follows since @xmath258 . we compare the two linear architectures to the joint ml decoder whose achievable rate is given by . for @xmath267 and @xmath260 , the rate of the joint decoder is @xmath268 finally , let us compare the three rates in the setting where @xmath269 and where the parameter @xmath147 in our channel model tends to zero according implies that @xmath270 . ] to @xmath271 in that special case , we can observe that @xmath272 hence , the loss from using the decorrelator instead of the integer - forcing receiver becomes unbounded in this regime as @xmath177 . furthermore , the integer - forcing receiver achieves the same scaling as the joint decoder . in this example , we illustrate the difference between the proposed integer - forcing architecture and the lattice - reduction receiver . first , we note that , unlike the integer - forcing receiver , the lattice - reduction receiver is not required to use a lattice code but it should use a constellation with regular spacing , such as pam or qam . however , the key difference is that the effective channel matrix for lattice - reduction receivers is restricted to be unimodular while the effective channel for integer - forcing receivers can be any full - rank integer matrix . in this example , we show that this restriction can result in an arbitrarily large performance penalty . we consider the @xmath273 mimo channel with channel matrix : @xmath274 a simple calculation shows that the inverse of this channel matrix is given by @xmath275 since @xmath276 has non - integer entries , @xmath33 is not unimodular . the coefficient matrix that maximizes the achievable rate for the exact integer - forcing receiver from corollary [ cor : ratesub ] is @xmath277 leading to an effective noise variance in each stream that satisfies @xmath278 by contrast , following the lattice - reduction receiver , we must ensure that the resulting effective channel matrix is unimodular . using the fact that @xmath279 is a basis for the body - centered cubic ( bcc ) lattice , it can be shown that the best choice of matrix is @xmath280 it follows that the effective noise variance in the worst stream is given by @xmath281 hence , as the number of antennas becomes large ( @xmath282 ) , restricting the effective matrix to be unimodular results in an arbitrarily large loss . finally , we illustrate the point that the integer - forcing receiver can sometimes be arbitrarily worse than optimal joint decoding . to see this , we consider a @xmath225 mimo channel with the channel matrix : @xmath283 where @xmath284 . the rate attainable via joint ml decoding is @xmath285 we note that the inverse of the channel matrix is given by @xmath286 we bound the achievable rate of exact integer - forcing from corollary [ cor : ratesub ] as follows @xmath287 let @xmath288 and consider the regime @xmath177 . the gap between the optimal joint receiver and the integer - forcing linear receiver can be arbitrarily large . however , as we will see in section [ sec : outage ] , the average behavior of the integer - forcing linear receiver is close to the joint decoder under a rayleigh fading distribution for medium to high @xmath15 . we now demonstrate that integer - forcing receiver nearly matches the performance of the joint ml decoder under a slow fading channel model . since the integer - forcing receiver can mimic the operation performed by a zero - forcing or mmse receiver ( as well as decode messages via equations ) , it is not surprising that it offers higher rates . however , these architectures are often coupled with some form of sic . we will show that the integer - forcing receiver outperforms the following standard sic architectures : * v - blast i : the receiver decodes and cancels the data streams in a predetermined order , irrespective of the channel realization . each data stream has the same rate . see for the rate expression . * v - blast ii : the receiver selects the decoding order separately for each channel realization in such a way as to maximize the effective snr for the data stream that sees the worst channel . each data stream has the same rate . see for the rate expression . * v - blast iii : the receiver decodes and cancels the data streams in a predetermined order . the rate of each data stream is selected to maximize the sum rate . the rate expression is given in section [ subsec : rateallocation ] . in sections [ subsec : outage ] and [ subsec : dmt ] , we compare these schemes through simulations as well as their diversity - multiplexing tradeoffs . for completeness , we also compare integer - forcing to an sic architecture that allows for both variable decoding order and unequal rate allocation in appendix [ ap : vblast4 ] . we adopt the standard quasi - static rayleigh fading model where each element of the complex channel matrix is independent and identically distributed according to a circularly symmetric complex normal distribution of unit variance . the transmitter is only aware of the channel statistics while the receiver knows the exact channel realization . as a result , we will have to cope with some outage probability @xmath289 . [ def : outage ] assume there exists an architecture that encodes each data stream at the same rate and achieves sum rate @xmath51 . for a target rate @xmath290 , then the outage probability is defined as @xmath291 for a fixed probability @xmath292 $ ] , we define the _ outage rate _ to be @xmath293 we have assumed that each data stream is encoded at the same rate . this is optimal for linear receivers under isotropic fading . however , when sic is used , rate allocation can be beneficial in an outage scenario . to compare the performance of integer - forcing to sic with rate allocation , we consider v - blast iii in this section . v - blast iii performs sic with a fixed decoding order and allows for rate allocation among the different data streams . without loss of generality for rayleigh fading , if we fix a decoding order , we may take it to be @xmath294 . from , the achievable rate for stream @xmath127 is @xmath295 since the streams are decoded in order , the later streams will achieve higher rates on average than the earlier streams . v - blast iii allocates lower rates to earlier streams and higher rates to later streams . we now generalize definition [ def : outage ] to include rate allocation . assume an architecture that achieves rate @xmath296 in stream @xmath127 . for a target rate @xmath290 , the the outage probability is given by : @xmath297 for a fixed probability @xmath292 $ ] , we define the _ outage rate _ to be : @xmath298 we now compare the outage rate and probabilities for the receiver architectures discussed above . it is easy to see that the zero - forcing receiver performs strictly worse than the mmse receiver and v - blast i performs strictly worse than v - blast ii . we have chosen to omit these two architectures from our plots to avoid overcrowding . figure [ f : out2p ] shows the @xmath151 percent outage rate and figure [ f : out5p ] shows the @xmath299 percent outage rate . in both cases , the integer - forcing receiver nearly matches the rate of the joint ml receiver while the mmse receiver achieves significantly lower performance . the sic architectures with either an optimal decoding order , v - blast ii , or an optimized rate allocation , v - blast iii , improve the performance of the mmse receiver significantly but still achieve lower rates than the integer - forcing receiver from medium @xmath15 onwards . our simulations suggest that the outage rate of the integer - forcing receiver remains within a small gap from the outage rate of the joint ml receiver . however , we recall from the example given in subsection [ ex:4 ] that it is not true that the integer - forcing receiver is uniformly near - optimal for all fading realizations . figure [ f : poutro3 ] shows the outage probability for the target sum rate @xmath300 . we note that the integer - forcing receiver achieves the same slope as the joint decoder . in the next subsection , we characterize the diversity - multiplexing tradeoff of the integer - forcing receiver and compare it with the diversity - multiplexing tradeoff of traditional architectures that are considered in @xcite . we show that the integer - forcing receiver attains the optimal diversity multiplexing - tradeoff in the case where each transmit antenna sends an independent data stream . complex - valued mimo channel with rayleigh fading.,width=345 ] complex - valued mimo channel with rayleigh fading.,width=345 ] complex - valued mimo channel with rayleigh fading for a target sum rate of @xmath300.,width=345 ] the diversity - multiplexing tradeoff ( dmt ) provides a rough characterization of the performance of a mimo transmission scheme at high @xmath15 @xcite . a family of codes is said to achieve spatial multiplexing gain @xmath301 and diversity gain @xmath302 if the total data rate and the average probability of error satisfy @xmath303 in the case where each transmit antenna encodes an independent data stream , the optimal dmt is @xmath304 where @xmath305 $ ] and can be achieved by joint ml decoding @xcite . the dmts achieved by the decorrelator and sic architectures are as follows @xcite : @xmath306 the decorrelator chooses the matrix @xmath81 to cancel the interference from the other data streams . as a result , the noise is heavily amplified when the channel matrix is near singular and the performance is limited by the minimum singular value of the channel matrix . in the integer - forcing linear receiver , the effective channel matrix @xmath19 is not limited to be the identity matrix but can be any full - rank integer matrix . this additional freedom is sufficient to recover the same dmt as the joint ml decoder . [ thm : dmt ] for a mimo channel with @xmath0 transmit , @xmath83 receive antennas , and rayleigh fading , the achievable diversity - multiplexing tradeoff for the integer - forcing receiver is given by@xmath307 where @xmath305 $ ] . the proof of theorem [ thm : dmt ] is given in appendix [ ap : dmt ] . figure [ fig : dmt ] illustrates the dmt for a @xmath308 mimo channel . the integer - forcing receiver achieves a maximum diversity of @xmath309 while the decorrelator and v - blast i achieve can only achieve a diversity of @xmath151 since their performance is limited by the worst data stream . v - blast ii achieves a higher dmt than v - blast i but a lower diversity than the integer - forcing receiver since its rate is still limited by the worst stream after the optimal decoding order is applied . v - blast iii achieves the optimal diversity at the point @xmath310 since only one data stream is used in transmission . for values of @xmath311 , the achievable diversity is suboptimal . as noted earlier , receiver architectures based on zero - forcing face a rate penalty when the channel matrix is ill - conditioned . integer - forcing circumvents this issue by allowing the receiver to first decode equations whose coefficients are matched with those of the channel . from one perspective , the resulting gains are of a similar nature as those obtained by lattice - reduction receivers . one important difference is that integer - forcing applies a modulo operation at the receiver prior to decoding , which retains the linear structure of the codebook . this allows us to derive closed - form rate expressions analogous to those for traditional linear receiver . typically , lattice reduction is used at the symbol level followed by a decoding step @xcite . while this form of lattice reduction can be used to obtain the full receive diversity @xcite , it does not seem to suffice in terms of rate . another key advantage of integer - forcing is that it completely decouples the spatial aspect of decoding from the temporal aspect . that is , the search for the best integer matrix @xmath19 to approximate the channel matrix @xmath33 is completed before we attempt to decode the integer combinations of codewords . thus , apart from the search for the best @xmath19 , which in a slow - fading environment does not have to be executed frequently , the complexity of the integer - forcing receiver is similar to that of the zero - forcing receiver . from our outage plots , it is clear that the integer - forcing receiver significantly outperforms the basic mmse receiver . moreover , integer - forcing beats more sophisticated sic - based v - blast architectures , even when these are permitted to optimize their rate allocation while integer - forcing is not . we note that it is possible to develop integer - forcing schemes that permit unequal rate allocations @xcite as well as a form of interference cancellation @xcite but this is beyond the scope of the present paper . integer - forcing also attains the full diversity - multiplexing tradeoff , unlike the v - blast architectures discussed above . earlier work developed lattice - based schemes that attain the full dmt @xcite but , to the best of our knowledge , ours is the first that decouples spatial decoding from temporal decoding . the caveat is that the dmt result presented in the current paper only applies if there is no spatial coding across transmit antennas , whereas the dmt results of @xcite apply in general . characterizing the dmt of integer - forcing when there is coding across transmit antennas is an interesting subject for future study . we have studied the performance of the integer - forcing linear receiver under the standard mimo channel and found that it achieves outage rates close those of the joint ml decoder as well as the same dmt . in this section , we show that integer - forcing architectures are also successful at dealing with a different kind of channel disturbance , namely interference . we assume that the interfering signal is low - dimensional ( compared to the number of receive antennas ) , and we are most interested in the case where the variance of this interfering signal increases ( at a certain rate ) with the transmit power . we show that the integer - forcing architecture can be used to perform `` oblivious '' interference mitigation . by oblivious , we mean that the transmitter and receiver are unaware of the codebook of the interferer ( if there is one ) . however , the receiver knows which subspace is occupied by the interference . by selecting equation coefficients in a direction that depends both on the interference space and on the channel matrix , the integer - forcing receiver reduces the impact of interference and attains a significant gain over traditional linear receivers . we will characterize the generalized degrees - of - freedom show that it matches that of the joint ml decoder . oblivious receivers have been thoroughly studied in the context of cellular systems @xcite and distributed mimo @xcite . for ease of notation and tractability , the discussion presented in this section is limited to channels whose channel matrix is square , i.e. , with equal number of transmit and receive antennas . recall that the real - valued representation of the @xmath273 complex - valued mimo channel ( see definition [ def : channel ] ) is given by @xmath312 where @xmath313 is the channel output , @xmath314 is the real - valued representation of the fading matrix , @xmath315 is the channel input , and the noise @xmath316 has i.i.d . gaussian entries with unit variance . in this section , we extend the standard mimo channel to include the case of interference . the generalized model has channel output @xmath317 where @xmath33 is the channel matrix , @xmath44 is the channel input , and @xmath318 is the noise , all as in the previous model . an external interferer adds @xmath319 in the direction represented by the column space of @xmath320 . we assume that each element of @xmath207 is i.i.d . gaussian with variance @xmath321 . we assume that @xmath322 is fixed during the whole transmission block and known only to the receiver . the definition for messages , rates , encoders , and decoders follow along similar lines as those for the standard mimo channel ( see definitions [ def : messages ] , [ def : encoders ] , [ def : decoder ] , and [ def : rate ] in section [ sec : problem ] ) . as in the case without interference , traditional linear receivers process the channel output @xmath113 by multiplying it by a @xmath323 matrix @xmath81 to arrive at the effective output @xmath324 and recover the message @xmath1 using only the @xmath86 row of the matrix @xmath325 by analogy to , the achievable sum rate can be expressed as @xmath326 where @xmath327 represents the achievable rate for the @xmath86 data stream ( using gaussian codebooks ) , @xmath328 again , let us discuss several choices of the matrix @xmath153 the decorrelator , given by @xmath262 , removes the interference due to other data streams but does not cancel the external interference @xmath322 ( except in the very special case where the subspace spanned by @xmath322 is orthogonal to the subspace spanned by @xmath276 ) . alternatively , if we choose @xmath329 where @xmath330 is the @xmath331 matrix whose rowspace is orthogonal to the columnspace of @xmath332 then the external interference term is indeed nulled . the resulting output @xmath333 can then be processed by a traditional linear receiver . this scheme achieves good performance in high @xmath321 regimes but does not perform well in high @xmath15 regimes since the interference due to the other data streams is mostly unresolved . the mmse receiver improves the performance of the both architectures by choosing @xmath334 . however , since there are @xmath34 data streams and the interference is of dimension @xmath335 it is impossible to cancel both the interference from other data streams and the external interference with any matrix @xmath81 . one way out of this conundrum is to reduce the number of transmitted streams to @xmath336 . for this scenario , the mmse receiver can be applied to mitigate both the external interference and the interference from other data streams . complexity permitting , we can again improve performance by resorting to successive interference cancellation architectures . the achievable rate for v - blast i in the standard mimo channel from becomes @xmath337 where @xmath338 and @xmath339 . the achievable rate for v - blast ii follows by maximizing the rate in over all decoding orders @xmath92 . we apply the integer - forcing linear receiver proposed in section [ sec : architecture ] to the problem of mitigating interference ( see figure [ fig : integerinf ] ) . the channel output matrix @xmath113 is first multiplied by a fixed matrix @xmath81 to form the matrix @xmath79 whose @xmath86 row is the signal fed into the @xmath86 decoder . each such row can be expressed as @xmath340 where @xmath155 is the effective channel to the @xmath86 decoder , @xmath341 is the effective interference with variance @xmath342 , and @xmath156 is the effective noise with variance @xmath157 . the next theorem and its following remarks generalize theorem [ thm : rate ] , corollary [ cor : ratesub ] , and corollary [ cor : bopt ] to include the case with interference . ( -24,-65)(190,40 ) ( -13,32 ) ( -1,0)@xmath61 ( 4,0)(10,0 ) ( 10,-5)(23,5 ) ( 16.5,0)@xmath62 ( 29,3.5)@xmath9 ( 23,0)(41,0 ) ( 3.5 , 12)@xmath63 ( -13,-16 ) ( -1,0)@xmath64 ( 6,0)(10,0 ) ( 10,-5)(23,5 ) ( 16.5,0)@xmath65 ( 30,3.5)@xmath66 ( 23,0)(41,0 ) ( 24 , -22.5 ) ( 4,0)(16,60 ) ( 10 , 30)@xmath33 ( 46 , -50.5 ) ( 4,-20)(16,30 ) ( -8,13)(4,13 ) ( -13,13)@xmath343 ( -13,6)@xmath63 ( -8,-5)(4,-5 ) ( -13,-5)@xmath344 ( 10 , 7)@xmath322 ( 16,-15)(28,32 ) ( 28,35 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 16,25)(28,79 ) ( 28,82 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 42,32 ) ( -2,0)(5,0 ) ( 8,0 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 0,8)(0,3 ) ( 0,12)@xmath67 ( 3,0)(21,0 ) ( 27,0)(44,0 ) ( 42,4.5)@xmath68 ( 94,-22.5 ) ( 0,0)(15,60 ) ( 8 , 30)@xmath81 ( 97,32 ) ( 12,0)(31,0 ) ( 22,4.5)@xmath106 ( 31,5)(46,-5 ) ( 38.5,0)@xmath107 ( 46,0)(63,0 ) ( 54,4)@xmath345 ( 50 , 12)@xmath63 ( 42,-16 ) ( -2,0)(5,0 ) ( 8,0 ) ( 0,0)3 ( -2,0)(2,0 ) ( 0,-2)(0,2 ) ( 0,8)(0,3 ) ( 0,12)@xmath346 ( 3,0)(21,0 ) ( 27,0)(44,0 ) ( 42.5,4)@xmath347 ( 135 , 12)@xmath63 ( 97,-16 ) ( 12,0)(31,0 ) ( 22,4.5)@xmath110 ( 31,5)(46,-5 ) ( 39,0)@xmath111 ( 46,0)(63,0 ) ( 54,4.5)@xmath348 ( 159,-22.5 ) ( 0,0)(18,60 ) ( 9,30)@xmath21 ( 177,10)(184,10 ) ( 188 , 24)@xmath122 ( 188 , 12)@xmath63 ( 188 , -6)@xmath123 ( 130,-40)@xmath349 \hspace{-0.05in}\mod q$ ] [ thm : rateinterference ] consider the mimo channel with channel matrix @xmath314 and interference matrix @xmath320 . for any full - rank integer matrix @xmath159 and any @xmath323 matrix @xmath350^t$ ] , the following sum rate is achievable using the integer - forcing linear receiver : @xmath351 where @xmath352 is given by @xmath353 exact integer - forcing selects @xmath354 . the achievable rate can be expressed more concisely as @xmath355 the optimal projection matrix that maximizes the achievable rate in theorem [ thm : rateinterference ] is given by @xmath356 in the case without interference , the equation coefficients @xmath357 should be chosen in the direction of the maximum eigenvector of @xmath191 to minimize the effective noise ( see figure [ fig : equations ] ) . when the interference is large , the equations coefficients should instead be chosen as close to orthogonal to the effective interference as possible . consider the ( suboptimal ) rate expression in . the effective " noise variance in the @xmath86 stream is @xmath358 let @xmath218 be the maximum singular value of @xmath276 and @xmath359 . the effective noise variance can be bounded by @xmath360 in the high interference regime ( @xmath361 ) , the equation coefficients should be chosen orthogonal to the direction of the effective " interference @xmath362 to minimize the effective noise variance . this is illustrated in figure [ fig : eqnsinf ] . in the case of traditional linear receivers , the equation coefficients are fixed to be the unit vectors : @xmath363^t , \mathbf{a}_2 = [ 0~1~\cdots0]^t , \ldots , \mathbf{a}_{2 m } = [ 0~0~\cdots1]^t$ ] . as a result , the interference space spanned by @xmath362 has significant projections onto at least some of the decoding dimensions @xmath133 . by contrast , in the case of the integer - forcing linear receiver , since @xmath364 need only be linearly independent , we can choose all of the decoding dimensions @xmath133 to be close to orthogonal to @xmath365 ( 2,40)(255,120 ) ( 0,0 ) ( 15,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 15,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 30,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 45,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 60,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 75,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 90,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 60,45)(72.5,55 ) ( 60,45)(20,90 ) ( 28,98)@xmath362 ( 60,45)(75,45 ) ( 60,45)(60,60 ) ( 60,65)@xmath245 ( 82,45)@xmath246 ( 105,0 ) ( 15,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 15,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 30,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 45,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 60,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 75,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 90,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 105,0 ) ( 0,30)(0,45)(0,60)(0,75)(0,90)(0,105 ) ( 60,45)(72.5,55 ) ( 60,45)(20,90 ) ( 28,98)@xmath362 ( 60,45)(105,75 ) ( 60,45)(75,60 ) ( 112,75)@xmath246 ( 75,65)@xmath245 [ fig : equationsinf ] to illustrate the impact of choosing equation coefficients in a fashion suitable to mitigate external interference , we consider the @xmath225 mimo channel with channel matrix @xmath33 and one - dimensional interference space @xmath322 given by @xmath366 where @xmath367 . in the case of the decorrelator , we invert the channel to arrive at the effective output : @xmath368 where @xmath369 . for @xmath361 , the effective noise variances scale as @xmath370 using the integer - forcing linear receiver with the choice of equations @xmath371^t$ ] and @xmath372^t$ ] , the effective channel output to the second decoder is @xmath373 where @xmath374 is the codeword sent by the @xmath375 antenna . it follows that the effective noise variances are @xmath376 where @xmath377 is a constant that does not scale with @xmath321 . in this example , the integer - forcing linear receiver is able to completely cancel the effect of interference in the second stream by choosing equation coefficients appropriately . we evaluate the generalized degrees of freedom for the @xmath273 complex mimo channel with @xmath378-dimensional interference . we specify the interference - to - noise ratio through the parameter @xmath379 where @xmath380 and consider the case where @xmath381 . the generalized degrees of freedom are defined as follows ( see @xcite ) : ( generalized degrees - of - freedom ) for a given channel matrix @xmath33 and interference matrix @xmath322 , the generalized degrees - of - freedom of a scheme is @xmath382 where @xmath383 is the achievable sum rate of the scheme . ( rational independence ) we call a matrix @xmath384 rationally independent if for all @xmath385 , we have that @xmath386 we consider the set of matrices @xmath387 such that @xmath388 is rationally independent . it can be seen that this set has lebesgue measure one . in the next theorem , we show that for this class of matrices , the integer - forcing linear receiver achieves the same number of generalized degrees of freedom as the joint decoder , and is thus optimal . [ thm : dof ] consider the @xmath273 complex mimo channel with @xmath378 dimensional interference . the integer - forcing linear receiver achieves the generalized degrees of freedom @xmath389 for a set of @xmath390 that have lebesgue measure one . a straightforward derivation shows that the optimal joint decoder with @xmath34 streams of data achieves the following generalized degrees of freedom @xmath391 for all full - rank channel matrices @xmath392 . the linear mmse receiver with @xmath34 data streams and the linear mmse receiver with @xmath393 data streams achieve the following degrees of freedom for all full - rank channel matrices @xmath392 : @xmath394 when @xmath34 data streams are transmitted ( on the real - valued representation of the complex mimo channel ) , the mmse receiver does not achieve the optimal number of degrees of freedom since it treats the interference as noise at high @xmath15 while the integer - forcing linear receiver mitigates the interference . when only @xmath395 data streams are transmitted , the linear mmse receiver can first cancel the interference and then separate the data streams to achieve a degree of freedom of @xmath395 . however , this is suboptimal for all regimes @xmath396 ( see figure 12 ) . a straightforward calculation shows that when the number of transmitted data streams is between @xmath395 and @xmath34 , the performance is strictly suboptimal in terms of degrees of freedom . mimo channel with @xmath397-dimensional interference @xmath398,width=345 ] [ fig : dof ] our proof of theorem [ thm : dof ] uses the following lemma . recall that a set of vectors is linearly independent if none of the vectors can be written as a linear combination of the others . [ lm : diophantine ] let @xmath399 be rationally independent and assume @xmath400 for all @xmath401 . there exists a @xmath402 such that for all @xmath403 , there exist @xmath34 linearly independent integer vectors @xmath404^t \in \mathbb{z}^{2 m - 2k } \times \mathbb{z}^{2k}$ ] for @xmath405 satisfying @xmath406 where @xmath377 is a constant that is independent of @xmath407 . lemma [ lm : diophantine ] bounds the approximation of @xmath408 with @xmath34 linearly independent integer vectors . its proof is given in appendix [ ap : diophantine ] . we now prove theorem [ thm : dof ] . ( theorem [ thm : dof ] ) to establish this result , we use the ( generally suboptimal ) choice of the matrix @xmath81 that we have referred to as exact integer - forcing , i.e. , from . the achievable rate of this version of the integer - forcing linear receiver is given by @xmath409 with @xmath410 , the effective noise in the worst data stream can be expressed as @xmath411 where @xmath412 is the maximum singular value of @xmath276 and we use the shorthand @xmath413 . let us partition @xmath414 into two parts , @xmath415 \ , \end{aligned}\ ] ] where @xmath416 and @xmath417 . observe that with probability one , @xmath414 has rank @xmath418 ( hence , is full - rank ) . this implies that we can permute the columns of @xmath414 in such a way as to ensure that its last @xmath418 columns are linearly independent . if we use the same permutation on the coefficients of the vector @xmath419 our upper bound on the effective noise variance given in will remain unchanged . therefore , without loss of generality , we may assume that @xmath420 has rank @xmath418 . we define @xmath421 . then , we can write @xmath422\\ & = [ -\mathbf{t } , \mathbf{i}_{2k}]\ . \label{eq : tildej}\end{aligned}\ ] ] let the coefficients for the @xmath86 equation be given by @xmath423^t$ ] where @xmath424 and @xmath425 . we use to bound @xmath426 : @xmath427 \mathbf{a}_m\|^2 \\ & \leq \lambda^2_{\text{\tiny{max}}}(\mathbf{s}_2 ) \|[-\mathbf{t } , \mathbf{i}_{2k}]\mathbf{a}_m\|^2\\ \label{eq : boundja } & = \lambda^2_\text{\tiny{max}}(\mathbf{s}_2 ) \|\mathbf{t}\mathbf{q}_m - \mathbf{p}_m\|^2\end{aligned}\ ] ] where @xmath428 the maximum singular value of @xmath420 . combining with , the effective noise variance is bounded as follows : @xmath429 we proceed to bound the quantity @xmath430 . we decompose @xmath408 into its integer and fractional parts : @xmath431 where @xmath432 represents the integer part of @xmath408 and @xmath433 represents the fractional part of @xmath408 . we define @xmath434^t\\ \tilde{\mathbf{a } } & = [ \tilde{\mathbf{a}}_1 \cdots \tilde{\mathbf{a}}_{2m}]^t\\ \mathbf{g } & = \left ( \begin{array}{ccc } \mathbf{i}_{2m-2k } & \mathbf{0 } \\ -\mathbf{t}_i & \mathbf{i}_{2k } \end{array } \right ) \ .\end{aligned}\ ] ] since @xmath435 is a @xmath323 lower triangular matrix with non - zero diagonal elements , it has rank @xmath34 . we note that @xmath436 since @xmath435 is invertible , it follows that if the matrix formed by the coefficient vectors @xmath437 is full - rank , then the matrix formed by @xmath438 is full - rank and vice versa . from , it have that @xmath439 and , from , we have that @xmath440 from , , and , the effective noise variance can be upper bounded by @xmath441 combining and , the effective noise variance from is bounded by @xmath449 let @xmath407 scale according to @xmath450 . it follows that @xmath451 setting @xmath452 , we find that the generalized degrees - of - freedom are @xmath453 which concludes the proof of theorem [ thm : dof ] . in this paper , we proposed a new receiver architecture for mimo channels that bridges the performance gap between traditional linear receivers and the optimal joint ml decoder . we studied the scenario without coding across the transmit antennas and found that the proposed integer - forcing linear receiver performs close to the optimal joint receiver while incurring only some additional complexity over traditional linear receivers . we characterized the diversity multiplexing tradeoff of the proposed architecture and showed that is it the same as that of the joint ml decoder . we considered a generalized mimo channel model with interference and found that the integer - forcing receiver can be used to effectively mitigate interference in an oblivious . furthermore , the proposed architecture achieves the same generalized degrees - of - freedom as the joint ml decoder . an interesting question for future work is how the integer - forcing architecture behaves when coding across transmit antennas is permitted . recall that v - blast ii performs decoding in the optimal order and v - blast iii allows for rate allocation . in this appendix , we introduce v - blast iv , which allows for both rate allocation and an optimized decoding order . under v - blast iv , the data streams are decoded with respect to the ordering @xmath454 where @xmath455 is given by . we compare the behavior of v - blast iv to that of the integer - forcing linear receiver in figures [ f : outv41 ] , [ f : outv45 ] , [ f : poutpv4 ] . the results show although v - blast iv achieves good performance for low to medium @xmath15 , the integer - forcing linear achieves higher outage rates and lower outage probabilities in the medium to high @xmath15 regime . complex - valued mimo channel with rayleigh fading.,width=355 ] complex - valued mimo channel with rayleigh fading.,width=355 ] complex - valued mimo channel with rayleigh fading for a target sum rate of @xmath300.,width=355 ] in order to establish theorem [ thm : dmt ] , we need a few key facts about lattices . a lattice @xmath456 is a set of points that satisfy the following properties : @xmath457 we call the rank-@xmath458 matrix @xmath435 a generator matrix for @xmath459 if @xmath460 we use the definition of dual lattices from @xcite . given a lattice @xmath456 with a rank-@xmath458 generator matrix @xmath435 , the dual lattice @xmath461 has generator matrix @xmath462 , @xmath463 to prove theorem [ thm : dmt ] , we consider _ successive minima _ for the involved lattices , a standard concept from the diophantine approximation literature ( see e.g. @xcite ) , defined as follows . [ def : smmse ] let @xmath464 be the unit ball . given a lattice @xmath456 with a rank-@xmath458 generator matrix , the @xmath86 successive minimum @xmath465 where @xmath466 is given by @xmath467 definition [ def : smmse ] implies that @xmath468 for any lattice @xmath459 . the following basic property linking the successive minima of a lattice with those of its dual lattice is key to our proof . [ lm : smdual ] let @xmath456 be an arbitrary lattice with a rank-@xmath458 generator matrix and @xmath461 be its dual lattice . the successive minima for @xmath459 and @xmath461 satisfy the following inequality : @xmath469 finally , we also need the following result concerning a random gaussian lattice . [ lm : smfirst ] let @xmath45 be the real - valued decomposition of a @xmath470 complex gaussian matrix with i.i.d . rayleigh entries . let @xmath471 be the lattice generated by @xmath33 . then @xmath472 where @xmath473 and @xmath474 are constants independent of @xmath475 . ( theorem [ thm : dmt ] ) . let @xmath476 be the target rate where @xmath305 $ ] . for a fixed set of equations @xmath215^t$ ] and a fixed preprocessing matrix @xmath477^t$ ] , the outage probability is given by @xmath478 for a fixed set of equations @xmath19 , we are free to choose any projection matrix @xmath81 , resulting in the following bound : @xmath479 we then choose the best set of full - rank equations by optimizing over all integer matrices @xmath159 with non - zero determinant : @xmath480 we use properties of dual lattices to bound . for a fixed @xmath33 , let @xmath481 be the lattice generated by @xmath33 and @xmath482 be the dual lattice generated by @xmath483 , @xmath484 using the definition of successive minima ( definition [ def : smmse ] ) , it follows that @xmath485 we now express in terms of the successive minima of @xmath482 , @xmath486 using lemma [ lm : smdual ] , we can bound the successive minima of @xmath482 in terms of the successive minima of @xmath481 , @xmath487 combining and , the outage probability is upper bounded by @xmath488 this probability can in turn be upper bounded using lemma [ lm : smfirst ] . for large @xmath15 , we find that @xmath489 where @xmath490 are constants independent of @xmath15 . the achievable diversity for multiplexing gain @xmath301 is thus @xmath491 in order to prove lemma [ lm : diophantine ] , we use a technique introduced in @xcite . we first construct semi - norms @xmath492 and @xmath493 as well as a norm @xmath494 . we then apply minkowski s theorem ( see theorem [ thm : sm ] below ) to find @xmath34 linearly independent integer vectors that achieve the successive minima ( with respect to the norm @xmath495 ) . afterwards , we will show that these integer vectors satisfy the conditions in lemma [ lm : diophantine ] . we will need the following definitions and theorems in the proof . let @xmath494 be a norm . the unit ball with respect to @xmath495 ( or @xmath495-unit ball ) is denoted by @xmath496 the volume of @xmath497 is denoted by @xmath498 . [ def : smgen ] let @xmath494 be a norm and @xmath497 be the @xmath495-unit ball . the @xmath86 successive @xmath495-minimum @xmath499 is given by @xmath500for @xmath501 . [ thm : sm ] for any norm @xmath494 , the successive @xmath495-minima satisfy @xmath502 [ dirichlet ] [ thm : dirichlets ] for any @xmath399 and @xmath503 , there exists a @xmath504^t \in \mathbb{z}^{2m-2k } \times \mathbb{z}^{2k } \backslash \mathbf{0}$ ] such that @xmath505 [ thm : kg ] fix a function @xmath506 . if @xmath507 then , for almost all @xmath399 satisfying @xmath400 for all @xmath401 , there are only finitely many solutions @xmath504^t \in \mathbb{z}^{2m-2k } \times \mathbb{z}^{2k}$ ] to the inequality @xmath508 theorem [ thm : sm ] can be found in ( * ? ? ? * equation 1.1 ) . theorems [ thm : dirichlets ] and [ thm : kg ] can be found in many standard diophantine approximation texts ( see , for instance , @xcite ) . ( lemma [ lm : diophantine ] ) . for any vector @xmath509 , we denote the first @xmath395 components by @xmath510 and the remaining @xmath418components by @xmath511 and will thus write @xmath512 from the lemma statement , @xmath513^t$ ] is a ( rationally independent ) @xmath514 real - valued matrix with @xmath400 for all @xmath401 . for a fixed @xmath408 , we define the semi - norms @xmath183 , @xmath515 as follows : @xmath516 for a fixed @xmath517 , we let @xmath518 denote the minimum value of @xmath519 under the constraint @xmath520 , @xmath521^t \neq \mathbf{0 } } } \|\mathbf{t}\mathbf{q } - \mathbf{p}\| , \label{eq : lambda1}\end{aligned}\ ] ] @xmath522 denote the integer vector that achieves @xmath518 @xmath523^t \neq \mathbf{0 } } } \|\mathbf{t}\mathbf{q } - \mathbf{p}\|,\end{aligned}\ ] ] and @xmath524 be the value of @xmath525 evaluated by @xmath515 , @xmath526 based on the seminorms @xmath183 and @xmath515 , we define the function @xmath494 as follows : @xmath527 in the sequel , we show that @xmath495 is a norm for @xmath528 . we define the @xmath323 matrix @xmath529 as follows : @xmath530\ ] ] note that we can rewrite the function @xmath495 using @xmath529 : @xmath531 since exchanging rows of a matrix only affects the sign of its determinant , we have that @xmath532 \right ) \right| \ . \end{aligned}\ ] ] now we use the fact that the determinant of a lower triangular matrix is just the product of its diagonal entries , latexmath:[\ ] ] and , we have that @xmath596 by definition , we have that @xmath597 using , , , and , and assuming that @xmath407 is sufficiently large , we bound @xmath598 as follows : @xmath599 where @xmath591 is a constant that does not depend on @xmath407 . similarly , we can bound @xmath600 as follows : @xmath601 which concludes the proof of lemma [ lm : diophantine ] . the authors are grateful to or ordentlich for pointing out the suboptimality of unimodular integer forcing , giuseppe caire for corrections in the proof of theorem [ thm : dof ] , and seyong park for many helpful discussions . , `` ieee standard for information technology telecommunications and information exchange between systems local and metropolitan area networks specific requirements part 11 : wireless lan medium access control ( mac ) and physical layer ( phy ) specifications amendment 5 : enhancements for higher throughput , '' _ ieee std 802.11n-2009 _ , pp . c1502 , 29 2009 . j. winters , j. salz , and r. gitlin . , `` the impact of antenna diversity on the capacity of wireless communication systems , '' _ ieee transactions on communications _ , vol . 42 , no . 234 , pp . 1740 1751 , feb / mar / apr 1994 . u. niesen and p. whiting , `` the degrees of freedom of compute - and - forward , '' in _ proceedings of the ieee international symposium on information theory ( isit 2011 ) _ , st . petersburg , russia , august 2011 . m. o. damen , h. el gamal , and g. caire , `` on maximum - likelihood detection and the search for the closest lattice point , '' _ ieee transactions on information theory _ , vol . 49 , no . 23892402 , october 2003 . a. burg , m. borgmann , m. wenk , m. zellweger , w. fichtner , and h. bolcskei , `` vlsi implementation of mimo detection using the sphere decoding algorithm , '' _ ieee journal of solid - state circuits _ , vol . 40 , no . 7 , pp . 1566 1577 , july 2005 . t. richardson , m. shokrollahi , and r. urbanke , `` design of capacity - approaching irregular low - density parity - check codes , '' _ ieee transactions on information theory _ , vol . 47 , no . 2 , pp . 619 637 , february 2001 . m. varanasi and t. guess , `` optimum decision feedback multiuser equalization and successive decoding achieves the total capacity of the gaussian multiple - access channel , '' in _ conference record of the 31st asilomar conference on signals , systems , and computers _ , pacific grove , ca , usa , november 1997 , pp . 1405 1409 . p. w. wolniansky , g. j. foschini , g. d. golden , and r. a. valenzuela , `` v - blast : an architecture for realizing very high data rates over the rich - scattering wireless channel , '' in _ ursi international symposium on signals , systems , and electronics ( issse 98 ) _ , pisa , italy , sept .- oct . 1998 . h. yao and g. w. wornell , `` lattice - reduction - aided detectors for mimo communication systems , '' in _ proceedings of the ieee global communications conference ( globecom 2002 ) _ , taipei , taiwan , november 2002 . m. taherzadeh , a. mobasher , and a. khandani , `` lll reduction achieves the receive diversity in mimo decoding , '' _ ieee transactions on information theory _ 53 , no . 12 , pp . 48014805 , december 2007 . h. el gamal , g. caire , and m. o. damen , `` lattice coding and decoding achieve the optimal diversity - multiplexing tradeoff of mimo channels , '' _ ieee transactions on information theory _ 50 , no . 6 , pp . 968985 , june 2004 . j. jalden and p. elia , `` dmt optimality of lr - aided linear decoders for a general class of channels , lattice designs and system models , '' _ ieee transactions on information theory _ , 56 , no . 10 , pp . 47654780 , october 2010 . o. ordentlich , j. zhan , u. erez , m. gastpar , and b. nazer , `` practical code design for compute - and - forward , '' in _ proceedings of the ieee international symposium on information theory ( isit 2011 ) _ , st . petersburg , russia , august 2011 . j. zhan , b. nazer , , o. ordentlich , u. erez , and m. gastpar , `` integer - forcing architectures for mimo : distributed implementation and sic integer - forcing architectures for mimo : distributed implementation and sic , '' in _ conference record of the 44th asilomar conference on signals , systems , and computers _ , monterey , ca , november 2010 . a. sanderovich , s. shamai , , h. v. poor , and y. steinberg , `` uplink macro diversity of limited backhaul cellular network , '' _ ieee transactions on information theory _ , 55 , no . 8 , pp . 34573478 , august 2009 . m. taherzadeh , a. mobasher , and a. khandani , `` communication over mimo broadcast channels using lattice - basis reduction , '' _ ieee transactions on information theory _ , 53 , no . 12 , pp . 45674582 , december 2007 .

the post - agb phase is a short period in the evolution of low- and intermediate - mass stars ( @xmath4 ) between the asymptotic giant branch ( agb ) and the planetary nebula ( pn ) phases . the theoretical evolutionary tracks of the central stars @xcite predict that this phase lasts @xmath5 yrs , depending on the core mass , yet even shorter kinematical age for nebulosities , of order of @xmath6 yrs @xcite . shorter kinematical life time is obvious from the observational point of view since the entire shell ( especially an extended cold shell of post - agb object ) is not necessarily detected at the optical wavelength . many significant changes occur during this phase to both the star and the nebula . the superwind ( suddenly increased mass loss at the end of the agb phase ) stops , large amplitude pulsations cease and the circumstellar envelope slowly expands and cools revealing the central star . the post - agb star evolves at constant luminosity on the hertzsprung - russell diagram towards higher temperatures . the morphology of the expanding envelope changes , resulting in the fascinating shapes of the nebula itself . despite many efforts the post - agb phase is still poorly understood . especially the departure from spherically symmetric agb circumstellar envelopes ( e.g. , * ? ? ? * ) and the formation of diverse , axisymmetric , nebulae ( e.g. , * ? ? ? * ) are of great interest to astronomers . determining the main cause(s ) of the breaking of mass - loss symmetry is an important issue in understanding the late stages of stellar evolution . previous morphological studies of post - agb objects ( e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * hereafter umb00 ) revealed many asymmetrical nebulae around central stars , where jets and concentric arcs or rings were not unusual . further , the ground - based mid - ir ( 821 @xmath7 m ) observations done by @xcite showed two morphological classes of extended axisymmetric nebulae among post - agb objects : `` core / elliptical '' with unresolved cores and elliptical nebulae and `` toroidal '' with limb - brightened peaks that suggest equatorial density enhancements . optical observations revealed more interesting details on structures seen in post - agb nebulae . umb00 observed 27 ppn candidates and divided them into two classes : the star - obvious low - level elongated ( sole ) nebulae which show star - dominated emission with faint extended nebulosity and dust - prominent longitudinally extended ( duplex ) nebulae which are dust dominated with a faint or completely obscured central star . they also noticed that those two optical classes corresponded to the mid - ir ones , i.e. core / elliptical @xmath8 duplex and toroidal @xmath8 sole , and that the main difference between sole and duplex nebulae was the degree of the equatorial enhancement ( i.e. , duplexes tend to have a higher equator - to - pole density ratio than soles ) and could not be attributed only to the inclination - angle effects . they also suggested that sole nebulae were optically thinner than duplex nebulae and that sole ppns would have evolved from low - mass progenitors and duplex from high - mass progenitors . later model calculations @xcite confirmed the physical differences between those two classes of objects . other high - resolution near - infrared @xmath3 observations ( e.g. , * ? ? ? * ; * ? ? ? * ) also supported optical results . follow - up studies by e.g. , @xcite were conducted to further investigate the asymmetry of observed objects . they also found that differences in observed morphologies are caused ( mainly ) by optical depth and ( to a lesser extent ) by inclination of the objects . all those observations strengthen the hypothesis of an equatorial mass loss enhancement at the end of the agb ( superwind ) phase @xcite and that morphological shaping of nebulae occurs before the post - agb phase . moreover , the resulting diverse shapes observed in proto - planetary as well as pns may be caused by some properties of the central stars such as chemical composition , mass and/or metallicity . in order to find further support to those results we have executed an @xmath3 snapshot survey of post - agb objects by increasing the range of masses and chemical composition ( c / o ratio ) and by enlarging the sample of studied objects . the goal of the project is to connect the observed diversity of nebular shapes with physical and chemical properties of the central stars in this paper we present the results for post - agb objects observed recently with @xmath3 . observation and data reduction are summarized in @xmath92 . the results and differences between selected groups of objects are shown in @xmath93 . in @xmath94 we discuss obtained results with references to physical and chemical properties of stars . the conclusions are included in @xmath95 . table 1 : target summary for @xmath3 acs / hrc observation of post - agb objects llrrrrrl & & & & & & & + 01005 + 7910 & ... & 10627 & 10.96 & 10.274 & 21000 & 0.55 & c + 04395 + 3601 & afgl 618 & 9430 & ... & 13.510 & 25000 & ... & c + 06034 + 1354 & dy ori & 10627 & ... & 8.073 & 5900 & @xmath10 0.55 & o + 08143@xmath114406 & pm 1@xmath1139 & 10627 & ... & 9.172 & 7150 & @xmath10 0.55 & c + 09256@xmath116324 & iw car & 10627 & 8.33 & 5.875 & 6700 & @xmath10 0.55 & c + 11385@xmath115517@xmath12 & v885 cen & 9436 & 7.04 & 5.947 & 8500 & 0.55 & o + 12067@xmath114508 & ru cen & 10627 & 9.13 & 7.616 & 6000 & @xmath10 0.55 & o + 12175@xmath115338 & v1024 cen & 10627 & 9.40 & 8.326 & 7350 & 0.62 & o + 12222@xmath114652 & hd 108015 & 10627 & 8.01 & 6.941 & 6800 & @xmath10 0.55 & o + 12538@xmath112611 & ln hya & 10627 & 6.82 & 5.251 & 6000 & @xmath10 0.55 & o + 13416@xmath116243 & ... & 10627 & ... & 10.302 & ... & ... & c + 13428@xmath116232@xmath12 & ... & 9463 & ... & 13.176 & ... & ... & c + 15039@xmath114806 & hd 133656 & 10627 & 7.58 & 6.680 & 8000 & 0.55 & o + 15469@xmath115311 & ... & 10627 & 10.82 & 7.190 & 7500 & @xmath10 0.55 & o + 15553@xmath115230@xmath12 & ... & 10627 & ... & 13.380 & ... & ... & ... + 16206@xmath115956 & ls 3591 & 10627 & 10.00 & 9.002 & 8500 & 0.66 & o + 17163@xmath113907 & hen 3@xmath111379 & 10185 & 12.45 & 4.635 & ... & ... & ... + 17243@xmath114348 & lr sco & 10627 & 10.49 & 8.035 & 6750 & 0.94 & o + 17279@xmath111119 & v340 ser & 10627 & 9.78 & 7.845 & 7300 & @xmath10 0.55 & c + 17516@xmath112525 & ... & 9436 & 17.76 & 8.695 & ... & ... & o + 17534 + 2603 & 89 her & 10627 & 5.51 & 4.998 & 6550 & 0.61 & o + 18135@xmath111456 & oh 15.7 + 0.8 & 9436 & 16.61 & ... & ... & ... & o + 19125 + 0343 & bd+03 3950 & 10627 & 10.46 & 7.903 & 7750 & 0.58 & o + 19157@xmath110247 & bd@xmath1102 4931 & 10627 & 10.87 & 8.872 & 7750 & 0.58 & o + 19306 + 1407@xmath12 & ... & 9436 & ... & 11.286 & b & ... & o + 19475 + 3119@xmath12 & hd 331319 & 9436 & 9.60 & 7.773 & 7750 & 0.58 & o + 20000 + 3239@xmath12 & ... & 9436 & ... & 8.021 & 5500 & ... & c + 20117 + 1634 & r sge & 10627 & 9.31 & 7.818 & 5000 & 0.93 & o + 20547 + 0247 & u equ & 9436 & ... & 11.561 & g & ... & o + 22036 + 5306@xmath12 & ... & 10185 & ... & 11.666 & ... & ... & o + 22223 + 4327@xmath12 & bd+42 4388 & 9436 & 10.00 & 7.812 & 6500 & 0.55 & c + 23304 + 6147@xmath12 & pm 2@xmath1147 & 9436 & ... & 8.501 & 6750 & 0.66 & c + 23541 + 7031 & m 2@xmath1156 & 9436 & ... & 13.857 & b & ... & o + @xmath13 9430 - pi s. trammell , 9436 & 10185 - pi r.sahai , 10627 - pi m. meixner . + @xmath14 magnitudes from gsc2.2 and 2mass catalogues . + @xmath15 spectral type is shown if t@xmath16 is not known from the model atmosphere analysis . + @xmath17 data collected by @xcite with few exceptions : @xmath18 @xcite and references therein , @xmath19 c / o ratio established by the inspection of the iso spectrum , @xmath20 @xcite and references therein , @xmath21 @xcite , @xmath22 @xcite and references therein . + @xmath12 images of those objects were also published in the newest paper by @xcite . post - agb stars do not form a homogeneous group . one can find objects with different masses ( including very low - mass objects that will never become pns ) and chemical composition . however , they are expected to be supergiants of spectral type b k with an evidence of the circumstellar envelope ( e.g. , infrared excess ) and no large photometric variability ( the classification criteria for an object to be a post - agb star were described in e.g. , * ? ? ? * ; * ? ? ? * ; * ? ? ? in addition , few unique classes of stars like rv tau objects @xcite , r crb stars and extreme helium stars @xcite or even `` born - again '' supergiants ( sakurai s object and fg sge , e.g. , * ? ? ? * ) are considered as evolved objects in the post - agb phase . all those objects are included in the toru catalogue of galactic post - agb and related objects ( * ? ? ? * http://www.ncac.torun.pl/postagb ) . from this database we selected for our snapshot survey those sources that were not yet imaged with @xmath3 . because the exposure time for an object from a snapshot survey can not be too long , our sample was biased towards rather bright stars . the selected post - agb objects were observed with the high resolution channel ( hrc ) of the advanced camera for surveys ( acs ) on - board @xmath3 @xcite , which has a @xmath23 x @xmath24 field of view and a plate - scale of 0.027@xmath25 pixel@xmath26 . observations for 19 objects from our proposal ( program i d 10627 , pi m. meixner ) were done between july 2005 and january 2007 . we also searched for other post - agb stars ( from the toru catalogue ) observed with acs / hrc . in the @xmath3 archive we found 17 objects from previous observation cycles . we included 14 post - agb sources in our reduction and analysis of the images ( selected objects from programs i d 9463 & 10185 , pi r. sahai and 9430 , pi s. trammell ) and excluded 3 objects ( the extended structures of egg nebula and frosty leo are bigger than the size of an image what prevents the correct measurements and iras 19024 + 0044 was already carefully analyzed by @xcite ) . together we have 33 ppns observed in broad @xmath27 ( f435w ) , @xmath28 ( f606w ) and @xmath29 ( f814w ) filters . selected candidates differ in central star masses , optical and infrared colors and effective temperatures ( * ? ? ? * and references therein ) . thereby we increased the number of studied post - agb objects and covered more extensively the diverse nature of post - agb stars . the properties of the analyzed post - agb objects are shown in table 1 . table 2 : properties of sole post - agb objects lrrllrrrr iras i d & & & & & @xmath30 & & + & & & & & & & & + 01005 + 7910 & 01 04 45.59 & 79 26 47.08&f606w&1.23e@xmath1113&11.17 & 26900 & 3.75 x 2.20 ( 2)&0.41 + & & & f814w&5.94e@xmath1114&11.97 & 14510 & 3.65 x 2.25 ( 4)&0.38 + 08143@xmath114406&08 16 02.98&@xmath1144 16 04.88&f606w&5.58e@xmath1114&12.03 & 7500 & 2.10 x 1.30 ( 10 ) & 0.38 + & & & f814w&7.61e@xmath1114&11.70 & 3700 & 2.10 x 1.35 ( 4)&0.36 + 11385@xmath115517&11 40 58.82&@xmath1155 34 25.52&f435w&8.10e@xmath1112 & 6.63 & 338000 & 8.30 x 6.60 ( 12 ) & 0.20 + & & & f606w&4.47e@xmath1112@xmath22&7.27@xmath22 & 81500 & 8.30 x 6.20 ( 8)&0.25 + 12175@xmath115338&12 20 15.08&@xmath1153 55 31.62&f606w&6.46e@xmath1113 & 9.37 & 76000 & 5.70 x 2.50 ( 7)&0.56 + & & & f814w&3.76e@xmath1113 & 9.96 & 36300 & 5.60 x 2.20 ( 10 ) & 0.61 + 16206@xmath115956&16 25 02.61&@xmath1160 03 31.35&f606w&3.33e@xmath1113&10.09 & 21500 & 3.85 x 2.60 ( 9)&0.32 + & & & f814w&1.87e@xmath1113&10.72 & 21300 & 3.85 x 2.55 ( 6)&0.34 + 19306 + 1407 & 19 32 55.15 & + 14 13 36.89&f606w&7.83e@xmath1115&14.17 & 9000 & 7.70 x 2.00 ( 4)&0.74 + & & & f814w&1.09e@xmath1114&13.81 & 6400 & 4.40 x 2.05 ( 4)&0.53 + 19475 + 3119 & 19 49 29.63 & + 31 27 15.32&f435w&6.91e@xmath1113 & 9.30 & 190000&10.45 x 5.10 ( 3)&0.51 + & & & f606w&6.72e@xmath1113 & 9.33 & 153000 & 9.95 x 5.10 ( 2)&0.49 + 20000 + 3239 & 20 01 59.56 & + 32 47 33.71&f606w&3.39e@xmath1114&12.57 & 3800 & 2.15 x 1.45 ( 5)&0.33 + & & & f814w&1.02e@xmath1113&11.38 & 2000 & 2.00 x 1.20 ( 7)&0.40 + 22223 + 4327 & 22 24 31.48 & + 43 43 10.67&f435w&3.47e@xmath1113&10.05 & 13000 & 3.40 x 2.10 ( 10 ) & 0.38 + & & & f606w&5.02e@xmath1113 & 9.65 & 10000 & 3.50 x 2.10 ( 6)&0.40 + 23304 + 6147 & 23 32 44.71 & + 62 03 48.92&f606w&3.30e@xmath1114&12.61 & 3800 & 2.25 x 1.60 ( 14 ) & 0.29 + & & & f814w&8.31e@xmath1114&11.60 & 2100 & 2.20 x 1.45 ( 9)&0.34 + @xmath18 in units of erg s@xmath26 @xmath31 @xmath26 . + @xmath19 hst system magnitudes ( stmag ) . + @xmath20 star surface intensity @xmath32 to nebula surface intensity @xmath33 ratio . + @xmath21 ellipticity @xmath34 , where @xmath35 and @xmath36 are major- and minor - axis lengths , respectively . + @xmath22 due to a saturation problem , @xmath37 and hst mag are only the lower limits . table 3 : properties of duplex post - agb objects lrrlrrrrr iras i d & & & & & @xmath30 & & + & & & & & & & & + 04395 + 3601 & 04 42 53.50 & + 36 06 51.71 & f606w&2.54e@xmath1115&15.39& ... &15.40 x 4.80 ( 4)&0.67 + & & & 1.98e@xmath1115&15.66&124 & 8.25 x 4.75 ( 4)&0.42 + & & & 5.16e@xmath1116&17.12 & 85 & 6.20 x 3.20 ( 4)&0.48 + 13428@xmath116232&13 46 20.96 & @xmath1162 47 58.10@xmath22 & f606w & ... & ... & ... & ... & ... + & & & f814w&3.86e@xmath1116&17.43 & 4 & 8.75 x 6.90 ( 1)&0.21 + 15553@xmath115230&15 59 10.66&@xmath1152 38 38.15 & f814w&1.90e@xmath1117&20.70 & ... & 2.50 x 1.10 ( 2)&0.56 + & & & 1.21e@xmath1117&21.20 & 10 & 1.35 x 1.00 ( 2)&0.26 + & & & 6.01e@xmath1118&21.95 & 4 & 1.20 x 0.85 ( 2)&0.29 + 22036 + 5306 & 22 05 30.35 & + 53 21 33.97 & f606w&1.32e@xmath1115&16.10&330 & 6.75 x 2.30 ( 3)&0.66 + & & & f814w&2.87e@xmath1115&15.26&120 & 5.60 x 1.85 ( 4)&0.67 + 23541 + 7031 & 23 56 36.31 & + 70 48 18.39 & f606w&5.33e@xmath1117&19.58 & 55 & 3.75 x 1.60 ( 2)&0.57 + & & & f814w&4.96e@xmath1117&19.66 & 88 & 3.30 x 0.90 ( 1)&0.73 + @xmath18 in units of erg s@xmath26 @xmath31 @xmath26 . + @xmath19 hst system magnitudes ( stmag ) . + @xmath20 peak surface intensity @xmath32 to nebula surface intensity @xmath33 ratio . + @xmath21 ellipticity @xmath34 , where @xmath35 and @xmath36 are major- and minor - axis lengths , respectively . + @xmath22 2mass counterpart s coordinates . only clearly visible part of the nebula was measured on f814w image ( see text ) . table 4 : properties of stellar post - agb objects lrrllrr iras i d & & & & & + & & & & & & + 12222@xmath114652&12 24 53.49&@xmath1147 09 08.36&f606w&2.27e@xmath1112 & 8.01 & 1.28e+1 + & & & f814w&1.21e@xmath1112 & 8.70 & 4.48e+0 + 12538@xmath112611&12 56 30.00&@xmath1126 27 36.55&f606w&5.39e@xmath1112 & 7.07 & 1.21e+1 + & & & f814w&4.08e@xmath1112 & 7.37 & 1.20e+1 + 13416@xmath116243&13 45 07.26&@xmath1162 58 15.67&f814w&1.77e@xmath1115&15.78 & 6.47e@xmath116 + 15039@xmath114806&15 07 27.52&@xmath1148 17 53.85&f606w&3.11e@xmath1112 & 7.67 & 1.27e+1 + & & & f814w&1.83e@xmath1112 & 8.25 & 6.34e+0 + 17163@xmath113907&17 19 49.26&@xmath1139 10 38.29&f435w&7.93e@xmath1116&16.65 & 6.89e@xmath113 + & & & f606w&7.92e@xmath1114&11.65 & 4.38e@xmath111 + 17279@xmath111119&17 30 46.82&@xmath1111 22 08.21&f606w&5.47e@xmath1113 & 9.55 & 3.59e+0 + & & & f814w&4.56e@xmath1113 & 9.75 & 5.27e+0 + 17516@xmath112525&17 54 43.35&@xmath1125 26 28.77&f606w&5.35e@xmath1116&17.08 & 3.59e@xmath113 + 17516@xmath112525a@xmath21 & & & & 4.39e@xmath1116&17.29 & + 17516@xmath112525b@xmath21 & & & & 6.12e@xmath1117&19.43 & + & & & f814w&9.48e@xmath1115&13.96 & 3.24e@xmath112 + 17516@xmath112525a@xmath21 & & & & 7.90e@xmath1115&14.16 & + 17516@xmath112525b@xmath21 & & & & 1.31e@xmath1116&18.61 & + 17534 + 2603 & 17 55 25.19 & + 26 03 01.49&f606w&2.07e@xmath1111@xmath22&5.61@xmath22&1.21e+1 + & & & f814w&1.12e@xmath1111@xmath22&6.28@xmath22&1.23e+1 + 18135@xmath111456&18 16 25.47&@xmath1114 55 16.17&f606w&5.56e@xmath1116&17.04 & 4.23e@xmath113 + & & & f814w&6.45e@xmath1116&16.88 & 3.37e@xmath113 + 20547 + 0247 & 20 57 16.29 & + 02 58 44.47&f606w&3.85e@xmath1115@xmath22&14.94@xmath22&8.32e@xmath113 + & & & f814w&8.20e@xmath1115&14.12 & 2.35e@xmath112 + + 06034 + 1354 & 06 06 15.04 & + 13 54 18.95&f606w&7.67e@xmath1114&11.69 & 4.23e@xmath111 + & & & f814w&1.10e@xmath1113&11.30 & 3.25e@xmath111 + 09256@xmath116324&09 26 53.52&@xmath1163 37 48.78&f606w&1.45e@xmath1112 & 8.49 & 7.80e+0 + & & & f814w&1.41e@xmath1112 & 8.53 & 5.27e+0 + 12067@xmath114508&12 09 23.72&@xmath1145 25 34.60&f606w&6.02e@xmath1113 & 9.45 & 2.89e+0 + & & & f814w&4.41e@xmath1113 & 9.79 & 1.93e+0 + 15469@xmath115311&15 50 43.75&@xmath1153 20 42.88&f606w&2.64e@xmath1113&10.35 & 1.37e+0 + & & & f814w&3.72e@xmath1113 & 9.97 & 1.23e+0 + 17243@xmath114348&17 27 53.60&@xmath1143 50 46.73&f606w&2.46e@xmath1113&10.42 & 1.64e+0 + & & & f814w&2.55e@xmath1113&10.38 & 1.07e+0 + 19125 + 0343 & 19 15 01.17 & + 03 48 43.52&f606w&2.71e@xmath1113&10.32 & 2.39e+0 + & & & f814w&2.83e@xmath1113&10.27 & 1.24e+0 + 19157@xmath110247&19 18 22.75&@xmath1102 42 10.79&f606w&1.93e@xmath1113&10.69 & 1.17e+0 + & & & f814w&1.68e@xmath1113&10.84 & 6.11e@xmath111 + 20117 + 1634 & 20 14 03.76 & + 16 43 35.77&f606w&9.79e@xmath1113 & 8.92 & 5.42e+0 + & & & f814w&7.87e@xmath1113 & 9.16 & 3.04e+0 + @xmath18 in units of erg s@xmath26 @xmath31 @xmath26 . + @xmath19 hst system magnitudes ( stmag ) . + @xmath20 peak intensity ( in units of erg s@xmath26 @xmath31 @xmath26 sr@xmath26 ) . + @xmath21 see text for details . + @xmath22 due to a saturation problem , @xmath37 and hst mag are only the lower limits . each source was observed with various exposure times in order to increase the dynamic range of the final image . faint nebulae were revealed in the long exposures , but the central star pixels were often saturated . the short exposures , free from saturated pixels , were combined with the longer ones to produce a nonsaturated final image . because each object was observed a couple of times we had 2 or more nonsaturated final images . then iraf / stsdas routines were used for calibration and reduction of images . the multidrizzle package was used to combine multiple frames of images into a single image ( with plate - scale of 0.025@xmath25 pixel@xmath26 ) , remove cosmic rays and subtract background radiation . the reduced images were used to measure photometric and geometric properties of our objects . we adopted @xmath3 photometric calibration of synphot ( v5.0 , * ? ? ? * ) to calibrate the flux density ( @xmath37 in units of erg s@xmath26 @xmath31 @xmath26 ) . the defined aperture was big enough to encircle the source together with the diffraction features , while the background emission was calculated from a ( up to , depending on the proximity of the other stars on the images ) 10-pixel wide annulus separated from the selected aperture by a buffer zone . the sky emission was then subtracted from the emission of the entire source . the derived quantity , which was in counts per second , was finally converted into @xmath37 and @xmath3 acs magnitudes ( stmag ) . we measured the sizes of nebulosities around central stars . our motivation was to determine the extension of the nebula as a whole , but not the sizes of its particular parts ( e.g. , each lobe ) . for this we needed only two axis of extension . in case of duplex sources , where the central star was not visible and the lobes were clearly separated , we measured the size of each lobe separately and , in addition , the extension of the whole object . the extent of the nebula was defined to be the outermost recognizable structure in which emission level was from 1@xmath38 up to 14@xmath38 ( depending on the image quality ) above the background level . the major and minor axis , as well as the ellipticity of the nebulae were derived by fitting ellipses to the selected isophotes . we measured also surface intensity of the nebulae at the edge and compared it to the star ( or peak if the central star was not visible , as in duplex sources ) surface intensity and got star - to - nebula surface intensity ratios ( @xmath39 ) . for objects without nebulae only the peak intensity was measured . names and the coordinates of all objects , as well as other derived parameters ( @xmath37 , @xmath3 magnitudes , @xmath40 and extension ) are shown in tables 2 , 3 and 4 for sole , duplex and stellar sources , respectively . figures 1 , 2 and 3 show the reduced images . the logarithmic scale is used to illustrate the contrast between nebulae and central stars . nevertheless , some faint nebulae are still barely visible . @xmath39 ranges from 4 to @xmath41 ( see tables 2 and 3 ) . the point spread function ( psf ) effects are seen on most of the images . to remove them and to make the nebulae more clearly seen we deconvolved our images using richardson - lucy ( rl ) algorithm with a stellar psf created especially for our images by the code tinytim ( v6.3 , * ? ? ? as a result , the contrast between nebula and the background emission was improved in general , but in case of very faint nebulosities where the nebulae emission was comparable to sky emission , the difference between reduced and deconvolved images is rather insignificant . there is also one inconvenience in using rl algorithm , which is known to produce false depressions around unusually bright pixels . those `` holes '' are seen on our deconvolved images , but are not considered real structures . the deconvolved images are shown together with the reduced ones in figures 1 , 2 and 3 in the same logarithmic scale . fig.1 . - images of sole objects ( north is up and east is to the left ) shown in order of increasing ra ( see table 2 ) . the leftmost frame shows the iras i d and size of an image . the filters types are shown at the bottom of each frame with `` + rl '' indicating richardson - lucy deconvolution . the images are in logarithmic scale to illustrate the contrast between central star and the nebula . [ imsole1 ] fig.1 . - continued . [ imsole2 ] fig.1 . - continued . [ imsole3 ] [ imduplex1 ] fig.2 . - images of duplex objects shown in order of increasing ra ( see table 3 ) . the displaying scheme is the same as in fig.1 . [ imduplex2 ] we detected nebulosities in 15 out of 33 sources and classified them as sole and duplex following umb00 . soles have an optically thin ( continuous ) envelope and the star is therefore very bright at optical and near - infrared wavelengths in comparison to the nebula ( in terms of @xmath39 we consider values of @xmath42 ) . the surface brightness peaks at the central star . because of multiaxis lobes , with different shapes and surface brightness , umb00 divided those objects in subgroups of simple ellipse , multiple ellipses , an ellipse with embedded bipolar structure and an ellipse with concentric shells . these structures are also present in our objects . in addition we observe snail - like structures . the psf effects are apparent in the images : the faint nebular structures are often buried under the psf of the bright central star . duplexes have an optically thick torus that completely or partially obscures the central star . @xmath39 values are much below 1000 and the peak of the surface brightness is far away from the center of the object . the nebula is bipolar and the ( multiple ) lobes are seen through scattered light . the psf diffraction feature is usually not seen in the images of duplex objects . these observational characteristics of the nebulae are consistent with the present understanding of the structure of these shells , in which sole and duplex nebulae respectively represent the low and high ends of the degree of equatorial enhancement for an intrinsically axisymmetric circumstellar shell . in other words , there is not much of a difference between sole and duplex nebulae as far as the shell structure goes , but the difference is the optical depth along the equator due to the difference in the equator - to - pole density ratio resulting from the different degree of equatorial density enhancement . this is what makes duplex nebulae appear bipolar with a well - defined dust lane , while making sole nebulae appear elliptical without a clear dust lane . because of this , their shell morphology tends to become less definitive when the optical depth along the equator is close to unity ( and/or the inclination angle of the shell is far from edge - on ) . we have 10 sole objects ( iras 01005 + 7910 , iras 08143@xmath114406 , iras 11385@xmath115517 , iras 12175@xmath115338 , iras 16206@xmath115956 , iras 19306 + 1407 , iras 19475 + 3119 , iras 20000 + 3239 , iras 22223 + 4327 and iras 23304 + 6147 ) and 5 duplex objects ( iras 04395 + 3601 , iras 13428@xmath116232 , iras 15553@xmath115230 , iras 22036 + 5306 and iras 23541 + 7031 ) in our sample . @xmath39 for our duplexes is at most 330 ( iras 22036 + 5306 ) and for soles reaches the value of a few hundred thousands ( e.g. , iras 11385@xmath115517 or iras 19475 + 3119 ) . the remaining 18 sources do not show any nebulosities around them and are referred to as stellar objects . due to a different morphology we may expect also the differences in spectral energy distribution ( sed ) shapes and the position in infrared color - color diagrams . all groups of sources are discussed separately in the following sections . images of sole , duplex and stellar objects are shown in figures 1a , 1b and 1c , respectively . iras 01005 + 7910 have a nebulosity of the similar size in both bands ( 3.75@xmath25 @xmath43 2.20@xmath25 in @xmath28 and 3.65@xmath25 @xmath43 2.25@xmath25 in @xmath29 ) . its morphology is irregular . it has two lobes of different shapes pointed in north - west and south - east and additional small lobe emerging in south - eastern direction visible in the @xmath28 band . the circular halo corresponds to the wings of the psf around the bright central star . iras 08143@xmath114406 shows an elliptical nebula of 2.1@xmath25 long and 1.4@xmath25 wide . it has two small lobes emerging in western and eastern direction . the circular psf halo is also present . the object looks similar in both bands with the only exception in @xmath39 , which is two times bigger in the @xmath28 band than in the @xmath29 band . this source is a carbon - rich object with the overabundance of s - process elements @xcite and is similar in appearance to iras 07430 + 1115 from umb00 sample , which is also rich in s - process elements ( e.g * ? ? ? the nebulosity of iras 11385@xmath115517 is very asymmetric and looks like pinwheel . this a - type oxygen - rich low - mass supergiant was thought to be a binary , but later observations suggested rather disk - like structure and high - velocity bipolar outflows ( e.g. , * ? ? ? * ) . it is possible that we see only one of this outflows on @xmath3 images because of the inclination effect . the snail - like shape of the outflow is quite similar to the spiral pattern found in c - rich agb star afgl 3068 @xcite . the spiral outflow in iras 11385@xmath115517 is also single - armed and its extension is 8.3@xmath25 @xmath43 6.6@xmath25 in the @xmath27 band . the object is very bright ( @xmath27=6.6mag and @xmath28=7.3mag ) and looks similar in both bands . @xmath44 is very high in the @xmath27 band with the value of @xmath45 being the highest among sole sources . iras 12175@xmath115338 has very complex morphology . the nebula is 5.7@xmath25 long and 2.5@xmath25 wide . it has at least 4 lobes , pointed in different directions . the lobes are of very different brightness , with the bright ones closer to the central star and very faint more extended structures . they are less clearly visible in the @xmath29 band . the circular psf halo is also seen around the star . the object is quite bright in both bands , @xmath469.5mag , and is known to be a metal - deficient pulsating variable ( e.g. , * ? ? ? iras 16206@xmath115956 has a nebulosity very similar to the one seen in iras 11385@xmath115517 . it is also asymmetric and has a snail - like shape , but is much smaller with an extension of 3.85@xmath25 @xmath43 2.6@xmath25 . there is also a faint lobe emerging to the south - west with a quite complex structure . there is no visible lobe on the other side of the objects , but it may be extremely faint and buried under the psf of the bright central star . the structures are better resolved in the @xmath28 band than in the @xmath29 band , but @xmath39 is the same , @xmath4621000 , in both bands . iras 16206@xmath115956 is also a - type oxygen - rich supergiants similar to iras 11385@xmath115517 , but is more massive with a core mass of @xmath47 ( see table 1 ) . iras 19306 + 1407 is a b - type oxygen - rich ppn . it has a bright central star and low surface brightness lobes ( @xmath39 is @xmath46 9000 and 6400 in the @xmath28 and @xmath29 band , respectively ) . the lobes have bipolar structure , also seen in previous observations @xcite . the lobes seem to be smaller in the @xmath29 band , with the extension of only 4.4@xmath25 @xmath43 2.0@xmath25 , in comparison to the @xmath28 band where they are longer , 7.7@xmath25 @xmath43 2.0@xmath25 . iras 19475 + 3119 has a quadrupolar nebula and a psf spherical halo . similar complex double - elongation structure can be seen in iras 04296 + 3429 from umb00 , except for the different sizes of the nebula : iras 19475 + 3119 has much bigger nebulosity of 10.45@xmath25 long and 5.1@xmath25 wide . also , iras 19475 + 3119 is known to be oxygen - rich ( its dust envelope was described in details by * ? ? ? * ) , whereas iras 04296 + 3429 is a carbon - rich source with s - process overabundance and 21 @xmath7 m emission feature ( e.g. , * ? ? ? iras 20000 + 3239 is a carbon - rich source with s - process elements and 21 @xmath7 m feature ( e.g. , * ? ? ? * ; * ? ? ? it has small lobes emerging to west and east with the extension of 2.15@xmath25 @xmath43 1.45@xmath25 in @xmath28 ( and slightly smaller in @xmath29 band ) and is very similar to iras 08143@xmath114406 described above . iras 22223 + 4327 is also a carbon - rich source with s - process elements and 21 @xmath7 m feature . it is quite bright in @xmath28 band ( @xmath469.6 ) and the nebula has a size of 3.4@xmath25 x 2.1@xmath25 . there are two pairs of lobes pointed to north and south and they are more clearly visible on the south side . the object is quite similar to iras 08143@xmath114406 and iras 20000 + 3239 . iras 23304 + 6147 also belongs to the group of objects with 21 @xmath7 m feature . its extension is comparable to the extension of iras 20000 + 3239 ( 2.25@xmath25 @xmath43 1.6@xmath25 ) . it has 2 lobes emerging in northern and southern directions and another inner pair of lobes in western and eastern directions . those inner lobes are not so clearly seen even on deconvolved images . the object s structure could be similar to one seen in iras 06530@xmath110213 from umb00 . while the appearance of iras 20000 + 3239 , iras 22223 + 4327 and iras 23304 + 6147 is similar and those three objects have the same 21 @xmath7 m feature , the sole morphology is not characteristic to the whole group with 21 @xmath7 m , e.g. , duplex source iras 22574 + 6609 ( umb00 ) or egg nebula ( e.g. , * ? ? ? iras 04395 + 3601 ( better known as afgl 618 ) is a carbon - rich ppn surrounding a b0 central star , rapidly evolving to a pn stage . the acs image shows bipolar structure of the nebula , with the extent of 15.4@xmath25 @xmath43 4.8@xmath25 . the eastern lobe is larger ( 8.25@xmath25 @xmath43 4.75@xmath25 ) and brighter ( @xmath28=15.7mag ) than western one ( 6.2@xmath25 @xmath43 3.2@xmath25 and @xmath28=17.1mag ) . the star itself is hidden in the thick dust torus and only huge outflows are seen in the optical @xcite . the structure of the object is very complex . the lobes consist of shock - excited gas and dust and are composed of several jetlike structures ( e.g. , * ? ? ? the compact hii region is present in the central part of the nebula and the circumstellar envelope contains various molecular species ( e.g. , * ? ? ? * ; * ? ? ? iras 13428@xmath116232 is a very faint object with no optical counterpart . the nebula is not visible in @xmath28 band and parts of it are revealed in @xmath29 band . the object is extended in the north - east and south - west directions with a size of more than 20@xmath25 . however , only its brightest central fragment with the size of 8.75@xmath25 @xmath43 6.9@xmath25 ( extended perpendicularly to the whole object ) is above 1@xmath38 level and could be measured . this source was studied previously by @xcite , who observed a nice bipolar shape of iras 13428@xmath116232 in the near - infrared with the angular size of 4.1@xmath25 @xmath43 12.0@xmath25 in @xmath48 band and 6.0@xmath25 @xmath43 11.0@xmath25 in @xmath49 band . iras 15553@xmath115230 is a poorly known post - agb star . this iras source was observed in near - infrared by @xcite and @xcite and there was no agreement about the correct counterpart , however , @xcite noticed that the object was elongated . iso spectrum shows no features . our observations confirmed coordinates indicated by @xcite and revealed the bipolar shape of this small ( 2.5@xmath25 x 1.1@xmath25 ) nebula . the object is very faint and the central star is invisible . the lobes differ in size and shape , which may be the effect of the nebula orientation towards observer . a small feature pointing out from the east lobe may be a faint outflow or a jet . fig.3 . - exemplary images ( reduced and deconvolved ) of 4 stellar objects . iras name , filter and size of images are shown . [ imstellar ] the extended and complex bipolar nebula of iras 22036 + 5306 was already known from the previous @xmath3 observations ( proposal i d 9101 , * ? ? ? the size of the object and magnitudes in the @xmath28 and @xmath29 bands determined in this paper are slightly different from the ones given by sahai and collaborators , but the wfpc2 camera used in their observations had different plate scale and different response curve than acs . the morphology of the object is very complex , with central ring structures , elongated lobes and faint knotty jetlike linear structures . the object is oxygen - rich , with a very characteristic feature in iso spectra of water ice absorption at @xmath46 3.1 @xmath7 m ( e.g. , * ? ? ? iras 23541 + 7031 ( also known as m 2@xmath1156 ) is a ppn with a strong shock emission seen in optical line spectrum . @xcite discovered its bipolar nature and found weak emission extending up to 10@xmath25 from the central bright region . from our measurements the nebula is only 3.75@xmath25 x 1.6@xmath25 in size in the @xmath28 band and smaller and much fainter in the @xmath29 band . however , there is a very faint emission ( below 1@xmath38 above the sky level ) outside the measured size and we were not able to estimate its extent . the central star of iras 23541 + 7031 , which is much fainter in the @xmath29 band is classified as type b , but cooler than the central star of afgl 618 . stellar sources are not associated with any obvious nebulosities . objects are both very bright , like iras 17534 + 2603 , and very faint , like iras 18135@xmath111456 . for bright sources diffraction features are very strong . various artificial features produced during deconvolution can be seen in figure 3 . all analyzed rv tau stars belong to this group . they are cool post - agb objects with effective temperatures below @xmath50 , usually low - mass central stars and oxygen - rich envelopes ( for few exceptions see e.g. , * ? ? ? * ) with photospheres depleted in the high condensation temperature elements as a result of gas - dust separation ( e.g. , * ? ? ? their ir excess indicates the presence of a warm dust which should be visible as an extended circumstellar envelope . however , there is no evidence of such envelope on analyzed images . this is caused by properties of both @xmath3 and rv tau stars . @xmath3 is sensitive to only nearby , relatively dense circumstellar dust , while rv tau stars ( because of their long - time evolution ) have very extended and therefore very low surface brightness envelopes . hence , rv tau stars appear as stellar sources . the rest of the objects in the stellar group are post - agb stars with infrared excess . iras 12538@xmath112611 and iras 15039@xmath114806 are supergiants with cool dust , while iras 12222@xmath114652 and iras 17279@xmath111119 show a near - ir excess which stands for hot dust ( compare seds on fig.6 . ) . it was suggested by @xcite that all objects with near - infrared excess are binaries . one of them is also iras 17534 + 2603 , better known as 89 her . it s a high - latitude f2 supergiant with a hot dust in a shell , large co shell @xcite and the h@xmath51 line showing the evidence for the on - going mass loss ( e.g. , * ? ? ? the model constructed for the circumstellar envelope also assumed that iras 17534 + 2603 must have evolved as a binary system . only for iras 17279@xmath111119 one can see a clump of denser material in the north - west part of an image . it must be connected with the star since there are no other objects in the near vicinity . it may be an ejected knot . the object is classified as stellar source since we do not see any evidence for continuous nebulosity around it ( as it is in case of sole objects ) , but the knot makes it peculiar . iras 13416@xmath116243 was connected with a highly reddened g1 supergiant by @xcite , but the conclusions from low resolution iras spectrum about 10 and 18 @xmath7 m silicate absorption bands were misleading . instead , iso spectrum shows moderate pah bands and absorption feature of c@xmath52h@xmath52 at 13.7@xmath7 m suggesting a carbon rich , dusty envelope . there is also a 6.9@xmath7 m feature which is seen only in the post - agb phase and only in objects with a 21@xmath7 m feature ( e.g. , * ? ? ? unfortunately , a 21@xmath7 m is not seen in this object and we can only speculate that the feature is quite weak and iso resolution was not enough to see it . iras 17516@xmath112525 is known to be an oxygen - rich source with the oh maser emission . it is a very faint star with @xmath5317.8mag , but analyzed images show that there are indeed 2 stars very close to each other ! the separation is 1.45@xmath25 and they are resolved only on @xmath3 images . it is not clear if the stars are related . iras 17516@xmath112525a is redder in the @xmath29 band than iras 17516@xmath112525b and most likely this is a correct iras counterpart . iras 20547 + 0247 is a peculiar variable star . iras 12/25 color suggests the optically thin envelope , however iras - lrs spectrum shows 10@xmath7 m silicate absorption feature typical for thick dusty envelope . to explain this contradiction @xcite proposed a dusty , thick disk around iras 20547 + 0247 but seen edge - on . its optical spectrum is very unusual with strong absorption and emission bands of metallic oxides , tio , alo , and vo , which should be not visible in g - type star . the infrared spectroscopy revealed the hot circumstellar gas located probably in a disk - like structure @xcite . we see no evidence for a disk in the images . in this section we describe various properties of the analyzed objects . table 5 compares the morphology of the objects with other properties . to get a wider view we include in our discussion other post - agb objects observed with @xmath3 and already analyzed by other authors . together there are 66 post - agb objects with @xmath3 images published in the literature . the morphologies of ppns are very complex for both sole and duplex objects , from an optically thin nebulosities to multiple dusty outflows with microstructures as we have reviewed in @xmath93 . they are different for sole and duplex objects , but they all show some degree of asymmetry . all nebulae are elongated . the sizes of the nebulae vary from small ( e.g. , 2.1@xmath25 @xmath43 1.3@xmath25 for iras 08143@xmath114406 ) to big ones ( e.g. , iras 04395 + 3601 with the extent of 15.4@xmath25 @xmath43 4.8@xmath25 ) , but the physical extension is not known due to a lack of known distances to the objects . small or faint nebulosities can be also buried under psf of bright central star and therefore hard to see . the surface intensity ratio between star and nebula in sole objects is always high , varying from 2000 to almost 340000 and the nebula appears faint in comparison to the central star . duplex sources are all faint ( @xmath28=1521mag in this study ) and the surface intensity ratio is below 330 . hence , even huge nebulae like in afgl 618 are not easy to detect . as mentioned already in this paper , the observational characteristics of the nebulae are consistent with the present understanding of the structure of these shells , in which sole and duplex nebulae respectively represent the low and high ends of the degree of equatorial enhancement for an intrinsically axisymmetric circumstellar shell . as far as the shell structure goes , there is not much of a difference between sole and duplex nebulae , but the difference is the optical depth along the equator due to the difference in the equator - to - pole density ratio resulting from the different degree of equatorial density enhancement . thus duplex nebulae appear bipolar with a well - defined dust lane , while sole nebulae appear elliptical without a clear dust lane . however , their shell morphology tends to become less definitive when the optical depth along the equator is close to unity and/or the inclination angle of the shell is far from edge - on . table 5 shows that there is no obvious connection between morphology and characteristics of the objects in our sample . among soles we have a comparable number of carbon- and oxygen - rich objects which implies that the shapes of the nebulae are not necessary correlated with the chemical composition of the objects . good examples are two oxygen - rich supergiants iras 11385@xmath115517 and iras 16206@xmath115956 and a carbon - rich agb star afgl 3068 ( see @xmath93.1 ) . fig.4 . - spectral energy distributions for sole objects shown in order of increasing ra . values are taken from gsc2.2 , 2mass , msx6c and iras catalogues . measured hst magnitudes are shown as crossed points . [ sed1 ] fig.5 . - spectral energy distributions for duplex objects shown in order of increasing ra . the displaying scheme is the same as in figure 4 . [ sed2 ] fig.6 . - spectral energy distributions for stellar sources shown in order of increasing ra . the displaying scheme is the same as in figure 4 . sed of iras 18135@xmath111456 is composed of two sources , see explanation in text , @xmath94.3 . [ sed3 ] fig.7 . - spectral energy distributions for rv tau objects shown in order of increasing ra . the displaying scheme is the same as in figure 4 . [ sed4 ] also , sole objects occupy a wide range of effective temperatures ( @xmath54 ) and masses ( @xmath55 ) . therefore , it is unlikely that sole morphology represents a specific evolutionary epoch and a specific mass object . the lack of known chemical composition and masses of most duplexes makes it impossible to connect the morphology with these properties . the range of @xmath56 of duplexes is large ( k7 - b , see table 5 ) and no correlation with morphology can be found . on the other hand , we find both morphologies among sources with similar chemical and physical properties , e.g. 21 @xmath7 m sources : iras 20000 + 3239 and iras 22223 + 4327 are soles and egg nebula and iras 06176@xmath111036 are duplexes . umb00 also did nt find any connections between shapes of the nebulae and properties of analyzed objects . derived @xmath27 , @xmath28 , and @xmath29 @xmath3 magnitudes are in general consistent with previously published measurements , although acs filters do not have exact counterparts in other photometric systems ( photometric transformations were discussed by * ? ? ? * ) . however , 2 duplex objects , iras 04395 + 3601 and iras 23541 + 7031 , are very bright in the @xmath57 band in comparison to other optical wavelengths ( compare also seds on fig.5 ) . afgl 618 was not measured previously in the @xmath28 band . our measurements show it is faint at @xmath58=15.39 . from sed ( fig.5 ) we see that it is also quite faint in the near - infrared . however , the object has significant line emission in h@xmath51 ( e.g. , * ? ? ? * ) and that may account for the anomalous increase in the @xmath57-band brightness . similarly , iras 23541 + 7031 is a bright source in h@xmath51 ( e.g. , * ? ? ? * ) and its @xmath57 magnitude is higher than expected . the h@xmath51 line is an evidence for the ongoing mass loss and can be found in many post - agb objects , even in rv tau stars ( e.g. , * ? ? ? we do not see such a big increase of brightness in the @xmath57 band in any other object in our sample . due to differences between soles and duplexes we expect them also to have different sed shapes , as was already discussed by umb00 . while soles show both of the optical and ir peaks clearly , duplexes show the prominent ir peak with a highly obscured optical peak . this is consistent with duplexes being optically thicker than soles due to their density distribution of a higher equator - to - pole ratio . seds for 33 objects which images are analyzed in this study are shown in figures 4 , 5 , 6 and 7 . photometric values are taken from gsc2.2 , 2mass , msx6c , iras catalogues and hst measurements . we adopted the classification scheme of @xcite and grouped objects into four groups : i - flat spectrum between 4 and 25 @xmath7 m and a steep fall - off toward shorter wavelengths , ii - maximum around 25 @xmath7 m and a gradual fall - off to shorter wavelengths , iii - maximum around 25 @xmath7 m and a steep fall - off to shorter wavelengths with a plateau between 1 and 4 @xmath7 m , iv - two distinct maxima : one around 25 @xmath7 m and second between 1 and 2 @xmath7 m ( iva ) or below 1 @xmath7 m ( ivb ) . in sole objects , the optical peak due to the central star and the infrared peak due to dust emission are clearly distinguishable . iras 08143@xmath114406 , iras 20000 + 3239 , and iras 23304 + 6147 have the optical maximum above 1 @xmath7 m and their nebulae are small ( compare the sizes in table 2 ) and they are of class iva . the rest of sole sources , except for iras 19306 + 1407 , are of class ivb with a peak below 1 @xmath7 m and they have extended nebulosities . in addition , sed of iras 11385@xmath115517 shows a prominent near - infrared excess that suggests a hot dust in circumstellar material and/or ongoing mass loss . sed diagram for iras 19306 + 1407 is slightly different from other sole objects , with a very strong cold dust emission in comparison to the stellar component , but no indication of a hot dust which could be stored in the vicinity of the central star . the size of the nebula is rather big , which is characteristic for class ivb , but the optical maximum seems to be above 1 @xmath7 m like a class iva object . it may be a case of an intermediate optical depth object with a massive circumstellar envelope that makes its dust peak higher than the stellar component . the lack of other objects with intermediate optical depth in our sample is caused by the fast evolution of post - agb objects with effective temperature between 10000 and 16000 ( e.g. , * ? ? ? * ) and hence a very small number of those objects to observe . seds for duplex objects are different from those for sole sources . the dust maximum is very prominent and the optical part is steep or flat without a clear peak . seds are of class ii ( for iras 22036 + 5306 and iras 23541 + 7031 ) and iii ( for iras 04395 + 3601 and iras 13428@xmath116232 ) depending on the steepness of fall - off at shorter wavelengths . we have classified iras 15553@xmath115230 as type ii / iii . stellar sources show various kinds of seds . the non rv tau sources can be divided into two groups . iras 12222@xmath114652 , iras 12538@xmath112611 , iras 17279@xmath111119 , and iras 17534 + 2603 have relatively flat optical / near - infrared spectrum decreasing in the far - infrared . one may see some resemblance of their seds to the sed of sole object iras 11385@xmath115517 . this indicates that these sources are actually sole objects without nebulosities , and hence , viewed pole - on and/or too distant and/or too compact for the shell to be resolved . iras 15039@xmath114806 may also belong to this category , but its sed is much steeper in the mid - infrared . the second group consists of iras 13416@xmath116243 , iras 17163@xmath113907 , iras 17516@xmath112525 , and iras 20547 + 0247 with a steep , rising near / mid - infrared spectrum and prominent but flat far - infrared part . they look similar to seds of duplexes , which may suggest that they are indeed duplex sources . however , we do not see any apparent presence of circumstellar envelope / disk on @xmath3 images . the lack of visible nebulosities may be caused by an orientation effect as the sources with the possible tori around are viewed almost pole - on . dust scattering is typically anisotropic and directions of scattering almost perpendicular to the direction of incident photons are less probable ( e.g. , * ? ? ? * ) . therefore , pole - on duplex sources may not show obvious nebulosities . those objects may also be too compact and/or too distant to be resolved and hence appearing as stellar sources . in fact , iras 13416@xmath116243 is known to have a dusty envelope and iras 20547 + 0247 is suspected to have dusty thick disk around central star seen edge - on ( see previous sections ) . sed of iras 17516@xmath112525 is similar to the ones of duplex objects , but it is constructed from the combined magnitudes of two stars unresolved up to now and the contribution from each star can not be separated . fig.8 . - infrared color - color diagrams for post - agb objects from table 5 . dashed lines separate parts of the diagrams occupied by different groups of objects . [ ir ] the sed of iras 18135@xmath111458 may also appear similar to the one typical for duplexes . however , while investigating this object we found that two distinct sources separated by @xmath59 were identified as this object . one is weakly visible at optical wavelengths with no infrared counterpart and observed with @xmath3 . the other is detected only in the far - ir . the latter is most likely the correct iras source ( also confirmed by msx observations ) and genuine post - agb object . the sed for this object shown in figure 6 is then the composition of measurements of two different objects and further observation should be undertaken to characterize the optical property of this post - agb object . the rv tau objects have similar seds with the significant infrared excess , but they can not be classified according to the sed classification scheme of @xcite . infrared data from iras and 2mass catalogues allow us to construct and analyze infrared color - color diagrams . fig.8 presents @xmath60 vs. @xmath61 $ ] and @xmath62 vs. @xmath63 diagrams for post - agb objects from the paper of umb00 and the post - agb stars analyzed in this study . infrared counterparts for each star were taken from the toru catalogue of galactic post - agb and related objects @xcite . the transformation of 2mass filters to johnson system @xcite was done and iras flux at 25@xmath7 m was converted to magnitude by @xmath64=-2.5log(f_{\nu}/6.73)$ ] ( iras explanatory supplement 1988 ) . the division between different groups of objects are drawn in fig.8 . clustering of sole , duplex and stellar objects at distinct locations was already discussed by umb00 . a larger object sample in this study allowed us to draw for the first time the specific division lines . duplex sources are the most red in @xmath60 color , and this is consistent with their optically thicker circumstellar shell . there is also a division of duplexes around @xmath61 \sim 12 $ ] between those with a partially visible central star ( @xmath61 < 12 $ ] ) and those with a fully obscured central star ( @xmath61 > 12 $ ] ) . sole objects are bluer than duplexes in @xmath60 color because of a thinner dust envelope . the slightly shifted positions of iras 11385@xmath115517 ( @xmath60 = 2.0 , @xmath61 $ ] = 7.2 ) is caused by the prominent near - ir excess , while the peculiar position of duplex source iras 09371 + 1212 = frosty leo ( @xmath60 = 0.6 , @xmath61 $ ] = 7.1 ) is a result of its very bright central star and therefore unusual ( for duplex source ) near - infrared colors . stellar objects are the most blue in both colors , except for few peculiar ones described in previous section . their position on the diagram overlapping with duplexes suggest they may be also duplex objects , but as we discussed earlier , seen pole - on . also near - infrared diagram shows the differences between sole and duplex classes . sole objects are bluer then duplex ones in both colors , with @xmath65 and @xmath66 , as it was expected for sources with clearly visible central stars and thin dust envelopes . the position of iras 11385@xmath115517 ( @xmath62=0.8 , @xmath63=1.4 ) is caused by a presence of hot dust . stellar objects are in general blue in @xmath62 color , @xmath67 , similar to sole sources , but redder in @xmath63 color , @xmath68 , more like duplex objects . it may be also a result of a hot dust very close to the star as it is in case of 89 her . again , 4 stellar objects discussed in section 4.2 are found among duplexes suggesting that they may be duplex objects for which we did not detect the nebulae because of the orientation effect . the differences in infrared colors between morphological classes are clearly visible and our conclusions follow and complement the results of umb00 . the study of @xcite did not recognize the ir color difference , however , there is a little overlap between their objects and ours . this is caused probably by selection . their objects are selected based on iras colors . they are bright in the infrared and located close to the galactic plane , so they are biased towards more massive objects . some of them may not be post - agb objects , e.g. , iras 15154@xmath115258 is a [ wr ] pn . we also analyzed iras color - color diagram for the objects from our study . the diagram is presented in figure 9 and regions defined by @xcite are drawn . the differences between morphological classes are also clearly seen . sole and duplex objects have similar [ 12]-[25 ] colors ( typical for evolved objects , located in regions iv , v and redder ) but they differ in [ 25]-[60 ] color with soles being bluer and duplexes being redder indicating colder dust in duplex sources . stellar objects are bluer in both colors and located in regions characteristic for less evolved objects ( regions iii and vii ) . four peculiar stellar objects described in the previous sections as possibly duplex sources but seen pole - on are redder in [ 25]-[60 ] color than other stellar sources and located in regions vib and iv occupied indeed by duplexes . the position of sole object iras 11385@xmath115517 ( in region viii ) is due to its cold dust . @xmath69 for frosty leo is not certain ( the quality of iras measurement is poor ) but anyway the object is very red in [ 25]-[60 ] color ( 2.97 ) and is located outside the presented diagram . large @xmath70 excess means that the object has a detached shell composed of cold matter or a reservoir of cold matter around central star . - iras color - color diagram for post - agb objects from table 5 with regions defined by @xcite . [ iras ] galactic distribution of all post - agb objects from table 5 is shown in fig.10 . duplex sources are closer to the galactic plane , while soles are usually further from the galactic plane . the mean galactic latitude for duplexes is @xmath71 with the standard deviation of 10.03 ( however , without peculiar frosty leo @xmath72 and standard deviation is @xmath73 ) and for soles @xmath74 with the standard deviation of 11.23 . previous studies of pn @xcite suggested that because of their position close to galactic plane bipolar pns have evolved from more massive progenitors than elliptical pns . following those findings , umb00 suggested that duplex objects have more massive progenitors than sole ones and will evolve into bipolar pns , while soles will form elliptical pns . however , the masses of duplex objects and some of soles are not known and this suggestion can not be fully confirmed . the distribution of stellar objects is quite wide with the mean galactic latitude of @xmath75 and the standard deviation of 18.07 . some of them , like iras 13416@xmath116243 , lie very close to the galactic plane suggesting again that they may be in fact duplex sources . on the other hand , stellar post - agb objects far from galactic plane most likely originate from low mass progenitors and may not become pns . fig.10 . - galactic distribution of post - agb objects from table 5 . [ galactic ] based on previous studies of pns and findings from observation of ppns , umb00 suggested that sole and duplex objects have different agb progenitors with low and high masses , respectively . more massive objects will have more material in circumstellar shells and the star will be quite obscured , while it wo nt be in a case of object with less material in the envelope . sources with dusty shells will therefore be fainter at the optical wavelengths but brighter in the infrared . this was already shown on color - color diagrams in fig.8 , with sole objects being bluer then duplex ones . knowing the masses of post - agb objects would be the best tool to confirm this dichotomy . derived physical parameters for 125 proto - planetary nebulae ( with references ) were gathered from the literature by @xcite . unfortunately , there are no information for duplex objects because of the lack of optical spectra from which those parameters can be obtained . the mean mass for soles from this study is @xmath76 , but they can be either very massive ( e.g. , iras 17436 + 5003 or iras 07134 + 1005 from umb00 with a mass of @xmath77 ) and very `` light '' ( e.g. , iras 08143@xmath114406 from this study with a mass below 0.55m@xmath78 ) . one has to remember also about errors on derived central star masses due to the uncertainties in effective temperature and gravity with mean values of @xmath79k and @xmath80 g = 0.3dex ( for some objects uncertainties are large and hence a range of possible masses is big ) . stellar objects in our sample are mostly of low mass , but there are some exceptions as well ( e.g. , iras 20117 + 1634 with a mass of @xmath81 ) . figure 11 shows the distribution of masses of sole and stellar objects from table 5 ( masses of duplexes are not know and can not be drawn on the diagram ) . masses of stellar objects are in general lower than masses of soles , especially when we do not consider rv tau stars ( plain columns on the diagram ) . - mass histogram for soles and stellar objects from table 5 . the most left column shows all objects with masses below @xmath82 . lower panel : plain columns correspond to rv tau and r crb stars . [ mass ] we searched also for a connection between different morphological groups and chemistry of ppns . in the sample of more than 50 objects from this study and umb00 we have 12 c - rich and 9 o - rich sole objects and 7 c - rich and 6 o - rich duplex objects . we see no obvious correlation between their chemistry and ppn morphology . however , different galactic distributions of soles and duplexes and therefore possibly differences in mass progenitors may suggest also the division of chemical composition . the low mass objects do not experience the 3rd dredge - up and evolve all the time as oxygen - rich sources . more massive objects experience thermal pulses and their composition undergoes significant changes , especially in carbon abundance , so they are expected to be carbon - rich . if duplexes are indeed more massive they should be rather c - rich and soles should be rather o - rich if they have indeed small masses . however , chemical composition depends on metallicity ( e.g. , * ? ? ? * ) and stars of equal masses and different metallicity can have different chemical abundances . therefore the simple chemical division in context of masses of stars is not possible . since we find both o- and c - rich chemistry in both soles and duplexes in our sample , massiveness of duplexes relative to soles based on the shell chemistry is not possible . however , after examining iso , iras and visir spectra ( http://www.ncac.torun.pl/postagb and + http://archive.eso.org/ ) we do find that all duplexes with the 9.8um feature show it in absorption while all soles with the 9.8um feature show it in emission . this finding possibly indicates that duplexes tend to be optically thicker than soles , suggesting the more massive nature of duplexes with respect to soles . while it is an intriguing trend , it is nevertheless based on small number statistics ( only 3 sole : iras 11385 - 5517 , iras 12175 - 5338 and iras 18095 + 2704 and 4 duplex : iras 08005 - 2356 , iras 10197 - 5750 , iras 17150 - 3224 and iras 23541 + 7031 objects ) . thus , more observations are needed to verify this . most of the stellar sources in our sample is oxygen - rich ( 15 o - rich and 5 c - rich ) and we see no correlation between their chemical composition and morphology and/or masses . thus it can not be said that chemistry of the central star is a key leading to differences between groups of post - agb objects . umb00 found almost equal number of sole and duplex objects , 11 and 9 respectively ( the 10th duplex source in their sample , iras 09452 + 1330 , is an agb star ) , while in our sample there are 9 sole and only 4 duplex post - agbs . the reason for this are selection criteria . the umb00 sample focused mostly on sources for which evidence for spatial extension existed . our objects were selected to complement the existing images of ppns and to cover more diverse properties of stars . most of them were already studied previously and hence we know their magnitudes , temperatures and chemistry . and usually they are quite bright , there are not many faint objects in our sample . this selection was also caused by requirements of snapshot survey preferring brighter stars with shorter observation time . all this provided to the lack of duplex objects , which are usually very faint ( @xmath8315.4 in our sample ) . we analyzed hst images of 31 post - agb objects . we increased the number of observed ppns and covered a wider variety of ppns . following previous results of umb00 we selected sole and duplex sources based on their morphology . our results are consistent with umb00 , and therefore , strengthen the validity of the general ppn structure suggested by umb00 , in which an intrinsically axisymmetric shell assumes a different morphology mainly due to the varying degree of the equatorial density enhancement which determines the presence of the dust lane in duplexes and the absence of it in soles . we selected a separate group of stellar post - agb objects without visible nebulosities . we searched for connections between morphology of the nebulosities and the chemical and physical properties of their central stars and did not find any obvious connection . we confirmed results of umb00 that sole objects with thin envelopes are bluer then dusty duplex sources . we also suggested that some of the stellar sources may be objects with the nebulosity around central star , but seen pole - on and/or too faint and distant to be resolved . we also found the differences in galactic distribution of different groups with the soles lying farther from the galactic plane than duplexes , which could suggest that duplex objects may have more massive progenitors . however , we were not able to confirm this suggestion directly because the masses of duplexes and some soles are not known . n. sidmiak and m. meixner acknowledge support from nasa / nag5 - 12595 , nasa / stsci - go-10627.01 and nasa / stsci - go-09377.05 . n. sidmiak also acknowledges support from grant n203 0661 33 of the science and high education ministry of poland . t. ueta acknowledges support from stsci go-10627.01 . r. szczerba acknowledges support from grant n203 019 31/2874 of the science and high education ministry of poland . balick , b. , frank , a. 2002 , , 40 , 439 barnbaum , c. , omont , a. , morris , m. 1996 , , 310 , 259 blcker , t. 1995 , , 299 , 755 carpenter , j.m . 2001 , , 121 , 2851 castro - carrizo , a. , bujarrabal , v. , snchez contreras , c. , alcolea , j. , neri , r. 2002 , , 386 , 633 cernicharo , j. , heras , a.m. , tielens , a.g.g.m . , pardo , j.r . , herpin , f. , gulin , m. , waters , l.b.f.m . 2001 , , 546 , l123 cohen , m. , van winckel , h. , bond , h.e . , gull , t.r . 2004 , , 127 , 2362 corradi , r.l.m . , schwarz , h.e . 1995 , , 293 , 871 de ruyter , s , van winckel , h. , dominik , c. , waters , l.b.f.m , dejonghe , h. 2005 , , 435 , 161 de ruyter , s , van winckel , h. , maas , t. , lloyd evans , t. , waters , l.b.f.m . , dejonghe , h. 2006 , , 448 , 641 draine , b.t . 2003 , , 598 , 1017 engels , d. 2002 , , 388 , 252 engels , d. , schmid - burgk , j. , walmsley , c.m . 1986 , , 167 , 129 fong , d. , meixner , m. , sutton , e.c . , zalucha , a. , welch , w.j . 2006 , , 652 , 1626 garca - lario , p. , manchado , a. , pych , w. , pottasch , s.r . 1997 , , 126 , 479 gauba , g. , parthasarathy , m. 2004 , , 417 , 201 geballe , t.r . , barnbaum , c. , noll , k.s . , morris , m. 2005 , , 624 , 983 gledhill , t.m . 2005 , , 356 , 883 gledhill , t.m . , chrysostomou , a. , hough , j.h . , yates , j.a . 2001 , , 322 , 321 gledhill , t.m . , yates , j.a . 2003 , , 343 , 880 gonzaga , s. et al 2005 , acs instrument handbook , v6.0 ( baltimore : stsci ) gonzalez , g. , lambert , d.l . , giridhar , s. 1997 , , 479 , 427 hony , s. , tielens , a.g.g.m . , waters , l.b.f.m . , de koter , a. 2003 , , 402 , 211 hrivnak , b.j . , kwok , s. , su , k.y.l . 2000 , in asp conf . ser . 1999 , asymmetrical planetary nebulae ii : from origins to microstructures , ed . j. h. kastner , n. soker & s. rappaport ( san francisco : asp ) , 155 hrivnak , b.j . , kwok , s. , su , k.y.l . 2001 , , 121 , 2775 hrivnak , b.j . , volk , k. , kwok , s. 2000 , , 535 , 275 hu , j.y . , slijkhuis , s. , de jong , t. , jiang , b.w . 1993 , , 100 , 413 iben , i.jr . , kaler , j.b . , truran , j.w . , renzini , a. 1983 , , 264 , 605 jura , m. 1986 , , 309 , 732 krist , j. , hook , r. 2004 , the tiny tim user s guide v6.3 ( baltimore : stsci ) kwok , s. 1993 , , 31 , 63 kwok , s. , hrivnak , b.j . , su , k.y.l . 2000 , , 544 , l149 kwok , s. , volk , k. , hrivnak , b.j . 2002 , , 573 , 720 laidler , v. et al . 2005 , synphot user s guide v5.0 ( baltimore : stsci ) lawlor , t.m . , macdonald , j. 2003 , , 583 , 913 maas , t. , van winckel , h. , lloyd evans , t. 2005 , , 429 , 297 maas , t. , van winckel , h. , waelkens , c. 2002 , , 386 , 504 marigo , p. , girardi , l. , bressan , a. 1999 , , 344 , 123 mauron , n. , huggins , p.j . 2006 , , 452 , 257 meixner , m. et al . 1999 , , 122 , 221 meixner , m. , ueta , t. , bobrowsky , m. , speck , a. 2002 , , 571 , 936 meixner , m. , zalucha , a. , ueta , t. , fong , d. , justtanont , k. 2004 , , 614 , 371 nyman , l .- . , hall , p.j . , olofsson , h. 1998 , , 127 , 185 olofsson , h. 2001 , in science with the atacama large millimeter array , ed . a. wootten ( san francisco : astronomical society of the pacific ) , 355 olofsson , h. , nyman , l .- . 1999 , , 347 , 194 reddy , b.e . , bakker , e.j . , hrivnak , b.j . 1999 , , 524 , 831 reddy , b.e . , lambert , d.l . , gonzalez , g. , yong , d. 2002 , , 564 , 482 renzini , a. 1979 , in stars and star systems ; proceedings of the fourth european regional meeting in astronomy ( dordrecht , d. reidel publishing co. ) , 155 reyniers , m. , van winckel , h. , gallino , r. , straniero , o. 2004 , , 417 , 269 sahai , r. 1999 , , 524 , l125 sahai , r. 2004 , in asp conf . 2004 , asymmetrical planetary nebulae iii : winds , structure and the thunderbird , ed . m. meixner , j.h . kastner , b. balick & n. soker ( san francisco : asp ) , 141 sahai , r. et al . 1998 , , 493 , 301 sahai , r. , bujarrabal , v. , castro - carrizo , a. , zijlstra , a. 2000 , , 360 , l9 sahai , r. , morris , m. , snchez contreras , c. , claussen , m. 2007 , eprint arxiv:0707.4662 sahai , r. , snchez contreras , c. , morris , m. 2005 , , 620 , 948 sahai , r. , te lintel hekkert , p. , morris , m. , zijlstra , a. , likkel , l. 1999 , , 514 , l115 sahai , r. , young , k. , patel , n.a . , snchez contreras , c. , morris , m. 2006 , , 653 , 1241 sahai , r. , zijlstra , a. , bujarrabal , v. , te lintel hekkert , p. 1999 , , 117 , 1408 sahai , r. , zijlstra , a. , snchez contreras , c. , morris , m. 2003 , , 586 , l81 snchez contreras , c. , sahai , r. , gil de paz , a. 2002 , , 578 , 269 sarkar , g. , sahai , r. 2006 , , 644 , 1171 schnberner , d. , blcker , t. 1993 , in luminous high - latitude stars ; the international workshop on luminous high - latitude stars , ed . sasselov ( san francisco : astronomical society of the pacific ) , 337 sirianni , m. , et al . 2005 , , 117 , 1049 stasiska , g. , szczerba , r. , schmidt , m. , sidmiak , n. 2006 , , 450 , 701 su , k.y.l . , hrivnak , b.j . , kwok , s. , 2001 , , 122 , 1525 su , k.y.l . , hrivnak , b.j . , kwok , s. , sahai , r. 2003 , , 126 , 848 su , k.y.l . , volk , k. , kwok , s. , hrivnak , b.j . 1998 , , 508 , 744 szczerba , r. , sidmiak , n. , stasiska , g. , borkowski , j. 2007 , , 469 , 799 szczerba , r. , stasiska , g. , sidmiak , n. , grny , s.k . 2003 in exploiting the iso data archive . infrared astronomy in the internet age , ed . c. gry , s. peschke , j. matagne , p. garcia - lario , r. lorente , & a. salama ( esa publications series ) , 149 trammell , s.r . 2000 , in asp conf . ser . 1999 , asymmetrical planetary nebulae ii : from origins to microstructures , ed . j. h. kastner , n. soker & s. rappaport ( san francisco : asp ) , 147 trammell , s.r . , goodrich , r.w . 1998 , in 193rd aas meeting , 1403 trammell , s.r . , goodrich , r.w . 2002 , , 579 , 688 ueta , t. , fong , d. , meixner , m. 2001 , , 557 , 117 ueta , t. , meixner , m. , bobrowsky , m. 2000 , , 528 , 861 ueta , t. , murakawa , k. , meixner , m. 2005 , , 129 , 1625 van de steene , g.c . , van hoof , p.a.m . , wood , p.r . 2000 , , 362 , 984 van de steene , g.c . , van hoof , p.a.m . 2003 , , 406 , 773 van der veen , w.e.c.j . , habing , h.j . , geballe , t.r . 1989 , , 226 , 108 van der veen , w.e.c.j . , habing , h.j . , van langevelde , h.j . , geballe , t.r . 1989 , , 216 , l1 van winckel , h. 1997 , , 319 , 561 van winckel , h. 2003 , , 41 , 391 van winckel , h. , reyniers , m. 2000 , , 354 , 135 van winckel , h. , waelkens , c. , waters , l.b.f.m . 2000 , in iau symposium 177 , the carbon star phenomenon , ed . r.f . wing ( kluwer academic publishers , dordrecht ) , 285 van winckel , h. , waelkens , c. , waters , l.b.f.m . , molster , f.j . , udry , s. , bakker , e.j . 1998 , , 336 , l17 volk , k. , hrivnak , b.j . , kwok , s. 2004 , , 616 , 1181 volk , k. , hrivnak , b.j . , su , k.y.l . , kwok , s. 2006 , , 651 , 294 volk , k. , kwok , s. , hrivnak , b.j . 1999 , , 516 , l99 volk , k. , kwok , s. , hrivnak , b.j . , szczerba , r. 2002 , , 567 , 412 waters , l.b.f.m . , trams , n.r . , waelkens , c. 1992 , , 262 , l37 waters , l.b.f.m . , waelkens , c. , mayor , m. , trams , n.r 1993 , , 269 , 242 , waelkens , c. , waters , l.b.m.f . 2004 , in asymptotic giant branch stars , ed . habing & h. olofsson , springer , 519 woods , p.m. , millar , t.j . , herbst , e. , zijlstra , a.a . 2003 , , 402 , 189 woods , p.m. , schier , f.l . , nyman , l .- . , olofsson , h. 2003 , , 402 , 617 lcccrcrl + 01005 + 7910 & 10627 & ivb & 21000 & 0.55 & c & 16.59 & 13,32 + z02229 + 6208 & 6364 & iva & 5500 & 0.56 & c & 1.50 & 13,15,16,32,38,40 + 04296 + 3429 & 6364 & iva & 7000 & 0.55 & c & -9.05 & 11,22,32,38,40,46 + 05341 + 0852 & 6364 & iva & 6500 & 0.55 & c & -12.14 & 13,16,32,38,46 + 06530@xmath110213 & 6364 & iva & 7250 & 0.56 & c & -0.14 & 11,16,20,32,38,40 + 07134 + 1005 & 6737 & ivb & 7250 & 0.84 & c & 10.00 & 12,13,15,18,32,38,40,46 + 07430 + 1115 & 6364 & iva & 6000 & @xmath10 0.55 & c & 17.07 & 11,16,32,38 + 08143@xmath114406 & 10627 & iva & 7150 & @xmath10 0.55 & c & -5.07 & 20,32 + 11385@xmath115517 & 9463 & ivb & 8500 & 0.55 & o & 5.94 & 19,32 + 12175@xmath115338 & 10627 & ivb & 7350 & 0.62 & o & 8.66 & 32,45 + 16206@xmath115956 & 10627 & ivb & 8500 & 0.66 & o & -7.49 & 7,32 + 17436 + 5003 & 6737 & ivb & 6600 & 0.91 & o & 30.87 & 9,10,16,32,38 + 18095 + 2704 & 6364 & iva & 6600 & 0.55 & o & 20.18 & 9,16,32,38 + 19114 + 0002 & 6737 & iva & 6750 & 0.66 & o & -4.96 & 9,32,38 + 19306 + 1407 & 9463 & iva & b & ... & o & -2.49 & 13,27,48 + 19475 + 3119 & 9463 & ivb & 7750 & 0.58 & o & 2.73 & 9,31,32 + 20000 + 3239 & 9463 & iva & 5500 & ... & c & 1.16 & 9,13,16,47 + 20462 + 3416 & 6364 & ivb & b1i & ... & o & -5.75 & 9,38 + 22223 + 4327 & 9463 & ivb & 6500 & 0.55 & c & -11.56 & 9,16,32,46 + 22272 + 5435 & 6364 & iva & 5750 & 0.57 & c & -2.52 & 9,32,38,39 + 23304 + 6147 & 9463 & iva & 6750 & 0.66 & c & 0.59 & 16,32,46 + + 04395 + 3601 & 9430 & iii & 25000 & ... & c & -6.53 & 3,30,37 + 06176@xmath111036 & 7297 & iii & 7500 & 0.62 & c / o & -11.76 & 4,5,16,32 + 08005@xmath112356 & 6364 & ii & f5 & ... & c / o & 3.58 & 16,38 + 09371 + 1212 & 9463 & ivb & k7ii & ... & ... & 42.73 & 16,25 + 10197@xmath115750 & 6816&ii / iii & a2i & ... & ... & -0.79 & 24 + 13428@xmath116232 & 9463 & iii & ... & ... & c & -0.59 & 41 + 15553@xmath115230 & 10627&ii / iii & ... & ... & ... & 0.36 & 41 + 16342@xmath113814 & 6364 & iii & ... & ... & o & 5.85 & 16,23 + 16594@xmath114656 & 8210 & iii & b7 & ... & c & -3.29 & 34,35,40,42,49 + 17106@xmath113046 & 8210 & iii & ... & ... & ... & 4.70 & 9,14,16,34 + 17150@xmath113224 & 6364 & iii & g2 & ... & o & 2.98 & 16,35,38 + 17245@xmath113951 & 8210 & iii & ... & ... & ... & -2.84 & 9,16,34,35 + 17423@xmath111755 & 6364 & ii & b & ... & o & 5.78 & 7,38 + 17441@xmath112411 & 6364 & iii & f5 & ... & c ? & 2.15 & 16,33,35,38 + 19024 + 0044 & 9463 & iii & ... & ... & ... & -2.65 & 28 + 19374 + 2359 & 6364 & ii & f5 & ... & o & 0.96 & 11,16,38 + 19477 + 2401 & 8210 & ii & ... & ... & ... & -1.06 & 9,13,16,34 + 20028 + 3910 & 6364 & iii & g4 & ... & c & 4.17 & 9,16,34,38,48 + 22036 + 5306 & 10185 & ii & ... & ... & o & -1.84 & 16,26,29 + 22574 + 6609 & 8210 & ii & ... & ... & c & 5.96 & 16,34,38 + 23321 + 6545 & 6364&ii / iii & ... & ... & c & 4.32 & 16,38 + 23541 + 7031 & 9463 & ii & b & ... & o & 8.42 & 2,18 + egg nebula & 9463&ii / iii & 6500 & @xmath84 0.94 & c / o & -6.50 & 21,32 + + 05113 + 1347 & 6364 & iva & 5250 & 0.60 & c & -14.29 & 16,32,38 + 10158@xmath112844 & 6364 & n / a & 7600 & 0.56 & c / o & 22.94 & 5,16,32,38 + 12222@xmath114652 & 10627 & n / a & 6800 & @xmath10 0.55 & o & -15.49 & 5,32,45 + 12538@xmath112611 & 10627 & n / a & 6000 & @xmath10 0.55 & o & 36.40 & 32,45 + 13416@xmath116243 & 10627 & ii & ... & ... & c & -0.73 & 36,41 + 15039@xmath114806 & 10627 & n / a & 8000 & 0.55 & o & 8.65 & 32,45 + 17163@xmath113907 & 10185 & ii & ... & ... & ... & -1.12 & 18 + 17279@xmath111119 & 10627 & n / a & 7300 & @xmath10 0.55 & c & 12.17 & 32,45 + 17516@xmath112525 & 9463 & ii & ... & ... & o & 0.06 & 36,44 + 17534 + 2603 & 10627 & n / a & 6550 & 0.61 & o & 23.19 & 5,16,32,50 + 18135@xmath111456 & 9463 & iii & ... & ... & o & 0.77 & 6,43 + 20136 + 1309 & 8210 & iva & g0 & ... & ... & -11.97 & 34 + 20547 + 0247 & 9463 & ii & g & ... & o & -26.11 & 1,8 + rv tau stars : + 06034 + 1354 & 10627 & n / a & 5900 & @xmath10 0.55 & o & -3.42 & 5,32 + 09256@xmath116324 & 10627 & n / a & 6700 & @xmath10 0.55 & c & -9.24 & 5,32 + 12067@xmath114508 & 10627 & n / a & 6000 & @xmath10 0.55 & o & 16.82 & 5,32 + 15469@xmath115311 & 10627 & n / a & 7500 & @xmath10 0.55 & o & 0.63 & 5,32 + 17243@xmath114348 & 10627 & n / a & 6750 & 0.94 & o & -4.99 & 5,32 + 19125 + 0343 & 10627 & n / a & 7750 & 0.58 & o & -3.49 & 5,32 + 19157@xmath110247 & 10627 & n / a & 7750 & 0.58 & o & -7.23 & 5,32 + 20117 + 1634 & 10627 & n / a & 5000 & 0.93 & o & -9.76 & 5,32 + r crb stars : + 15465 + 2818 & 6364 & n / a & 6705 & 0.66 & c & 50.98 & 16,32,38 + references : all objects are included in the toru catalogue of galactic post - agb and related objects ( * ? ? ? * http://www.ncac.torun.pl/postagb ) . 1 . @xcite , 2 . @xcite , 3 . @xcite , 4 . @xcite , 5 . @xcite , 6 . @xcite , 7 . @xcite , 8 . @xcite , 9 . @xcite , 10 . @xcite , 11 . @xcite , 12 . @xcite , 13 . @xcite , 14 . @xcite , 15 . @xcite , 16 . @xcite , 17 . @xcite , 18 . @xcite , 19 . @xcite , 20 . @xcite , 21 . @xcite , 22 . @xcite , 23 . @xcite , 24 . @xcite , 25 . @xcite , 26 . @xcite , 27 . @xcite , 28 . @xcite , 29 . @xcite , 30 . @xcite , 31 . @xcite , 32 . @xcite , 33 . @xcite , 34 . @xcite , 35 . @xcite , 36 . @xcite , 37 . @xcite , 38 . @xcite , 39 . @xcite , 40 . @xcite , 41 . @xcite , 42 . @xcite , 43 . @xcite , 44 . @xcite , 45 . @xcite , 46 . @xcite , 47 . @xcite , 48 . @xcite , 49 . @xcite , 50 .

the low - mass microquasar gx 339 - 4 was discovered in 1972 by the satellite oso7 ( markert , canizares & clark 1973 ) . since then , it has been extensively observed at all wavelengths from radio to x - rays , and detected in all the canonical spectral states of x - ray binaries . little is known , however , about the remaining characteristics of the binary system . based on modulations in the optical photometry , callanan et al . ( 1992 ) inferred an orbital period of 14.8 hs , later confirmed by buxton & vennes ( 2003 ) using optical spectroscopy . further optical spectroscopic measurements and analysis of long - term x - ray light curves showed no evidence of this modulation , revealing instead a periodicity of @xmath01.75 days ( hynes et al . 2003 ; levine & corbet 2006 ) . the first estimates of the distance to gx 339 - 4 placed the system at @xmath1 kpc , see zdziarski et al . ( 1998 ) and references therein . this result was later revised by zdziarski et al . ( 2004 ) , who concluded that the minimum distance lay in the range @xmath2 kpc @xmath3 kpc . they favoured a location in the galactic bulge at @xmath4 kpc . however , a study of absorption lines performed by hynes et al . ( 2004 ) suggests that a location in the far side of the galaxy at a distance @xmath5 kpc can not be completely ruled out . the emission in the optical band is dominated by the accretion flow ( imamura et al . 1990 ) , preventing direct observation of the secondary star even when the system is going through the very low x - ray luminosity state . the first detection of the star was made by hynes et al . ( 2003 ) during the x - ray outburst observed in 2002 . the mass and spectral type of the donor star have not been firmly established yet . according to hynes et al . ( 2004 ) , an orbital period of @xmath01.7 days implies a low density for the companion of @xmath6 g @xmath7 . this probably corresponds to a low - mass subgiant of spectral type g or f , depending on the assumed distance . following this idea , muoz - darias , casares & martnez - pais ( 2008 ) suggested that the star is a `` stripped - giant '' , in which mass loss is due to the burning of an hydrogen shell . in this model , the mass of the secondary must be in the range @xmath8 . from this result , the authors constrained the mass of the compact object to be @xmath9 or even @xmath10 , for a mass of the secondary near the lower or upper limit , respectively . these values strongly support the idea that the compact object is a black hole ( see also hynes et al . 2003 ) . gx 339 - 4 has been observed in radio , infrared , optical and x - ray wavelengths , sometimes simultaneously or quasi simultaneously ( hannikainen et al . 1998 ; wilms et al . 1999 ; nowak , wilms & dove 2002 ; homan et al . the source goes through all the spectral states of x - ray binaries : low - hard , high - soft , very high state , intermediate state and quiescence ( mcclintock & remillard 2006 ) . it frequently displays outbursts associated with state transitions , episodes during which the x - ray luminosity can reach peaks of @xmath11 erg s@xmath12 for an assumed distance of 6 kpc ( homan et al . 2005 ; yu et al . 2007 ) . it was after the x - ray outburst of 2002 that gallo et al . ( 2004 ) imaged for the first time a relativistic radio jet on @xmath13 au scales in the system ( see also corbel et al . the detection of the jet confirmed that gx 339 - 4 is a microquasar . the hard x - ray emission during the low - hard state of microquasars and black hole binaries is generally thought to have its origin in a hot corona that surrounds the black hole . the observed power - law spectrum is explained through compton up - scattering of accretion disc photons by hot electrons in the corona . some of these up - scattered photons in turn excite iron nuclei in the disc material , giving rise to the appearance of a k@xmath14 line at @xmath15 kev , superimposed on the power - law continuum ; for evidence supporting the presence of the fe k@xmath14 line in gx 339 - 4 , see for example dunn et al . . however , corbel et al . ( 2003 ) showed that in gx 339 - 4 the radio and x - ray fluxes are tightly correlated , @xmath16 . this suggests that the emission in both bands might have a common origin in synchrotron radiation produced by non - thermal electrons in the jet , and not in the corona ( corbel & fender 2002 ; corbel et al . this idea was explored by markoff et al . ( 2003 ) and markoff , nowak & wilms ( 2005 ) , who applied a jet model to fit the observations . they showed that synchrotron radiation of relativistic electrons in the base of a jet can explain both the radio and x - ray spectra and reproduce the observed correlation . this model , however , is purely leptonic : the contribution to the radiation output of relativistic protons in the jet is not taken into account . in this work we present a lepto - hadronic model for the broadband electromagnetic spectrum of gx 339 - 4 . in our model , protons and electrons are accelerated up to relativistic energies . the hadronic radiative contribution extends well into the gamma - ray band , since typical maximum proton energies are much higher than those of electrons . we expect that our predictions can be tested in the near future with the data collected by instruments like the fermi gamma - ray satellite and by atmospheric cherenkov arrays like the high energy stereoscopic system ii ( hess ii ) . along with electromagnetic radiation , the creation of electron - positron pairs is a necessary result of relativistic particle interactions . they are injected , for example , through photon - photon annihilation and as a by - product of hadronic interactions . recently , measurements carried out with the integral spectrometer ( spi ) instrument of the integral satellite have allowed to complete a detailed map of an extended region of emission line at 511 kev in the galaxy ( weidenspointner et al . these observations confirm the diffuse ( rather than point - like ) distribution of the line , with bright emission around the galactic center ( flux @xmath17 @xmath18 s@xmath12 ) , and a clear , asymmetric ( toward negative longitude values ) disc component . the disc total flux is @xmath19 of the bulge flux . the disc emission is detected up to scales of @xmath20 from the galactic center . different types of positron sources have been suggested in the literature , including pulsars , the massive black hole at the galactic center , microquasars , nucleosynthesis events , and extended processes like cosmic ray nuclear reactions and dark matter decay . the fact that both bulge and disc emission are clear , seems to disfavour the latter two possibilities as the main positron sources . microquasars , in particular those with a low - mass donor star , seem to be a particularly appealing possibility , given the spatial correlation of the line emission with the overall distribution of low - mass x - ray binaries in the galaxy ( weidenspointner et al . 2008 ) and energetic arguments ( guessoum , jean , & prantzos 2006 ) . up to the present , none of the microquasar models available presents quantitative predictions of the positron production rate , although in romero & vila ( 2008 , 2009 ) the radiative output of positrons was taken into account . the recent findings described above makes it timely an exploration of the possibilities for electron - positron production in self - consistent models for microquasar jets . in the present paper we have devoted a section to estimate the positron production rate in systems like gx 339 - 4 . the article is organized as follows : in section [ model ] we present the basis of the model , describing in detail the hypothesis made about the geometry and energetics of the system . we also describe the calculation of the relativistic particle distributions ( for primary and secondary particles ) and the radiative processes . in section [ general results ] , we discuss the observational data sets and the constraints they impose on the model parameters . we also present the best - fitting spectral energy distributions obtained and analyse the results . our estimations for the positron production rate are described in section [ positrons ] . finally , in section [ conclusions ] , we summarise our results and we compare them with those of previous works . we close discussing the predictions for the very high - energy emission . the jet model applied here is based on the model developed in romero & vila ( 2008 ) . a sketch of the system is shown in figure [ jet - sketch ] . the jet is supposed to be conical , and launched at a distance @xmath21 from the compact object , where @xmath22 is the gravitational radius of the black hole . the initial radius of the jet is taken to be a fraction @xmath23 of the value of the injection distance , @xmath24 . following falcke & biermann ( 1995 ) , we relate the jet power to the eddington luminosity of the black hole as @xmath25 where @xmath26 erg s@xmath12 . the factor @xmath27 accounts for the existence of a counterjet of equal power . + of the black hole ( bh ) , and acceleration of particles takes place further away at @xmath28 . the width of the acceleration region is indicated as @xmath29 . the jet axis makes an angle @xmath30 with the line of sight . ] at an arbitrary distance @xmath31 from the compact object , the total energy budget of the jet can be roughly divided into magnetic energy , bulk kinetic energy and particle internal energy since part of @xmath32 is dissipated as radiation . ] : @xmath33 if the plasma is ejected by some kind of magnetocentrifugal mechanism , we can assume that at the base of the jet the entire jet power is in the form of magnetic energy , @xmath34 . this allows to estimate the value of the magnetic field @xmath35 by equipartition between the magnetic energy density and the total energy density , @xmath36 where @xmath37 is the bulk velocity of the outflow . after the launching point the magnetic field decreases with the distance @xmath31 to the compact object as @xmath38 with @xmath39 ( krolik 1999 ) . magnetic energy is then converted into bulk kinetic energy ( and a fraction of this is dissipated at shocks and converted into internal energy of relativistic particles , see below ) , in such a way that equipartition holds only at the base of the jet . a few remarks are in order here . in the region @xmath40 the outflow is a poynting - dominated flux , formed by a thermal plasma that accelerates as it gets energy from the ordered magnetic field . thermal electrons and ions in the `` pre - acceleration '' region move approximately with the jet bulk velocity and have an energy @xmath41 . the plasma is neutral , but almost all the bulk kinetic energy of the jet is carried by the ions . no significant electron synchrotron radiation is expected from this region . at @xmath42 shocks perturb the field , and a fraction of the energy associated with the bulk motion of the plasma is converted into internal energy of relativistic non - thermal particles . in this region synchrotron cooling becomes efficient , especially for electrons . this situation is different from the case of high - mass microquasars , where `` cold '' relativistic electrons can cool by inverse compton interactions with the stellar photon field . a similar scenario arises in binary systems formed by a pulsar and a luminous be star , such as psr b1259 - 63/ss2883 , see khangulyan et al . ( 2007 ) . most of the bulk kinetic energy of the jet is carried by a thermal plasma . if shock waves propagate through some region of the outflow , the suprathermal tail of the maxwellian particle distribution can be accelerated up to relativistic energies by diffusion across the shock front ( see , for instance , drury 1983 and references therein ) . the physical conditions for an efficient acceleration are not clear . for the plasma to be mechanically compressible and allow the formation of shocks , the magnetic energy density @xmath43 must be in sub - equipartition with the bulk kinetic energy density @xmath44 of the plasma ( see komissarov et al . 2007 for a discussion on this topic ) . therefore , the acceleration region has to be located at a distance @xmath28 from the black hole such that @xmath45 the kinetic energy density of the jet can be written as @xmath46 where @xmath47 is the cold particle density , and @xmath48 is the relativistic kinetic energy of a proton that moves with the jet bulk velocity , @xmath49 the presence of shocks may not suffice to accelerate particles efficiently . according to gaisser ( 1990 ) , for diffusive shock acceleration to work , the ram pressure in the acceleration region must dominate over the magnetic pressure . this condition can be written as @xmath50 where @xmath51 is the internal matter energy density . for a cold - proton - dominated jet , @xmath51 can be calculated as in bosch - ramon , romero & paredes ( 2006 ) , @xmath52 here @xmath53 is the jet mass flow , @xmath54 and @xmath55 is the classical kinetic energy of a thermal proton , @xmath56 the mean velocity of the particles was taken to be equal to the lateral expansion velocity of the jet , @xmath57 , that is of the order of the speed of sound in the plasma . condition ( [ sub - equipartition_2 ] ) is stronger than ( [ sub - equipartition ] ) , in the sense that if the former is fulfilled , so is the latter . in either case , the location of the innermost acceleration region can be determined demanding that the appropriate condition is satisfied , @xmath58 with @xmath59 . we assume that the jet is dynamically dominated by cold matter . the total power @xmath60 injected in relativistic particles is only a small fraction of the jet power , @xmath61 this energy is shared between relativistic protons and leptons , @xmath62 . we relate the energy budget of both species as @xmath63 the parameter @xmath64 remains free in our model but , as we are interested in jets with a relevant hadronic content , we kept @xmath65 throughout . the differential spectrum of particles accelerated by diffusion through shock waves is a power - law in energy , and the particle injection function can be written as @xmath66 the normalization constant @xmath67 is calculated from the total injected power , @xmath68 where @xmath69 is the volume of the acceleration region . particles can gain energy up to a certain value @xmath70 for which the sum of the cooling rates @xmath71 over all the relevant processes of energy loss @xmath72 , equals the acceleration rate ( aharonian 2004 ) , @xmath73 the parameter @xmath74 characterises the efficiency of the acceleration mechanism . the steady - state energy distributions of relativistic particles @xmath75 is then calculated solving the transport equation in the one zone approximation in the jet co moving reference frame ( where the distributions are supposed to be isotropic ) , @xmath76+\frac{n(e)}{t}=q(e ) . \label{transp_eq}\ ] ] the time @xmath77 is the characteristic time of `` catastrophic '' non - radiative particle losses : processes through which particles are removed from the system , for example by escape or decay . as protons and electrons are stable particles , @xmath77 is simply equal to the escape time - scale from the acceleration region , @xmath78 equation ( [ transp_eq ] ) neglects effects of convection and diffusion , and therefore it is only valid in a thin region of the jet . its exact analytical solution can be found , for example , in khangulyan et al . ( 2007 ) . we consider several mechanisms of interaction of the relativistic particles with the magnetic field , photons and matter in the jet . both protons and leptons interact with the magnetic field emitting synchrotron radiation . electrons also radiate by relativistic bremsstrahlung as they are accelerated in the mean electrostatic field of the ions in the jet plasma , and by inverse compton scattering ( ic ) against the synchrotron photon field ( synchrotron self compton process , ssc ) . interaction of protons with the synchrotron field ( @xmath79 ) proceeds through photopair production @xmath80 and photomeson production @xmath81 @xmath82 an approximation for the cross section for photopair production is given in maximon ( 1968 ) . if @xmath83 is the photon energy in the proton rest frame , then for @xmath84 @xmath85\rm{mb } , \label{sigma_photopair1}\ ] ] whereas for @xmath86 @xmath87\rm{mb}. \label{sigma_photopair2}\ ] ] in the case of photomeson interactions , the cross section @xmath88 can be roughly approximated as a step function ( atoyan & dermer 2003 ) , @xmath89 120\,\mu\mathrm{barn } \quad \epsilon^\prime>500\,\mathrm{mev } , \end{array } \right . \label{sigma_phtomeson}\ ] ] relativistic protons also produce pions by interaction with the jet matter field through proton - proton @xmath90 collisions , @xmath91 @xmath92 and @xmath93 for high proton energies , the probability of creation of the three pion species is almost the same . the total inelasticity of the process is @xmath94 ; most of the kinetic energy lost by the proton is carried away by only one or two leading pions . the inelastic @xmath95 cross section can be accurately approximated as ( kelner , aharonian & bugayov 2006 ) @xmath96 ^ 2\,\rm{mb } , \label{pp3b}\end{aligned}\ ] ] where @xmath97 gev is the proton threshold energy for @xmath98 production . the pion multiplicities @xmath64 and @xmath99 have power - law dependences on the relativistic proton energy , @xmath100 with @xmath101 ( mannheim & schlickeiser 1994 ) . + the gamma - ray output from proton - proton and photomeson interactions is due to the decay of neutral pions , @xmath102 besides radiative cooling , particles also lose energy through adiabatic losses , since they exert work on the jet boundary surface . the expression for the adiabatic cooling rate and for all the radiative mechanisms of energy loss are compiled in romero & vila ( 2008 ) and reynoso & romero ( 2009 ) . we calculate the contribution of each radiative process to the luminosity . for synchrotron radiation , relativistic bremsstrahlung , and inverse compton scattering ( in both thomson and klein - nishina regimes ) we use the formulae given in blumenthal & gould ( 1970 ) . the gamma - ray luminosities from proton - proton and photomeson interactions are calculated following kelner , aharonian & bugayov ( 2006 ) and kelner & aharonian ( 2008 ) , respectively . all the luminosities , except those from process of interactions with matter , are better calculated in the jet co moving reference frame . the corresponding luminosity in the observer frame is then obtained applying the doppler boosting factor @xmath103 , @xmath104^{-1 } , \label{boost - factor}\ ] ] where @xmath30 is the viewing angle ( the angle between the jet axis and the line of sight ) and @xmath105 . denoting with primes the magnitudes in the co moving frame , for an approaching jet the luminosity transforms as @xmath106 whereas the photon energy in the observer frame is @xmath107 charged pions , muons and electron - positron pairs produced in proton - proton and proton - photon interactions can also contribute to the electromagnetic emission spectrum . if the magnetic field and/or the photon and matter fields in the jet are strong enough , @xmath108 and @xmath109 can significantly cool through synchrotron radiation , inverse compton scattering and even pion - proton collisions @xmath110 before decaying . charged pions decay creating muons and neutrinos , @xmath111 @xmath112 muons in turn decay yielding electrons , positrons and more neutrinos , @xmath113 @xmath114 the pion injection function @xmath115 for pions created in @xmath95 collisions is given in kelner , aharonian & bugayov ( 2006 ) ; for pions injected through photomeson interactions we apply the @xmath116-functional approximation , see atoyan & dermer ( 2003 ) . from @xmath115 , we calculate the muon injection function @xmath117 using the formalism presented in lipari , lusignoli & meloni ( 2007 ) . electron - positron pairs created by muon decay and those directly injected through photopair interactions are also a source of electromagnetic radiation . we calculate the pair injection function from muon decays following schlickeiser ( 2002 ) . to estimate the pair output @xmath118 from photopair production , we apply the formulae given in chodorowski , zdziarski & sikora ( 1992 ) and mcke et al . finally , there is another source of pair injection due to photon - photon ( @xmath119 ) annihilation , @xmath120 the corresponding pair source function is given in bttcher & schlickeiser ( 1997 ) . the process of @xmath119 annihilation is also a photon sink , and can eventually modify the production spectrum . its effect on the escape of photons is discussed in the next section . the steady state distribution of secondary pions , muons and electron - positron pairs is calculated solving the transport equation ( [ transp_eq ] ) , in the same way as for the primary particles . in the case of @xmath108 and @xmath109 , the decay time must be included in @xmath77 along with the escape time , @xmath121 in the jet co moving reference frame , the decay time @xmath122 is given by @xmath123 here @xmath124 is lorentz factor of the particles and @xmath125 is the decay time in their rest frame , @xmath126 s and @xmath127 s. as remarked above , gamma - ray photons can annihilate against low - energy radiation to create electron - positron pairs . for a gamma ray of energy @xmath128 to interact with a photon of energy @xmath129 , their energies must exceed the threshold @xmath130 the angle @xmath131 is that between the momenta of the colliding photons . according to equation ( [ gamma - gamma - threshold ] ) , tev gamma rays can be strongly absorbed by infrared photons produced , for example , by electron synchrotron radiation in the jet . thus , this process of internal absorption can modify the shape of the high - energy region of the production spectrum . the probability of absorption of a gamma ray can be quantified through the opacity @xmath132 , defined as @xmath133 here @xmath134 is the angle averaged cross section ( gould & schrder 1967 ) , @xmath135 is the distribution of target photons , @xmath136 , and @xmath137 is the length of the path traversed by the gamma ray from the emission site to the observer . the shape of the modified luminosity @xmath138 can be estimated from the primary production spectrum as @xmath139 this provides a zeroth - order approximation to the final emission spectrum , mainly indicating those energy ranges where the emission is suppressed . a more refined treatment of the problem requires solving a set of coupled equations to calculate self - consistently the equilibrium distribution of particles and photons , or to solve the equations of an electromagnetic cascade , if it develops . however , the strong magnetic field suppresses ic cascades through fast cooling by the synchrotron channel ( see khangulyan , aharonian & bosch - ramon 2008 ) . cascades are expected to play a more important role in high - mass x - ray binaries , see for example orellana et al . gx 339 - 4 was extensively observed simultaneously in radio and x - rays during the low - hard state ( lhs ) in 1997 , 1999 and 2002 . for some of these observations , simultaneous near infrared ( nir ) and optical data are also available . the 1997 and 1999 radio observations were carried out with the australia telescope compact array ( atca ) and the molonglo observatory synthesis telescope ( most ) , and are described in detail in corbel et al . ( 2000 ) . the radio , nir and optical data from 2002 is presented in homan et al . the x - ray data were collected with the rossi x - ray timing explorer ( rxte ) , and are compiled in wilms et al . ( 1999 ) , nowak , wilms & dove ( 2002 ) , corbel et al . ( 2003 ) and homan et al . we refer the reader to these works for model assumptions and other details of the data extraction in each case . no further data reductions were performed for this paper . additional information on the particular data sets used here is presented in table [ obs_data ] . [ cols="^,^,^,^,^,^,^,^ " , ] + the values of @xmath140 , @xmath64 , @xmath141 , @xmath142 and @xmath143 were left free during the fitting procedure ; @xmath144/d.o.f . is the value of @xmath144 per degree of freedom . the spectral energy distributions ( seds ) presented in figure [ seds1 ] all correspond to obs1 . the best fit is obtained for @xmath145 . the x - ray data range is always covered by the synchrotron emission of primary electrons , but as @xmath146 grows , synchrotron radiation of secondary pairs begins to dominate at radio wavelengths . this diminishes the quality of the fit . also as @xmath146 increases the slope of the x - ray spectrum gets worse modeled , indicating that the injection index should be harder than the assumed @xmath147 . between @xmath148 gev and @xmath148 tev , the emission is dominated by synchrotron self - compton ( ssc ) radiation and synchrotron emission of protons and secondary particles ; at higher energies the contributions of proton - proton and proton - photon interactions are the dominant ones . all these processes become more relevant when @xmath28 is nearer the jet base , since the magnetic field is stronger and enhances the synchrotron radiation of @xmath149 , @xmath108 and @xmath109 . also the matter and photon densities are larger , providing denser targets for @xmath95 and @xmath79 collisions , and ssc scattering . the contribution of secondary pairs from @xmath119 annihilation is significantly increased for large values of @xmath146 due to this effect as well . in all cases the best fits favour large minimum particle energies , @xmath150 . a powerful jet ( @xmath151 ) and equipartition of energy between primary protons and leptons ( @xmath152 ) is also required , since the power injected in electrons needs to be as large as possible to account for the x - ray observations . models e and f in figure [ seds2 ] correspond to fits of obs1 and obs2 , respectively . in both cases , @xmath28 was calculated demanding that @xmath153 . for the same @xmath146 and @xmath143 , this condition gives larger values of @xmath28 and weaker magnetic fields . now the best fits are obtained for large values of @xmath146 . these sets of parameters reproduce the slope of the x - ray spectrum for the same value of the injection index better than the models of figure [ seds1 ] . figure [ seds3 ] shows two model fits to low - luminosity low - hard state observations of gx 339 - 4 , carried out in 1999 ( obs3 and obs4 ) . the radio and x - ray emission is due to primary electrons ; all radiative contributions of protons and secondary particles are negligible . the jet power required to account for the data is now only a fraction @xmath154 of the eddington luminosity . the best fit models are obtained for low values of the acceleration efficiency and minimum particle energy . for each model we calculated the synchrotron emission of thermal electrons at the base of the jet . for an electron energy @xmath155 and a magnetic field @xmath156 g , the peak of the spectrum is at @xmath157 ev . however , the luminosity of this component is below or just above the jet emission . this contribution is not significant in the relevant energy bands . therefore , the results of the fits are not affected . the analysis of simultaneous radio and x - rays observations from 1997 - 1999 , led corbel et al . ( 2003 ) to find out that the fluxes in both energy bands are tightly correlated . they showed that the radio flux at 8.6 mhz is related the 3 - 9 kev integrated x - ray flux as . this correlation suggests a common origin in the jet ( synchrotron radiation ) . according to markoff et al . ( 2003 ) , if all model parameters except the jet power are kept frozen , the correlation index @xmath116 is given by @xmath158 here @xmath159 is the spectral index of the x - ray region of the synchrotron spectrum @xmath160 , and @xmath161 that of the synchrotron radio flux @xmath162 . in our model radio and x - ray emission is due to synchrotron radiation of electrons . we find a value of the radio spectral index @xmath163 , which corresponds to the optically thick part of the spectrum from a particle distribution with a low - energy cutoff . the value of the x - ray spectral index is @xmath164 , as expected for an injection function @xmath165 ( notice that electrons are then strongly cooled due to synchrotron losses ) . these values yield @xmath166 . figure [ correlations ] shows the correlation curves predicted by our model for cases a , e , f and g , together with the data from corbel et al . the model results are in reasonable agreement with the observations . , [ seds2 ] and [ seds3 ] . the model correlation index is @xmath166 , whereas that of the 1997 - 1999 observational data is @xmath167 ( corbel et al . 2003 ) . ] simultaneously with the radio and x - ray observations of obs3 and obs5 , gx 339 - 4 was also detected at nir and optical wavelengths ( markoff et al . 2003 ; homan et al . 2005 ) . as in the case of the radio emission , the nir / optical flux also displays a strong correlation with the x - ray flux . from an analysis of low - hard state data from the 2002 outburst of gx 339 - 4 ( same epoch as obs5 ) , homan et al . ( 2005 ) shown that the flux density in the nir _ h_-band and the @xmath168 kev bolometric x - ray flux correlate as @xmath169 , with @xmath170 . a similar correlation was found between the optical _ v_-band and _ i_-band flux densities and the integrated x - ray flux , with correlation indices @xmath171 and @xmath172 , respectively . these correlations disappear when the source leaves the low - hard state . the _ h_-band emission , however , rises and decays faster than the optical during the state transition , while the slope between the _ i _ and _ v _ bands remains constant . as suggested by homan et al . ( 2005 ) , this may indicate a different origin for the nir and optical emission during the low - hard state . the correlations between the radio / x - ray and nir / x - ray fluxes suggest that the emission in the three ranges must originate in the jet . this is further supported by the fact that the nir flux extrapolates back to the radio data ( see also corbel et al . direct or reprocessed emission from an accretion disc can be ruled out due to the shape of the nir / optical spectrum and the short decay time scales . furthermore , the nir and radio fluxes are quenched when the disc begins to contribute significantly to the x - ray emission . homan et al . ( 2005 ) conclude that the nir emission probably originates in the jet , and approximately coincides with the position of the break of the synchrotron spectrum . the optical flux may be due to thermal and/or non - thermal reprocessed radiation from the accretion disc or star , or from a region of the jet different from where the nir emission is produced . these ideas are supported by the recent results of coriat et al . ( 2009 ) , who presented an analysis of five years of observations of gx 339 - 4 ( from 2002 to 2007 , a period that comprises five outbursts ) . they found a strong ir / x - ray correlation over four decades in flux during the low - hard state . the correlation index , however , is not unique : a break appears at bolometric ( 3 - 100 kev ) x - ray fluxes @xmath173 erg s@xmath12 @xmath18 ( @xmath174 for @xmath175 and @xmath176 kpc ) . coriat et al . ( 2009 ) argue that this break can be explained attributing the x - ray emission to ssc radiation from the jet ( see also nowak et al . 2005 , where it is suggested that models more complex than a single jet synchrotron component maybe needed to explain the correlations ) . they find no clear evidence of a similar break in the _ v_-band / x - ray correlation , and suggest that the optical emission in the low - hard state is dominated by the outer part of the accretion disc , and not by the jet . this correlation is not peculiar of gx 339 - 4 , but it seems to be a signature of low - mass black hole x - ray binaries . russell et al . ( 2006 ) analysed radio , nir , optical and x - ray data from 16 sources ( including extragalactic systems in the lmc ) . their results agree with those of homan et al . ( 2005 ) for gx 339 - 4 . they estimate that the jet contribution to the nir emission during the low - hard state is @xmath177 , but only @xmath178 to the _ i _ and _ v _ bands . figure [ seds4 ] shows the best fit models obtained for obs3 ( including the lowest energy point in the optical ) and obs5 . in the case of obs5 , the radio , nir , optical and x - ray data are reasonably well reproduced using a hard particle injection spectral index @xmath147 . in the case of obs3 , the data at optical frequencies can not be accounted for with a single synchrotron component . however , using a softer particle injection spectral index @xmath179 , it is possible to obtain models models where the synchrotron turnover occurs in the optical . the rise in the spectrum at higher energies , however , can not be fitted . this emission must have a different origin , for example in an accretion disc ( markoff et al . 2003 ) . due to @xmath119 annihilation in the total photon field of the emission region . absorption is negligible in all cases and does not modify the production spectrum . ] in order to asses the effect of photon self - absorption by @xmath119 annihilation , we calculated the attenuation parameter @xmath180 in equation ( [ attenuated - luminosity ] ) . as it can be seen from figure [ attenuation ] , internal attenuation is almost negligible due to the low photon density in the emission region . the production spectrum is not appreciably modified in any case , contrary to some models calculated by romero & vila ( 2008 ) . according to the results of figures [ seds1 ] , [ seds2 ] and [ seds4 ] , during the high luminosity low - hard state , gx 339 - 4 would be a source detectable by fermi in the energy range 100 mev - 10 gev . in this range the emission is due to electron synchrotron radiation and ssc . however , it will be far from being a bright source , possibly appearing with a significance of 5 - 6 @xmath181 after one year of exposure . at higher energies , the emission of hadronic origin will be harder to detect with the present instruments . for some models , the predicted luminosities at @xmath148 tev are above the sensitivity of hess for a point source at @xmath182 kpc through exposures of more than 50 hours . future instruments like the cherenkov telescope array ( cta ) will make it easier to obtain a firm detection . such a detection at high energies is crucial to evaluate the hadronic content of the jet . on the other hand , when the source is in a low luminosity phase , according to the results in figure [ seds3 ] it would be undetectable at high energies with the present instruments . recently , an analysis of several years of data collected by the integral satellite led weidenspointner et al . ( 2008 ) to suggest that the observed asymmetry in the distribution of the electron - positron annihilation line traces the distribution of hard low - mass x - ray binaries ( lmxbs ) in the galaxy . these authors discovered that the 511 kev line emission from the inner galactic disc is clearly asymmetric : the flux from the region of negative galactic longitudes is 1.8 times larger than that from the corresponding region of positive longitudes . the same type of asymmetry is observed in the spatial distribution of hard lmxbs , those that show appreciable emission above 20 kev . the authors argue that positrons escaping the injection region with an energy of @xmath148 mev will not diffuse more than @xmath183 pc before annihilating . this distance is short enough for the line emission to be still correlated with the large - scale distribution of the sources . weidenspointner et al . ( 2008 ) estimate that the positron production rate required to account for the observed flux is @xmath184 erg s@xmath12 . to investigate this possibility we calculated the positron injection rate in our model . according to heinz ( 2008 ) , the number of injected positrons per unit time can be roughly estimated as @xmath185 in this expression @xmath186 is the total luminosity injected in pairs and @xmath187 is the mean lorentz factor of the positrons when they leave the source . it is reasonable to expect that positrons have almost completely cooled when they reach the end of the jet , and thus @xmath187 is of the order of the jet bulk lorentz factor , @xmath188 . in our case , the most relevant process of pair production is @xmath119 annihilation . the calculated positron injection rates are shown in table [ fit - parameters ] . for those models that correspond to the high x - ray luminosity states , @xmath189 is very near or even exceeds the necessary lower limit given by weidenspointner et al . there are about @xmath190 lmxb in the galaxy and , although not detected yet , possibly most of them produce jets . even if many of them are considerably less powerful than the jet in gx 339 - 4 , the added contributions might well account for the observed flux at 511 kev . our estimations show that the proposed association between lmxb and the annihilation line emission is indeed feasible at least in energetic terms . in this way , there may be no need to resort to other more exotic explanations , such as annihilation of dark matter . a more detailed discussion on this topic will be presented elsewhere . we have shown that , under certain general conditions , the model developed here is capable of explaining the observed radio and x - ray spectrum of the low - mass microquasar gx 339 - 4 . the parameter that mainly determines the characteristics of the obtained spectra is the magnetic field decay index @xmath146 . this parameter , in turn , depends on unknown factors as the geometry and turbulence level of the magnetic field in the inner jet . the best fits to the high luminosity 1997 and 1999 data are obtained in those cases where the acceleration / emission region is placed relatively far from the compact object and the jet apex , @xmath191 and @xmath192 cm ( models a and e , respectively ) . in both cases the value of the magnetic field is similar , @xmath193 g , although @xmath145 in model a and @xmath194 in model e. the slope of the x - ray spectrum was difficult to reproduce , even assuming a hard particle injection with spectral index @xmath147 . nevertheless , a harder particle injection could be achieved in principle through diffusive acceleration at relativistic shocks , see for example the works of stecker , baring & summerling ( 2007 ) and keshet & waxman ( 2005 ) . in any case , the high x - ray luminosities require a powerful jet with a large leptonic content . in fact , all the fits yield @xmath195 , what means that as much energy is given to the primary relativistic electrons as it is allowed by the constraints imposed . the hadronic contribution to the spectrum in cases a and e is undetectable with the present gamma - ray instruments . in the other models , synchrotron radiation of protons and secondary muons and pions , and at higher energies the contribution of @xmath95 interactions , could be detectable by fermi and hess ( and in the future by cta ) , respectively . for the low - luminosity observations ( models g and h ) , the best fits were obtained for low values of the acceleration efficiency and minimum particle energy . the required jet luminosity is now only @xmath196 . the predicted emission above @xmath183 mev is too faint to be detected with the present gamma - ray telescopes . we have also calculated fits to simultaneous radio , nir / optical and x - ray observations from 1999 and 2002 ( models i and j ) . for these sets of parameters , the break in the synchrotron spectrum occurs approximately in the nir , and the lowest - energy data was reasonably fit . the rising shape of the spectrum at optical wavelengths , however , could not be reproduced . this component is likely to originate mostly outside the jet , probably in the accretion disc . in all models , the spectrum is essentially of leptonic origin . in this sense , the results do not differ from those of previous works like those of markoff et al . ( 2003 , 2005 ) . our model , however , besides making predictions for the emission in the very high - energy regime , introduces some refinements over the previous scenarios adopted for this source . the particle distributions are calculated self - consistently taken into account the effect of energy losses on the injection spectrum . we also calculate the radiation emitted by secondary particles produced in hadronic interactions , and that of the electron - positron pairs product of photon - photon annihilation . the importance of photon self - absorption is assessed as well , although it turns out not to be relevant since the emission region is in a zone of low radiation density . we have also shown that the pair injection rate is significant enough , if this kind of model is solid in general for low - mass microquasars , to account for the observed line emission at 511 kev , according to the lower limit given by weidenspointner et al . if the proposed association between hard low - mass x - ray binaries and the electron - positron annihilation line flux can be proved , other explanations such as annihilation of dark matter could result unnecessary . detailed models for dealing with this topic are in preparation . g.s.v . thanks nicols casco for his help with fitting algorithms and matas reynoso for useful discussion . we thank an anonymous referee for very useful comments . g.e.r . thanks sera markoff for insightful remarks and suggestions . this research was supported by anpcyt through grant pict-2007 - 00848 bid 1728/oc - ar and by the ministerio de educacin y ciencia ( spain ) under grant aya 2007 - 68034-c03 - 01 , feder funds . aharonian f.a . , 2004 , very high energy cosmic gamma radiation . world scientific publishing , singapore atoyan a.m. , dermer , c.d . , 2003 , apj , 586 , 79 blandford r.d . , begelman m.c . , 1999 , mnras , 303 , l1-l5 blumenthal g.r . , gould r.j . , 1970 , rev . phys . , 42 , 237 bogovalov s.v . , kelner s.r . , 2005 , astronomy reports , 49 , 57 bosch - ramon v. , romero g.e . , paredes j.m . , 2006 , a&a , 447 , 263 bttcher m. , schlickeiser r. , 1997 , a&a , 325 , 866 buxton m. , vennes s. , 2003 , mnras , 342 , 105 callanan p.j . , charles p.a . , honey w.b . , thorstensen j.r . , 1992 , mnras , 259 , 395 chodorowski m.j . , zdziarski a.a . , sikora m. , 1992 , apj . 400 , 181 corbel s. , fender r.p . , tzioumis a.k . , nowak m. , mcintyre v. , durouchoux p. , sood r. , 2000 , a&a , 359 , 251 corbel s. , fender r.p . , 2002 , apj , 573 , l35 corbel s. , nowak m.a . , fender r.p . , tzioumis a.k . , markoff s. , 2003 , a&a , 400 , 1007 coriat m. , corbel s. , buxton m. m. , bailyn c. d. , tomsick j. a. , krding e. , kalemci e. , 2009 , mnras , 400 , 123 drury l. , 1983 , rep . , 46 , 973 dunn r.j.h . , fender r.p . , krding e.g. , cabanac c. , belloni t. , 2008 , mnras , 387 , 545 falcke h. , biermann p. l. , 1995 , a&a , 293 , 665 fender r.p . , belloni t.m . , gallo e. , 2004 , mnras , 355 , 1105 fragile p.c . , meier d.l . , 2009 , apj , 693 , 771 gaisser t. k. , 1990 , cosmic rays and particle physics . cambridge university press , cambridge gallo e. , corbel s. , fender r.p . , maccarone t.j . , tzioumis a.k . , 2004 , mnras , 347 , l52 gould r.j . , schrder g.p . , 1967 , , 155 , 1404 guessoum n. , jean p. , prantzos n. , 2006 , a&a , 2006 , 457 , 753 hannikainen d.c . , hunstead r.w . , campbell - wilson d. , sood r.k . , 1998 , a&a , 337 , 460 heinz s. , 2008 , int . j. mod . d , 17 , 1947 heinz s. , sunyaev r.a , 2003 , mnras , 343 , l59 homan j. , buxton m. , markoff s. , bailin c.d . , nespoli e. , belloni t. , 2005 , apj , 624 , 295 hynes r.i . , steeghs d. , casares j. , charles p.a . , obrien k. , 2003 , apjl , 583 , l95 hynes r.i . , steeghs d. , casares j. , charles p.a . , obrien k. , 2004 , apj , 609 , 317 imamura j.n . , kristian j. , middleditch j. , steiman - cameron t.y . , 1990 , apj , 365 , 312 kay s.m . , 1993 , fundamentals of statistical signal processing , volume 1 : estimation theory . prentice hall signal processing series , upper saddle river , new jersey kelner s.r . , aharonian f.a . , bugayov v.v . , 2006 , d , 74 , 034018 kelner s.r . , aharonian f.a . , 2008 , phys . d , 78 , 034013 keshet u. , waxman e. , 2005 , phys . lett . , 94 , 111102 khangulyan d. , aharonian f.a . , bosch - ramon v. , 2008 , mnras , 383 , 467 khangulyan d. , hnatic s. , aharonian f.a . , bogovalov s. , 2007 , mnras , 380 , 320 komissarov s. , barkov m. , vlahakis n. , knigl a. , 2007 , mnras , 380 , 51 krolik j.h . , 1999 , active galactic nuclei . princeton university press , princeton levine a.m. , corbet r. , 2006 , atel , 940 lipari p. , lusignoli m. , meloni d , 2007 , phys . d , 75 , 123005 mannheim k. , schlickeiser r. , 1994 , a&a , 286 , 983 markert t.h . , canizares c.r . , clark g.w . , 1973 , apj , 184 , l67 markoff s. , nowak m.a . , corbel s. , fender r. , falcke h. , 2003 , a&a , 397 , 645 markoff s. , nowak m.a . , wilms j. , 2005 , apj , 635 , 1203 maximon l.c . , 1968 , j. res . natl , 72b , 79 mcclintock j.e . , remillard r.a . 2006 , in : compact stellar x - ray sources , walter lewin & michiel van der klis ( eds ) . cambridge astrophysics series , no . 39 . cambridge university press , cambridge , p. 157 mcke a. , engel r. , rachen j.p . , protheroe r.j . , stanev t. , 2000 , comm . phys . comp . , 124 , 290 muoz - darias t. , casares j. , martnez - pais i.g . , 2008 , mnras , 385 , 2205 narayan r. , yi i. , 1995 , apj , 444 , 231 nowak m.a . , wilms j. , dove j.b . , 2002 , mnras , 332 , 856 nowak m.a . , wilms j. , heinz s. , pooley g. , pottschmidt k. , corbel s. , 2005 , apj , 626 , 1006 orellana m. , bordas p. , bosch - ramon v. , romero , g. e. , paredes j. m. , 2007 , a&a , 476 , 9 reynoso m.m . , romero g.e . , 2009 , a&a , 493 , 1 romero g.e . , vila g.s , 2008 , a&a , 485 , 623 romero g.e . , 2009 , a&a , 494 , l33 russell d. m. , fender r. p. , hynes r. i. , brocksopp c. , homan j. jonker p. g. , buxton m. m. , 2006 , mnras , 371 , 1334 shakura n.i . , sunyaev r. , 1973 , a&a , 24 , 337 schlickeiser r. , 2002 , cosmic ray astrophysics . astronomy and astrophysics library series , springer , berlin stecker f.w . , baring m.g . summerling e.j . , 2007 , apj , 667 , l29-l32 weidenspointner g. et al . , 2008 , nature , 451 , 159 wilms j. , nowak m.a . , dove j.b . , fender r.p . , di matteo t. , 1999 , apj , 522 , 460 yu w. , lamb f.k . , fender r. , van der klis m. , 2007 , apj , 663 , 1309 zdziarski a.a . , poutanen j. , mikolajewska j. , gierliski m. , ebisawa k. , johnson w.n . , 1998 , mnras , 301 , 435 zdziarski a.a . , gierliski m. , mikolajewska j. , wardziski g. , smith d. , alan harmon b. , kitamoto s. , 2004 , mnras , 351 , 791

continuous state differential equations are a popular modeling choice for the chemical concentrations of biochemical reaction networks in several disciplines , including industrial chemistry and systems biology . however , differential equations should only be used to model chemical concentrations when the counts of the reactant species are high @xcite . when the multiplicity of the individual species is low , as is often the case in enzymatic and genetic systems , it is important to use a model with a discrete state space which tracks individual molecular counts . predictions pertaining to the long - term behavior of a particular system can change dramatically depending upon whether the system is modeled with differential equations or with a discrete state space . in particular , discrete - space models may exhibit an extinction event where none exists in the corresponding continuous state model . for example , consider the following chemical reaction network : ( c1 ) @xmath0 ; ( c2 ) @xmath1 ; ( c3 ) @xmath2 ; ( c1 ) edge node [ above left = -0.15 cm ] @xmath3 ( c2 ) ; ( c2 ) edge node [ below right = -0.15 cm ] @xmath4 ( c1 ) ; ( c2 ) edge node [ above = -0.15 cm ] @xmath5 ( c3 ) ; where the labels correspond to the enumeration of the reactions . the deterministic mass action model predicts an asymptotically stable steady state for a wide range of parameter values . however , for the discrete - space model with stochastic mass - action kinetics and @xmath6 , the state @xmath7 is the inevitable absorbing state regardless of parameter values . this extinction event can be achieved by reaction 3 occurring until the count of species @xmath8 is zero , at which point no further reactions may occur . several frameworks exist for tracking trajectories of discrete state chemical reaction systems , including those of continuous time markov chains @xcite and stochastic petri nets @xcite . in these settings , the admissible transitions between states are assumed to occur randomly at a known rate and the occurrence of each reaction instantaneously updates the system according to the stoichiometry of the associated reaction . analysis of such systems is typically conducted by generating sample trajectories ( through a stochastic simulation algorithm , e.g. _ gillespie s algorithm _ @xcite or the next reaction method @xcite ) , by analyzing the evolution of the probability distribution via kolmogorov s forward equations ( i.e. the _ chemical master equation _ ) , by characterizing the stationary distributions of the models @xcite , or by studying the stochastic equations for the model @xcite . the study of extinction events in discrete interaction models is well - established in population dynamics and epidemic modeling , but the corresponding study in systems biology has only recently gained widespread attention . in @xcite anderson _ et al . _ described a large class of systems for which an extinction event necessarily occurs in the discrete model . interestingly , this class of models had previously been shown to have a particular `` robustness '' when modeled with deterministic ordinary differential equations @xcite . in @xcite r. brijder utilized tools from petri net theory to further extend the scope of networks known to have extinction behavior , by relating a kernel condition introduced in @xcite to the @xmath9-invariants of the corresponding petri net . related recent work analyzing transient and post - extinction behavior in discrete chemical reaction systems can be found in @xcite . in this paper , we further develop a network - based approach to determining when discrete - space chemical reaction systems may exhibit an extinction event . our main results , theorem [ mainresult ] and corollary [ maincorollary ] , state that a chemical reaction network with a discrete state space exhibits an extinction event if there is a modified network , called the domination - expanded reaction network , on which a particular set of inequalities on the edges _ can not _ be satisfied . the conditions we present may be summarized as systems of equalities and inequalities and , like corollary 2 of @xcite , suggests computational implementation . such an implementation will be explored in further depth in a follow - up paper @xcite . we demonstrate the effectiveness of corollary [ maincorollary ] on several models , including a model of the envz - ompr signaling pathway in _ escherichia coli _ @xcite . the notation of the paper is drawn from _ chemical reaction network theory _ which has proven effective for relating topological properties of a network s reaction graph to its admissible qualitative dynamical behaviors @xcite . the notions introduced here may be equivalently defined in the context of petri nets , which we summarize in appendix [ appendixd ] @xcite . we also adopt the following common notation throughout the paper : * @xmath10 and @xmath11 , * for @xmath12 , we define @xmath13 , * for a set @xmath14 of indexed elements and a subset @xmath15 , we define @xmath16 , * for a subset @xmath15 , we define the complement @xmath17 , * for @xmath18 , we define @xmath19 if @xmath20 for each @xmath21 . we outline the background notation and terminology relevant to the study of _ chemical reaction network theory _ ( crnt ) . ( for further background , see martin feinberg s online lecture notes @xcite . ) the fundamental object of interest in crnt is the following . [ crn ] a * chemical reaction network * ( crn ) is given by a triple of finite sets @xmath22 where : 1 . the * species set * @xmath23 contains the species of the crn . 2 . the * reaction set * @xmath24 consists of ordered pairs @xmath25 where @xmath26 and where the values @xmath27 are the * stoichiometric coefficients*. we will also write reactions @xmath28 as @xmath29 . the * complex set * @xmath30 consists of the linear combinations of the species in . specifically , @xmath31 . the number of distinct complexes is denoted @xmath32 . + allowing for a slight abuse of notation , we will let @xmath33 denote both the complex itself and the complex vector @xmath34 . we assume an arbitrary but fixed ordering of the species , reactions and complexes . it is common to impose that a crn does not contain any self - loops ( i.e. reactions of the form @xmath35 ) . since this assumption is not used in our results , and since it is common to allow self - loops in petri net theory , we will not make this assumption here . the interpretation of reactions as directed edges naturally gives rise to a reaction graph @xmath36 where the set of vertices is given by the complexes ( i.e. @xmath37 ) and the set of edges is given by the reactions ( i.e. @xmath38 . the following terminology will be used . a. a complex @xmath33 is _ connected _ to a complex @xmath39 if there exists a sequence of complexes @xmath40 such that either @xmath41 or @xmath42 for all @xmath43 . b. there is a _ path _ from @xmath33 to @xmath39 if there is a sequence of distinct complexes such that @xmath44 . c. a maximal set of mutually connected complexes is called a _ linkage class _ ( lc ) while a maximal set of mutually path - connected complexes is called a _ strong linkage class _ ( slc ) . the set of linkage classes will be denoted @xmath45 while the set of slcs will be denoted @xmath46 . d. an slc @xmath47 is called _ terminal _ if there are no outgoing reactions , i.e. @xmath48 and @xmath49 implies @xmath50 . the set of terminal slcs will be denoted @xmath51 . a complex @xmath52 is called terminal if it belongs to a terminal slc , and a reaction @xmath53 is terminal if @xmath33 is terminal . e. a set @xmath54 is called an _ absorbing complex set _ if it contains every terminal complex and has no outgoing edges , i.e. @xmath55 and @xmath53 implies @xmath56 . a complex @xmath55 is called @xmath57-interior , and a reaction @xmath53 is called @xmath57-interior if @xmath33 is @xmath57-interior ; otherwise they are @xmath57-exterior . absorbing complex sets are a generalization of the set of terminal complexes of a crn , since they must contain , but may be strictly larger than , this set . note that the set of terminal complexes is a closed complex set of the crn , as is the set @xmath58 . we will be particularly interested in the case where @xmath57 is the set of terminal complexes , as this provides the foundation upon which our main results are built . to each reaction @xmath53 we associate a _ reaction vector _ @xmath59 which tracks the net gain and loss of each chemical species as a result of the occurrence of this reaction . the _ stoichiometric subspace _ is defined by @xmath60 the _ stoichiometric matrix _ @xmath61 is the matrix with the reaction vectors as columns . a crn is said to be _ conservative _ ( respectively , _ subconservative _ ) if there exists a @xmath62 such that @xmath63 ( respectively , @xmath64 ) . the vector @xmath65 is called a _ conservation vector_. conservative crns have the property that a particular linear combination of all species remains constant as a result of each reaction , while subconservative crns have a combination of species which is nonincreasing in every reaction . a common example is conservation in the overall amount of enzyme or substrates in a closed enzymatic system , but there need not be such a physical interpretation . we present three examples in order to illustrate definitions . [ example234 ] consider the following crn : ( c1 ) @xmath1 ; ( c2 ) @xmath2 ; ( ghost ) @xmath66 ; ( c3 ) @xmath67 ; ( c4 ) @xmath8 ; ( c1 ) edge node [ above left = -0.15 cm ] @xmath3 ( c2 ) ; ( c2 ) edge node [ below right = -0.15 cm ] @xmath4 ( c1 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath5 ( c4 ) ; this crn has the sets @xmath68 , @xmath69 , and @xmath70 . the linkage classes are @xmath71 while the slcs are @xmath72 note that slcs may consist of singletons . the terminal slcs are @xmath73 the stoichiometric matrix is as follows : @xmath74.\ ] ] the stoichiometric subspace is given by @xmath75 , and there is the conservation vector @xmath76 . this conservation vector represents the fact that @xmath77 is constant . [ example235 ] consider the following crn : ( c1 ) @xmath8 ; ( c2 ) @xmath2 ; ( c3 ) @xmath67 ; ( c4 ) @xmath0 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath3 ( c2 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath4 ( c4 ) ; the set of terminal complexes is @xmath78 . there are several additional choices for absorbing complex sets , including @xmath79 and @xmath80 . the stoichiometric matrix is as follows : @xmath81.\ ] ] the stoichiometric subspace is given by @xmath82 . there is no vector @xmath83 for which @xmath84 , so the crn is not conservative or subconservative . [ example236 ] consider the following crn : ( c1 ) @xmath1 ; ( c2 ) @xmath8 ; ( c3 ) @xmath67 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath3 ( c2 ) ; ( c2 ) edge node [ above left = -0.15 cm ] @xmath4 ( c3 ) ; ( c3 ) edge node [ below right = -0.15 cm ] @xmath5 ( c2 ) ; the stoichiometric matrix is as follows : @xmath85.\ ] ] there is no vector @xmath83 such that @xmath63 , so the crn is not conservative ; however , the vector @xmath76 has the property that @xmath86 so that the crn is subconservative . a discrete _ state _ @xmath87 is an element of @xmath88 and denotes the molecular counts of each species . we let @xmath89 denote the state where @xmath90 corresponds to the count of species @xmath91 at time @xmath92 . these discrete states evolve as follows : @xmath93 where @xmath94 and , for all @xmath95 , @xmath96 is the number of times the @xmath97th reaction has occurred up to time @xmath92 . there are several established frameworks for modeling the time - evolution of crns on discrete state spaces , including that of continuous time markov chains ( ctmcs ) and stochastic petri nets . we will not concern ourselves with precise dynamical details ; rather , we will focus on where trajectories may evolve in @xmath88 . for a similar treatment , see the paper of l. paulev _ _ @xcite . we will say that a complex @xmath98 is _ charged _ at state @xmath99 if @xmath100 for all @xmath101 . we will then say that reaction @xmath53 is _ charged _ at state @xmath99 if the `` source complex '' @xmath33 is charged at @xmath87 . note that a reaction is therefore charged at a state @xmath87 if the species counts are sufficient for the source complex of that reaction . we will be primarily interested in how trajectories @xmath102 move through the state space @xmath88 of subconservative crns . in particular , we will be interested in the long - term behavior . we therefore introduce the following terminology , which is adapted from the conventions of stochastic processes . [ recurrence ] consider a crn on a discrete state space . then a state @xmath103 * reacts to * a state @xmath104 ( denoted @xmath105 ) if there is a reaction @xmath53 such that @xmath106 and @xmath33 is charged at state @xmath87 . a state @xmath104 is * reachable * from a state @xmath103 ( denoted @xmath107 if there exists a sequence of states such that @xmath108 . a state @xmath103 is * recurrent * if , for any @xmath104 , @xmath109 implies @xmath110 ; otherwise , the state is * transient*. note that the state space of a subconservative crn is finite ( see theorem 1 , @xcite ) . for this classification of crns , therefore , the notion of recurrence introduced above therefore agrees with the notion of positive recurrence from the language of ctmc ( see @xcite ) . we now extend the properties of recurrence and transience of states to the complexes and reactions of a crn . further considerations on the recurrence properties of the slcs of a crn are contained in appendix [ appendixc ] . [ turnedon ] consider a crn on a discrete state space . then a complex @xmath52 is * recurrent * from state @xmath103 if @xmath109 implies that there is a @xmath111 for which @xmath112 and @xmath33 is charged at @xmath111 ; otherwise , @xmath33 is * transient * from @xmath87 . a reaction @xmath53 is * recurrent * from state @xmath103 if the source complex @xmath33 is recurrent from @xmath87 ; otherwise , @xmath53 is * transient * from @xmath87 . in plain english , a complex @xmath33 is recurrent from a state @xmath87 if , whenever the process can go from the state @xmath87 to the state @xmath113 , then the process can move from the state @xmath113 to some state @xmath111 where @xmath33 is charged . the following clarifies the type of behavior for crns on discrete state spaces in which we will be interested . consider a crn on a discrete state space . we will say that the crn exhibits : 1 . an * extinction event on @xmath54 from @xmath114 * if every complex @xmath55 is transient from @xmath87 . 2 . a * guaranteed extinction event on @xmath54 * if it has an extinction event on @xmath57 from every @xmath114 . consider the crn in example [ example234 ] . through repeated application of reaction @xmath5 , we can arrive at the state @xmath115 where @xmath116 . since this is a possible outcome from any initial @xmath117 , we have that this crn has a guaranteed extinction event on @xmath118 . notice that no reaction may occur after the extinction event . consider the crn in example [ example236 ] . notice that the reaction @xmath119 can not occur indefinitely since all other reactions in the crn preserve @xmath120 . it follows that the model has a guaranteed extinction event on @xmath121 . notice , however , that so long as @xmath122 the reactions @xmath123 and @xmath124 are both recurrent . an extinction event therefore does not necessarily imply that all reactions must cease . in this section , we motivate and present the main new constructions and theory of the paper ( theorem [ mainresult ] and corollary [ maincorollary ] ) . we introduce the following . [ complexdomination ] let @xmath125 denote two distinct complexes of a crn . we say that @xmath33 * dominates * @xmath39 if @xmath126 . we define the * domination set * of a crn to be @xmath127 the notion of complex domination was introduced by d. anderson _ et al . _ in @xcite as an adaptation of the notion of `` differing in one species '' introduced by g. shinar and m. feinberg in @xcite . the domination property was extended to slcs by r. brijder in @xcite where it was also shown that , for conservative crns , the domination properties give rise to a binary relation on the slcs of a crn whose transitive closure is a partial ordering on the slcs of the crn ( lemma 2 , @xcite ) . we consider further properties of transience and recurrence of slcs in appendix [ appendixc ] . we note that the definition of complex domination in definition [ complexdomination ] is consistent with @xcite but reversed from @xcite . [ example919 ] consider the crns from examples [ example234 ] , [ example235 ] , and [ example236 ] respectively . for the crn in example [ example234 ] , we set @xmath128 , @xmath129 , @xmath130 , and @xmath131 and have @xmath132 , @xmath133 , and @xmath134 . for the crn in example [ example235 ] , we set @xmath135 , @xmath129 , @xmath130 , and @xmath136 , and have @xmath137 and @xmath133 . for the crn in example [ example236 ] , we set @xmath128 , @xmath138 , and @xmath130 , and have @xmath139 and @xmath132 . the key construction of this paper is the following , which uses the domination relations @xmath140 to expand crns into larger crns we call _ domination - expanded reaction networks_. [ dominationnetwork ] we say that @xmath141 is a * domination - expanded reaction network * ( dom - crn ) of the crn @xmath22 if @xmath142 . furthermore , we say a dom - crn is * @xmath57-admissible * if , given an absorbing complex set @xmath54 of the dom - crn , we have : ( i ) @xmath143 , and ( ii ) @xmath144 implies @xmath145 . a dom - crn consists of the original crn with additional directed edges corresponding to some ( potentially all ) of the domination relations @xmath126 . note that the reaction arrows flow from the dominating complex to the `` smaller '' complex in the domination relation , i.e. @xmath126 implies we add @xmath29 . consequently , like reactions , we will denote domination relations as either @xmath28 or @xmath29 . a dom - crn is admissible if we do not add any reactions which lead to the absorbing complex set @xmath57 of the dom - crn . when applying definition [ dominationnetwork ] , we will commonly let the absorbing complex set @xmath57 coincide with the set of terminal complexes of the dom - crn . in such cases , we will say a dom - crn is simply _ admissible _ with the understanding that @xmath57 is the set of terminal complexes . note that a dom - crn is a crn in itself and therefore has associated to it all of the quantities and structural matrices given section [ crntsection ] . while a dom - crn in general may have different structural properties than the original crn , an important restriction is given by the following result , which is based on lemma 2 of @xcite . the proof is contained in appendix [ appendixa ] . [ terminallemma ] if a crn is subconservative , then for any dom - crn : ( i ) the slcs of the crn and the dom - crn coincide , and ( ii ) every terminal slc of the dom - crn is a terminal slc of the crn . we can interpret lemma [ terminallemma ] as saying that , for a subconservative crn , the addition of domination edges does not create new cycles between slcs since this would create new slcs . [ example920 ] consider the crn from examples [ example234 ] and [ example919 ] . recall that the crn is conservative , and therefore subconservative , so that lemma [ terminallemma ] applies . the maximal dom - crn is given by the following : ( c1 ) @xmath1 ; ( c2 ) @xmath2 ; ( c3 ) @xmath67 ; ( c4 ) @xmath8 ; ( c1 ) edge node [ above left = -0.15 cm ] @xmath3 ( c2 ) ; ( c2 ) edge node [ below right = -0.15 cm ] @xmath4 ( c1 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath5 ( c4 ) ; ( c1 ) edge [ right ] node @xmath146 ( c3 ) ; ( c1 ) edge [ left ] node @xmath147 ( c4 ) ; ( c2 ) edge [ right ] node @xmath148 ( c3 ) ; where we have indexed the domination relations for clarity . as guaranteed by lemma [ terminallemma ] , the slcs of the crn and dom - crn coincide . notice that the terminal complex @xmath8 in the dom - crn above is terminal in the original crn , but that the terminal complexes @xmath77 and @xmath2 in the crn are not terminal in the dom - crn . notice also that this dom - crn is not admissible since the domination relations @xmath149 leads to the terminal complex @xmath8 . consider instead the subset @xmath150 which corresponds to the following dom - crn : ( c1 ) @xmath1 ; ( c2 ) @xmath2 ; ( c3 ) @xmath67 ; ( c4 ) @xmath8 ; ( c1 ) edge node [ above left = -0.15 cm ] @xmath3 ( c2 ) ; ( c2 ) edge node [ below right = -0.15 cm ] @xmath4 ( c1 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath5 ( c4 ) ; ( c1 ) edge [ right ] node @xmath146 ( c3 ) ; ( c2 ) edge [ right ] node @xmath147 ( c3 ) ; this dom - crn is admissible since @xmath151 contains no domination edges which lead to the terminal complex @xmath8 . consider the crn from example [ example235 ] and [ example919 ] . recall that the crn is neither conservative nor subconservative . thus , lemma [ terminallemma ] stands silent . the maximal dom - crn is given by the following : ( c1 ) @xmath8 ; ( c2 ) @xmath2 ; ( c3 ) @xmath67 ; ( c4 ) @xmath0 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath3 ( c2 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath4 ( c4 ) ; ( c2 ) edge [ right ] node @xmath152 ( c3 ) ; ( c4 ) edge [ left ] node @xmath152 ( c1 ) ; we have that there is only one slc in the dom - crn , which is given by @xmath153 , so that the slcs of the crn and dom - crn do not coincide . we can see , therefore , that lemma [ terminallemma ] does not hold in general if we remove the subconservative assumption . the following concept is adapted from numerous sources in graph theory . trees have been used extensively in crnt @xcite and the related notion of arborescences factored in @xcite . [ forest1 ] consider a crn @xmath22 and a @xmath57-admissible dom - crn @xmath141 where @xmath54 is an absorbing complex set on the dom - crn . then @xmath154 where @xmath155 and @xmath156 is called an * @xmath57-exterior forest * if , for every complex @xmath157 , there is a unique path in @xmath158 from @xmath33 to @xmath57 . a @xmath57-exterior forest is a forest in the usual sense in graph theory after restricting to the @xmath57-exterior portion of the reaction graph of the dom - crn . note that definition [ forest1 ] places no restrictions on @xmath57-interior reactions . by convention , we will include such reactions in every @xmath57-exterior forest . if @xmath57 consists solely of the terminal complexes of the dom - crn , we say @xmath154 is simply an _ exterior forest_. we will be interested in particular in @xmath57-exterior forests which satisfy the following property . [ balancing ] consider a crn @xmath22 and a @xmath57-admissible dom - crn @xmath141 where @xmath54 is an absorbing complex set on the dom - crn . let @xmath159 . then a @xmath57-exterior forest @xmath154 is said to be * balanced * if there is a vector @xmath160 with @xmath161 for at least one @xmath57-exterior reaction which satisfies : 1 . supp@xmath162 supp@xmath163 and supp@xmath164 supp@xmath165 ; 2 . @xmath166 ; and 3 . for every @xmath167 where @xmath157 , we have @xmath168 where @xmath169 . otherwise , the @xmath57-exterior forest is said to be * unbalanced*. the third condition of definition [ balancing ] can be interpreted as saying that , for every @xmath157 , the weight of the outgoing edge in the @xmath57-exterior forest must be at least as large as the sum of all incoming edges . when taken together , the three conditions of definition [ balancing ] generate a set of equalities and inequalities on the edges of the dom - crn . this suggests a computational implementation , which is investigated in the companion paper @xcite . [ example921 ] recall the crn taken from examples [ example234 ] , [ example919 ] , and [ example920 ] and the admissible dom - crn from example [ example920 ] . this dom - crn admits several exterior forests , for example the following substructures in bold red : ( c11 ) @xmath1 ; ( c21 ) @xmath2 ; ( c31 ) @xmath67 ; ( c41 ) @xmath8 ; ( c12 ) @xmath1 ; ( c22 ) @xmath2 ; ( c32 ) @xmath67 ; ( c42 ) @xmath8 ; ( c11 ) edge node [ above left = -0.15 cm ] @xmath170 ( c21 ) ; ( c21 ) edge node [ below right = -0.15 cm ] @xmath4 ( c11 ) ; ( c31 ) edge node [ above = -0.15 cm ] @xmath171 ( c41 ) ; ( c11 ) edge [ right ] node @xmath146 ( c31 ) ; ( c21 ) edge [ right ] node @xmath172 ( c31 ) ; ( c12 ) edge node [ above left = -0.15 cm ] @xmath3 ( c22 ) ; ( c22 ) edge node [ below right = -0.15 cm ] @xmath173 ( c12 ) ; ( c32 ) edge node [ above = -0.15 cm ] @xmath171 ( c42 ) ; ( c12 ) edge [ right ] node @xmath174 ( c32 ) ; ( c22 ) edge [ right ] node @xmath147 ( c32 ) ; note that every nonterminal complex has a unique path to @xmath8 . we now check whether these exterior forests are balanced by definition [ balancing ] by checking equalities and inequalities on the vector of edges of the following form : @xmath175 note also that the stoichiometric matrix is given by @xmath176.\ ] ] 1 . in order for the left exterior forest to be balanced , it is required that we find a vector @xmath177 , @xmath178 , satisfying : @xmath179 we can choose @xmath180 so that this is balanced exterior forest . in order for the right exterior forest to be balanced , it is required that we find a nontrivial vector @xmath177 , @xmath178 , satisfying : @xmath181 substituting condition 1 into condition 2 gives @xmath182 which is inconsistent with the requirement from condition 3 that @xmath183 and at least one entry be nonzero . it follows that this is an unbalanced exterior forest . we now present the main results of this paper , which are inspired by theorem 1 and corollary 2 of @xcite . the proof of theorem [ mainresult ] is contained in appendix [ appendixb ] . [ mainresult ] consider a subconservative crn and a @xmath57-admissible dom - crn where @xmath54 is an absorbing complex set on the dom - crn . suppose that there is a complex @xmath157 of the dom - crn which is recurrent from a state @xmath184 in the discrete state space crn . then every @xmath57-exterior forest of the dom - crn is balanced . this result places restrictions on the structure of a subconservative crn that does not experience a guaranteed extinction event . we will be more frequently interested in when discrete extinction occurs , and therefore present the following corollary which follows immediately as the contrapositive of theorem [ mainresult ] . [ maincorollary ] consider a subconservative crn and a @xmath57-admissible dom - crn where @xmath54 is an absorbing complex set on the dom - crn . suppose there is a @xmath57-exterior forest of the dom - crn which is unbalanced . then the discrete state space crn has a guaranteed extinction event on @xmath185 . recall that an exterior forest is unbalanced if there is a set of equalities and inequalities on the edges of the dom - crn which can not be satisfied . the question of determining sufficient conditions for discrete extinction is therefore reduced to determining the feasibility of particular sets of equalities and inequalities . notice also that , even if a crn permits many @xmath57-exterior forests , it is sufficient for a single one to be unbalanced for an extinction event to follow . furthermore , the set of transient complexes corresponds to the set of complexes not in @xmath57 . note that this may contain terminal complexes in the original crn ( see example [ example920 ] ) . by convention , when applying corollary [ maincorollary ] , if no mention of an absorbing complex set @xmath54 is made , it is assumed to be the set of terminal complexes in the dom - crn . reconsider the crn analyzed in example [ example234 ] , [ example919 ] , and [ example920 ] . this crn is conservative , and in example [ example921 ] we showed that there is an admissible dom - crn with an unbalanced exterior forest . it follows from corollary [ maincorollary ] that the discrete state space crn has a guaranteed extinction event on the set of nonterminal complexes of the dom - crn . that is , from all states @xmath103 , there is guaranteed to be a time after which the count of the species is insufficient for any reaction from the complexes @xmath77 , @xmath2 , and @xmath67 to occur . this is consistent with our earlier observation that the state @xmath115 where @xmath116 absorbs all trajectories through repeated application of the reactions @xmath186 and @xmath124 . notice that this pathway consists of the true reactions in the unbalanced exterior forest . in this section , we consider a crn which was proposed as underlying the envz / ompr signaling pathway in _ escherichia coli _ in @xcite . this crn has been studied previously with a discrete state space in the papers @xcite where it was shown to exhibit a guarantee extinction event . we reconsider the crn here to demonstrate the process of applying corollary [ maincorollary ] and also to demonstrate the advantages of our approach . in particular , the graphical method of constructing @xmath57-exterior forests suggests the pathways to extinction in the crn . [ example1 ] consider the following reaction mechanism , which was proposed by g. shinar and m. feinberg as underlying the envz / ompr signaling pathway in _ escherichia coli _ in the supplemental material of @xcite : ( x1 ) @xmath8 ; ( x2 ) @xmath67 ; ( x3 ) @xmath187 ; ( x4 ) @xmath188 ; ( x4x5 ) @xmath189 ; ( x6 ) @xmath190 ; ( x2x7 ) @xmath191 ; ( x3x7 ) @xmath192 ; ( x8 ) @xmath193 ; ( x3x5 ) @xmath194 ; ( x1x7 ) @xmath195 ; ( x9 ) @xmath196 ; ( x1x5 ) @xmath197 ; ( x1 ) edge node [ above left = -0.15 cm ] @xmath3 ( x2 ) ; ( x2 ) edge node [ below right = -0.15 cm ] @xmath4 ( x1 ) ; ( x2 ) edge node [ above left = -0.15 cm ] @xmath5 ( x3 ) ; ( x3 ) edge node [ below right = -0.15 cm ] @xmath198 ( x2 ) ; ( x3 ) edge node [ above = -0.15 cm ] @xmath199 ( x4 ) ; ( x4x5 ) edge node [ above left = -0.15 cm ] @xmath200 ( x6 ) ; ( x6 ) edge node [ below right = -0.15 cm ] @xmath201 ( x4x5 ) ; ( x6 ) edge node [ above = -0.15 cm ] @xmath202 ( x2x7 ) ; ( x3x7 ) edge node [ above left = -0.15 cm ] @xmath203 ( x8 ) ; ( x8 ) edge node [ below right = -0.15 cm ] @xmath204 ( x3x7 ) ; ( x8 ) edge node [ above = -0.15 cm ] @xmath205 ( x3x5 ) ; ( x1x7 ) edge node [ above left = -0.15 cm ] @xmath206 ( x9 ) ; ( x9 ) edge node [ below right = -0.15 cm ] @xmath207 ( x1x7 ) ; ( x9 ) edge node [ above = -0.15 cm ] @xmath208 ( x1x5 ) ; where @xmath209 , @xmath210 , @xmath211 , @xmath212 , @xmath213 , @xmath214 , @xmath215 , @xmath216 , @xmath217 . consider the admissible dom - crn with @xmath218 . the dom - crn may be graphically represented as : ( x1x7 ) @xmath195 ; ( x9 ) @xmath196 ; ( x1x5 ) @xmath197 ; ( x1 ) @xmath8 ; ( x2 ) @xmath67 ; ( x3 ) @xmath187 ; ( x4 ) @xmath188 ; ( x3x7 ) @xmath192 ; ( x8 ) @xmath193 ; ( x3x5 ) @xmath194 ; ( x4x5 ) @xmath189 ; ( x6 ) @xmath190 ; ( x2x7 ) @xmath191 ; ( x1 ) edge node [ above left = -0.15 cm ] @xmath3 ( x2 ) ; ( x2 ) edge node [ below right = -0.15 cm ] @xmath4 ( x1 ) ; ( x2 ) edge node [ above left = -0.15 cm ] @xmath5 ( x3 ) ; ( x3 ) edge node [ below right = -0.15 cm ] @xmath198 ( x2 ) ; ( x3 ) edge node [ above = -0.15 cm ] @xmath199 ( x4 ) ; ( x4x5 ) edge node [ below right = -0.15 cm ] @xmath200 ( x6 ) ; ( x6 ) edge node [ above left = -0.15 cm ] @xmath201 ( x4x5 ) ; ( x6 ) edge node [ above = -0.15 cm ] @xmath202 ( x2x7 ) ; ( x3x7 ) edge node [ above left = -0.15 cm ] @xmath203 ( x8 ) ; ( x8 ) edge node [ below right = -0.15 cm ] @xmath204 ( x3x7 ) ; ( x8 ) edge node [ above = -0.15 cm ] @xmath205 ( x3x5 ) ; ( x1x7 ) edge node [ right = -0.1 cm ] @xmath206 ( x9 ) ; ( x9 ) edge node [ left = -0.1 cm ] @xmath207 ( x1x7 ) ; ( x9 ) edge node [ right = -0.1 cm ] @xmath208 ( x1x5 ) ; ( x1x7 ) edge [ above right = -0.15 cm ] node @xmath147 ( x1 ) ; ( x1x5 ) edge [ below right = -0.15 cm ] node @xmath146 ( x1 ) ; ( x3x7 ) edge [ left ] node @xmath219 ( x3 ) ; ( x3x5 ) edge [ right ] node @xmath220 ( x3 ) ; ( x2x7 ) edge [ right = -0.15 cm ] node @xmath148 ( x2 ) ; consider furthermore the following exterior forest : ( x1x7 ) @xmath195 ; ( x9 ) @xmath196 ; ( x1x5 ) @xmath197 ; ( x1 ) @xmath8 ; ( x2 ) @xmath67 ; ( x3 ) @xmath187 ; ( x4 ) @xmath188 ; ( x3x7 ) @xmath192 ; ( x8 ) @xmath193 ; ( x3x5 ) @xmath194 ; ( x4x5 ) @xmath189 ; ( x6 ) @xmath190 ; ( x2x7 ) @xmath191 ; ( x1 ) edge node [ above left = -0.15 cm ] @xmath170 ( x2 ) ; ( x2 ) edge node [ below right = -0.15 cm ] @xmath4 ( x1 ) ; ( x2 ) edge node [ above left = -0.15 cm ] @xmath171 ( x3 ) ; ( x3 ) edge node [ below right = -0.15 cm ] @xmath198 ( x2 ) ; ( x3 ) edge node [ above = -0.15 cm ] @xmath221 ( x4 ) ; ( x4x5 ) edge node [ below right = -0.15 cm ] @xmath222 ( x6 ) ; ( x6 ) edge node [ above left = -0.15 cm ] @xmath201 ( x4x5 ) ; ( x6 ) edge node [ above = -0.15 cm ] @xmath223 ( x2x7 ) ; ( x3x7 ) edge node [ above left = -0.15 cm ] @xmath203 ( x8 ) ; ( x8 ) edge node [ below right = -0.15 cm ] @xmath224 ( x3x7 ) ; ( x8 ) edge node [ above = -0.15 cm ] @xmath205 ( x3x5 ) ; ( x1x7 ) edge node [ right = -0.1 cm ] @xmath206 ( x9 ) ; ( x9 ) edge node [ left = -0.1 cm ] @xmath225 ( x1x7 ) ; ( x9 ) edge node [ right = -0.1 cm ] @xmath208 ( x1x5 ) ; ( x1x7 ) edge [ above right ] node @xmath172 ( x1 ) ; ( x1x5 ) edge [ below right ] node @xmath226 ( x1 ) ; ( x3x7 ) edge [ left ] node @xmath227 ( x3 ) ; ( x3x5 ) edge [ right ] node @xmath228 ( x3 ) ; ( x2x7 ) edge [ right = -0.15 cm ] node @xmath229 ( x2 ) ; in the highlighted structure ( bold red ) , there is a unique path from every complex to the terminal complex @xmath188 . it can be seen directly that this exterior forest is unbalanced by noting that we need a vector @xmath230 , @xmath178 , which has support on a subset of the red highlighted structure above . to satisfy condition 2 of definition [ balancing ] , we need to satisfy @xmath166 . we can check that @xmath231 has the generators : @xmath232 \ker(\gamma ) \cap \mathbb{r}_{\geq 0}^r = \ { & ( 1 , & 1 , & 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & \;0 , & \ ; 0 , & \ ; 0 , & \;0 , & \ ; 0),\\ & ( 0 , & 0 , & 1 , & 1 , & 0 , & 0 , & 0 , & 0 , & 0 , & \;0 , & \ ; 0 , & \ ; 0 , & \;0 , & \ ; 0),\\ & ( 0 , & 0 , & 0 , & 0 , & 0 , & 1 , & 1 , & 0 , & 0 , & \;0 , & \ ; 0 , & \ ; 0 , & \;0 , & \ ; 0),\\ & ( 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & 1 , & \;1 , & \ ; 0 , & \ ; 0 , & \;0 , & \ ; 0),\\ & ( 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & 0 , & \;0 , & \ ; 0 , & \ ; 1 , & \;1 , & \ ; 0),\\ & ( 0 , & 0 , & 1 , & 0 , & 1 , & 1 , & 0 , & 1 , & 1 , & \;0 , & \ ; 1 , & \ ; 0 , & \;0 , & \ ; 0),\\ & ( 0 , & 0 , & 1 , & 0 , & 1 , & 1 , & 0 , & 1 , & 0 , & \;0 , & \ ; 0 , & \ ; 1 , & \;0 , & \ ; 1 ) \ ; \ ; \;\ } \end{array}\ ] ] the first five vectors correspond to reversible reaction pairs in the crn and so may be ignored . in order to obtain a nontrivial vector @xmath233 , we require @xmath234 . to build such a vector using the sixth vector yields a vector with support on @xmath235 while building it out of the seventh vector yields a vector with support on @xmath236 . neither of these options is consistent with condition 1 of definition [ balancing ] so that the exterior forest is unbalanced . it follows by corollary [ maincorollary ] that the discrete state space crn has a guaranteed extinction event , and that every complex except @xmath188 is transient . in fact , all trajectories are absorbed by a state where @xmath237 , @xmath238 , and @xmath239 for @xmath240 . this result was previously obtained in @xcite and also proved for a simplified crn in @xcite . the construction of the dom - crn , and computational implementation , is unique . this method also suggests a pathway toward extinction through the reactions in the unbalanced exterior forest . such a pathway was not apparent by the methods of either @xcite or @xcite . in this section , we provide further examples which demonstrate how to apply corollary [ maincorollary ] , and also demonstrate the necessity of several of the technical conditions required of the result . example [ example000 ] presents a crn which can be shown to have an extinction event for an absorbing complex set @xmath54 which is not the set of terminal complexes in the dom - crn . example [ example001 ] presents a crn which does not have a guaranteed extinction event , but which can be shown to have an unbalanced exterior forest if we do not insist on the underlying dom - crn being admissible . example [ example999 ] demonstrates that including condition 3 of definition [ balancing ] allows further classification of crns with extinction events than would be possible otherwise . examples [ example100 ] and [ example101 ] provide crns which show that the conditions of corollary [ maincorollary ] are sufficient , but not necessary , for a guaranteed extinction event to occur . [ example000 ] it is natural to wonder whether , when applying corollary [ maincorollary ] , there is an advantage to generalizing the set of terminal complexes to an absorbing complex set @xmath54 . to show that there is , consider the following crn : ( c1 ) @xmath0 ; ( c2 ) @xmath241 ; ( c3 ) @xmath242 ; ( c4 ) @xmath2 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath3 ( c2 ) ; ( c2 ) edge node [ above = -0.15 cm ] @xmath4 ( c3 ) ; ( c3 ) edge node [ above left = -0.15 cm ] @xmath5 ( c4 ) ; ( c4 ) edge node [ below right = -0.15 cm ] @xmath198 ( c3 ) ; there are no domination relations so that the only dom - crn corresponds to the crn shown , and it is trivially admissible . the only exterior forest consists of all reactions . notable , it contains reactions 1 and 2 on the nonterminal component . we can easily determine that @xmath243 satisfies the conditions of definition [ balancing ] and therefore that this exterior forest is balanced . therefore , corollary [ maincorollary ] does not apply and we may not conclude that an extinction event occurs . consider instead taking @xmath244 . this set is absorbing and contains every terminal complex of the crn . the only exterior forest again contains all reactions but only reaction 1 is @xmath57-exterior . since there is no balancing vector @xmath245 for which @xmath246 , we may conclude by corollary [ maincorollary ] that there is a guaranteed extinction event on @xmath247 . in fact , we can see this directly since repeated application of reaction 1 will deplete @xmath8 and there are no pathways by which to replenish it . [ example001 ] it is natural to wonder whether it is necessary to insist on dom - crns being admissible . to show that removing this assumption from corollary [ maincorollary ] can lead to misclassification , consider the following crn : ( c1 ) @xmath0 ; ( c2 ) @xmath2 ; ( c3 ) @xmath67 ; ( c4 ) @xmath187 ; ( c1 ) edge node [ above left = -0.15 cm ] @xmath3 ( c2 ) ; ( c2 ) edge node [ below right = -0.15 cm ] @xmath4 ( c1 ) ; ( c3 ) edge node [ above left = -0.15 cm ] @xmath5 ( c4 ) ; ( c4 ) edge node [ below right = -0.15 cm ] @xmath198 ( c3 ) ; the crn has only the single domination relation @xmath248 . since the corresponding domination relation @xmath249 leads to a terminal component in any resulting dom - crn , we may not add it , so that the only admissible dom - crn corresponds to the original crn . suppose , however , that we do not insist on dom - crns being admissible . specifically , suppose we allow the following dom - crn : ( c1 ) @xmath0 ; ( c2 ) @xmath2 ; ( c3 ) @xmath67 ; ( c4 ) @xmath187 ; ( c1 ) edge node [ above left = -0.15 cm ] @xmath3 ( c2 ) ; ( c2 ) edge node [ below right = -0.15 cm ] @xmath4 ( c1 ) ; ( c3 ) edge node [ above left = -0.15 cm ] @xmath5 ( c4 ) ; ( c4 ) edge node [ below right = -0.15 cm ] @xmath198 ( c3 ) ; ( c2 ) edge node [ above = -0.15 cm ] @xmath152 ( c3 ) ; the only exterior forest is given in bold red as follows : ( c1 ) @xmath0 ; ( c2 ) @xmath2 ; ( c3 ) @xmath67 ; ( c4 ) @xmath187 ; ( c1 ) edge node [ above left = -0.15 cm ] @xmath170 ( c2 ) ; ( c2 ) edge node [ below right = -0.15 cm ] @xmath4 ( c1 ) ; ( c3 ) edge node [ above left = -0.15 cm ] @xmath171 ( c4 ) ; ( c4 ) edge node [ below right = -0.15 cm ] @xmath250 ( c3 ) ; ( c2 ) edge node [ above = -0.15 cm ] @xmath251 ( c3 ) ; notice that we have included the terminal reactions in the exterior forest . in order to be balanced , we must find a vector @xmath252 which is nonzero on at least one of the nonterminal reactions @xmath253 and @xmath254 , such that @xmath255 conditions 1 and 2 imply that @xmath256 so that @xmath245 is does not have support on the nonterminal portion of the dom - crn . it follows that the exterior forest is unbalanced . note , however , that corollary [ maincorollary ] remains silent since the presented dom - crn is not admissible . since all reactions of the discrete state space crn are recurrent whenever @xmath257 , this example highlights the importance of the assumption that dom - crns be admissible . [ example999 ] it is natural to wonder whether condition 3 of definition [ balancing ] is useful in classifying discrete state space crns with extinction events . to see that it can be , consider the following crn : ( c4 ) @xmath1 ; ( c5 ) @xmath0 ; ( c6 ) @xmath2 ; ( c4 ) edge node [ above left = -0.15 cm ] @xmath3 ( c5 ) ; ( c5 ) edge node [ below right = -0.15 cm ] @xmath4 ( c4 ) ; ( c5 ) edge node [ above = -0.15 cm ] @xmath5 ( c6 ) ; there are no domination relations so the dom - crn coincides with the original crn . we have only the following exterior forest in bold red : ( c1 ) @xmath1 ; ( c2 ) @xmath0 ; ( c3 ) @xmath258 ; ( c1 ) edge node [ above left = -0.15 cm ] @xmath170 ( c2 ) ; ( c2 ) edge node [ below right = -0.15 cm ] @xmath4 ( c1 ) ; ( c2 ) edge node [ above = -0.15 cm ] @xmath171 ( c3 ) ; in order for this exterior forest to be balanced , we need to have a vector @xmath259 , @xmath178 , which satisfies the following equalities and inequalities : @xmath260 condition 1 reduces condition 2 to @xmath261 , so that , combining with condition 3 , we have @xmath262 this can only be satisfied by @xmath256 and @xmath263 , which is a violation . it follows that the exterior forest is unbalanced and therefore , by corollary [ maincorollary ] , the discrete state space crn has a guaranteed extinction event on the nonterminal complexes @xmath77 and @xmath0 . note , however , that the vector @xmath264 satisfies conditions 1 and 2 of definition [ balancing ] . it follows that condition 3 of definition [ balancing ] allows furthermore classification of crns with extinction events than conditions 1 and 2 allow by themselves . note also that this crn is also not classified as having a guaranteed extinction event by corollary 2 of @xcite . [ example100 ] it is natural to wonder whether the conditions of theorem [ mainresult ] and corollary [ maincorollary ] are necessary as well as sufficient for a discrete state space crn to have an extinction event . to show that they are sufficient only , consider the following crn , which is simplified from the crn in eqn . ( 66 ) of @xcite and reproduced as eqn . ( 49 ) in the supplemental material for @xcite : ( c1 ) @xmath8 ; ( c2 ) @xmath67 ; ( c3 ) @xmath265 ; ( c4 ) @xmath266 ; ( c5 ) @xmath267 ; ( c6 ) @xmath268 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath3 ( c2 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath4 ( c4 ) ; ( c5 ) edge node [ above = -0.15 cm ] @xmath5 ( c6 ) ; ( c6 ) edge node [ below = -0.15 cm ] @xmath198 ( c5 ) ; the crn has a guaranteed discrete extinction event , since @xmath187 may convert into @xmath8 through reaction 3 , then @xmath8 may convert into @xmath67 through reaction 1 . this shuts down all reactions . to show that corollary [ maincorollary ] is incapable of affirming this extinction event , it is necessary to show that every @xmath57-exterior forest of every @xmath57-admissible dom - crn is unbalanced . we start by considering the terminal complexes and the set @xmath269 . this gives the following dom - crn : ( c3 ) @xmath265 ; ( c4 ) @xmath266 ; ( c5 ) @xmath267 ; ( c6 ) @xmath268 ; ( ghost ) ; ( c1 ) @xmath8 ; ( c2 ) @xmath67 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath3 ( c2 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath4 ( c4 ) ; ( c5 ) edge node [ above = -0.15 cm ] @xmath5 ( c6 ) ; ( c6 ) edge node [ below = -0.15 cm ] @xmath198 ( c5 ) ; ( c4 ) edge node [ above = -0.15 cm ] @xmath146 ( c1 ) ; ( c6 ) edge node [ below = -0.15 cm ] @xmath147 ( c1 ) ; this dom - crn is admissible and admits only a single exterior forest in bold red : ( c3 ) @xmath265 ; ( c4 ) @xmath266 ; ( c5 ) @xmath267 ; ( c6 ) @xmath268 ; ( ghost ) ; ( c1 ) @xmath8 ; ( c2 ) @xmath67 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath170 ( c2 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath173 ( c4 ) ; ( c5 ) edge node [ above = -0.15 cm ] @xmath171 ( c6 ) ; ( c6 ) edge node [ below = -0.15 cm ] @xmath198 ( c5 ) ; ( c4 ) edge node [ above = -0.15 cm ] @xmath226 ( c1 ) ; ( c6 ) edge node [ below = -0.15 cm ] @xmath172 ( c1 ) ; this forest is balanced if we have a nontrivial vector @xmath270 + @xmath271 , @xmath272 , which satisfies the following : @xmath273 this can be satisfied by the vector @xmath274 . it follows that the forest is balanced , and since this is the only exterior forest for the given dom - crn , no conclusion may be reached as a result of corollary [ maincorollary ] . we now consider more general absorbing complex sets @xmath54 . notice that any potential @xmath57 which contains a subset of @xmath275 can be balanced by the @xmath245 above , with perhaps different support on @xmath276 . if @xmath277 , however , we must have @xmath278 in order for the dom - crn to be @xmath57-admissible . otherwise , we would have an edge in @xmath151 which would lead to @xmath57 . for @xmath278 , however , we have that @xmath279 and @xmath280 are terminal in the dom - crn and therefore @xmath267 and @xmath268 must be included in @xmath57 . this leaves @xmath58 which has an empty exterior forest . there are no other cases to consider , so we are done . it follows that every @xmath57-exterior forest of every @xmath57-admissible dom - crn is balanced . since the crn has a guaranteed extinction event , however , it follows that the conditions of corollary [ maincorollary ] are not necessary for extinction events in discrete state space crns . [ example101 ] to show that the gap raised in example [ example100 ] may not be easily overcome by structural considerations alone , consider the following crn : ( c1 ) @xmath8 ; ( c2 ) @xmath67 ; ( c3 ) @xmath281 ; ( c4 ) @xmath280 ; ( c5 ) @xmath194 ; ( c6 ) @xmath197 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath3 ( c2 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath4 ( c4 ) ; ( c5 ) edge node [ above = -0.15 cm ] @xmath5 ( c6 ) ; ( c6 ) edge node [ below = -0.15 cm ] @xmath198 ( c5 ) ; this is the crn in example [ example100 ] with @xmath187 replaced with @xmath188 in the reaction 2 , and @xmath188 replaced with @xmath282 in reactions 3 and 4 . examples [ example100 ] and [ example101 ] share significant structural data , including connectivity of paths , domination relations between complexes , and ker@xmath283 . taking @xmath284 gives the following admissible dom - crn : ( c3 ) @xmath281 ; ( c4 ) @xmath280 ; ( c5 ) @xmath194 ; ( c6 ) @xmath197 ; ( ghost ) ; ( c1 ) @xmath8 ; ( c2 ) @xmath67 ; ( c1 ) edge node [ above = -0.15 cm ] @xmath3 ( c2 ) ; ( c3 ) edge node [ above = -0.15 cm ] @xmath4 ( c4 ) ; ( c5 ) edge node [ above = -0.15 cm ] @xmath5 ( c6 ) ; ( c6 ) edge node [ below = -0.15 cm ] @xmath198 ( c5 ) ; ( c4 ) edge node [ above = -0.15 cm ] @xmath146 ( c1 ) ; ( c6 ) edge node [ below = -0.15 cm ] @xmath147 ( c1 ) ; we arrive at the same balancing equalities and inequalities ( [ conditions ] ) as example [ example100 ] , so that every @xmath57-exterior forest on this dom - crn is balanced . since the connectivity and domination relations are shared with example [ example100 ] , we can exhaust nontrivial @xmath57-admissible dom - crns in the same way , and we conclude that corollary [ maincorollary ] is inconclusive . in contrast to example [ example100 ] , this example does not exhibit an extinction event for most initial conditions . provided @xmath237 , @xmath285 , and any one of @xmath286 , @xmath287 , and @xmath288 is positive , every complex is recurrent . this analysis suggests that comprehensive conditions for extinction events must depend on further structural information than that considered in this paper . in this paper , we have presented novel conditions ( theorem [ mainresult ] and corollary [ maincorollary ] ) on the structure of a crn that are sufficient to guarantee that the corresponding crn exhibits an extinction event . the conditions presented generalize the dependence on terminal slcs in @xcite and @xcite , and also produces a system of equalities and inequalities which can be directly verified . our conditions have the additional advantage of being fundamentally graphical in nature and suggesting pathways to extinction . this work raises several promising avenues for future work : 1 . while corollary [ maincorollary ] gives sufficient conditions for discrete extinction , they are not necessary ( see examples [ example100 ] and [ example101 ] ) . this raises the question of whether there are structural conditions which are both sufficient and necessary for discrete extinction and , if so , which further structural components of the crn might be utilized in such a result . the conditions of corollary [ maincorollary ] consist of a system of equalities and inequalities . this suggests a computational implementation amenable , in particular , to the methods of linear programming . linear programming has already been used widely in crnt for verifying crns with desirable structural properties @xcite . this will be explored and utilized to characterize crns with extinction events in the companion paper @xcite . mdj and dfa were supported by army research office grant w911nf-14 - 1 - 0401 . dfa was also supported by nsf - dms-1318832 and mdj was also supported by the henry woodward fund . gc was supported by nsf - dms-1412643 . rb is a postdoctoral fellow of the research foundation flanders ( fwo ) . 10 david f. anderson . a modified next reaction method for simulating chemical systems with time dependent propensities and delays . , 127(21):214107 , 2007 . david f. anderson , daniele cappelletti , and thomas g. kurtz . finite time distributions of stochastically modeled chemical systems with absolute concentration robustness . available on the arxiv at arxiv:1604.03388 , 2016 . david f. anderson , gheorghe craciun , and thomas g. kurtz . product - form stationary distributions for deficiency zero chemical reaction networks . , 72(8):19471970 , 2011 . david f. anderson , german enciso , and matthew d. johnston . stochastic analysis of chemical reaction networks with absolute concentration robustness . , 11(93):20130943 , 2014 . david f. anderson and thomas g. kurtz . continuous time markov chain models for chemical reaction networks . in h. koeppl et al . , editor , _ design and analysis of biomolecular circuits : engineering approaches to systems and synthetic biology _ , pages 342 . springer , 2011 . david f. anderson and thomas g. kurtz . . springer , 2015 . falko bause and pieter s. kritzinger . . vieweg verlag 2 . aufl . , 2002 . balzs boros . on the dependence of the existence of the positive steady states on the rate coefficients for deficiency - one mass action systems : single linkage class . , 51(9):24552490 , 2013 . robert brijder . dominant and t - invariants for petri nets and chemical reaction networks . , 9211:115 , 2015 . gheorghe craciun , alicia dickenstein , anne shiu , and bernd sturmfels . toric dynamical systems . , 44(11):15511565 , 2009 . german a. enciso . transient absolute robustness in stochastic biochemical networks . , 13(121):20160475 , 2016 . martin feinberg . lectures on chemical reaction networks . unpublished written versions of lectures given at the mathematics research center , university of wisconsin . martin feinberg . complex balancing in general kinetic systems . , 49:187194 , 1972 . martin feinberg . chemical reaction network structure and the stability of complex isothermal reactors : i. the deficiency zero and deficiency one theorems . , 42(10):22292268 , 1987 . martin feinberg . chemical reaction network structure and the stability of complex isothermal reactors : ii . multiple steady states for networks of deficiency one . , 43(1):125 , 1988 . michael a. gibson and jehoshua bruck . efficient exact stochastic simulation of chemical systems with many species and many channels . , 104:18761889 , 2000 . daniel gillespie . a general method for numerically simulating the stochastic time evolution of coupled chemical reactions . , 22(4):403434 , 1976 . fritz horn . necessary and sufficient conditions for complex balancing in chemical kinetics . , 49:172186 , 1972 . fritz horn and roy jackson . general mass action kinetics . , 47:81116 , 1972 . matthew d. johnston . translated chemical reaction networks . , 76(5):10811116 , 2014 . matthew d. johnston . a linear programming approach to dynamical equivalence , linear conjugacy , and the deficiency one theorem . , 54(8):16121631 , 2016 . matthew d. johnston . a computational approach to extinction events in chemical reaction networks with discrete state spaces . available on the arxiv at arxiv:1701.02014 , 2017 . matthew d. johnston , casian pantea , and pete donnell . a computational approach to persistence , permanence , and endotacticity of chemical reaction networks . , 72(1):467498 , 2016 . matthew d. johnston , david siegel , and gbor szederknyi . a linear programming approach to weak reversibility and linear conjugacy of chemical reaction networks . , 50(1):274288 , 2012 . tom g. kurtz . the relationship between stochastic and deterministic models for chemical reactions . , 57:29762978 , 1972 . gregory f. lawler . chapman & hall , 2006 . grard memmi and grard roucairol . linear algebra in net theory . in wilfried brauer , editor , _ net theory and applications _ , volume 84 of _ lecture notes in comput . _ , pages 213223 . springer , 1975 . jost neigenfind , sergio grimbs , and zoran nikoloski . biochemical network decomposition reveals absolute concentration robustness . available on the arxiv at arxiv:1105.0624v1 . james l. peterson . petri nets . , 9(3):223252 , 1977 . wolfgang reisig and grzegorz rozenberg , editors . , volume 1491 of _ lecture notes in comput . _ springer , 1998 . loc paulev , gheorghe craciun , and keinz koeppl . dynamical properties of discrete reaction networks . , 69(1):5572 , 2014 . guy shinar and martin feinberg . structural sources of robustness in biochemical reaction networks . , 327(5971):13891391 , 2010 . gabor szederknyi . computing sparse and dense realizations of reaction kinetic systems . , 47:551568 , 2010 . * lemma [ terminallemma ] . * _ if a crn is subconservative , then for any dom - crn : ( i ) the slcs of the crn and the dom - crn coincide , and ( ii ) every terminal slc of the dom - crn is a terminal slc of the crn . _ _ proof of ( i ) : _ consider a subconservative crn and dom - crn . since the reactions of the crn are contained in the reactions of the dom - crn , it follows that the slcs of crn remain strongly connected in the dom - crn and therefore are contained in the slcs of the dom - crn . now suppose that there is an slc of the dom - crn which is not contained in any slc of the crn . it follows that there are slcs @xmath289 of the crn such that is a path in the dom - crn from some complex @xmath290 to some complex @xmath291 , and there is a path in the dom - crn from some complex @xmath292 to some complex @xmath293 . since @xmath294 and @xmath295 are strongly connected , we can create a cycle in the dom - crn by constructing a path from @xmath296 to @xmath297 to @xmath298 to @xmath299 back to @xmath296 . furthermore , since this is not a cycle in the crn ( otherwise , @xmath294 and @xmath295 would not be maximally strongly connected in the crn ) , we have that there is at least one reaction in this cycle which is from @xmath151 . we now index the complexes in the cycle so that , if there are @xmath300 reactions from @xmath151 in the cycle , we have the following segments in between these reactions : @xmath301 we take the segments above to be connected by reactions in @xmath302 . we also take @xmath303 and @xmath304 so that the endpoints of successive segments are joined together by @xmath305 . let @xmath306 denote the vector of counts of the reactions in ( [ cycle ] ) , and take @xmath307 . it follows that @xmath308 by the domination relations @xmath303 . since the crn is subconservative , it follows that there is a @xmath62 such that @xmath309 . it follows that we have @xmath310 \alpha = \mathbf{c}^t [ \gamma \alpha ] > 0\ ] ] where the last strict inequality follows from @xmath311 for @xmath312 and the observation that at least one component in ( [ 10 ] ) must be strictly greater than zero since the complexes of the crn are stoichiometrically distinct . this is a contradiction . it follows that such a cycle does not exist in the dom - crn so that @xmath294 and @xmath295 are slcs of the crn . the slcs of the crn and dom - crn therefore coincide and ( i ) is shown . + _ proof of ( ii ) : _ note that ( i ) guarantees that the crn and dom - crn share the same set of slcs which we will denote @xmath46 . suppose that @xmath47 is terminal in the dom - crn but not in the crn . this implies that there is a reaction @xmath25 where @xmath313 and @xmath314 ; however , this reaction is included in the dom - crn so that @xmath294 may not be terminal in the dom - crn . it follows that every terminal slc of the dom - crn is a terminal slc of the crn , and ( ii ) is shown . * theorem [ mainresult ] . * _ consider a subconservative crn and a @xmath57-admissible dom - crn where @xmath54 is an absorbing complex set on the dom - crn . suppose that there is a complex @xmath157 of the dom - crn which is recurrent from a state @xmath184 in the discrete state space crn . then every @xmath57-exterior forest of the dom - crn is balanced . _ the following proof is inspired by the proof of theorem 1 in @xcite . the notation has been adapted to that of crnt . consider a subconservative crn and a @xmath57-admissible dom - crn where @xmath54 is an absorbing complex set on the dom - crn . suppose that there is a complex @xmath157 of the dom - crn which is recurrent from a state @xmath184 . we will show that every @xmath57-exterior forest is balanced ; that is , every @xmath57-exterior forest admits a vector @xmath315 satisfying the requirements of definition [ balancing ] . we will accomplish this by constructing a sequence of reactions which may be executed indefinitively , and then demonstrating that this sequence repeats . we will define @xmath245 based on a specific repeating portion of this sequence and show that it is balanced . let @xmath103 denote our initial state . by assumption , there is a state @xmath316 and a complex @xmath317 such that ( i ) @xmath318 , ( ii ) @xmath319 is charged at @xmath320 , and ( iii ) no complex @xmath157 is charged at any state along the sequence of reactions from @xmath318 . that is , @xmath319 is the first @xmath57-exterior complex which becomes charged as a result of the reaction sequence . ( if @xmath319 is charged at @xmath87 then the sequence of reactions is empty . ) by construction , there is a unique path in the exterior forest from @xmath319 to @xmath57 . let @xmath321 denote the complex at the end of this path and @xmath322 denote the state obtained by the sequential occurrence of the true reactions in the path ( i.e. include reactions in @xmath323 but exclude domination relations @xmath324 ) . note that ( i ) @xmath325 , and ( ii ) @xmath326 is charged at @xmath327 . we now iterate this procedure for @xmath328 starting from the state @xmath329 rather than @xmath87 . this generates the following sequence of transitions , which may be continued indefinitely by the recurrence assumption : @xmath330 since the crn is subconservative , we have that there is a finite number of accessible states ( theorem 1 , @xcite ) . it follows that there is a state in @xmath331 which is repeated . we let @xmath332 and @xmath333 where @xmath334 denote the first and second indices for the set @xmath331 such that @xmath335 . this gives the following subsequence of ( [ stuff1 ] ) @xmath336 since @xmath335 , ( [ stuff2 ] ) defines a sequence of reactions which can be repeated indefinitely . we now define the vector @xmath160 in the following way : ( i ) @xmath233 consists of the counts of the reactions in the sequence of reactions in ( [ stuff2 ] ) , and ( ii ) @xmath276 consists of the counts of the domination relations in the paths taken to construct the reaction sequences in ( [ stuff2 ] ) . we now show that @xmath245 if balanced according to definition [ balancing ] . it is clear , first of all , that @xmath245 only has support on @xmath323 and @xmath324 so that condition @xmath3 is satisfied . in order to show that @xmath337 , we note from eqn . ( [ markov ] ) of the main text , and the definition of @xmath233 , that @xmath338 it follows that @xmath337 and therefore @xmath245 satisfies condition @xmath4 of definition [ balancing ] . to verify condition @xmath5 , we note that , since @xmath339 is always chosen to be the first complex exterior to @xmath57 which becomes charged , the only contribution to @xmath245 from the nonterminal component comes from the segments corresponding to @xmath340 , i.e. the paths from @xmath339 to @xmath57 . it follows that , at every complex exterior to @xmath57 , the count of the reaction out is at least as great as the sum of the reactions in , and @xmath245 therefore satisfies condition @xmath5 of definition [ balancing ] . the result is therefore shown . we define the following , which extends definition [ recurrence ] . consider a crn on a discrete state space . an slc @xmath47 is said to be * recurrent * from state @xmath103 if every @xmath313 is recurrent from @xmath87 ; otherwise , we will say @xmath294 is * transient * from @xmath87 . in other words , an slc is recurrent if every complex @xmath33 in the slc is recurrent . we now consider how the transient and recurrence of complexes may be distributed throughout a crn . [ lemma10 ] consider a crn on a discrete state space . then a complex @xmath313 is recurrent if and only if the slc @xmath47 is recurrent . let @xmath47 denote an slc of a crn . suppose @xmath313 is recurrent from @xmath103 and @xmath341 . from the recurrence of @xmath33 , it follows that , for every @xmath104 such that @xmath109 , there is a @xmath342 such that @xmath112 and @xmath33 is charged at @xmath111 . since @xmath33 and @xmath39 belong to the same linkage class , it follows that there is a path from @xmath33 to @xmath39 . it follows that there is a @xmath343 such that @xmath344 and @xmath39 is charged at @xmath345 . we have used the observation that reactions in a path in the reaction graph may occur in sequence since each reaction necessarily produces sufficient molecularity for the next reaction to proceed . we therefore have that @xmath346 so that , for every @xmath104 , such that @xmath109 , there is a @xmath343 such that @xmath347 and @xmath39 is charged at @xmath345 . it follows that @xmath39 is recurrent from @xmath87 , and we are done . this result shows that complex and slcs recurrent is equivalent in the sense that we may not have one without the other . we may further relate recurrence to the reaction graph of crn with the following . [ lemma1 ] consider a crn on a discrete state space and a dom - crn . suppose there is a path in the maximal dom - crn from a complex @xmath313 to a complex @xmath348 where @xmath349 are two slcs of the crn . then the following hold : 1 . if @xmath294 is recurrent from @xmath103 , then @xmath295 is recurrent from @xmath87 . if @xmath295 is transient from @xmath103 , then @xmath294 is transient from @xmath87 . the set of recurrent complexes is an absorbing complex set of the maximal dom - crn consisting of the union of slcs . let @xmath289 denote an slc of a crn and suppose @xmath294 is recurrent from @xmath103 . suppose there is a path in the dom - crn from a complex @xmath313 to a complex @xmath348 . it follows that , for every @xmath104 such that @xmath109 , there is a @xmath342 such that @xmath112 and @xmath33 is charged at @xmath111 . now consider the path in the dom - crn from @xmath33 to @xmath39 . the path may be composed of reactions in @xmath302 or @xmath151 . we have that , for any sequence of reactions in the path which are only from @xmath302 , if the path starts with a recurrent complex , recurrence is transferred to every complex in the path , including the final one . this can be realized by noting that the occurrence of each reaction necessarily confers sufficient molecularity for the next reaction in the path to take place . also notice that , for any reaction in @xmath151 , say @xmath350 we have @xmath351 for all @xmath352 . it follows that , if @xmath353 is charged at a state , then @xmath354 is charged at the state . combining these two results , we have that there is a @xmath343 such that @xmath344 and @xmath39 is charged at @xmath345 . it follows that @xmath39 is recurrent from @xmath87 . it follows that @xmath295 is recurrent from lemma [ lemma10 ] . this proves claim @xmath3 and , since claim @xmath4 is the contrapositive of claim @xmath3 , this is also shown . claim @xmath5 follows by noting that , if @xmath54 is the set of recurrent complexes but is not an absorbing complex set in the maximal dom - crn then there is a reaction @xmath355 such that @xmath356 and @xmath357 . it then follows from claim @xmath3 that @xmath358 is recurrent , which contradicts the construction of @xmath57 . since the set of recurrent complexes must consist of the union of slcs by lemma [ lemma10 ] , the result is shown . this result gives restrictions on the distribution of transient and recurrent complexes and slcs within a crn . recurrence travels with the direction of the paths in the dom - crn while transience travels against the direction of these paths . claim @xmath5 furthermore suggests that absorbing complex sets are the correct object of study when considering recurrence and transience in discrete state space crns . consider the following example . consider the crn structure contained in figure [ figure1 ] , where the boxes represent slcs . we write @xmath359 if there is a path from some @xmath360 to @xmath361 in the crn , and we write @xmath362 if @xmath363 for some @xmath360 and @xmath361 . in ( b ) and ( c ) , potential patterns for recurrent slcs consistent with lemma [ lemma1 ] are highlighted . if @xmath364 is recurrent , @xmath365 and @xmath366 must be recurrent as well ( red ) . if @xmath367 is recurrent , @xmath368 and @xmath369 must be recurrent as well ( green ) . note that in ( c ) this recurrence implication flows through the domination relationship . note also that lemma [ terminallemma ] guarantees that a subconservative crn may not contain any cycles in the representation figure [ figure1 ] . slcs . in ( b ) and ( c ) , two possible configurations of recurrent slcs ( red and green , respectively ) are presented which are consistent with lemma [ lemma1].,width=453 ] for subconservative crns , another way to interpret lemma [ lemma1 ] is by introducing the order operation @xmath370 on the slcs of a crn , where @xmath371 if there are @xmath313 and @xmath348 such that @xmath372 . in @xcite , r. brijder showed that the transition closure of this operator is a partial order on the slcs of a subconservative crn . we extend this slightly by defining the relation @xmath373 to be such that @xmath374 if either @xmath371 or there is a path from a complex @xmath348 to a complex @xmath313 in the crn , and then let @xmath375 denote the transitive closure of @xmath373 . since the relation @xmath375 corresponds to path - connectedness in the dom - crn , lemma [ terminallemma ] is equivalent to the property of @xmath375 being a partial order on the slcs of a subconservative crn . in this interpretation , the minimal slcs of the crn under the partial order @xmath375 correspond to the terminal slcs of the dom - crn . we may then interpret lemma [ lemma1 ] as stating that , for a subconservative crn with slcs @xmath349 such that @xmath376 , ( a ) if @xmath294 is recurrent from @xmath103 , then @xmath295 is recurrent from @xmath87 , and ( b ) if @xmath295 is transient from @xmath103 , then @xmath294 is transient from @xmath87 . that is , relative to the partial order @xmath375 on slcs of a subconservative crn , recurrence flows downward while transience flows upward . petri nets form a well - studied model of concurrent computation , see , e.g. , @xcite . petri nets are essentially . however , this difference is irrelevant for this paper . ] equivalent to crns on discrete state spaces . as a consequence , results concerning crns on discrete state spaces can be equivalently stated in terms of petri nets and vice versa . in a petri net ( without initial marking ) @xmath377 , species are called _ places _ ( i.e. , @xmath378 ) , reactions are called _ transitions _ ( i.e. , @xmath379 ) , and the stoichiometric coefficients of each reaction is encoded by a function @xmath380 . moreover , molecules are called _ tokens _ and states @xmath87 are called _ markings _ @xmath381 . furthermore , the stoichiometric matrix @xmath382 is known as the _ incidence matrix _ of a petri net , conservation vectors @xmath383 are known as _ @xmath384-invariants _ , and vectors @xmath385 such that @xmath386 are known as _ graphical depictions are different for petri nets compared to crns . for example , consider the following example . reconsider the crn from example [ example234 ] , which is represented graphically in figure [ fig : petri ] . in the setting of petri nets , species / places are denoted by circles and reactions / transitions by boxes . moreover , the reactants of a reaction / transition are the incoming edges of that transition ( including multiplicity as edge labels ) and the products of a reaction / transition are the outgoing edges of that transition ( including multiplicity as edge labels ) . the petri net has the incidence matrix : @xmath387 a @xmath384-invariant is given by the vector @xmath388 and @xmath9-invariants by the vectors @xmath389 and @xmath390 .

the intuitive meaning of _ heat _ and _ work _ in thermodynamics is that of two types of energetic resources , one fully controllable and useful , the other uncontrolled and wasteful . an impressive effort has been devoted to provide a consistent mathematical characterisation of these notions within a quantum mechanical description of physics @xcite this is a challenge since in contrast to other thermodynamic quantities , such as internal energy and entropy , heat and work are not properties of individual states of a system . they are defined for _ continuous processes _ connecting different states @xcite , implying that their statistical fluctuations can not be described in terms of a single system observable . two - point correlation functions characterising the correlations along process paths are required - a problematic territory for quantum mechanics where definite trajectories can not be fixed unless the system is continuously measured . resolving these issues has been the topic of a number of publications that have formulated quantum trajectory approaches @xcite . in contrast here we focus on the _ mean _ values of heat and work where the analysis simplifies but still requires careful thought . we will adopt the identification of the system s internal energy with @xmath0 $ ] where @xmath1 is the density matrix describing the state of the system at given time , and @xmath2 is its instantaneous hamiltonian . clearly , a proper definition of this hamiltonian is in general problematic ! if the system is coupled to an environment the non - equilibrium behaviour of a general open system makes the definition of the system s hamiltonian ambiguous @xcite both , mathematically and experimentally . ultimately the choice of the hamiltonian one assigns to the system must rely on the set of operations and observables one can access experimentally . in many situations of physical interest @xmath2 can be identified with the bare system hamiltonian or an effective system hamiltonian that incorporates the effect of the environment . while the environment degrees are in principle uncontrolled , full control can be exerted over the temporal variation " of the system hamiltonian . for instance , the size of a container in which steam is pumped can be freely chosen and a piston can be attached to the container that can push the wheels of a train . a formal definition of mean heat and work is then obtained by considering an infinitesimal change of the internal energy @xmath3 = { { \rm tr}}[\rho \ , { d}h ] + { { \rm tr } } [ { d}\rho \ , h]\ ; , \end{aligned}\ ] ] associated with the time evolution of the system which brings its density matrix from @xmath1 to @xmath4 while the hamiltonian varies from @xmath2 to @xmath5 . the origin for the change of @xmath1 may here be due to both , the variation of @xmath2 induced by the experimenter and by the dynamics due to the coupling with the environment . the possibility of externally controlling @xmath2 suggests to identify the first term on the r.h.s . of eq . ( [ firsteq ] ) with the average work done by the experimenter during the evolution . the second term describes the internal energy change due to a reconfiguration of the system , i.e. a variation of the system s density matrix . this is an energy contribution over which the experimenter has no direct control and this is why it is associated with heat . the infinitesimal average _ heat absorbed by the system _ and the infinitesimal average _ work done on the system _ @xcite are therefore defined as @xmath6 \quad \mbox{and } \quad { { \delta}}w : = { { \rm tr } } [ \rho \ , { d}h]\;,\end{aligned}\ ] ] with the symbol @xmath7 indicating that heat and work are in general no full differentials , i.e. they do not correspond to observables . while the first law of thermodynamics states that the sum of the two average energy types is the average internal energy , @xmath8 the _ split _ into these two types of energies is crucial as it allows the formulation of the second law of thermodynamics . a fundamental law of physics , it sets limits on the work extraction of heat engines and establishes the notion of _ irreversibility _ in physics . the second law can be phrased in form of clausius inequality : @xmath9 stating that the change in a system s entropy must be larger than the average heat absorbed by the system during a process . while the first law of thermodynamics is fundamental for any process , the second law was originally stated for processes that _ start and end in equilibrium_. recently , the non - equilibrium work relation due to jarzynski has been used to argue that the second law should also hold for any process starting from equilibrium , at temperature @xmath10 , but ending in an arbitrary non - equilibrium state @xcite . however , no conclusive argument has yet established the most general set of dynamical processes that obey the clausius inequality @xcite . extending the infinitesimal scenario to _ finite , continuous processes _ in which the temporal evolution of @xmath11 and @xmath12 in time @xmath13 is known , the mean heat and work can be found by integrating over the trajectory taken from @xmath14 and @xmath15 to @xmath16 and @xmath17 , i.e. @xmath18 \ ; , \label{eq : q } \\ w & : = & \int_{0}^{\tau } \ , { d}t \ , { { \rm tr } } [ \rho(t ) \ , \dot { h}(t)]\ ; , \label{eq : q+w}\end{aligned}\ ] ] while the first law becomes @xmath19 = q + w\;.\end{aligned}\ ] ] the mathematical consistency of the above expressions and their compatibility with the predictions of thermodynamics have been verified for many models , for example , for processes that are induced by markovian master equations @xcite . there are two paradigmatic examples of _ all work _ and _ all heat _ processes that we introduce here and which will become important in the later part of the manuscript . the first process is a unitary process , which we will also refer to as _ closed _ , where the ( non - equilibrium ) evolution of the state is given by the schrdinger equation , @xmath20\;. \label{inuni}\end{aligned}\ ] ] mean heat and work are then @xmath21 \ , h(t ) ] = 0\ ; , \\ w_{unitary } & = & \delta u\ ; , \label{example1}\end{aligned}\ ] ] consistent with the physical intuition that _ no _ heat has been provided to the system during the evolution . the second example is a system that evolves through the action of a dissipative , i.e. _ open _ , markov process via a master equation @xcite , @xmath22 + { \cal l}(\rho(t))\ ; , \label{ddf}\end{aligned}\ ] ] with @xmath23 being the dissipative lindblad term . under the assumption that the typical time scales associated with the time - independent @xmath2 are much shorter than those associated with @xmath23 we can treat the system as almost isolated and use eq.([firsteq ] ) to compute its internal energy . in this limit eq . ( [ eq : q+w ] ) is valid with the hamiltonian just being the time - independent @xmath2 , @xmath24 - { { \rm tr}}[\rho(0 ) \ , h ] = \delta u , \,\ , \label{identity111}\end{aligned}\ ] ] which is in full agreement with the physical intuition that no work has been performed on the system . while these examples constitute special cases of continuous processes the heat and work in a general process depend intimately on the exact details of the process . however , the caveat with this viewpoint is that in most real life applications one does _ not _ know what the dynamics of the state of the system is nor what the appropriate local hamiltonian is at all times . importantly , this is not just due to our ignorance of what happens at the quantum level . quantum physics has strong fundamental limitations on what we _ can know _ without choosing a measurement apparatus , measuring the system and interpreting the data . moreover , if the system is indeed measured then the experimenter s choice of what degrees of freedom she actually measures will effect what the _ measured _ heat and work will be . in other words , we propose that there is no _ one _ average heat and work for a particular process , there are different sensible outcomes to this question and the answer depends on the choice of system hamiltonians in time , @xmath12 , that corresponds to specific measurement choices . the aim of this paper is to show that it is possible to formulate a general second law independently of these choices . to achieve this we will depart entirely from the traditional continuous trajectory approach and propose a rather drastic but pragmatic change of perspective . we develop a consistent framework of mean heat and work for _ discrete thermodynamic processes_. the rationale for this approach is that while the true process is continuous , observations we make on the system are almost always discrete . ( we will neglect here the possibility of monitoring through continuous weak measurements . ) for discrete snapshots of the dynamics , we find that by decomposing the transition into possible sequences of two fundamental primitives , it is possible to define heat and work for the discrete process in a way that is experimentally and mathematically clear . this allows us to establish a general second law for discrete processes between equilibrium and non - equilibrium states and the analysis of a discrete carnot cycle , where we uncover the usual carnot efficiency . the paper is structured as follows . in sec . [ sec : entropy ] we review the traditional perspective on the second law and the definition of entropy . in sec . [ sec : configuration ] we define the dynamical configuration space of a system that allows us to formulate a notion of two primitives for discrete processes in sec . [ sec : duts+dtts ] , the discrete unitary and discrete thermalising transformations ( duts and dtts ) . [ sec:2ndlaw ] contains the main results of the paper . first we show that entropic inequalities when applied to discrete trajectories formed by concatenating duts and dtts yields the second law of thermodynamics in the clausius formulation . we then derive two consequences : we find the minimum and maximum heat for a single dut and dtt sequence and prove the existence of a discrete trajectory , formed by sequences of duts and dtts , that connects two given thermodynamical configurations while asymptotically saturating the clausius inequality . finally we identify a discrete trajectory that connects the same initial and final configurations as the continuous trajectory through a sequence of duts and dtts , and which approximates the continuous heat . in sec . [ sec : efficiency ] we derive the thermal efficiency of a discrete cycle , the carnot efficiency , and conclude in sec . [ sec : conc ] . in 1865 clausius established that the overall heat flow in any cyclic , reversible process vanishes , implying that the integral over any non - cyclic process must be path independent . this led him to define the state function _ entropy _ , @xmath25 , and the entropy change , @xmath26 , between the final and initial point of a reversible process , @xmath27 clausius also showed that any cyclic process , reversible or irreversible , obeys @xmath28 this relation is the basis for a formulation of the second law of thermodynamics , known as the _ clausius - inequality_. it is a statement for all thermodynamic processes , not just cyclic ones , that start from equilibrium at temperature @xmath10 , @xmath29 and it simplifies to @xmath30 when the system interacts with a bath at constant temperature , @xmath10 . in this form clausius inequality establishes the existence of an upper bound to the heat received by the system . clearly , clausius s goal was to characterise different forms of _ energy _ and their interconversion . however , by formulating the second law of thermodynamics he defined a new quantity : entropy . in contrast , in modern information theory the focus is on the _ state _ of a system . entropy is here used as the central physical quantity to measure the amount of information of a state , while heat and work , and energy in general , have no well - defined purpose for the interpretation of information processing . this opens the possibility of turning clausius original argument around ! it allows us to use the entropy change in discrete quantum processes to _ define _ the average heat , and work . before we proceed , let us first highlight that non - trivial entropy bounds exists for any process between two states . a state @xmath1 describes an amount of information , quantified by the _ von neumann entropy _ , @xmath31 $ ] . the evolution of a quantum system from an initial state @xmath32 to a final state @xmath33 through an arbitrary process , or _ quantum channel _ , has a meaningful associated entropy change , @xmath34 + { { \rm tr}}[\rho_i \ln \rho_i ] , \end{aligned}\ ] ] which quantifies the change of the encoded amount of information . the entropy change is non - trivially bounded from above and below by virtue of the positivity of the _ relative entropy _ ( classically kullback - leibler divergence @xcite ) . the relative entropy , @xmath35 , between two states @xmath36 and @xmath37 characterises the number of additional bits required to encode @xmath36 when using the diagonal basis of @xmath37 , rather than the diagonal basis of @xmath36 . it is defined as @xcite @xmath38 - { { \rm tr}}[\rho_1 \ln \rho_2]\;,\end{aligned}\ ] ] and is a positive quantity @xmath39 intuitively , the relative entropy is similar to a distance measure , however , it is important to keep in mind that it is asymmetric @xmath40 . rewriting the entropy change , eq . ( [ eq : entropychange ] ) , in two ways @xmath41 + s(\rho_i\| \rho_f ) \\ & = & - { { \rm tr } } [ \delta \rho \ , \ln \rho_i ] - s(\rho_f\| \rho_i ) , \label{eq : v - entropy}\end{aligned}\ ] ] a lower and upper bound on the entropy change emerge @xmath42 \geqslant \delta s(\rho_i,\rho_f ) \geqslant - { { \rm tr } } [ \delta \rho \ , \ln \rho_f ] .\end{aligned}\ ] ] from the information theory point of view , bounds on the entropy change are important in their own right as they characterise how much information is lost or gained . if we now assume the special case that @xmath33 is a thermal state for the hamiltonian @xmath43 at an inverse temperature @xmath44 then the lower bound becomes @xmath45.\ ] ] interpreting @xmath46 $ ] as the heat of the discrete process , the above expression would constitute exactly the second law of thermodynamics . this is exactly what we will pursue in sec . [ sec : duts+dtts ] , e.g. in eq . ( [ dttheat ] ) . from ( [ eq : entropybounds ] ) it is apparent that also an upper bound on the entropy change exists that is rarely discussed in the literature . this maximum value of the entropy change is enforced to ensure that any reverse process , from @xmath33 to @xmath32 , also obeys the second law . \geqslant \delta s(\rho_i,\rho_f ) $ ] becomes @xmath47,$ ] with @xmath48 being the state increment for the reverse process . ] to assist our discussion of discrete quantum processes we introduce the concept of configuration space , following the spirit of @xcite , and propose a graphical representation for that space , see fig . [ fig : config ] . let @xmath25 be the quantum system under investigation , with @xmath49 its hilbert space , @xmath50 the set of linear operators on @xmath49 , and @xmath51 the set of density matrices on @xmath49 . we define the _ dynamical configuration space _ @xmath52 of @xmath25 as the set formed by the pairs @xmath53 with @xmath54 a density matrix = 1 $ ] . ] and @xmath55 a hermitian operator on @xmath49 whose spectrum is bounded from below . points in the dynamical configuration space @xmath56 are called `` configurations '' to distinguish them from `` states '' , @xmath1 . the evolution of the system is described by discrete trajectories in @xmath52 : a _ discrete trajectory _ @xmath57 is defined as an ordered list of elements of @xmath52 that describes the succession of configurations , with each element @xmath58 containing both the density matrix @xmath1 of @xmath25 and the local hamiltonian @xmath2 of @xmath25 at that specific instance of the evolution . we stress that both @xmath1 and @xmath2 of a configuration point @xmath59 have a clear experimental meaning . @xmath1 is the density matrix that one would reconstruct by state tomography , i.e. the preparation of many copies of the same state @xmath1 and the full tomographic measurement of its properties . the hamiltonian of the system , @xmath2 , is determined by the set of projective measurements @xmath60 the experimenter performs on the system to `` measure the energy '' together with the interpretation of the corresponding energy eigenvalues , @xmath61 , so that @xmath62 . ( the choice of the measurement and interpretation can be motivated by a process tomography on the hamiltonian at any point in time . for this the system needs to be decoupled from the rest of the universe at that instance and evolve for a complete set of states for a short time interval @xmath63 through the action of @xmath2 . by measuring the final states of the evolution the unitary @xmath64 and hence @xmath2 can be uncovered . ) it is then straightforward to establish the internal energy and the entropy for each point in dynamical configuration space . for each configuration @xmath65 in @xmath52 we define the internal energy as @xmath66,\end{aligned}\ ] ] and the entropy as the von neumann entropy @xmath67 of @xmath1 , @xmath68.\end{aligned}\ ] ] a central notion in thermodynamics is the canonical gibbs state , often also referred to as _ thermal state _ or _ equilibrium _ state . since a thermal state , @xmath1 , at temperature @xmath10 is well - defined only with respect to a certain hamiltonian , @xmath2 , it is actually the _ configuration _ @xmath69 that is thermal . an element @xmath70 describes a _ thermal equilibrium configuration _ ( or briefly _ thermal configuration _ ) if @xmath1 is a gibbs state of the hamiltonian , @xmath2 , for some finite inverse temperature @xmath71 , i.e. @xmath72 with @xmath73 $ ] being the associated partition function . to be finite . in the zero temperature limit the associated density matrix @xmath1 of the thermal configuration @xmath74 approaches the projector on the ground state of @xmath2 . the limiting case does however not belong to the set of the thermal configurations but to their closure . this definition ensures that the density matrices @xmath1 of the thermal configurations @xmath75 are always full rank and hence strictly positive , i.e. the state has no zero eigenvalue . ] in the following the thermal configurations will be indicated by @xmath76 with the subscript @xmath77 specifying the configuration s temperature . thermal configurations @xmath78 are very special . firstly , for a given hamiltonian , @xmath2 , from all possible states that have a fixed value of the internal mean energy , @xmath79 $ ] , the thermal state maximises the entropy @xmath80 $ ] . in other words @xmath81 is the most unbiased configuration one can assert to the system given only the knowledge of @xmath82 @xcite . another insightful characterisation of thermal configurations in terms of a property called _ complete passivity _ was achieved by lenard @xcite , building on ideas of pusz and woronowicz @xcite . complete passivity captures the intuitive notion of thermal equilibrium . a configuration @xmath58 is said to be _ passive _ if no work can be extracted from the system , i.e. @xmath83 cf . ( [ eq : q+w ] ) , when subjected to any unitary transformation for a time @xmath63 generated by an arbitrary time - dependent hamiltonian with the sole constraint that @xmath84 . a configuration @xmath85 is _ completely passive _ if all its regularised configurations @xmath86 are passive for @xmath87 . here the unitary operations entering in this definition are generated by _ arbitrary _ time - dependent hamiltonians @xmath88 that satisfy the constraint @xmath89 . i.e. during @xmath900,\tau[$ ] , @xmath88 is allowed to introduce any sort of interactions between the various copies of @xmath1 . it turns out that while all @xmath69 with commuting @xmath1 and @xmath2 are passive configurations , only thermal configurations @xmath75 and the ground state are _ completely passive _ @xcite . to stress the special role of thermal configurations graphically , they are denoted as red circles while all configuration that are not thermal will be called _ non - equilibrium _ configurations and are depicted as blue squares , see fig . [ fig : config ] . [ [ a - note - on - gauge . ] ] a note on gauge . + + + + + + + + + + + + + + + + given a generic state @xmath91 which is full rank , there always exists a hermitian operator @xmath92 and a @xmath93 such that @xmath78 is thermal is not full rank there is no thermal configuration @xmath75 whose density matrix exactly coincides with @xmath1 . however , it is still possible to find gibbs configurations whose density matrices are arbitrarily close to @xmath1 . ] at the inverse temperature @xmath77 . in fact the problem admits infinite solutions , since there are two gauge freedoms for the choice of @xmath2 and @xmath77 . firstly , the zero - point of the energy scale can be chosen arbitrarily by a constant @xmath94 . the second gauge , @xmath95 , is the temperature itself , which sets a spacing of the energy scale . the pair @xmath96 is equivalent to @xmath97 in that they have the same set of thermal configurations . in particular the internal energy ( [ defu ] ) and the entropy ( [ defs ] ) of such configurations do not depend on the values of @xmath94 and @xmath95 . in the following we will assume that both gauges have been chosen to some fixed values . among all possible discrete transformations in dynamical configuration space @xmath52 we identify two classes that admit a clear analysis of the energetic balance and can be used as primitives for general discrete dynamical evolutions . these transformations map an initial configuration @xmath98 to a final configuration @xmath99 , denoted as @xmath100 , with the only constraint that @xmath101 for some _ unitary _ @xmath102 . no constraint is posed on the relationship between @xmath103 and @xmath43 . the definition of duts is inspired by continuous unitary transformations , see eq . ( [ eq : unitary ] ) . there the system is thermally isolated while evolving through the action of some external force that modifies the hamiltonian in time , @xmath12 , and generates arbitrary unitary evolutions @xmath104 where @xmath57 indicates time ordering . for the discrete mapping @xmath100 no assumption is made on the time duration @xmath63 nor the specific form of @xmath12 which realises the unitary @xmath102 . in analogy to the continuous situation , we define the work done on the system due to a dut identical to the total variation of the internal energy , @xmath105 , i.e. @xmath106 , \end{split}\ ] ] while no heat is associated with duts , i.e. @xmath107 a special class of duts are the _ discrete unitary quenches _ ( duqs ) . experimentally , a quench is an abrupt , instantaneous change of the system hamiltonian which leaves the system density matrix unchanged , i.e. @xmath108 , @xmath109 . we also note that for full rank states @xmath32 a duq can be found that brings @xmath110 to a final configuration that is thermal , @xmath111 , with the hamiltonian defined as @xmath112 . three duts are shown , depicted as blue arrows , each ending at a final configuration @xmath113 for @xmath114 . the special case where @xmath115 is transformed into @xmath116 is an example of a duq . also shown are two dtts , depicted as red fuzzy arrows , originating from @xmath115 and ending at thermal configurations @xmath117 and @xmath118 where @xmath119 and @xmath120 are inverse temperatures associated with the thermal configurations . , title="fig : " ] duts will be denoted as blue arrows in the graphical representation of the configuration space , see fig . [ fig : duts&dtts ] . these work arrows " indicate the closed nature of the transformation . the duts characteristic properties are summarised here : 1 . duts can start from and end at either non - thermal or thermal configurations . 2 . duts change the state by a unitary and no entropy change is induced by a dut , i.e. @xmath121 . duts can be concatenated to produce another dut . any dut has an inverse that is also a dut . dtts are defined as those transformations which take a generic @xmath122 into a gibbs state at some inverse temperature @xmath77 , @xmath123 , without modifying the system hamiltonian @xmath2 . the prototypical example of a dtt is an arbitrary thermalisation process in which the system is put into a weak thermal contact with a reservoir at inverse temperature @xmath77 and left until its state becomes time - independent . physically this is realised by the system weakly interacting with a large external environment . the requirement of a _ small _ coupling ensures a clear definition of a local system hamiltonian . for example , the dissipative evolution @xmath124 defined in eq . ( [ ddf ] ) with the additional assumption that the lindblad term @xmath23 commutes with @xmath2 will for @xmath125 converge to @xmath36 . in analogy with this continuous process , we assume that the internal energy change due to a dtt is a result solely of the heat absorbed by the system @xmath126 , \end{split}\ ] ] while the work of a dtt vanishes , @xmath127 this non - trivial expression of the heat is exactly of the form that we expected from the bounds on the entropy in eq . ( [ eq : entropy ] ) , and it will be the basis for deriving a general second law for discrete quantum trajectories in the next section . dtts will be denoted as horizontal red arrows in the graphical representation , see fig . [ fig : duts&dtts ] . the fuzziness of these heat arrows " indicates the open nature of the transformation . the characteristic properties of dtts are summarised here : 1 . dtts always end in thermal configurations . dtts do not change the hamiltonian . the entropy change associated with a dtt is in general nonzero , i.e. @xmath128 4 . dtts can be concatenated to produce another dtt . the inverse of a dtts is in general not a dtt . only if the initial configuration @xmath115 was already thermal , can the action of a dtt be reversed by another dtt . having identified two fundamental process primitives in configuration space , we now focus on more complex discrete trajectories . these can start from equilibrium or non - equilibrium configurations , however , we restrict ourselves to discrete trajectories that can be obtained by concatenating dut and dtts . within this scenario we will be able to formulate a general second law for discrete quantum processes , that does not require detailed knowledge of the continuous state and local hamiltonian evolution . let us begin with the simplest non trivial discrete transformation which can be used to connect two equilibrium configurations . we consider a trajectory that starts from a thermal configuration @xmath129 and ends at a final thermal configuration @xmath130 via the action of a single dut followed by a dtt . the heat of the discrete process can then be determined as the sum of the heats of each component , for both of which the heat is a well - defined quantity . the dut first unitarily rotates the input density matrix to @xmath131 while the hamiltonian changes from @xmath103 to @xmath43 , ending in an intermediate ( not necessarily thermal ) configuration @xmath132 . a dtt follows that brings @xmath133 to @xmath134 , resulting in the discrete overall trajectory @xmath135 shown in fig . [ fig : dut+dtttraj ] . ( black arrow ) and discrete decomposition into a dut , @xmath136 ( blue arrow ) , to an intermediate point @xmath132 followed by a dtt , @xmath137 ( red arrow).,title="fig : " ] while heat of process ( [ eq : dut+dtt ] ) is only exchanged during the dtt , the amount of exchanged heat depends on the duts unitary @xmath102 @xmath138.\end{aligned}\ ] ] clearly , the value of the heat depends on the choice of the unitary @xmath102 with the maximum and minimum heat given by @xmath139 where @xmath140 and @xmath141 are the eigenvalues of @xmath43 and @xmath103 ordered in decreasing order and @xmath142 the partition functions of the initial and final configuration . the derivation of this expression and the corresponding unitaries @xmath143 and @xmath144 , are given in appendix [ sec : min&max ] . however , for any possible choice of the dut connecting @xmath145 to the intermediate step @xmath133 , i.e. for any unitary transformation @xmath102 , a second law can be established by linking the heat and the entropy change . the entropy change is bounded according to eq . ( [ eq : entropybounds ] ) , by @xmath146 \\ & = \beta_f \ ; { { \rm tr}}[(\rho_f - v \ , \rho_i \ , v^{\dag } ) \ ; h_f ] . \end{aligned}\end{aligned}\ ] ] implying @xmath147 thus the process ( [ eq : dut+dtt ] ) obeys a clausius - type inequality , cf . ( [ eq : clausius ] ) , which states that the heat absorbed by the system is upper bounded by the entropy change . we now turn to discrete non - equilibrium processes for which establishing the clausius inequality in the continuous case has only recently been addressed @xcite . in our approach this can be done by observing that given two generic configurations @xmath98 and @xmath148 in @xmath52 , it always possible to connect them via a discrete trajectory composed by three primitive steps which differs from the one given in eq . ( [ eq : dut+dtt ] ) only by a final duq transformation . specifically we can write @xmath149 with intermediate configurations @xmath150 and @xmath151 . notice that the first and the last step of ( [ real1333 ] ) do not alter the entropy of the system , nor contribute to the heat exchange since they are duts . this implies the identity @xmath152 and allows us to identify the heat associated with @xmath57 with the quantity @xmath153 .\ ] ] the lower bound of eq . ( [ eq : entropybounds ] ) can then be used to establish a clausius - type inequality for the discrete transformation ( [ real1333 ] ) , i.e. @xmath154 = \beta_2 \ ; q({\cal t}),\end{aligned}\ ] ] where @xmath120 is the temperature of the intermediate configuration @xmath155 . + as a sequence of two dut+dtt transformations , via an intermediate thermal configuration , @xmath156 . panel * b * shows the realisation of a general transformation @xmath157 through a sequence of dut+dtt transformations , via many intermediate thermal configurations , @xmath158.,title="fig : " ] + * b * + as a sequence of two dut+dtt transformations , via an intermediate thermal configuration , @xmath156 . panel * b * shows the realisation of a general transformation @xmath157 through a sequence of dut+dtt transformations , via many intermediate thermal configurations , @xmath158.,title="fig : " ] the trajectories defined in eq . ( [ eq : dut+dtt ] ) and eq . ( [ real1333 ] ) are just specific choices of discrete trajectories connecting two configurations @xmath159 and @xmath160 . we will now show that a clausius inequality , e.g. inequalities of the type ( [ eq : gg1113 ] ) , holds for general discrete processes as long as they can be decomposed in a _ sequence _ of dut+dtts steps . to show this , we first consider the discrete trajectory , @xmath161 , pictured in panel * a * of fig . [ fig : multiple ] where @xmath162 is transformed into @xmath163 via two dut+dtt transformations and a third thermal configuration @xmath164 , i.e. @xmath165 in this scenario the following inequality for the entropy holds , @xmath166 where @xmath167 and @xmath168 , and where eq . ( [ eq : gg1113 ] ) was used for the two dut+dtt transformations . inequality ( [ hhvv ] ) can immediately be generalised to an arbitrary number of intermediate dut+dtt steps connecting @xmath169 to @xmath170 . specifically , consider a generic discrete trajectory @xmath57 composed of @xmath171 consecutive dut+dtt steps that pass through the thermal configurations @xmath172 as shown in panel * b * in fig . [ fig : multiple ] . then by expressing the total entropy increment @xmath173 as a sum of terms @xmath174 associated with the various steps of @xmath57 and applying eq . ( [ eq : gg1113 ] ) to each one of them , the clausius inequality becomes @xmath175 the generality of this derivation implies that sequences of discrete unitary and discrete thermalising transformations always fulfil a clausius type equation . to formulate this as a lemma , we introduce a useful discrete process quantity , @xmath176 , for a dut+dtt sequence , @xmath177 , \\ & = \beta_{k+1 } \ , q(c_{k } ( \beta_{k } ) \stackrel{dut+dtt}{\longrightarrow } c_{k+1 } ( \beta_{k+1 } ) ) . \end{split}\end{aligned}\ ] ] the quantity @xmath176 is the discrete analog to the continuous expression @xmath178 . for the trajectory @xmath179 the overall @xmath180 is obtained by summing over the @xmath176 contributions of the various steps , i.e. @xmath181 with @xmath182 and @xmath183 corresponding to the initial and final configurations , @xmath184 and @xmath185 , respectively . then the lower entropy bound , eq . ( [ yyy1 ] ) , can be neatly expressed : any trajectory @xmath57 made of sequences of dut+dtts fulfils a clausius inequality of the form @xmath186 while in general a single dut+dtt process can not saturate the equality , see appendix [ sec : min&max ] , we will now show that augmenting intermediate steps will always increase the r.h.s . of eq . ( [ yyy1 ] ) . moreover , we find that in the limit of infinitely long sequences the asymptotic saturation of the inequality ( [ yyy1 ] ) is always possible . to show that the entropy bound eq . ( [ yyy1 ] ) can be saturated we construct a class of trajectories @xmath187 from a generic dut+dtt trajectory @xmath57 , as depicted in panel * b * of fig [ fig : multiple ] , for which the functional @xmath188 is always larger than @xmath180 . one class of trajectories @xmath187 is the trajectory identical to @xmath57 however with the step @xmath189 replaced with the sequence @xmath190 as shown in fig . [ fig : augmenting ] , where @xmath191 has an intermediate density matrix @xmath192 with mixing probability @xmath1930,1[$ ] . is a well - defined thermal configuration as its state is by construction full rank . ] the increment on the r.h.s . of ( [ eq : sl ] ) for the new trajectory @xmath194 , @xmath195 , see fig . [ fig : augmenting ] , is then _ strictly positive _ for any @xmath196 @xmath197 -{{\rm tr } } [ \ln \rho_{k+1 } \ , ( \rho_{k+1 } - \rho_{m } ) ] + { { \rm tr } } [ \ln \rho_{k+1 } \ , ( \rho_{k+1 } - \tilde{\rho}_{k } ) ] \nonumber\\ & > - p s ( \tilde{\rho}_{k } \| \tilde{\rho}_{k } ) - ( 1-p ) \ ; s ( \rho_{k+1 } \| \tilde{\rho}_{k } ) - p s(\rho_{k+1 } \| \tilde{\rho}_{k } ) - ( 1-p)\ ; s(\rho_{k+1 } \| \rho_{k+1 } ) + s(\rho_{k+1 } \| \tilde{\rho}_{k } ) \\ & = 0 , \end{aligned}\ ] ] where we have assumed @xmath198 , and used eq . ( [ middlepoint ] ) and the joint convexity of the relative entropy @xcite , @xmath199 with equality _ iff _ @xmath200 . we summarise this result in the following lemma : adding intermediate thermal configurations @xmath201 ( see eq . [ middlepoint ] ) to any trajectory @xmath57 results in a new trajectory @xmath202 with increased @xmath176 , i.e. @xmath203 having confirmed that it is possible for any given discrete trajectory to introduce intermediate steps such that the r.h.s . ( [ eq : sl ] ) increases , the task is now to show that the entropy bounds can be saturated by reiterating the procedure . the proof relies on lower bounding @xmath176 and showing that the bound converges to the upper bound on @xmath176 , eq . ( [ eq : sl ] ) , in the limit of infinite steps . the detailed derivation is given in appendix [ sec : satproof ] proving the following theorem : [ lem : clausiussat ] let @xmath57 be a discrete trajectory connecting the initial gibbs configuration @xmath204 to the final gibbs configuration @xmath205 via a sequence of @xmath171 concatenated dut+dtt steps @xmath206 connecting the thermal configurations @xmath207 as in panel * b * in fig . [ fig : multiple ] . then a sequence of trajectories @xmath208 exists , obtained from @xmath57 by adding @xmath209 intermediate thermal steps , which saturates the clausius bound ( [ yyy1 ] ) in the asymptotic limit , i.e. @xmath210 by a single dtt with heat @xmath211 , and by a sequence of two duq+dtt transformations via an intermediate thermal configurations , @xmath212 . the two dtts have heats @xmath213 and @xmath214 , respectively.,title="fig : " ] in the introduction we have seen that for continuous processes where consistent definitions of @xmath11 and of the local hamiltonian @xmath12 can be assigned for all @xmath13 , eq . ( [ eq : q ] ) defines the heat absorbed by the system . we have already discussed the difficulties of knowing @xmath11 and identifying a proper local hamiltonian @xmath12 for the system . however , in what follows we will assume that some valid " continuous trajectory @xmath215 is given for which eqs . ( [ eq : q ] ) and ( [ eq : q+w ] ) apply . we now wish to identify a discrete trajectory that connects the same initial and final configurations as the continuous trajectory through a sequence of duts and dtts , and which approximates the continuous heat . the analysis leads to the following theorem : for a continuous process between two configurations @xmath216 and @xmath217 that obeys the clausius inequality , a discrete trajectory @xmath218 exists that connects the same configurations and has exactly the same heat as the continuous process . _ proof : _ consider an infinitesimal heat increment along the continuous trajectory , @xmath219,\end{aligned}\ ] ] with @xmath11 and @xmath220 being the density matrices of two infinitesimally separated configurations on the trajectory . define the initial and final configuration for a discrete trajectory to be @xmath221 and @xmath222 . to compare the continuous heat ( [ quantity ] ) with a discrete heat we need to identify a discrete trajectory , @xmath161 , connecting the same initial and final configuration as the continuous trajectory . one example is the sequence @xmath161 shown as a solid line in fig . [ fig : triangle ] , @xmath223 where @xmath224 and @xmath225 are equilibrium configurations requires @xmath33 to be full rank . if this is not the case one can still define a trajectory ( [ trj ] ) with @xmath226 replaced by a full rank configuration which can be chosen to be arbitrarily close to @xmath226 . ] and @xmath227 is a non - equilibrium configuration . the inverse temperatures of the equilibrium configurations @xmath228 and @xmath229 are both chosen @xmath230 . heat @xmath231 , @xmath232 , \end{split}\ ] ] is exchanged when passing from @xmath233 via a duq+dtt through the intermediate configuration @xmath133 , see fig . [ fig : triangle ] . heat @xmath214 , @xmath234,\ ] ] is exchanged when passing from @xmath235 . therefore @xmath236 and @xmath226 , of the continuous process . a discrete trajectory , @xmath161 , that connects @xmath115 and @xmath226 is also indicated ( solid ) . also shown is the closed loop discrete trajectory @xmath237 ( dotted ) . the two thermal configurations @xmath238 and @xmath225 are both chosen with inverse temperature @xmath239 , as indicated in the circles . , title="fig : " ] on the other hand , the continuous heat increment @xmath240 can be decomposed into two heat contributions , @xmath241 = { { \rm tr}}[(\rho_2 - \rho_1 ) \tilde{h}_2 ] \\ & = { { \rm tr } } [ ( { \rho}_2 -\tilde{\rho}_2 ) \tilde{h}_2 ] + { { \rm tr } } [ ( \tilde{\rho}_2 -\rho_1 ) \tilde{h}_2 ] \\ & = - { { \rm tr } } [ ( { \rho}_2 -\tilde{\rho}_2 ) \ln \tilde{\rho}_2 ] - { { \rm tr } } [ ( \tilde{\rho}_2 -\rho_1 ) \ln \tilde{\rho}_2 ] \\ & = - { } q_f + { } q_i . \end{aligned}\ ] ] where @xmath242 would be the heat absorbed by the system if it passed from @xmath226 to @xmath228 via a dtt , i.e. @xmath243 .\ ] ] to compare the continuous heat ( [ defdq ] ) with the discrete heat ( [ defqgamma ] ) , we introduce a second discrete trajectory @xmath237 . this is a closed loop sequence of duq and dtt transformations , see fig . [ fig : triangle ] , @xmath244 in trajectory @xmath237 heat is exchanged from @xmath245 and from @xmath246 , so @xmath247 . as discussed in previous sections the discrete heat always obeys the clausius inequality ( [ eq : gg1113 ] ) , and with @xmath248 for both steps this implies @xmath249 for trajectory @xmath237 . using this in eq . ( [ defqgamma ] ) we find that the heat associated with the discrete trajectory @xmath161 from @xmath115 to @xmath226 is a _ lower bound _ to the infinitesimal continuous heat for the same initial and final configuration , ( [ quantity ] ) , i.e. @xmath250 this can immediately be extended to the full continuous process : for any arbitrary continuous process between @xmath216 and @xmath217 there is always a discrete trajectory @xmath251 between the same two configurations that has a lower heat than the continuous heat , eq . ( [ eq : q ] ) . moreover , by augmenting intermediate steps in the discrete trajectory @xmath251 , resulting in the trajectory @xmath218 that passes through an infinite sequence of points @xmath252 , it is possible to increase the associated heat , as shown in sec . [ sec : multiple ] . from theorem [ lem : clausiussat ] follows that if the continuous trajectory fulfils the clausius inequality , then a discrete trajectory connecting the same initial and final configuration can be found that has the same heat as the continuous trajectory . @xmath253 the last piece in our analysis of the energy balance in discrete quantum processes is to determine the efficiency of a discrete cyclic process , such as the one depicted in fig . [ fig : cycle ] where @xmath117 , @xmath254 are equilibrium configurations , while @xmath255 , and @xmath256 are not . this results in the following lemma : the thermal efficiency of the discrete cycle depicted in fig . [ fig : cycle ] is bounded by the carnot efficiency , @xmath257 and the optimal efficiency is achievable . _ proof : _ for the overall loop , @xmath258 the entropy change is zero . however , the entropy change nevertheless bounds the heat of the two heat - producing dtt processes , @xmath259 and @xmath260 , @xmath261 ) . , title="fig : " ] this implies that at least one of the two heats must be negative . let us assume for instance that the heat exchanged with the thermal reservoir at temperature @xmath262 is positive , @xmath263 , while the other heat is negative @xmath264 ( other scenarios can be treated analogously , see below ) . the total heat absorbed per cycle is @xmath265 with the energy balance implying that the overall absorbed heat must be equal to the negative work done on the system during the cycle , @xmath266 the thermal efficiency is defined as the ratio between the work performed and the heat absorbed , leading to @xmath267 where we used eq . ( [ eq : heatbound ] ) . if @xmath268 the system absorbs heat from a higher temperature bath and gives heat to a lower temperature bath . the efficiency @xmath269 is then positive and smaller than unity with the optimal efficiency reproducing the carnot efficiency . the optimal efficiency can be reached by augmenting the discrete trajectory to saturate the equality in the clausius - inequality , see theorem [ lem : clausiussat ] . @xmath253 _ remark : _ if instead @xmath270 , i.e. heat is absorbed from a lower temperature bath and given to one at a higher temperature , the system operates as a refrigerator . in this case the total work _ absorbed _ by the system , @xmath271 , is positive . the efficiency of the process can be measured by the coefficient of performance , cop@xmath272 , defined as the ratio between the heat absorbed from the cold reservoir @xmath273 ( i.e. @xmath274 ) and the total work done on the system @xmath275 which again is always smaller than one . finally , if the signs for the heats @xmath274 and @xmath276 are interchanged the above argument still holds with eqs . ( [ eta ] ) and ( [ cooling ] ) being replaced by the inequalities @xmath277 and @xmath278 , respectively . the early development of thermodynamics culminating in the formulation of the second law also gave birth to a new quantity , the entropy , whose physical meaning was at first opaque . only later was its meaning elucidated by the works of boltzmann and others . in this paper we proposed to turn the original argument around and use the well - established notion of entropy that characterises the information content in a ( quantum ) state to motivate the definition of a notion of heat for discrete quantum processes . the approach circumvents the large cluster of problems surrounding the idea of a unique definition of heat and work in processes where ( i ) the hamiltonian of the system is not well - defined due to the open nature of the system and ( ii ) there are fundamental limitations on the notion of trajectories where full knowledge is only given at discrete points in time when a measurement with a specific hamiltonian occurred . by introducing thermodynamic configurations , identifying two primitives for discrete processes , duts and dtts , and defining heat to pertain only to dtts we were able to uncover a general second law valid for any discrete process consisting of sequences of dut+dtts between both , equilibrium and non - equilibrium configurations . moreover , we showed that an infinite sequence of dut+dtt processes exists that saturates the clausius inequality . in other words , saturation occurs when a discrete trajectory is mapped out into a continuous one by a sequence of measurements that are infinitely close together . this provides a link between _ reversibility _ - here the reversibility of a discrete process - and the equality in the second law for discrete processes , reminiscent of the clausius statement of equality for thermodynamic reversible continuous processes , eq . ( [ eq : reversible ] ) . on the other hand , we also showed that for any continuous process between two configurations that obeys the clausius inequality , there exists a discrete process between the same configurations with the same heat . finally , we showed that for the discrete version of a thermodynamic cycle , formed by a discrete trajectory passing through four configurations carnot s efficiency is recovered . the strength of our approach is to give meaning to heat and work , reversibility and efficiency following from just a few sensible and simple definitions . in some respect this is analogous to the axiomatic approach to thermodynamics first developed by carathodory @xcite . we hope that the presented analysis will inspire discussions and future work on characterizing heat and work in quantum processes . of course many open questions remain . one direction of particular relevance is clearly the identification of a proper _ metric _ in configuration space , that would allow to quantify , in a precise and ( hopefully ) operationally well defined way , how distant two generic discrete trajectories are . vg thanks rosario fazio and seth lloyd for fruitful discussions and comments . vg is supported by miur through firb - ideas project no . ja is supported by the royal society . here we discuss the impact of the unitary @xmath102 of eq . ( [ eq : heatdut+dtt ] ) on the heat of the dut+dtt process ( [ eq : dut+dtt ] ) . specifically , we want to identify the duts that maximize and minimize the heat on the r.h.s of eq . ( [ eq : gg1113 ] ) and establish if a dut exists that leads to a saturated clausius equality . the last question can be easily solved by noticing that saturation occurs only when the inequality holds in eq . ( [ eq : s - bound ] ) . however , this is only true iff @xmath279 , i.e. @xmath32 and @xmath33 must be unitarily equivalent for some unitary @xmath280 , @xmath281 . in other words equivalence in ( [ eq : dut+dtt ] ) requires that no dtt enters in the process so that @xmath282 . for any non - trivial dtt a finite gap between the l.h.s . and the r.h.s . of eq . ( [ eq : gg1113 ] ) exists . ( this is not true however for sequences of dut - dtt transformations as considered in sec . [ sec : sat ] . ) to the gibbs state @xmath130 via some dut , specified by a unitary @xmath102 , followed by an appropriate dtt . , title="fig : " ] to determine the minimum / maximum gap we require the maximum / minimum heat with respect to all possible duts of a dut+dtt process connecting two thermal configurations @xmath129 and @xmath130 , see fig [ fig : max - minheat ] , i.e. @xmath283 , \\ q_{\min } & : = \min_{v } { { \rm tr}}[(\rho_f - v \ , \rho_i \ , v^\dag ) \ ; h_f ] , \end{split}\ ] ] where the maximisation / minimisation is taken over all unitary transformations @xmath102 . this task is solved with the following lemma . given @xmath284 and @xmath285 hermitian operators on a finite dimensional hilbert space @xmath286 , with @xmath287 and @xmath288 being their corresponding eigenvalues which are ordered in decreasing order ( i.e. @xmath289 , @xmath290 ) . then the minimum of @xmath291 $ ] over the set of unitary transformations is achieved by the unitary @xmath143 which maps the eigenvector @xmath292 of @xmath293 into the eigenvectors @xmath294 of @xmath295 in such a way that @xmath296 i.e. the maximum eigenvector of @xmath293 is mapped into the minimum eigenvector of @xmath295 . as a consequence the minimum expectation value is @xmath297 = { { \rm tr } } [ a \ , v_{\min } \ , b \ , v_{\min}^\dag ] = \sum_j \alpha_j \beta_{n - j+1}.\ ] ] similarly the maximum of @xmath291 $ ] over the set of unitary transformations is achieved by the unitary @xmath144 that maps the eigenvectors @xmath292 of @xmath293 into the eigenvectors @xmath294 of @xmath295 in such a way that @xmath298 consequently the maximum expectation value is @xmath299 = { { \rm tr } } [ a \ , v_{\max } \ , b \ , v_{\max}^\dag ] = \sum_j \alpha_j \beta_{j}.\ ] ] _ proof : _ these minimum and maximum expectation values are a trivial consequence of the theorem 2 of ref . @xcite . @xmath253 + application of eqs . ( [ eq : min ] ) and ( [ eq : max ] ) gives the minimum and maximum heat for the dut+dtt process ( [ eq : dut+dtt ] ) @xmath300 where @xmath140 and @xmath141 are the eigenvalues of @xmath43 and @xmath103 ordered in decreasing order . consider the @xmath301-th step of the trajectory @xmath57 in panel * b * in fig . [ fig : multiple ] , which connects the thermal points @xmath302 and @xmath303 . we define a new trajectory @xmath304 which is identical with the original trajectory @xmath57 except that the @xmath301-th step is now replaced with a sequence of @xmath305 intermediate _ thermal configurations _ @xmath306 , @xmath307 , @xmath308 that are linked through a duq - dtt sequence @xmath309 the configurations @xmath310 have density matrices @xmath311 , @xmath307 , @xmath312 defined by the mixtures @xmath313 where @xmath314 and @xmath315 . applying the clausius inequality ( [ yyy1 ] ) in the form of ( [ eq : sl ] ) to sequence ( [ real22222 ] ) yields @xmath316 by definition ( [ eq : lambdapiece ] ) , the transformations being _ discrete unitaries _ and the fact that intermediate inverse temperatures are all set to 1 , @xmath317 can be expressed as @xmath318 \\ & = { { \rm tr}}\left [ ( \rho_{{k } } - \rho_{k+1})\ ; \frac{1}{n } \sum_{\ell=0}^{n-1 } \ ; \ln \rho'_{\ell+1 } \right ] . \end{split}\ ] ] on the other hand the entropy change @xmath174 can be lower bounded by @xmath319 \\ & = { { \rm tr}}\left [ ( \rho_{{k } } - \rho_{k+1})\ ; \frac{1}{n } \sum_{\ell=0}^{n-1 } \ ; \ln \rho'_{\ell } \right]\\ & = \lambda({\cal t}_{k;n } ' ) + { { \rm tr}}\left [ ( \rho_{k } - \rho_{k+1})\ ; \frac{\ln \rho_k - \ln \rho_{k+1}}{n } \right]\\ & = \lambda({\cal t}_{k;n } ' ) + \frac{s(\rho_{k+1}\| \rho_{k})+ s(\rho_{k}\| \rho_{k+1})}{n } , \end{split}\ ] ] which implies @xmath320 @xmath321 and @xmath322 are density matrices of gibbs configurations and thus of full rank . consequently , the quantity @xmath323 is finite . explodes if and only if the kernel of @xmath324 or @xmath322 admits a non - trivial overlap with the support of the other state @xcite . since @xmath321 and @xmath322 are full rank , neither conditions can be fulfilled and the symmetric sum of relative entropies is always finite . ] from eq . ( [ yyy1121 ] ) and ( [ tttt ] ) it then follows that @xmath325 converges to @xmath174 for @xmath326 , i.e. @xmath327 in other words by augmenting the intermediate points of which connects @xmath328 and @xmath329 we can saturate the associated clausius inequality for the @xmath301-th step of the trajectory @xmath57 . by repeating the same procedure for each of the steps of @xmath57 a new trajectory emerges as the union of the individual sequences , @xmath330 where @xmath209 is the multidimensional variable @xmath331 . lemma [ lem : clausiussat ] follows from the additivity of @xmath176 ( [ functionals ] ) and taking the limit of each @xmath332 ( [ limit ] ) . @xmath253

to probe dynamical properties of a system , one often applies an external stimulus to the system and investigate its response . particularly , periodically stimulated biological oscillators have attracted much attention in various systems such as the embryonic chick heart - cell aggregates @xcite and the squid giant axon @xcite . these periodically forced systems have been found to exhibit rich regular and chaotic behaviors @xcite . recently , similar lockings and chaotic responses have also been found in neocortical networks of pyramidal neurons under periodic synaptic input @xcite . on the contrary , quasiperiodically forced case has received little attention @xcite . hence , intensive investigation of quasiperiodically forced biological oscillators is necessary for understanding their dynamical responses under the quasiperiodic stimulus . strange nonchaotic ( sn ) attractors typically appear between the regular and chaotic attractors in quasiperiodically forced dynamical systems @xcite . they exhibit some properties of regular as well as chaotic attractors . like regular attractors , their dynamics is nonchaotic in the sense that they do not have a positive lyapunov exponent ; like usual chaotic attractors , they have a geometrically strange fractal structure . here , we are interested in dynamical responses of neural oscillators subject to quasiperiodic stimulation . sn oscillations are expected to occur in quasiperiodically forced neural oscillators . this paper is organized as follows . in sec . [ sec : sn ] , we study dynamical responses of the quasiperiodically forced hodgkin - huxley ( hh ) neuron which was originally introduced to describe the behavior of the squid giant axon @xcite . as a dc stimulus passes a threshold value , a self - sustained oscillation ( corresponding to a spiking state ) is induced . effect of periodic forcing on this hh oscillator was previously studied @xcite . thus , regular ( such as phase locking and quasiperiodicity ) and chaotic responses were found . we note that similar dynamical responses were also observed in experimental works of the periodically forced squid giant axon @xcite . in this way , the model study on the hh neuron may be examined in real experiments of the squid giant axon . here , we numerically study the case that the hh oscillator is quasiperiodically forced at two incommensurate frequencies and compare the dynamical responses with those for the periodically forced case . in the periodically forced case ( _ i.e. _ , in the presence of only one ac stimulus source ) , a transition from a periodic to a chaotic oscillation was found to occur via period - doubling bifurcations @xcite . effect of the quasiperiodic forcing on this period - doubling route to chaotic oscillation is particularly investigated by adding another independent ac stimulus source . thus , unlike the case of periodic forcing , new type of sn oscillating states are found to occur between the regular and chaotic oscillating states as intermediate ones . as a result of their strange geometry , sn oscillating states give rise to the appearance of aperiodic complex spikings , as in the case of chaotic oscillations . diverse dynamical routes to sn oscillations are identified , like the case of the quasiperiodically forced logistic map @xcite . it is also expected that these sn spikings might be observed in experimental works on the quasiperiodically forced squid giant axon . finally , a summary is given in sec . [ sec : sum ] . we consider the conductance - based hh neuron model which serves as a canonical model for tonically spiking neurons . the dynamics of the hh neuron , which is quasiperiodically forced at two incommensurate frequencies @xmath0 and @xmath1 , is governed by the following set of differential equations : @xmath2 [ eq : hh1 ] where the external stimulus current density ( measured in units of @xmath3 ) is given by @xmath4 , @xmath5 is a dc stimulus , @xmath6 and @xmath7 are amplitudes of quasiperiodic forcing , and @xmath8 is irrational ( @xmath0 and @xmath1 : measured in units of khz ) . here , the state of the hh neuron at a time @xmath9 ( measured in units of ms ) is characterized by four variables : the membrane potential @xmath10 ( measured in units of mv ) , the activation ( inactivation ) gate variable @xmath11 of the @xmath12 channel [ _ i.e. , _ the fraction of sodium channels with open activation ( inactivation ) gates ] , and the activation gate variable @xmath13 of the @xmath14 channel ( _ i.e. , _ the fraction of potassium channels with open activation gates ) . in eq . ( [ eq : hha ] ) , @xmath15 represents the membrane capacitance per surface unit ( measured in units of @xmath16 ) and the total ionic current @xmath17 consists of the sodium current @xmath18 , the potassium current @xmath19 , and the leakage current @xmath20 . each ionic current obeys ohm s law . the constants @xmath21 , @xmath22 , and @xmath23 are the maximum conductances for the ion and the leakage channels , and the constants @xmath24 , @xmath25 , and @xmath26 are the reversal potentials at which each current is balanced by the ionic concentration difference across the membrane . the three gate variables obey the first - order kinetics of eq . ( [ eq : hhb ] ) . the rate constants are given by @xmath27}{\exp [ \{25-(v - v_r)\}/10 ] - 1 } , \;\ ; \beta_m(v ) = 4 \exp [ -(v - v_r)/18 ] , \\ \alpha_h(v ) & = & 0.07 \exp [ -(v - v_r)/20 ] , \;\ ; \beta_h(v ) = \frac{1}{\exp [ \{30-(v - v_r)\}/10]+1 } , \\ \alpha_n(v ) & = & \frac{0.01[10-(v - v_r)]}{\exp [ \{10-(v - v_r)\}/10]-1 } , \;\ ; \beta_n(v ) = 0.125 \exp [ -(v - v_r)/80],\end{aligned}\ ] ] [ eq : rc ] where @xmath28 is the resting potential when @xmath29 . for the squid giant axon , typical values of the parameters ( at 6.3 @xmath30 ) are @xcite : @xmath31 , @xmath32 , @xmath33 , @xmath34 , @xmath35 mv , @xmath36 mv , @xmath37 mv , and @xmath38 mv . to obtain the poincar map of eq . ( [ eq : hh1 ] ) , we make a normalization @xmath39 , and then eq . ( [ eq : hh1 ] ) can be reduced to the following differential equations : @xmath40 , \\ \frac{dm}{dt } & = & f_2({\bf{x}},\theta ) = { \frac { 1 } { f_1 } } [ \alpha_m ( v ) ( 1-m ) - \beta_m ( v ) m ] , \\ \frac{dh}{dt } & = & f_3({\bf{x}},\theta ) = { \frac { 1 } { f_1 } } [ \alpha_h ( v ) ( 1-h ) - \beta_h ( v ) h ] , \\ \frac{dn}{dt } & = & f_4({\bf{x}},\theta ) = { \frac { 1 } { f_1 } } [ \alpha_n ( v ) ( 1-n ) - \beta_n ( v ) n ] , \\ \frac { d\theta } { dt } & = & \omega~~{\rm { ( mod~1)}},\end{aligned}\ ] ] [ eq : hh2 ] where @xmath41 \equiv ( v , m , h , n)$ ] and @xmath42 . the phase space of the quasiperiodically forced hh oscillator is six dimensional with coordinates @xmath10 , @xmath43 , @xmath44 , @xmath13 , @xmath45 , and @xmath9 . since the system is periodic in @xmath45 and @xmath9 , they are circular coordinates in the phase space . then , we consider the surface of section , the @xmath10-@xmath43-@xmath44-@xmath13-@xmath45 hypersurface at @xmath46 ( @xmath13 : integer ) . the phase - space trajectory intersects the surface of section in a sequence of points . this sequence of points corresponds to a mapping on the five - dimensional hypersurface . the map can be computed by stroboscopically sampling the orbit points @xmath47 @xmath48 $ ] at the discrete time @xmath13 ( corresponding to multiples of the first external driving period @xmath49 ) . we call the transformation @xmath50 the poincar map , and write @xmath51 . numerical integration of eqs . ( [ eq : hh1 ] ) and ( [ eq : hh2 ] ) is done using the fourth - order runge - kutta method . dynamical analysis is performed in both the continuous - time system ( _ i.e. _ , flow ) and the discrete - time system ( _ i.e. _ , poincar map ) . for example , the time series of the membrane potential @xmath52 and the phase flow are obtained in the flow . on the other hand , the lyapunov exponent @xcite and the phase sensitivity exponent @xcite of an attractor are calculated in the poincar map . to obtain the lyapunov exponent of an attractor in the poincar map , we choose 20 random initial points @xmath53 with uniform probability in the range of @xmath54 , @xmath55 , @xmath56 , @xmath57 and @xmath58 . for each initial point , we get the lyapunov exponent , and choose the average value of the 20 lyapunov exponents . ( the method of obtaining the phase sensitivity exponent will be explained below . ) in the presence of only the dc stimulus ( @xmath59 @xmath60 ) , a transition from a resting state to a periodic spiking state occurs for @xmath61 @xmath62 @xmath63 via a subcritical hopf bifurcation when the resting state absorbs the unstable limit cycle born via a fold limit cycle bifurcation for @xmath64 @xmath65 @xcite . thus , a self - sustained oscillation ( corresponding to a spiking state ) is induced in the hh neuron model for @xmath66 . here , we set @xmath67 @xmath65 and @xmath68 to be the reciprocal of the golden mean [ _ i.e. _ , @xmath69 , and numerically investigate dynamical responses of the ( self - sustained ) hh oscillator under the ac external stimulus . we first study the case of periodic forcing ( _ i.e. _ , @xmath70 ) by varying @xmath6 for @xmath71 hz . figures [ fig : pd](a)-[fig : pd](c ) show the time series of @xmath52 for @xmath72 , @xmath73 , and @xmath74 @xmath65 , respectively , and the bifurcation diagram in the poincar map @xmath75 is also given in fig . [ fig : pd](d ) ; stroboscopically sampled points in @xmath75 are represented by solid circles in figs . [ fig : pd](a)-[fig : pd](c ) . as @xmath6 is decreased , successive period - doubling bifurcations occur . for example , periodic oscillations of @xmath52 in figs . [ fig : pd](a ) and [ fig : pd](b ) correspond to period-1 and period-2 states in @xmath75 , respectively . when @xmath6 passes a threshold @xmath76 @xmath77 @xmath65 ) a chaotic transition occurs . thus , for @xmath78 chaotic oscillations with positive lyapunov exponents appear , as shown in fig . [ fig : pd](c ) . from now on , we investigate the effect of quasiperiodic forcing on the period - doubling route to chaotic oscillation by changing @xmath6 and @xmath7 for @xmath71 hz . figure [ fig : sd ] shows a state diagram in the @xmath79 plane . each state is characterized by the largest ( nontrivial ) lyapunov exponent @xmath80 , associated with dynamics of the variable @xmath81 [ besides the ( trivial ) zero exponent , related to the phase variable @xmath45 of the quasiperiodic forcing ] and the phase sensitivity exponent @xmath82 . the exponent @xmath82 measures the sensitivity of the variable @xmath83 with respect to the phase @xmath45 of the quasiperiodic forcing and characterizes the strangeness of an attractor @xcite . regular quasiperiodic oscillations occur on smooth tori . a smooth torus that has a negative largest lyapunov exponent ( _ i.e. _ , @xmath84 ) and has no phase sensitivity ( _ i.e. _ , @xmath85 ) exists in the region denoted by @xmath86 and shown in light gray . when crossing a solid line , the smooth torus becomes unstable and bifurcates to a smooth doubled torus in the region represented by @xmath87 . smooth quadrupled tori , bifurcated from doubled tori , also exist in the region denoted by @xmath88 . on the other hand , chaotic oscillating states with positive largest lyapunov exponents ( @xmath89 ) exist in the region shown in black . between these regular and chaotic regions , sn oscillating states that have negative largest lyapunov exponents ( @xmath90 ) and positive phase sensitivity exponents @xmath91 exist in the region shown in gray . due to their high phase sensitivity , sn oscillating states have a strange fractal phase space structure . various dynamical routes to sn oscillations via gradual fractalization , collision with a smooth unstable torus , and collision with a nonsmooth ring - shaped unstable set will be discussed below . when passing a heavy solid boundary curve in fig . [ fig : sd ] , a transition from a smooth torus to an sn attractor occurs via gradual fractalization @xcite . as an example , we study such transition to sn oscillations along the route @xmath92 by decreasing @xmath6 for @xmath93 @xmath65 . figures [ fig : frac](a)-[fig : frac](c ) show the time series of @xmath52 for the quasiperiodic oscillation , the sn oscillation , and the chaotic oscillation when @xmath94 , @xmath95 , and @xmath96 @xmath97 , respectively . these regular , sn , and chaotic states are analyzed in terms of the largest lyapunov exponent @xmath80 and the phase sensitivity exponent @xmath82 in the poincar map . projections of their corresponding attractors onto the @xmath98 plane are shown in figs . [ fig : frac](d)-[fig : frac](f ) . for the regular state , a smooth torus exists in the @xmath98 plane [ see fig . [ fig : frac](d ) ] . as @xmath6 is decreased , the smooth torus becomes more and more wrinkled and transforms to an sn attractor without apparent mediation of any nearby unstable invariant set @xcite . as an example , see an sn attractor in fig . [ fig : frac](e ) . this kind of gradual fractalization is the most common route to sn attractors . with further decrease in @xmath6 , such an sn attractor turns into a chaotic attractor , as shown in fig . [ fig : frac](f ) . a dynamical property of each attractor is characterized in terms of the largest lyapunov exponent @xmath80 ( measuring the degree of sensitivity to initial conditions ) . the lyapunov - exponent diagram ( _ i.e. _ , plot of @xmath80 vs. @xmath6 ) is given in fig . [ fig : frac](g ) . when passing a threshold value of @xmath99 @xmath65 , an sn attractor appears . the graph of @xmath80 for the sn attractor is shown in black , and its value is negative as in the case of smooth torus . however , as @xmath6 passes the chaotic transition point of @xmath100 @xmath65 , a chaotic attractor with a positive @xmath80 appears . although sn and chaotic attractors are dynamically different , both of them have strange geometry . to characterize the strangeness of an attractor , we investigate the sensitivity of the attractor with respect to the phase @xmath45 of the external quasiperiodic forcing @xcite . this phase sensitivity may be characterized by differentiating @xmath101 with respect to @xmath45 at a discrete time @xmath46 . using eq . ( [ eq : hh2 ] ) , we may obtain the following governing equation for @xmath102 @xmath103 , @xmath104 where @xmath105 and @xmath106 s @xmath103 are given in eq . ( [ eq : hh2 ] ) . starting from an initial point @xmath107 and an initial value @xmath108 for @xmath109 , we may obtain the derivative values of @xmath110 @xmath111 at all subsequent discrete time @xmath46 by integrating eqs . ( [ eq : hh2 ] ) and ( [ eq : pse ] ) . one can easily see the boundedness of @xmath112 by looking only at the maximum @xmath113 we note that @xmath114 depends on a particular trajectory . to obtain a `` representative '' quantity that is independent of a particular trajectory , we consider an ensemble of randomly chosen initial points @xmath115 , and take the minimum value of @xmath116 with respect to the initial orbit points @xcite , @xmath117 figure [ fig : frac](h ) shows a phase sensitivity function @xmath118 , which is obtained in an ensemble containing 20 random initial orbit points @xmath119 which are chosen with uniform probability in the range of @xmath54 , @xmath120 , @xmath56 , @xmath57 and @xmath58 . for the case of the smooth torus in fig . [ fig : frac](d ) , @xmath118 grows up to the largest possible value of the derivative @xmath121 along a trajectory and remains for all subsequent time . thus , @xmath118 saturates for large @xmath122 and hence the smooth torus has no phase sensitivity ( _ i.e. _ , it has smooth geometry ) . on the other hand , for the case of the sn attractor in fig . [ fig : frac](e ) , @xmath123 grows unboundedly with the same power @xmath82 , independently of @xmath124 , @xmath125 here , the value of @xmath126 is a quantitative characteristic of the phase sensitivity of the sn attractor , and @xmath82 is called the phase sensitivity exponent . for obtaining satisfactory statistics , we consider 20 ensembles for each @xmath6 , each of which contains 20 randomly chosen initial points and choose the average value of the 20 phase sensitivity exponents obtained in the 20 ensembles . figure [ fig : frac](i ) shows a plot of @xmath82 versus @xmath127 @xmath128 . note that the value of @xmath82 monotonically increases from zero as @xmath6 is decreased away from the sn transition point @xmath76 @xmath129 @xmath130 . as a result of this phase sensitivity , the sn oscillating state has strange fractal geometry leading to aperiodic complex spikings , as in the case of chaotic oscillations [ _ e.g. , _ see figs . [ fig : frac](b ) and [ fig : frac](c ) ] . as a dashed boundary curve in fig . [ fig : sd ] is crossed , another route to sn attractors appears through collision between a stable smooth doubled torus and its unstable smooth parent torus @xcite . as an example , we study this transition to sn oscillations along the route @xmath131 by decreasing @xmath6 for @xmath132 @xmath65 . figure [ fig : hh ] shows a stable two - band torus ( denoted by a solid curve ) and an unstable smooth one - band parent torus ( denoted by a short - dashed curve ) for @xmath133 @xmath65 . the unstable parent torus is located in the middle of the two bands of the the stable torus . as @xmath6 is decreased , the bands of the stable torus becomes more and more wrinkled , while the unstable torus remains smooth [ see fig . [ fig : hh](b ) ] . when @xmath6 passes a threshold value of @xmath134 @xmath65 , the two bands of the stable torus touch its unstable parent torus at a dense set of @xmath45 values ( not at all @xmath45 values ) . as a result of this phase - dependent ( nonsmooth ) collision between the stable doubled torus and its unstable parent torus , an sn attractor is born , as shown in fig . [ fig : hh](c ) . this sn attractor , containing the former bands of the torus as well as the unstable parent torus , has a positive phase sensitivity exponent ( @xmath59 @xmath135 ) , inducing the strangeness of the sn attractor . however , its dynamics is nonchaotic because the largest lyapunov exponent is negative ( @xmath59 @xmath136 ) . as another threshold value of @xmath137 @xmath65 is passed , the sn attractor transforms to a chaotic attractor with a positive largest lyapunov exponent @xmath80 [ see fig . [ fig : hh](d ) ] . a main interesting feature of the state diagram in fig . [ fig : sd ] is the existence of `` tongues '' of quasiperiodic motion that penetrate into the chaotic region . the first - order ( second - order ) tongue lies near the terminal point ( denoted by a cross ) of the first - order ( second - order ) torus - doubling bifurcation curve . when crossing the upper boundary of the tongue ( denoted by a dotted line ) , an intermittent sn attractor appears via phase - dependent collision of a stable torus with a nonsmooth ring - shaped unstable set @xcite . we first study the transition to an intermittent sn attractor along the route @xmath138 in the first - order tongue by increasing @xmath7 for @xmath139 @xmath97 . figure [ fig : int](a ) shows a smooth torus for @xmath140 @xmath65 . when passing a threshold value of @xmath141 @xmath65 , a sudden transition to an intermittent sn attractor occurs , as shown in fig . [ fig : int](b ) for @xmath142 . due to high phase sensitivity , this sn attractor with @xmath143 has a strange fractal structure , while its dynamics is nonchaotic because of a negative largest lyapunov exponent ( @xmath144 ) . a typical trajectory on the intermittent sn attractor spends a long stretch of time in the vicinity of the former torus , then it bursts out from this region and traces out a much larger fraction of the state space , and so on . to characterize the intermittent bursting , we use a small quantity @xmath145 for the threshold value of the magnitude of the deviation from the former torus . when the deviation is smaller ( larger ) than @xmath145 , the intermittent attractor is in the laminar ( bursting ) phase . for each @xmath7 , we follow a long trajectory until @xmath146 laminar phases are obtained in the poincar map @xmath75 and get the average of characteristic time @xmath147 between bursts . as shown in fig . [ fig : int](c ) , the average value of @xmath147 exhibits a power - law scaling behavior , @xmath148 where the overbar represents time averaging and @xmath149 @xmath150 . the scaling exponent @xmath151 seems to be the same as that for the case of the quasiperiodically forced map @xcite . as @xmath7 passes another threshold value of @xmath152 @xmath65 , the sn attractor transforms to a chaotic attractor because the largest lyapunov exponent @xmath80 becomes positive , as shown in fig . [ fig : int](d ) . furthermore , using the rational approximation , the mechanism for the intermittent route to sn attractors will be investigated below . thus , a smooth torus is found to transform to an intermittent sn attractor via phase - dependent collision with a nonsmooth ring - shaped unstable set . we also study another intermittent route to sn attractors along the route @xmath153 in the second - order tongue by increasing @xmath7 for @xmath154 @xmath65 . figure [ fig : int](e ) shows a smooth two - band torus for @xmath155 @xmath65 . as @xmath7 passes a threshold value of @xmath156 @xmath157 , a band - merging transition from a smooth doubled torus to a single - band sn attractor occurs @xcite . thus , an intermittent single - band sn attractor appears [ _ e.g. , _ see the intermittent sn attractor with @xmath158 and @xmath159 in fig . [ fig : int](f ) for @xmath160 . a typical trajectory of the second iterate of the poincar map @xmath75 ( _ i.e. _ @xmath161 ) spends a long stretch of time near one of the two former attractors ( _ i.e. , _ smooth tori ) , then it bursts out of this region and comes close to the same or other former torus where it remains again for some time interval , and so on . as in the above case of intermittent route to sn attractors , we also obtain the @xmath146 laminar phases from a long trajectory in @xmath161 , and get the average of characteristic time @xmath147 between bursts . as shown in fig . [ fig : int](g ) , the average characteristic time shows a power - law scaling behavior , @xmath148 where @xmath162 @xmath163 . the scaling exponent @xmath151 seems to be the same as that for the case of the intermittent route to sn attractors occurring near the first - order tongue . since the dynamical mechanism for the appearance of intermittent sn attractors near the first - order and second - order tongues are the same ( _ i.e. , _ an intermittent sn attractor appears via a phase - dependent collision between a smooth torus and a nonsmooth ring - shaped unstable set ) , the intermittent sn attractors for both cases seem to exhibit the same scaling behaviors . as @xmath7 is further increased and passes another threshold value of @xmath164 @xmath65 , the sn attractor turns into a chaotic attractor with a positive largest lyapunov exponent @xmath80 , as shown in fig . [ fig : int](h ) . the dynamical mechanisms for the appearance of intermittent sn attractors near the tongues are the same , irrespectively of the tongue order . here , we consider the case of the main first - order tongue , and using the rational approximation to the quasiperiodic forcing , we search for an unstable orbit that causes the intermittent transition via collision with the smooth torus for @xmath165 @xmath65 . for the inverse golden mean @xmath166 $ ] , its rational approximants are given by the ratios of the fibonacci numbers , @xmath167 , where the sequence of @xmath168 satisfies @xmath169 with @xmath170 and @xmath171 . instead of the quasiperiodically forced system with @xmath68 , periodically forced systems with @xmath172 are studied in the rational approximation . as an example , we consider the rational approximation of level @xmath173 . the rational approximation to the smooth torus ( denoted by a black curve ) , composed of stable orbits with period @xmath174 , is shown in fig . [ fig : ra](a ) for @xmath175 @xmath65 . we note that a ring - shaped unstable set , composed of @xmath176 small rings , lies near the smooth torus . at first , each ring is composed of the stable ( shown in black ) and unstable ( shown in gray ) orbits with period @xmath176 [ see the inset in fig . [ fig : ra](a ) ] . however , as @xmath6 is changed such rings make evolution , as shown in fig . [ fig : ra](b ) for @xmath177 @xmath65 . for fixed values of @xmath6 and @xmath7 , the phase @xmath45 may be regarded as a `` bifurcation parameter . '' as @xmath45 is varied , a chaotic attractor appears via an infinite sequence of period - doubling bifurcations of stable periodic orbits in each ring , and then it disappears via a boundary crisis when it collides with the unstable @xmath176-periodic orbit [ see the inset in fig . [ fig : ra](b ) ] . thus , the attracting part ( shown in black ) of each ring is composed of the union of the originally stable @xmath176-periodic attractor , the higher @xmath178-periodic @xmath179 and chaotic attractors born through the period - doubling cascade . on the other hand , the unstable part ( shown in gray ) of each ring consists of the union of the originally unstable @xmath176-periodic orbit [ _ i.e. , _ the lower gray line in the inset in fig . [ fig : ra](b ) ] and the destabilized @xmath176-periodic orbit [ _ i.e. , _ the upper gray line in the inset in fig . [ fig : ra](b ) ] via a period - doubling bifurcation . as the parameters , @xmath6 and @xmath7 , are further changed , both the size and shape of the rings change , and eventually each ring is composed of a large unstable part ( shown in gray ) and a small attracting part ( shown in black ) , as shown in fig . [ fig : ra](c ) for @xmath139 @xmath65 and @xmath180 @xmath65 [ a magnified view is given in fig . [ fig : ra](d ) ] . we also note that new rings appear inside or outside the `` old '' rings . finally , in terms of the rational approximation of level 7 , we explain the mechanism for the intermittent transition occurring in the first - order tongue for @xmath139 @xmath65 [ see figs . [ fig : int](a)-[fig : int](b ) ] . as we approach the border of the intermittent transition in the state diagram , the ring - shaped unstable set comes closer to the smooth torus , as shown in fig . [ fig : ra](c ) for @xmath180 @xmath65 [ see a magnified view in fig . [ fig : ra](d ) ] . as @xmath7 passes a threshold value of @xmath181 @xmath65 , a phase - dependent ( nonsmooth ) collision occurs between the smooth torus and the unstable part ( shown in gray ) of the nonsmooth ring - shaped unstable set . then , the new attractor of the system contains the attracting part ( shown in black ) of the ring - shaped unstable set and becomes nonsmooth , which is shown in fig . [ fig : ra](e ) for @xmath182 @xmath65 [ see a magnified view in fig . [ fig : ra](f ) ] . as @xmath7 is further increased , the chaotic component in the rational approximation to the attractor increases , and eventually for @xmath183 @xmath65 , it becomes suddenly widened via an interior crisis when it collides with the nearest ring [ _ e.g. , _ see fig . [ fig : ra](g ) ] . then , @xmath174 `` gaps , '' where no attractors with period @xmath176 exist , are formed . a magnified gap is shown in fig . [ fig : ra](h ) . we note that this gap is filled by intermittent chaotic attractors . thus , the rational approximation to the whole attractor consists of the union of the periodic and chaotic components . for this case , the periodic component is dominant , and hence the average largest lyapunov exponent @xmath184 becomes negative , where @xmath185 denotes the average over the whole @xmath45 . hence , the rational approximation to the attractor becomes nonchaotic . we note that the 7th rational approximation to the attractor in fig . [ fig : ra](g ) resembles the ( original ) intermittent sn attractor in fig . [ fig : int](b ) , although the level of the rational approximation is low . in this way , the intermittent transition to an sn attractor occurs through two steps in the rational approximation : the phase - dependent ( nonsmooth ) collision and the interior crisis . we have numerically studied dynamical responses of the quasiperiodically forced hh neural oscillator and compared them with those for the periodically forced case . for the case of periodic forcing , a transition from a periodic to a chaotic oscillation has been found to occur via period doublings in both numerical and experimental works . effect of the quasiperiodic forcing on this period - doubling route to chaotic oscillation has been investigated . in contrast to the case of periodic forcing , new type of sn oscillating states have been found to exist between the regular and chaotic oscillating states as intermediate ones . due to their strange geometry , these sn oscillations lead to the occurrence of aperiodic complex spikings , as in the case of chaotic oscillations . hence , sn oscillating states might be a dynamical origin for the complex spikings which are usually observed in cortical neurons . various routes to sn oscillations via fractalization , collision with a smooth unstable torus , and collision with a nonsmooth ring - shaped unstable set have been identified , as in the quasiperiodically forced logistic map @xcite . these sn responses are also found to occur in other neurons exhibiting period - doubling route to chaos ( _ e.g. , _ the morris - lecar neuron and the fitzhugh - nagumo neuron ) under quasiperiodic stimulus @xcite . finally , we suggest an experiment on the quasiperiodically forced squid giant axon and expect that sn spikings to be observed . however , the real biological environment is a noisy one . hence , it is necessary to further investigate the effect of noise on the sn response for real experiment . this type of investigation is beyond the scope of the the present paper , and hence it is left as a future work . m. r. guevara , l. glass , and a. shrier , science * 214 * , 1350 ( 1981 ) ; l. glass , m. r. guevara , a. shrier , and r. perez , physica d * 7 * , 89 ( 1983 ) . k. aihara , t. numajiri , g. matsumoto , and m. kotani , phys . a * 116 * , 313 ( 1986 ) ; n. takahashi , y. hanyu , t. musha , r. kubo , and g. matsumoto , physica d * 43 * , 318 ( 1990 ) ; d. t. kaplan , j. r. clay , t. manning , l. glass , m. r. guevara , and a. shrier , phys . * 76 * , 4074 ( 1996 ) . k. aihara , scholarpedia 3(5 ) : 1786 ( 2008 ) ; see also references therein . l. glass and m. c. mackey , _ from clocks to chaos _ ( princeton university press , princeton , 1988 ) . r. stoop , k. schindler , and l.a . bunimovich , neurosci . 36 * , 81 ( 2000 ) ; nonlinearity * 13 * , 1515 ( 2000 ) . m. ding and j. a. s. kelso , int . j. bifurcation chaos appl . * 2 * , 295 ( 1992 ) . u. feudel , s. kuznetsov , and a. pikovsky , _ strange nonchaotic attractors _ ( world scientific , singapore , 2006 ) . a. prasad , s. s. negi , and r. ramaswamy , int . j. bifuraction chaos appl . * 11 * , 291 ( 2001 ) ; u. feudel , c. grebogi , and e. ott , phys . rep . * 290 * , 11 ( 1997 ) . c. grebogi , e. ott , s. pelikan , and j. a. yorke , physica d * 13 * , 261 ( 1984 ) . k. kaneko , prog . . phys . * 72 * , 202 ( 1984 ) ; t. nishikawa and k. kaneko , phys . e * 54 * , 6114 ( 1996 ) . j. f. heagy and s. m. hammel , physica d * 70 * , 140 ( 1994 ) . a. s. pikovsky and u. feudel , chaos * 5 * , 253 ( 1995 ) . a. prasad , v. mehra , and r. ramaswamy , phys . rev.lett . * 79 * , 4127 ( 1997 ) . kim , w. lim , and e. ott , phys . e * 67 * , 056203 ( 2003 ) ; s .- y . kim and w. lim , j. phys . a * 37 * , 6477 ( 2004 ) . w. lim and s .- y . kim , phys . a * 335 * , 383 ( 2005 ) . w. lim and s .- y . kim , phys . a * 355 * , 331 ( 2006 ) ; s .- y . kim and w. lim , _ ibid . _ * 334 * , 160 ( 2005 ) . kim , s .- y . kim , b. hunt , and e. ott , phys . e * 67 * , 036211 ( 2003 ) . w. l. ditto , m. l. spano , h. t. savage , s. n. rauseo , j. heagy , and e. ott , phys . lett . * 65 * , 533 ( 1990 ) ; w. x. ding , h. deutsch , a. dinklage , and c. wilke , phys . e * 55 * , 3769 ( 1997 ) ; b. p. bezruchko , a. p. kuznetsov , and y. p. seleznev , phys . e * 62 * , 7828 ( 2000 ) ; k. thamilmaran , d. v. senthikumar , a. venkatesan , and m. lakshmanan , phys . e * 74 * , 036205 ( 2006 ) . a. l. hodgkin and a. f. huxley , j. physiol.(london ) * 117 * , 500 ( 1952 ) . k. aihara , g. matsumoto , and y. ikegaya , j. theor.biol . * 109 * , 249 ( 1984 ) . d. hansel , g. mato , and c. meunier , europhys.lett . * 23 * , 367 ( 1993 ) . a. j. lichtenberg and m. a. lieberman , _ regular and stochastic motion _ ( springer - verlag , new york , 1983 ) , p. 283 . e. m. izhikevich , int . j. bifurcation chaos appl . eng . * 10 * , 1171 ( 2000 ) . s .- lee , a. neiman , and s. kim , phys . e * 57 * , 3292 ( 1998 ) . w. lim , s .- y . kim , and y. kim ( unpublished ) . @xmath65 and @xmath71 hz ( @xmath70 ) . time series of @xmath52 for ( a ) @xmath72 @xmath157 , ( b ) @xmath186 @xmath65 , and ( c ) @xmath187 @xmath65 , which correspond to the period-1 , period-2 , and chaotic states in the poincar map @xmath75 ( solid circles represent stroboscopically sampled points in @xmath75 ) , respectively . the largest lyapunov exponent for the chaotic oscillation in ( c ) is @xmath188 . ( d ) bifurcation diagram ( _ i.e. _ , plot of @xmath10 versus @xmath6 ) in @xmath75 . [ fig : pd ] ] plane for @xmath67 @xmath65 and @xmath71 hz in the quasiperiodically forced hh oscillator . regular , sn , and chaotic regions are shown in light gray , gray , and black , respectively . in the regular region , the torus , the doubled torus , and the quadrupled torus exist in the regions denoted by @xmath86 , @xmath87 , and @xmath88 , respectively , and the solid lines represent torus doubling bifurcation curves with terminal points denoted by crosses . sn attractors appear via fractalization , collision with a smooth unstable torus , and collision with a nonsmooth ring - shaped unstable set when passing the heavy solid line , the dashed line , and the dotted line , respectively . [ fig : sd ] ] in fig . [ fig : sd ] for @xmath93 @xmath157 . time series of @xmath52 for ( a ) the quasiperiodic spiking state ( @xmath94 @xmath65 ) , ( b ) the sn spiking state ( @xmath189 @xmath65 ) , and ( c ) the chaotic spiking state ( @xmath190 @xmath65 ) . projections of attractors onto the @xmath98 plane in the poincar map are shown for ( d ) the smooth torus [ corresponding to ( a ) ] , ( e ) the sn attractor [ corresponding to ( b ) ] , and ( f ) the chaotic attractor [ corresponding to ( c ) ] . ( g ) lyapunov - exponent diagram ( _ i.e. _ , plot of @xmath80 vs. @xmath6 ) ; @xmath80 for the sn attractor is shown in black . ( h ) phase sensitivity functions @xmath118 are shown for the smooth torus ( @xmath86 ) [ ( d ) ] and the sn attractor ( sna ) [ ( e ) ] . for the case of the sn attractor , the graph is well fitted with a dashed straight line with slope @xmath191 . ( i ) plot of the phase sensitivity exponent @xmath82 versus @xmath127 @xmath128 for the sn attractor ; @xmath192 . [ fig : frac ] ] in fig . [ fig : sd ] for @xmath132 @xmath65 . stable two - band torus ( denoted by a solid curve ) and its unstable smooth torus ( represented by a short - dashed curve ) for ( a ) @xmath133 @xmath65 and ( b ) @xmath193 @xmath65 . ( c ) sn attractor with @xmath194 and @xmath195 for @xmath196 @xmath65 . ( d ) chaotic attractor with @xmath197 for @xmath139 @xmath65 . [ fig : hh ] ] in fig . [ fig : sd ] for @xmath139 @xmath65 . ( a ) smooth torus with @xmath198 for @xmath140 @xmath157 . ( b ) intermittent sn attractor with @xmath199 and @xmath143 for @xmath200 @xmath157 . ( c ) plot of @xmath201 vs. @xmath202 . the graph is well fitted with a dashed straight line with slope @xmath203 . here @xmath204 is the average characteristic time between bursts and @xmath162 @xmath205 @xmath63 . for each @xmath206 , @xmath204 is calculated from @xmath146 laminar phases in the poincar map @xmath75 . ( d ) chaotic attractor with @xmath207 for @xmath208 @xmath65 . ( e)-(h ) band - merging transition from a two - band torus to an intermittent single - band sn attractor along the route @xmath153 in fig . [ fig : sd ] for @xmath154 @xmath157 . ( e ) smooth two - band torus with @xmath209 for @xmath155 @xmath65 . ( f ) intermittent single - band sn attractor with @xmath158 and @xmath159 for @xmath210 @xmath65 . ( g ) plot of @xmath211 vs. @xmath212 . the graph is well fitted with a dashed straight line with slope @xmath213 . here @xmath204 is the average characteristic time between bursts and @xmath162 @xmath214 @xmath65 ) . for each @xmath206 , @xmath204 is calculated from @xmath146 laminar phases in @xmath161 . ( h ) chaotic attractor with @xmath215 for @xmath216 @xmath157 . [ fig : int ] ] @xmath65 in the rational approximation of level @xmath173 . ( a ) smooth torus ( denoted by a black curve ) and ring - shaped unstable set for @xmath175 @xmath65 . each ring is composed of the attracting part ( shown in black ) and the unstable part ( shown in gray ) . a magnified view of a ring is given in the inset . ( b ) smooth torus ( represented by a black curve ) and an evolved ring - shaped unstable set for @xmath177 @xmath65 . as @xmath6 is decreased , the unstable part ( shown in gray ) in each ring becomes dominant . ( c)-(h ) mechanism for the intermittent transition to an sn attractor along the route @xmath138 in fig . [ fig : sd ] for @xmath139 @xmath65 . ( c ) smooth torus and a ring - shaped unstable set for @xmath180 @xmath65 [ a magnified view near @xmath217 is given in ( d ) ] . new rings appear inside or outside the old rings . ( e ) nonsmooth attractor born via a phase - dependent ( nonsmooth ) collision between the smooth torus and a ring - shaped unstable set for @xmath182 @xmath65 [ ( f ) is a magnified view near @xmath217 ] . ( g ) intermittent sn attractor with @xmath176 gaps ( born via an interior crisis ) for @xmath219 @xmath65 . ( h ) a magnified gap near @xmath217 . [ fig : ra ] ]

wireless access networks are the typical last - mile networks connecting multiple users to a high speed backbone . in these networks , a base station ( bs ) serves a large group of users . traditionally , either the time or the frequency are divided to ensure users do not interfere with each other . modern coding techniques , however , allow for multiple users to either transmit or receive simultaneously , and be decoded successfully using the appropriate multiple access channel ( mac ) codes or broadcast channel ( bc ) codes , respectively . nevertheless , consider a bs serving a very large number of users . in practical scenarios , not all users can be served simultaneously , and the problem of user selection or scheduling arises . in the downlink setting , where a bs transmits to a group of users , it is common to assume channel state information ( csi ) is available at the bs , hence intelligent user selection can be employed . e.g. , the bs can select a subset of the users with both strong channel norms ( high snr ) and relatively orthogonal directions ( to avoid interference ) . see section [ sec . related ] for a literature survey . such a selection can exploit the _ multi - user diversity _ inherent in these models . in the uplink setting , however , since the users transmit to the bs , it is highly desirable to avoid the process of collecting all information at the bs beforehand , and notifying the users which should transmit . this process is prohibitively complex when the number of users is very large . in fact , this is one of the main reasons why an emerging standard , ieee 802.11ac , includes multi - user multiple input multiple output ( mu - mimo ) only at the downlink @xcite . that is , the allegedly complex process of selecting a subset of users in the uplink is refrained from . hence , efficient and distributed algorithms for selecting the appropriate group of users are desirable . this way , there is hope to harness the benefits of multi - user diversity without the need to collect csi from all users . in this paper , we show that this is indeed possible , by showing the order - optimality of distributed algorithms . we consider a mimo mac channel with @xmath0 receiving antennas and @xmath1 users . we suggest distributed algorithms for selecting a group of users to transmit in each slot . in the first , a threshold value for the norm of the channel vector is set , and only users above the threshold transmit . hence , there is no need to collect csi from all users . nor is any cooperation required . in the second , an iterative process is suggested , where _ multiple thresholds _ are set on the norms of the _ projections of the channel vectors _ on the spaces of previously selected users . in this case , csi is shared , but _ only among the selected group_. an analysis of the resulting sum capacity in the limit of large @xmath2 and the respective scaling laws are given for both zero - forcing ( zf ) and minimal mean square error ( mmse ) receivers . this analysis employs recent tools from both point process approximation and asymptotic random matrix theory , that to the best of our knowledge , were not used in this setting before . via this analysis , the simple distributed , threshold based algorithm , is shown to achieve the optimal scaling laws . consequently , the benefit compared to traditional ( centralized ) techniques is demonstrated both analytically and via simulation results . the rest of this paper is organized as follows : section [ sec . related ] includes the most relevant related work . section [ sec . prelim ] includes the required preliminary material . section [ sec . dist alg ] introduces the distributed algorithm , gives its analysis and scaling law under zf decoding . section [ sec . sic ] gives an improved , multi - stage algorithm , which , while only semi - distributed ( as it requires messaging between the selected users ) , gives better performance for a small number of users . section [ sec . mmse rec ] gives the analysis under an mmse receiver . while conceptually similar , this analysis is the more technically challenging . section [ sec : proofs ] gives the previously omitted proofs . section [ sec . conc ] concludes the paper . the essence of multi - user diversity was introduced in @xcite , where selecting the strongest user in each time slot was first suggested . the work was followed by numerous scheduling algorithms for various scenarios . we list here only the most relevant . in @xcite , the authors considered the impact of multi - user diversity on the mimo downlink channel ( bc ) . assuming channel state information at the bs , the authors used order statistics to evaluate the effective snr when scheduling the strongest user in each slot . however , _ only one user was scheduled _ in each slot , and the results were given in terms of the @xmath2-fold statistics , without an extreme value analysis for large @xmath2 . in @xcite , a similar downlink model was considered , however , when users are scheduled simultaneously . the authors considered zero - forcing beamforming ( zfbf ) , and suggested a greedy algorithm to schedule the strongest and most orthogonal users . additional scheduling algorithms for downlink communication were given in @xcite . in fact , in the downlink scenario , it was shown later that zfbf and optimal user selection can indeed achieve the dirty paper coding ( dpc ) region @xcite , and is hence optimal in the gaussian case @xcite . additional surveys and scaling laws can be found in @xcite . a closely related scheme , yet still for the downlink model , was suggested in @xcite . using block diagonalization ( bd ) , a capacity - based greedy algorithm was suggested , in which first the strongest user is scheduled , and then additional users are added , one by one , based on their marginal contribution to the total capacity . in the same context , @xcite considered the special case of two transmit antennas and one receive antenna per user , and showed that a greedy , two - stage algorithm , which first selects the strongest user and then the second to form the best pair is asymptotically optimal . in the context of heterogeneous users , @xcite proposed a scheduling scheme which selects a small subset of the users with favorable channel characteristics . the above works focus on the downlink setting . in this scenario , it is reasonable to assume that at least some information is available at the bs , and a centralized decision can be made . in the uplink ( mac ) model , however , if one wishes to select a group of users without centralized processing at the bs , distributed algorithms are required . in this paper , we suggest both a single - stage distributed algorithm , and a multi - stage semi - distributed one for the uplink scenario , and , in addition , analyze their sum capacity in the limit of large number of users and give the resulting scaling laws . a pioneering study of the uplink model was done in @xcite , where a decentralized mac protocol for orthogonal frequency division multiple access ( ofdma ) channels was suggested . in this scheme , each user estimates the channel gain and compares it to a threshold . only above - the - threshold users can transmit . @xcite extended the scheme to a multi - channel setup , where each user competes on @xmath3 channels . in @xcite , the authors used a similar approach for power allocation in the multi - channel setup , and suggested an algorithm that asymptotically achieves the optimal water filling solution . however , the works above do not consider a _ mimo setting _ , nor do they consider the interaction within a group of users , when all are scheduled to use _ the same resources_. space - time coding for fading multi - antenna mac was considered in @xcite . the focus therein , however , was on joint code design for a given point in the rate region and the resulting error probability , rather than user scheduling and its resulting capacity . recently , we proposed a point process approximation which facilitates the analysis of various distributed threshold - based scheduling algorithms in the non - homogeneous scenario @xcite . this work , however , assumed only a single user can be successfully decoded in each time slot . a key contribution of the current work is the non - trivial extension of the work in @xcite to truly multiple - access protocols , where several users transmit simultaneously and should be decoded successfully , hence the questions that arise are how to distributively select a good subset of users to transmit and what the mutual influence between the users in the selected group is . for example , a closely related work is @xcite . therein , various decoding procedures were discussed , and the corresponding best user selection _ for the uplink setting _ was given . however , while reinforcing the necessity of proper user selection , the work in @xcite considered only the scenario where _ one user _ can access the radio channel at a given time . as for more complex topologies , spatial diversity in the context of multiple relays was considered in @xcite . therein , communication between a source and a destination is done through a group of relays . however , unlike conventional relay schemes , only the relays with the _ strongest received signal _ decode the message and cooperate via space time coding to successfully relay it to the destination . an asymptotically optimal scheme for multiple base stations ( with joint optimization ) was given in @xcite . extreme value theory ( evt ) is a key tool in proving capacity results under scheduling and multi - user diversity . in @xcite , the authors suggested a sub - carrier assignment algorithm , and used order statistics to derive an expression for the resulting link outage probability . in @xcite , the authors used evt to derive the scaling laws for scheduling systems using beamforming and linear combining . @xcite analyzed the scaling laws of base station scheduling , and showed that by scheduling the strongest among @xmath2 stations one can gain a factor of @xmath4 in the expected capacity ( compared to random or round - robin scheduling ) . in this section , we describe the system model and relevant results which will be used throughout this paper . throughout this paper , random matrices and random vectors are denoted in bold upper - case and bold lower - case letters , respectively . we consider a multiple - access model with k users , each with a single transmit antenna . the bs is equipped with @xmath0 receiving antennas . when @xmath5 users utilize the channel simultaneously , the received signal at the base station can be described as : @xmath6 where @xmath7 is the transmitted signal ( scalar ) . @xmath8 is constrained in its total power to @xmath9 , i.e. , @xmath10\leq p$ ] . however , in most cases , we will assume a _ constant _ power constraint @xmath9 . @xmath11 denotes the uncorrelated gaussian noise . @xmath12 is a complex random gaussian channel vector . when all users are identically and independently distributed , it is common to assume that all entries of @xmath13 are independent and have zero mean and variance @xmath14 imaginary and real parts , for all users . we assume that the channel is memoryless , that is , for each channel use ( slot ) , independent realizations of @xmath15 are drawn . furthermore , we assume full csi is available at the transmitter . that is , @xmath13 is known to the @xmath16th user . this can be accomplished by sending a pilot signal form each of the @xmath0 antennas at the base station . evt is a key tool in evaluating the capacity under scheduling in multi - user systems . we review here the most relevant result . moreover , we develop new normalizing constants for the problem at hand ( evt for the @xmath17 distribution ) , which will later aid at speeding up convergence results . the capacity obtained by letting an arbitrary user @xmath16 utilize the channel is given by @xmath18 . however , as mentioned , it is beneficial to schedule the strongest user in each slot . denote by @xmath19 the received channel vector with the largest norm . scheduling the strongest user clearly results in @xmath20 , while letting a _ group of @xmath0 orthogonal users _ , with the largest channel gains to utilize an uplink channel , results in a sum - rate that has the following upper bound:@xmath21 for the downlink channel , the sum - rate upper bound has the form @xcite @xmath22 the difference between the uplink to the downlink originates in the power constraint applied to the transmitter . that is , in the downstream , when transmitting to a group of receivers , each receiver gets a share of the available power , while in the upstream , there is a group of transmitters that transmit to a single receiver . it should be noted that usually the power constraint @xmath9 for the downlink and uplink are not equal , since the base station has a strong and steady power supply , whereas the user has a limited battery power supply . as the sum rate is mainly influenced by the channel vectors gains and directions , our goal is to explore this behavior for large number of users . specifically , we first wish to explore the behavior of the maximal gain . since the entries of @xmath23 are complex gaussian , the channel s gain follows a @xmath17 distribution with @xmath24 degrees of freedom , denoted @xmath25 . we utilize the following evt theorem . [ theorem . evt ] let @xmath26 be a sequence of i.i.d . random variables with distribution @xmath27 , and let @xmath28 . if there exists a sequence of normalizing constants @xmath29 and @xmath30 such that as @xmath31 , @xmath32 for some non - degenerate distribution g , then g is of the generalized extreme value ( gev ) distribution type @xmath33 and we say that @xmath27 is in the domain of attraction of @xmath34 , where @xmath35 is the shape parameter , determined by the ancestor distribution @xmath27 . the normalizing constants and the shape parameter of the gev can be obtained as follows . let @xmath36 be the reciprocal hazard function @xmath37 where @xmath38 and @xmath39 are the lower and upper endpoints of the ancestor distribution , respectively . the shape parameter @xmath35 is obtained as the following limit @xcite : @xmath40 when @xmath41 is a sequence of i.i.d . @xmath42 variables , the asymptotic distribution of @xmath43 is a gumbel distribution @xcite . specifically , @xmath44 where @xmath45 and @xmath46 is the gamma function . in the sequel , we will also use the upper incomplete gamma function , @xmath47 . however , for i.i.d . @xmath25 random variables , the convergence of the maxima to the gumbel distribution using the above normalizing constants is quite slow . that is , the approximation of the maximal value will not be tight for moderate values of @xmath48 . hence , a more appropriate set of normalizing constants for the @xmath25 distribution , which takes into account both @xmath0 and @xmath49 should be derived . letting @xmath30 to be the @xmath50 quantile , i.e. , @xmath51 , and choosing @xmath52 @xcite , we have the following . for the @xmath25-distribution , the following normalizing constants apply . @xmath53 where @xmath54 is the inverse of the regularized upper incomplete gamma function , that is , @xmath55 , and the inverse is defined with respect to @xmath56 . the @xmath17 distribution is a special case of the gamma distribution . i.e. , if @xmath57 then @xmath58 , where @xmath59 is the gamma distribution with shape parameter @xmath0 and rate parameter @xmath60 . accordingly , for the @xmath61 constant we consider the @xmath50 quantile of the gamma distribution , which can be obtained by using the inverse of the regularized upper incomplete gamma function . in particular , @xmath62 yields the @xmath50 quantile of the gamma distribution . to attain the @xmath63 constant , let us examine the hazard function @xmath36 of the gamma distribution . @xmath64 accordingly , for @xmath65 we obtain , @xmath66 figures [ fig : max_norm_dist30 ] and [ fig : max_norm_dist ] depict the simulation results versus the analytical results of the evt with the new normalizing constants derived herein ( @xmath63 and @xmath61 given in and , respectively ) . the maximum of @xmath2 @xmath17-distributed random variables is compared to the gumbel distribution predicted by evt . figure [ fig : max_norm_dist30 ] depicts the case where @xmath67 while in figure [ fig : max_norm_dist ] for @xmath68 . in both cases , figures are plotted for several values of receiving antennas ( @xmath0 ) . tight convergence is clearly visible , with reasonable approximation even for @xmath69 users . to see how the new normalizing constants relate to the previous ones reported in the literature , figure [ fig : an_convergence ] depicts the value of @xmath63 for several values of @xmath0 , as @xmath49 increases . while the new constant converges to @xmath70 slowly , it is constant for moderate values of @xmath49 , hence the tight convergence of _ the distribution of the maximum to the gumbel distribution_. degrees of freedom , respectively . the graphs are given for @xmath67 users . ] degrees of freedom , respectively . the graphs are given for @xmath68 users . ] for @xmath71 receiving antennas to the limit @xmath72 . note that this slow convergence does not affect the convergence of the actual distribution , as the values of @xmath63 are within a constant factor from @xmath70 even for very moderate @xmath49 . ] as we aim at analysing the capacity under practical constraints such as scheduling only a subset of the users in each time slot , we focus our attention on linear decoding at the bs . such decoders , such as the zf decoder or the mmse decoder , are indeed widely used in practice . hence , the analysis in this paper will be based on either linear decorrelation ( section [ sec . dist alg ] ) or the mmse receiver ( section [ sec . mmse rec ] ) , assuming optimal coding of the resulting single user gaussian channels , given the effective signal to noise ratio ( snr ) . specifically , for the zf receiver , focusing on the signal received from the @xmath73th user , rewrite as : @xmath74 let @xmath75 be a unitary matrix representing the null space of the subspace spanned by @xmath76 . since the entries of the channel vectors are i.i.d . , when @xmath5 users transmit the subspace spanned by the vectors @xmath76 has rank @xmath77 with probability one ( * ? ? ? * chapter 8) . thus , to decode , the receiver projects the received vector @xmath78 on the subspace spanned by @xmath75 , and nulls the inter - stream interference . finally , the signal of user @xmath73 can be demodulated using a matched filter ( i.e. , maximal ratio combiner ) . the algorithm is given in figure [ alg . note that a full degrees - of - freedom gain is attained when @xmath0 users transmit . in this case , @xmath79 , and @xmath80 , e.g. , @xcite . accordingly , when using a zf receiver , we aim at algorithms which select at most @xmath0 users ( of the available @xmath2 ) in each time slot . as mentioned , since we focus on the scenario in which @xmath1 , the set of selected users has a crucial effect on the system capacity . optimally , a bs would receive csi from all users , and schedule the @xmath0 best users for transmission . under the linear decorrelation above , the resulting expected capacity is @xmath81 however , we wish to avoid the overhead and complexity of such a centralized process , and select a group of users , approximating the optimal selection , distributively . while simple and intuitive , the zf receiver is limited in its performance . the mmse receiver , however , although still linear , maximizes the mutual information and hence achieves better performance ( e.g. @xcite ) . in this receiver , to decode the @xmath82 data stream , the receiver treats the rest of the streams as noise . it then whitens the resulting colored noise and uses a matched filter to obtain maximum sinr . let @xmath83 be a matrix whose columns are the channel vectors of the transmitting users . similarly , let @xmath84 be the matrix @xmath83 with its @xmath82 column removed and define @xmath85 then , the corresponding output @xmath86 on the stream @xmath16 can be expressed by @xcite : @xmath87 this sinr value will be at the basis of our analysis in section [ sec . mmse rec ] . a common approach to select a single user distributively , is a threshold - based procedure , in which a capacity threshold is set , and only a user who exceeds it transmits ( @xcite ) . of course , the events in which none of the users or several users exceed the threshold should be taken into account . in this paper , however , we wish to select a group of users , and analyze the resulting capacity . at the heart of the algorithms we suggest herein , stands a similar threshold - based procedure . however , the challenge is twofold . first , in selecting a threshold such that a favorable _ group of users _ exceed it . second , in analyzing the results under the various decoding procedures and at the limit of large @xmath2 . when doing this , a few important questions arise : on which variable should a threshold be set and how many users will pass it ? how can one assess the mutual interference between the users which passed ? what will be the loss in this distributed procedure compared to the optimal , centralized one ? in the next three sections , we answer the above questions . we set a threshold on the _ channel norms _ , and analyze the resulting exceedance rate . we further analyze the mutual interference , in terms of the _ angles _ between the exceeding users , and conclude by analysing the resulting _ sum capacity _ , showing that a distributed algorithm can achieve the _ same scaling laws _ as a centralized one . in particular , we suggest two distributed algorithms . in the first , described in this section , a single threshold is utilized when decoding is done using linear decorrelation . in the second , described in section [ sec . sic ] , we offer a _ set of thresholds _ to match a successive interference cancellation ( sic ) procedure . given the number of users @xmath2 , we set a threshold @xmath88 on the norm @xmath89 , such that @xmath90 strongest users exceed it on average . in each slot , each user estimates its channel s norm . a user with a norm greater than the threshold , transmits . we assume the transmission includes the channel vector as a low - rate preamble so the bs has the csi _ of the transmitting users_. the algorithm for user @xmath16 is given in figure [ alg . channel access ] . note that the receiver can not recover more than @xmath0 data streams . that is , since for more than @xmath0 users the performance ( both under zf decoding and mmse decoding in section [ sec . mmse rec ] ) deteriorates significantly , if more than @xmath0 users begin transmission simultaneously , we assume a collision occurs and the whole slot is lost ( zero capacity ) . similarly , since users act independently , a slot might be idle , if no user exceeded the threshold . thus , we say that a slot is utilized if at least one user , but no more than @xmath0 users , are transmitting . the first result , proposition [ c zf uniform users exact ] below , gives the sum capacity under the above distributed user selection algorithm and zf decoding . note that this simple proposition still includes an expectation on the channel vectors seen by the users , hence can not give the understanding we wish regarding the sum capacity under the suggested algorithm . still , it will be the starting point , from which we will derive the bounds which give the right insight and scaling laws . [ c zf uniform users exact ] for @xmath1 , the expected sum capacity of algorithm channel - access with zf decoding is given by @xmath91 + o\left(\frac{\log \log k}{k}\right),\end{aligned}\ ] ] where @xmath5 , to be optimized , is the expected number of users to exceed the threshold @xmath88 , the @xmath92 are the channel vectors of the users who exceeded the threshold and @xmath93 are the corresponding null spaces . according to the law of total probability , we express the expected capacity in a slot by summing over the number of users who exceed the threshold , and the sum capacity these users see , given that they exceeded the threshold . as mentioned , if more than @xmath0 users are transmitting in a slot , the receiver can not successfully null the inter - stream interference , and the capacity in that slot is zero . hence , the expected capacity has the form : @xmath94.\ ] ] when the users are i.i.d . , the probability of @xmath73 threshold exceedances follows the binomial distribution with probability @xmath95 to exceed the threshold . since we consider large @xmath2 and small values of @xmath5 , the number of users to exceed threshold can be approximated by the poisson distribution with an approximation error in the order of @xmath96 . this is based on the poisson point process approximation developed in @xcite for the single - user scenario . in short , this method gives a poisson approximation for the number of users exceeding a high threshold @xmath88 . however , as this approximation error is within the sum , it is multiplied by the individual capacities , which scales like the optimal scaling law of the multi - user diversity when a single , strongest user , is scheduled , i.e. , @xmath97 ( see e.g. , @xcite and the references therein ) . hence , the @xmath98 approximation error . finally , note that the number of exceeding users @xmath73 affects the effective snr seen by the attending users . in particular , when @xmath73 users exceed threshold , the dimension of @xmath99 is @xmath100 . thus , as @xmath73 decreases , the signal of the attending users is projected on a less restrictive null - space . accordingly , each stream may spread on more receiving antennas in the zf process , which leads to a higher power gain ( for details on zf decoding , see @xcite ) . nonetheless , the reader should not be confused . the highest capacity is attained when @xmath0 users utilize the channel simultaneously to achieve a full degrees - of - freedom gain . to ease notation , the @xmath98 approximation error is omitted from now on . to evaluate the result in proposition [ c zf uniform users exact ] , the behavior of @xmath101 should be understood , especially considering the fact that the number of users exceeding the threshold is random . to this end , the following upper and lower bounds are useful . these bounds will be the basis of the scaling laws we derive . [ c zf uniform users upper ] the expected sum capacity of algorithm channel - access with zf decoding satisfies the following upper bound @xmath102 where @xmath103 is given by and @xmath88 is the threshold set such that @xmath5 users exceed it on average . the bound in lemma [ c zf uniform users upper ] , while not giving the exact capacity , still depicts the essence of the system behavior . to understand its implications , we note the following : we set a threshold such that @xmath5 out of the @xmath2 users exceed it on average . i.e. , the average exceedance rate is @xmath5 . indeed , the expression @xmath104 in the sum over @xmath73 gives the probability for exactly @xmath73 users exceeding . each of the @xmath73 users , under zero forcing , experiences a single user channel , with its power @xmath9 scaled according to two factors : ( i ) a multiplication by @xmath105 , as this is the average norm of its channel vector , where @xmath88 is the threshold exceeded , and @xmath106 is the average distance _ above _ the threshold . ( ii ) a multiplication by @xmath107 , as in case only @xmath108 users exceeded the threshold , the zero forcing algorithm does not have to cancel @xmath109 users , only @xmath110 , hence the null space has a larger dimension , yet the number of receive antennas is @xmath0 . as the threshold @xmath111 will be shown to be @xmath112 , the optimal scaling law will follow . a complete discussion will be given after the lower bound is introduced . indeed , as it turns out in the simulation results , the bound in lemma [ c zf uniform users upper ] is tight even for relatively small number of antennas and users . note , however , that if less than @xmath0 users exceed , as a higher snr can be attained at the receiver , in order to achieve the capacity in this case , a user must know how many users exceeded , so it can exploit the high snr for , e.g. , higher transmission rate . hence , we require that the number of users that actually exceeded the threshold will be announced . note that in practice it is beneficial to choose @xmath5 slightly smaller than the number of antennas @xmath0 . this is since if less than @xmath0 users exceed , the snr seen by each user is only larger , yet if more than @xmath0 users exceed , the slot is lost . we start with proposition [ c zf uniform users exact ] . by jensen inequality , @xmath113 \nonumber\\ & \leq & \sum_{j=1}^{r}\frac{k^{j}e^{-k}}{j ! } \sum_{i=1}^{j } \log\big(1 + p \text{e}\left[\left . \arrowvert { \mathbf{v}}_i { \mathbf{h}}_i \arrowvert^2 \right| \vert { \mathbf{h}}_i \vert^2 > u_k \right ] \big).\label{eq . in upper bound proof}\end{aligned}\ ] ] consider the norm @xmath101 , where @xmath99 has @xmath114 rows . denoting by @xmath115 the @xmath3th row of @xmath99 , we have @xmath116 & = & \text{e}\left[\left.\sum_{m=1}^{r - j+1 } \vert \langle{\mathbf{v}}_i^{(m)},{\mathbf{h}}_i\rangle \vert^2 \right| \vert { \mathbf{h}}_i \vert^2 > u_k\right]\\ & \stackrel{(a)}{= } & \sum_{m=1}^{r - j+1 } \text{e}\left[\left . \arrowvert { \mathbf{h}}_i \arrowvert^2 \frac{\vert \langle{\mathbf{v}}_i^{(m)},{\mathbf{h}}_i\rangle \vert^2}{\arrowvert { \mathbf{h}}_i \arrowvert^2 \arrowvert { \mathbf{v}}^{(m)}_i \arrowvert^2 } \right| \vert { \mathbf{h}}_i \vert^2 > u_k\right]\\ & \stackrel{(b)}{= } & \text{e}[\left.\arrowvert { \mathbf{h}}_i \arrowvert^2 \right| \vert { \mathbf{h}}_i \vert^2 > u_k ] \sum_{m=1}^{r - j+1 } \text{e } \left [ \frac{\vert \langle{\mathbf{v}}_i^{(m)},{\mathbf{h}}_i\rangle \vert^2}{\arrowvert { \mathbf{h}}_i \arrowvert^2 \arrowvert { \mathbf{v}}^{(m)}_i \arrowvert^2}\right]\\ & \stackrel{(c)}{= } & \text{e}[\left.\arrowvert { \mathbf{h}}_i \arrowvert^2 \right| \vert { \mathbf{h}}_i \vert^2 > u_k](r - j+1 ) \int_{0}^{1}(1-\alpha)^{r-1}d \alpha\\ & \stackrel{(d)}{= } & ( u_k + a_k ) ( r - j+1 ) \frac{1}{r}.\end{aligned}\ ] ] in the above chain of equalities , ( a ) is since @xmath117 ( b ) is since @xmath13 is a random i.i.d . complex normal vector , and the squared - normalized inner product @xmath118 is its angle from @xmath115 , a vector in the null space of @xmath119 . since these vectors are independent of @xmath13 , this angle is independent of the norm of @xmath13 ( c ) is since the distributions of the norms and angles are independent of @xmath3 , and since , by ( * ? ? ? * lemma 3.2 ) , the angle has the same distribution as the minimum of @xmath109 independent uniform @xmath120 $ ] random variables ( i.e. , with cdf @xmath121 , @xmath122 ) ( d ) is the result of computing the expected norm of an i.i.d . complex normal random vector , _ given that it is above a threshold which is exceeded by only @xmath5 out of @xmath2 norms on average_. the details are in corollary [ coro : expected capacity above est by gev thr ] , section [ sec : proofs ] . substituting in , we have @xmath123 which completes the proof . we now present a corresponding lower bound . [ c zf uniform users lower ] the expected sum capacity of algorithm channel - access with zf decoding satisfies the following lower bound . @xmath124 where @xmath88 is a threshold set such that @xmath5 users exceed it on average . it is important to note that the integral in lemma [ c zf uniform users lower ] above has a finite series expansion with @xmath0 summands . this finite series has @xmath125 at the leading term , resulting in the expected scaling law . we describe it in claim [ claim . integral result ] below , within the proof of the main result in this section - theorem [ theorem : c zf uniform users - scaling ] . following the derivations of the upper bound , we have @xmath126 \\ & \stackrel{(a)}{\ge } \sum_{j=1}^{r}\frac{k^{j}e^{-k}}{j ! } \sum_{i=1}^{j } \text{e}\log\left(1 + p u_k \sum_{m=1}^{r - j+1 } \frac{\langle{\mathbf{v}}_i^{(m)},{\mathbf{h}}_i\rangle^2}{\arrowvert { \mathbf{h}}_i \arrowvert^2 \arrowvert { \mathbf{v}}^{(m)}_i \arrowvert^2 } \right ) \\ & \stackrel{(b)}{= } \sum_{j=1}^{r}\frac{k^{j}e^{-k}}{j!}j \text{e}\log\left(1 + p u_k \sum_{m=1}^{r - j+1 } \frac{\langle{\mathbf{v}}_{i'}^{(m)},{\mathbf{h}}_{i'}\rangle^2}{\arrowvert { \mathbf{h}}_{i ' } \arrowvert^2 \arrowvert { \mathbf{v}}^{(m)}_{i ' } \arrowvert^2 } \right ) \\ & \ge \sum_{j=1}^{r}\frac{k^{j}e^{-k}}{j ! } j \text{e}\log\left(1 + p u_k \frac{\langle{\mathbf{v}}_{i'}^{(1)},{\mathbf{h}}_{i'}\rangle^2}{\arrowvert { \mathbf{h}}_{i ' } \arrowvert^2 \arrowvert { \mathbf{v}}^{(1)}_{i'}\arrowvert^2 } \right ) \\ & \stackrel{(c)}{= } \sum_{j=1}^{r}\frac{k^{j}e^{-k}}{j ! } j \int_{0}^{1}(r-1)(1-\alpha)^{r-2}\log\left(1+p u_k \alpha\right ) d \alpha\end{aligned}\ ] ] where ( a ) is since the norms of _ all users participating _ are above the threshold @xmath88 ; ( b ) is since the angles in the inner sum are identically distributed and independent of @xmath16 , hence an arbitrary @xmath127 can be used ; ( c ) is by explicitly computing the expectation over the angle between @xmath13 and @xmath128 , remembering that it has a density @xmath129 for @xmath122 . this completes the proof . the results above lead to the following scaling law , which is the main result in this section . it asserts that the scaling law of @xmath130 for the sum rate in a multi - user system can in fact _ be achieved distributively _ , without collecting all channel states from all users and scheduling them in a centralized manner . in other words , the threshold based algorithm suggested selects an optimal _ set of users _ ( asymptotically in the number of users ) distributively and without any cooperation . this is summarized in the next theorem . [ theorem : c zf uniform users - scaling ] the expected sum capacity of algorithm channel - access with zf decoding scales as @xmath131 for large enough number of users @xmath2 . by lemma [ c zf uniform users upper ] , @xmath132 on the other hand , consider the lower bound given in lemma [ c zf uniform users lower ] . since @xmath5 is a parameter to be optimized , the optimum is at least as large as when choosing @xmath133 . we have @xmath134 where the last inequality is by evaluating the sum at @xmath135 . note that larger values of @xmath0 give only slightly larger values , with a limit of @xmath136 as @xmath137 . , calculated at @xmath109 . the limiting behavior can be found in @xcite . ] now , consider the integral over @xmath138 in lemma [ c zf uniform users lower ] . in section [ sec : proofs ] , we prove the following claim . integral result ] the integral over @xmath138 in lemma [ c zf uniform users lower ] has the following finite series expansion : @xmath139 this gives a finite series expansion for the integral , in terms of the power @xmath9 and the threshold @xmath140 . for example , for @xmath141 we have @xmath142 thus , the integral can be easily approximated by @xmath143 . since we consider the regime of large enough number of users @xmath2 , yet a finite number of antennas @xmath0 , we have @xmath144 and as a result @xmath145 . in fact , since the distribution of the projected channel gain seen by a user in this decoding scheme is the exponential distribution with rate @xmath14 , it can be shown that @xmath146 ( we discuss the threshold value in detail in section [ sec : thr est ] ) . this gives rise to the @xmath147 scaling law . it is well known that linear decorrelation is asymptotically optimal at high snr . the suggested algorithms pick the users with the highest snr , accordingly , they operate at high snr and loose only a constant fraction of the optimal centralized scheduling scheme . extensive simulations were conducted to compare the analytical bounds derived above to real world situations with a finite number of users . in figure [ fig : mac_sim248 ] , we compare the bounds on the expected capacity of the threshold - based scheduling scheme under zf - receiver ( i.e. , lemma [ c zf uniform users upper ] and lemma [ c zf uniform users lower ] ) , to the simulation results . the bounds and simulation are for @xmath68 users , with @xmath148 and @xmath149 receiving antennas at the bs . the tightness of the upper bound is clearly visible . while the lower bound is looser , it still gives the correct behaviour as a function of the number of users passing the threshold on average , @xmath5 . indeed , it is clear from the figure that a key factor affecting the system performance is the number of users passing the threshold . this distribution gives the graphs their poisson - like shape . to see that the trend holds even for a relatively small number of users , figure [ fig : mac_sim24830 ] includes the same plots for @xmath67 . note that even for @xmath69 users the algorithm manages to achieve a significant rate ( compared , e.g. , to the one achieved with @xmath150 users ) . this means the essence of the multi - user diversity is exploited by the algorithm even for a relatively small number of users . note , however , that as @xmath0 approaches @xmath2 then the accuracy of the bounds decreases . the results in figure [ fig : mac_sim248 ] and figure [ fig : mac_sim24830 ] give the expected sum capacity as a function of @xmath5 . the capacity distribution , compared to different algorithms , will be given in the subsequent sections . . bars are simulation results , while the solid and dashed lines represent the upper and lower bounds , respectively . the threshold @xmath88 is set such that @xmath5 users exceed it on average . the green ( upper ) , red ( middle ) and blue ( lower ) lobes are for @xmath151 and @xmath70 receive antennas , respectively . note that the optimal @xmath5 is smaller than @xmath0 . this is since it is better to aim at slightly less than @xmath0 users passing , as the performance deteriorates significantly for more than @xmath0 users transmitting concurrently . ] . bars are simulation results , while the solid and dashed lines represent the upper and lower bounds , respectively . the threshold @xmath88 is set such that @xmath5 users exceed it on average . the green ( upper ) , red ( middle ) and blue ( lower ) lobes are for @xmath151 and @xmath70 receive antennas , respectively . note that the essence of the analysis holds for even a relatively small number of users as @xmath69 , as the bounds still depict the key elements affecting the performance . ] in the previous section , only a single threshold was used , and a user s data was decoded by projecting the received signal on the null space of the sub - space spanned by the channels of the interfering users . while this algorithm managed to capture the key performance - enhancing aspects of the multi - user system , and achieve the optimal scaling laws distributively , better results can be expected for low to moderate number of users . in that regime , harnessing additional multiple access techniques and a more sophisticated thresholding algorithm can achieve better performance . in this section , we see that this is indeed so . when using sic ( * ? ? ? * ch.6 ) , the decoder uses the decoded signal of a previous user to decode the next one ( by subtracting it from the received stream ) . after @xmath109 iterations of sic , all data streams are decoded . with this in mind , we derive a second , iterative algorithm , to further utilizing the benefits of sic in a threshold based algorithm . at first glance , for large number of users , a _ set of thresholds _ @xmath152 should be chosen , such that in each iteration only a single user exceeds on average . this way , one can adapt the thresholds set according to the users admitted " thus far ( that is , users who already passed ) . for example , one can set the next threshold based on the angles of the previously admitted users . however , it is very likely that in some of the @xmath0 iterations , more than one user , or no user , will exceed the threshold . specifically , the probability that exactly one user will exceed in an iteration is approximately @xmath153 . yet , if a carrier sense , or collision detection mechanism is available , collisions can be resolved by allocating a few mini - slots devoted to finding the strongest user in each iteration @xcite . this gives rise to the following algorithm : after the strongest user is found , it begins its transmission by announcing its channel vector . in the @xmath154th iteration , the rest of the users project their channel vector on the orthonormal basis @xmath155 , which spans the null space of the channels vectors announced thus far . now , the scheduled user for iteration @xmath154 is the one with the _ strongest projection_. the process ends when @xmath0 users are selected . note that this involves announcements of the channel vectors only from the @xmath156 selected users . the algorithm is described in figure [ alg . @xmath157 ; @xmath158 when a user projects its channel on @xmath159 , the resulting gain distribution follows the @xmath17-distribution with @xmath160 degrees of freedom @xcite . that is , @xmath161 . let @xmath162 be the channel with the strongest norm in iteration @xmath154 , i.e. , @xmath163 note that in practice , the bs _ does not have to search for the maximum above _ , but , instead , a threshold is set , only the strongest users pass , and one is selected using a collision resolution mechanism . utilizing the evt for the @xmath164-distribution , the norm @xmath165 can then be characterized . we have the following . [ prop : c uniform users sic ] for sufficiently large @xmath2 , the expected sum capacity of algorithm sic - channel - access with zf - sic decoding is given by : @xmath166 \nonumber , \label{eq : expected capacity sic}\ ] ] where @xmath167 is the channel vector of with the largest norm after the @xmath168 projection . note that the above result assumes each transmitting user can adapt its transmission rate to be decoded successfully at the base station . this is possible as each user _ knowns the channel vectors on the interfering users _ and his own channel vector , thus can calculate the snr the bs will experience when decoding . furthermore , a splitting algorithm is used @xcite , hence the poisson coefficients can be omitted . again , to evaluate the performance of proposition [ prop : c uniform users sic ] , the distribution of @xmath169 should be examined . we now derive upper and lower bounds to better evaluate the performance unders the suggested algorithm . [ c sic uniform users upper ] the expected sum capacity of algorithm sic - channel - access with zf - sic decoding satisfies the following upper bound . @xmath170 where @xmath171 are the normalizing constant given in ( [ eqn : a_n normalized ] ) and ( [ eqn : b_n normalized ] ) respectively , and @xmath172 is the euler gamma constant . note that @xmath173 ( as it converges to @xmath174 in equation ) , hence the same scaling law is achieved . by jensen inequality , @xmath175 \nonumber\\ & \leq & \sum_{l=1}^{r } \log\big(1 + p \text{e } \left [ \arrowvert { \mathbf{h}}^{(l ) } \arrowvert^2 \right ] \big)\nonumber\\ & = & \sum_{l=1}^{r } \log\big(1 + p \text{e } \left [ \max_{1\leq i\leq k}\vert { \mathbf{v}}^{(l-1)}{\mathbf{h}}_i \vert^2 \right ] \big).\nonumber \ ] ] to analyse the expectation some care is needed . without conditioning in the @xmath154-th iteration on the event that the remaining users did not come out ahead in the previous iterations , that is , allowing " all users to compete again , the norm @xmath176 follows @xmath177-distribution . hence , by theorem [ theorem . evt ] , @xmath178 follows the gumbel distribution , with normalizing constants @xmath179 and @xmath180 given and , respectively . however , when such a conditioning is applied , the expectation can only decrease , as the previous users are not allowed to participate . thus , the expected maximal norm is upper bounded by the expected value of the gumbel distribution with the above parameters . lemma [ c sic uniform users upper ] then follows . to bound the performance from below , we only derive probabilistic lower bound . that is , a set of fixed relatively low threshold values are computed , such that in each iteration , we can guarantee that the norm of the strongest channel is greater than that value with high probability . however , in order to attain a non - trivial bound , these values need to be as high as possible , yet , maintain the lower bound with high probability . specifically , in each iteration , we chose a threshold value that is greater than the strongest norm with probability @xmath181 ( hence the probability that no user exceeds it is @xmath181 ) . this corresponds to a threshold that @xmath182 users exceed it on average . nonetheless , depending on the strictness of the bound , a different threshold sets can be chosen . accordingly , we have the following . [ c sic uniform users lower ] the expected sum capacity of algorithm sic - channel - access with zf - sic decoding satisfies the following lower bound with high probability . @xmath183 where @xmath184 is a threshold such that @xmath182 projected channel norms exceeds it on average , in iteration @xmath154 . let us consider the probability that the no user exceeds the threshold @xmath184 on iteration @xmath154 . this can be expressed as @xmath185 since @xmath186 is monotone , @xmath187 let @xmath188 , namely , the event in which the strongest channel in iteration @xmath154 is greater than threshold @xmath184 . we have @xmath189 that is , the probability that at least one user exceeds the threshold in each iteration , is greater than @xmath190 . as a result , since each l.h.s . element is greater than threshold with high probability , we conclude that @xmath191 with high probability consequently , @xmath192 \geq \sum_{l=1}^r \log(1+u_{\log k}^{(l)}).\ ] ] note that the sic scheme above achieves the same scaling law as the former linear decorrelation scheme ( clearly , one can not do better asymptotically ) , yet , it also achieves higher power gain , i.e. , there is a boost in performance since the effective snr is increasing in each iteration . this is clearly seen in the simulations below , where the results are clearly better for the finite population we tested . specifically , figure [ fig : mac_distributions4 ] depicts the gain of multi - user scheduling with and without sic , and compares it to scheduling schemes which schedule a single user at a time . receiving antennas and @xmath68 users , each with a single transmitting antenna . green ( left ) : a single arbitrary user is scheduled ( approximately gaussian ) . yellow ( second left ) : the strongest user is scheduled ( approximately gumbel ) . pink ( second right ) : a group of 4 strongest users are scheduled , with zf receiver . blue ( right ) : a group of 4 users are scheduled , with the suggested multi - threshold algorithm and zf - sic receiver . ] note that the capacity distribution in each iteration under sic can be approximated by the gumbel distribution . further , the sum - capacity distribution under sic is also approximated by the gumbel distribution , since the gumbel distribution is infinitely divisible @xcite . this is also consistent with figure [ fig : mac_distributions ] , which zooms in on the comparison between the capacity distributions : when the receiver uses zf versus zf - sic to decode the data streams . receiving antennas and @xmath68 users , each with a single transmitting antenna . the left ( pink ) lobe depicts the capacity when a group of @xmath141 users are scheduled , and the receiver demodulates the data streams by null forcing ( all users are above a threshold , according to the algorithm in section [ sec . dist alg ] ) . the right ( blue ) lobe depicts the capacity when a group of @xmath141 users are scheduled ( using the multi - threshold algorithm ) , and the receiver demodulates the data streams using sic . ] note that in the suggested sic algorithm , the users are naturally ordered from the strongest to the @xmath154-strongest user . hence , to improve the sum capacity , one might devise a distributed water - filling algorithm ( e.g. @xcite ) , such that each user will invest a power amount proportional to its order . the zf receiver discussed thus far is on the one hand simple enough to facilitate rigorous analysis , yet , as shown in the previous sections , powerful enough in the sense that with intelligent user selection ( in this paper , distributed ) can achieve the optimal scaling laws . still , this is not the optimal linear receiver . in this section , we explore the scaling laws of the expected capacity of the mmse receiver . as mentioned , in this case we let @xmath83 denote a matrix whose columns are the channel vectors of the _ transmitting users_. that is , when using the channel - access algorithm , vectors with norm greater than the threshold . @xmath84 is the matrix @xmath83 with its @xmath82 column removed . under these definitions , the sinr seen at the @xmath16th stream was given in . let @xmath193 denote the set of channels with norm greater than a threshold . then , the expected capacity under the threshold based scheduling algorithm in section [ sec . dist alg ] is as follows . [ c mmse uniform users][prop . c mmse ] for @xmath1 , the expected sum capacity with mmse decoding is @xmath194 = & \sum_{j=1}^{r}\frac{k^{j}e^{-k}}{j ! } \sum_{i=1}^j \text{e}\left[\left.\log\left(1 + p { \mathbf{h}}_i^{\dagger}{\mathbf{r}}{\mathbf{h}}_i \right)\right| \vert { \mathbf{h}}_s \vert^2 > u_k , \forall s \in \mathcal{s } \right ] + o\left(\frac{\log \log k}{k}\right)\end{aligned}\ ] ] where @xmath88 is the threshold set such that @xmath5 users exceed it on average . this expected capacity should be optimized over @xmath5 . the proof is similar to the zf setting . the only change is in the snr seen by the users , as reflected by the term within the @xmath195 . when zf decoding was used , the effective snr seen by a user was @xmath196 . as the matrix @xmath99 is unitary , it is clear why scaling up @xmath13 resulted in scaling up the snr . in the mmse case , however , the analysis is intricate . the snr is @xmath197 . while @xmath13 scales up , @xmath198 may appear to scale down as the eigenvalues of @xmath199 scale up . however , note that @xmath199 is not full rank , hence has at least one zero eigenvalue . as a result , @xmath200 has at least one eigenvalue which does not scale down . to evaluate the expected capacity in propositions [ prop . c mmse ] , the characteristics of the random variables @xmath201 should be understood , especially when the number of transmitting users is random . further , the influence of the norms @xmath202 on @xmath201 should be evaluated . the key technical challenge , however , is due to the norm condition _ inducing dependence on the matrix elements _ , hence the random matrix theory usually used in the mimo literature does not hold . part of the contribution in this section , is by bringing new tools to tackle this problem . note that in this section , regular type letters represent random variables as well as scalar variables . the difference will be clear from the context . as a first tool to handle the dependence within the matrix entries , we start with the following claim : [ claim . entries var of above thr vec ] assume the norms @xmath203 are above a given threshold @xmath88 . then the following properties hold : * the entries of the channel vector remain zero mean . * the variance of each entry in the vector scales . in particular , is equals to @xmath204 = ( u_k + a_{\{k , r\}})/r . \ ] ] * the vector elements remain uncorrelated in pairs . note that , to begin with , the entries of @xmath83 are i.i.d . the claim states that conditioned on exceeding a threshold , while not i.i.d . , they sustain the zero correlation . this property will be useful throughout the remainder of this paper . the proof of claim [ claim . entries var of above thr vec ] is deferred to section [ sec : proofs ] . now we are ready to derive the scaling - law of the mmse receiver , using the following upper and lower bounds on the threshold based expected capacity . c mmse upper ] the expected sum capacity of algorithm channel - access with mmse decoding satisfies the following upper bound . @xmath205 \leq \sum_{j=1}^{r}\frac{k^{j}e^{-k}}{j ! } j \\ \log\left(1 + p\bigg ( 1 - \frac{(j-1 ) u_k^2}{r\left((1+\frac{j}{r})(u_k + a_{\{k , r\}})^2 + a_{\{k , r\ } } ( a_{\{k , r\}}+1)+ u_k\right)}\bigg ) ( u_k + a_{\{k , r\ } } ) \right ) . \ ] ] where @xmath103 is given by ( [ eqn : a_n normalized ] ) and @xmath88 is the threshold set such that @xmath5 users exceed it on average . before we prove the lemma , it is interesting to compare the scaling law under mmse decoding to that achieved with zf decoding . in both lemma [ c zf uniform users upper ] and lemma [ lem . c mmse upper ] , the capacity seen by each user is approximately @xmath206 , for some constant @xmath207 . asymptotically , it follows that @xmath208 , and in both cases , the capacity gain comes from the threshold value @xmath88 , which is , as mentioned , @xmath209 . this gives the growth rate of @xmath210 per user . however , note that while @xmath211 in lemma [ c zf uniform users upper ] , for large enough @xmath88 the gain in lemma [ lem . c mmse upper ] is asymptotically @xmath212 , that is , a larger gain for any @xmath213 . this is not surprising , as an mmse decoder does give a better power gain , but does not improve the already optimal scaling law . the capacity seen by user @xmath16 is bounded by : [ eq : mmse upper bound ] [ c_i ( u_k ) ] & ( 1 + p ) + & ( 1 + p ) + & ( 1 + p _ n=1^r_m=1^r ) + & ( 1 + p _ n=1^r ) + & ( 1 + p ) in the above chain of equalities , ( a ) follows from jensen inequality . ( b ) is the equivalent quadric form of @xmath214 . ( c ) is since the vector @xmath13 is independent in the column vectors of the matrix @xmath200 . ( d ) is since by claim [ claim . entries var of above thr vec ] part ( iii ) , the elements of the vector @xmath13 are uncorrelated . ( e ) follows from claim [ claim . entries var of above thr vec ] part ( ii ) . now , to address @xmath215 , conditioned on @xmath216 , we first use the following upper bound on the trace of the inverse of a matrix . for a symmetric @xmath217 positive definite matrix a @xcite , @xmath218 where @xmath219 is the smallest eigenvalue of the matrix @xmath220 . in our case , take @xmath221 . to obtain @xmath222 , we recall that @xmath223 and that @xmath224 is an @xmath217 matrix with rank @xmath110 , with @xmath225 , thus , by the eigenvalues decomposition theorem , @xmath224 has @xmath114 of its eigenvalues equal to zero . hence , the smallest eigenvalue of @xmath226 is @xmath227 . thus , using ( [ eq : trace inequality ] ) , it follows that @xmath228 - \text{tr}^2\left({\mathbf{r}}^{-1}\right ) - r^2}{\text{tr}\left({\mathbf{r}}^{-\dagger}{\mathbf{r}}^{-1}\right ) - \text{tr}\left({\mathbf{r}}^{-1}\right)}\\ & = & \frac{r\left[\text{tr}\left({\mathbf{h}}_{(-i)}{\mathbf{h}}_{(-i)}^{\dagger } { \mathbf{h}}_{(-i)}{\mathbf{h}}_{(-i)}^{\dagger}\right ) + \text{tr}\left({\mathbf{h}}_{(-i)}{\mathbf{h}}_{(-i)}^{\dagger}\right ) \right ] - \text{tr}^2\left({\mathbf{h}}_{(-i)}{\mathbf{h}}_{(-i)}^{\dagger}\right)}{\text{tr}\left({\mathbf{h}}_{(-i)}{\mathbf{h}}_{(-i)}^{\dagger } { \mathbf{h}}_{(-i)}{\mathbf{h}}_{(-i)}^{\dagger}\right ) + \text{tr}\left({\mathbf{h}}_{(-i)}{\mathbf{h}}_{(-i)}^{\dagger}\right)}\\ & = & r - \frac{\text{tr}^2\left({\mathbf{h}}_{(-i)}^{\dagger}{\mathbf{h}}_{(-i)}\right)}{\text{tr}\left({\mathbf{h}}_{(-i)}^{\dagger}{\mathbf{h}}_{(-i)}{\mathbf{h}}_{(-i)}^{\dagger } { \mathbf{h}}_{(-i)}\right ) + \text{tr}\left({\mathbf{h}}_{(-i)}^{\dagger}{\mathbf{h}}_{(-i)}\right)}. \label{eq : inverse trace bound}\end{aligned}\ ] ] note that the expression in ( [ eq : mmse upper bound ] ) had a power loss of factor @xmath0 , that is , @xmath9 is divided by @xmath0 . however , it is multiplied by @xmath229 $ ] . in the last equality above we actually see that in this multiplication we gain back a factor of the form @xmath230 . this suggests that the mmse decoder is superior to the zf decoder in terms of power gain , as the expression in the zf decoder performance does not have this multiplicative factor ( this in consistent with , e.g. , @xcite ) . we will later evaluate the value of @xmath231 explicitly . furthermore , note that @xmath232 is actually the squared frobenius norm of @xmath226 , which is equal to the sum of squares of it entries . thus , we have , & & + & r - & + & = r - & + & = r - & where @xmath233 is the squared inner product . note that the numerator inside the expectation is at least @xmath234 . thus , & & + & r + & now , we have a concave function of the form @xmath235 inside the expectation . accordingly , using jensen inequality for concave functions , we obtain that & & + & r - ( j-1)^2 u_k^2 ( ) ^-1 & + & = r - ( j-1)^2 u_k^2 ( _ ^j . & + & + 2_^j_^j & + & + . _ ^j)^-1 & + where the last equality follows from the linearity of the expectation and since @xmath236 , which is the angle between two independent vectors , @xmath237 and @xmath238 , is independent of the vectors norms . moreover , we can now remove the conditioning on the norm when taking the expectation of the angle . accordingly , an upper bound can be evaluated by computing the expectations in the denominator as follows . & r - ( j-1)^2 u_k^2 ( _ ^j . & + & + 2_^j_^j & + & + . _ ^j)^-1 & + & r - ( j-1)^2 u_k^2 & + & ( ( j-1)(u_k^2 + 2a_\{k , r}u_k + 2a_\{k , r}^2 ) + j(j-1 ) ( u_k + a_\{k , r})^2 + ( j-1)(u_k + a_\{k , r } ) ) ^-1 & + & = r - , where ( a ) follows since the limit distribution of the channel norm tail is exponentially distributed with rate parameter @xmath239 given that it is above high threshold . namely , @xmath240 , @xcite . thus , @xmath241 $ ] can be interpreted as a second moment of exponential random variable . similar to section [ sec . dist alg ] , by ( * ? ? ? * lemma 3.2 ) , the angle has the same distribution as that of the minimum of @xmath109 independent uniform @xmath120 $ ] random variables ( i.e. , with cdf @xmath121 , @xmath122 ) . the expectations of the norms , again , follow from the evt . details are in corollary [ coro : expected capacity above est by gev thr ] , section [ sec : proofs ] . to completely characterize the performance of the suggested threshold - based algorithm under mmse decoding , we proceed to derive a corresponding lower bound . to this end , the following claims will be useful . the proofs are deferred to section [ sec : proofs ] . r is unitarily invariant ] let @xmath83 be a complex gaussian matrix with i.i.d . entries . then , * @xmath242 is unitary invariant . in particular , @xmath200 can be decomposed as @xmath243 , where @xmath244 is a unitary matrix , independent of the diagonal matrix @xmath245 . * the property above holds conditioned on all norms of the columns @xmath13 of @xmath83 being above a threshold @xmath140 . in particular , conditioned on @xmath246 , @xmath247 and @xmath13 have the same distribution . lower bound on upper gamma s geq 2 ] for @xmath248 , we have @xmath249 . lower bound on expectation ] let @xmath250 and let @xmath251 , independent of @xmath252 , for some integer @xmath253 . then , @xmath254 > \left(\frac{\gamma(r ) \gamma(r - j+1,u/2)}{\gamma(r , u/2)\gamma(r - j+1 ) } + \frac{e^{-u/2}u^{r-1}}{2^{r-1}\gamma(r , u/2)}\right)\log(u ) \\ \quad \quad + \frac{e^{-u/2}u^{r-1}}{2^{r-1}\gamma(r , u/2 ) } \left ( \psi(r - j+1 ) - \psi(r ) \right ) + e \frac{\gamma(r ) \gamma(r - j,1+u/2)}{\gamma(r , u/2)\gamma(r - j+1 ) } , \ ] ] where @xmath255 is the digamma function . to understand claim [ claim . lower bound on expectation ] , note that @xmath256 . hence , it is not hard to show that @xmath257 moreover , using claim [ claim . lower bound on upper gamma s geq 2 ] above , one can also show that @xmath258 hence , the importance of claim [ claim . lower bound on expectation ] is in showing that @xmath259=\theta(\log u_k)$ ] . similar to the upper bound , since @xmath145 , the scaling law of @xmath260 will result . to conclude , a lower bound on the performance of our threshold - based algorithm with mmse receiver is as follows . for simplicity of the presentation , we assume @xmath261 . c mmse lower ] for sufficiently large @xmath2 , the expected sum capacity under the threshold - based algorithm and mmse decoding satisfies the following lower bound . @xmath205 > \sum_{j=2}^{r}\frac{k^{j}e^{-k}}{j ! } j \left[\left(\frac{\gamma(r ) \gamma(r - j+1,u_k/2)}{\gamma(r , u_k/2)\gamma(r - j+1 ) } + e^{-u_k/2}\frac{u_k^{r-1}}{2^{r-1}\gamma(r , u_k/2)}\right)\log(u_k ) \right . \\ \quad \quad \left . + e^{-u_k/2}\frac{u_k^{r-1}}{2^{r-1}\gamma(r , u_k/2 ) } \left ( \psi(r - j+1 ) - \psi(r ) \right ) + e \frac{\gamma(r ) \gamma(r - j,1+u_k/2)}{\gamma(r , u_k/2)\gamma(r - j+1 ) } \right].\end{gathered}\ ] ] we assume @xmath253 users begin their transmission simultaneously and let @xmath193 denote the set of channels with norm greater than threshold . note that when only one user passed the threshold , while its capacity is easy to compute , as it is the single - user miso capacity , within this bound it is negligible as it is multiplied by the probability ( under the poisson distribution ) that only one user passed the threshold . hence , it is omitted for simplicity . in this case , the capacity of stream @xmath16 under mmse receiver satisfies the following lower bound . [ c_i ( u_k ) ] & = & + & & + & & + & & + & & + & & + & = & in the above chain of inequalities , ( a ) is since by claim [ claim . r is unitarily invariant ] , there is a unitary matrix @xmath244 such that @xmath262 with @xmath263 . as a result , we have the following quadric form : @xmath264 ( b ) is since the eigenvalues of @xmath200 are the reciprocals of those of @xmath226 . ( c ) and ( d ) follow since , first , the dimensions of @xmath224 is @xmath217 and its rank is @xmath110 . thus , @xmath114 of the eigenvalues of @xmath226 ( corresponding to the zero eigenvalues of @xmath224 ) are equal to one . then , as @xmath224 is positive semidefinite , the non - zero eigenvalues are non negative . in fact , the eigenvalues of @xmath265 which are not unity , are relatively large as all columns of @xmath83 are above a threshold . ( e ) is by claim [ claim . r is unitarily invariant ] ( ii ) . now , setting @xmath266 and @xmath267 , claim [ claim . lower bound on expectation ] completes the proof . in figure [ fig : mmse_sim ] we simulate the expected capacity of the threshold - based scheduling scheme , with mmse - receiver and compare the simulation results to the upper and lower capacity bounds given in lemma [ lem . c mmse upper ] and lemma [ lem . c mmse lower ] , respectively . in figure [ fig : zflower_vs_mmselower ] we compare the mmse receiver lower bound ( dot - dashed line ) and the zf receiver lower bound ( dashed line ) . as expected the mmse achieves better performance when the threshold is lower , which corresponds to a lower snr , and further , to a lower idle slot probability . on the other hand , as the threshold gets higher , which translates to high snr , the zf yields same results as mmse . from this observation , since the mmse receiver requires lower threshold to achieve higher sum - rate , we comprehend that the mmse receiver is preferable to the zf receiver , both in terms of sum - rate and idle slot probability , yet , still has a linear decoding complexity . . the threshold @xmath88 is set such that @xmath5 users exceed it on average . bars are the capacity under the threshold based scheme simulation results under mmse receiver , while the solid and dashed lines represent the upper and lower bounds , respectively . the orange ( upper ) , green ( 2nd from above ) , red ( 3rd from above ) and blue ( lower ) lobes are for @xmath268 and @xmath70 receive antennas , respectively . note that the optimal @xmath5 is smaller than @xmath0 . ] and @xmath70 receive antennas , respectively . note that since one should choose @xmath5 for best results , the mmse lower bound predicts better results , achieved for a smaller @xmath5 per @xmath0 ( the number of receive antennas ) . ] the methods discussed in the previous sections are based on a common baseline procedure . first , given that the norm threshold was exceeded , we wish to get a handle on the capacity distribution . then , we wish to express the rate at which users exceed the threshold , in order to understand how likely is that a given number of users exceed a certain threshold . finally , a threshold is estimated . this threshold is set according to the fraction of users which are required to exceed it on average . in this paper , the threshold is set on the channel vector norm , either directly or after projecting it on the null - space of the already chosen users . note that we used a similar approach to analyze the expected rate , when the channel capacity is approximately gaussian in @xcite . in this section , we discuss the following problems directly for the problem at hand . once a threshold is set , it is important to evaluate both the distribution of the number of users which exceed it as well as the conditional distribution of the norm given that the threshold was exceeded . herein , we utilize the point process approximation @xcite and its specific usage for threshold arrival rates in the _ single user case _ @xcite in order to derive these distributions for the problem at hand . assume that @xmath269 is a sequence of i.i.d . random variables with a distribution function @xmath27 , such that @xmath27 is in the domain of attraction of some gev distribution g , with normalizing constants @xmath270 and @xmath30 . we construct a sequence of points @xmath271 on @xmath120 \times { \mathbb{r}}$ ] by @xmath272 and examine the limit process , as @xmath273 . consider @xmath274 on the set @xmath120 \times ( b_l + \epsilon , \infty)$ ] , where @xmath275 . by @xcite , @xmath276 , where @xmath9 is a non - homogeneous poisson process with intensity density @xmath277 . @xmath56 is the sample value , and @xmath278 is the index of occurrence . in fact , in the i.i.d . case , the process intensity density @xmath279 is independent of @xmath278 . let @xmath280 be the expected number of points in the set @xmath281 . @xmath280 can be obtained by integrating the intensity of the poisson process over @xmath281 , that is @xmath282 . as we are interested in sets of the form @xmath283 \times ( v,\infty)$ ] , where @xmath284 ( threshold exceedance ) , we have @xmath285 , where @xmath286 denotes @xmath287 ( e.g. , @xcite ) . that is , the number of users whose channel norms exceed a threshold can be modeled by a poisson process , with parameter @xmath288 . note that for @xmath289 , as in the case of the @xmath17 distribution , we have @xmath290 , hence setting a threshold @xmath291 results in a per - user arrival rate of @xmath292 , and , as a result , a total arrival rate of @xmath5 . to compute the conditional distribution , we proceed similar to @xcite . for any fixed @xmath293 , let @xmath294 , and let @xmath295 . then @xmath296^{-1/\xi}\end{aligned}\ ] ] where @xmath297 and @xmath298 are the corresponding excess value @xmath299 at index @xmath278 , and the corresponding excess value at time @xmath278 in the limit process , respectively . namely , the result is a generalized pareto distribution , @xmath300 . note that for the @xmath17 distribution , @xmath301 , and we have @xmath302 for all @xmath303 . that is , the tail distribution can be approximated using an exponential distribution with rate @xmath304 . as a result , taking the expected value , we obtain the following corollary . [ coro : expected capacity above est by gev thr ] given a random complex gaussian vector , and a threshold such that @xmath5 users out of @xmath2 exceed it on average , then @xmath305 = u_k + a_{\{k , r\}},\ ] ] where @xmath103 is given in . let @xmath88 be a threshold such that only the @xmath5 users with the strongest channel norm @xmath306 will exceed it on average . since @xmath306 follows @xmath42 distribution , @xmath88 can be easily estimated using the inverse - gamma function . that is , @xmath307 note , however , that the expression above does not give any intuition on the actual threshold value , or , more importantly , how it scales with @xmath2 for fixed @xmath5 . yet , note that the threshold @xmath88 is closely related to the normalizing constant @xmath308 , since both aim to capture the last quantiles of the ancestor distribution . the exact relation can be obtain using the point process approximation . specifically , the exceedance rate for the @xmath17-distribution is @xmath309 accordingly , choosing @xmath310 gives a rate of @xmath5 in total . thus , letting @xmath103 and @xmath308 to be set according to and , respectively , @xmath88 is set . now , to see the scaling law , note that @xmath308 in converges to @xmath174 in ( [ eqn : b_n normalized slow ] ) , hence @xmath311 . figure [ fig : convergence of b ] depicts the above convergence result . in and the value in ( [ eqn : b_n normalized slow ] ) . the curves are plotted for @xmath268 and @xmath70 , top to bottom . ] we wish to solve the following definite integral : @xmath312 we first make the substitution @xmath313 . as a result , we obtain the following integral : @xmath314 now , using integration by parts , we have @xmath315 first , note that @xmath316 now , performing a polynomial long division , the term inside the integral can be expressed as @xmath317 finally , we exchange the integration with the ( finite ) summation , then integrate each term in the sum according to @xmath318 . we have : @xmath319 thus , claim [ claim . integral result ] follows . it can be shown by the total law of expectation , that the entries of the channel vector remain zero , given that the channel norm is greater than a threshold , as follows . @xmath320 & = & \text{e}\left [ h_{i , n } \big| \sum_{n=1}^r \vert h_{i , n } \vert^2 > u_k\right]\\ & = & \text{e}\left[\left . \text{e}\left [ h_{i , n } \big| \sum_{n=1}^r \vert h_{i , n } \vert^2 > u_k , \sum_{\substack{m=1 , \\ m\neq n}}^r \vert h_{i , m } \vert^2 = \tilde{u } \right]\right| \sum_{n=1}^r \vert h_{i , n } \vert^2 > u_k \right]\\ & = & \text{e}\left[\left . \underbrace{\text{e}\left [ h_{i , n } \big| \vert h_{i , n } \vert^2 > u_k - \tilde{u } \right]}_{(\ast)}\right| \sum_{n=1}^r \vert h_{i , n } \vert^2 > u_k \right ] \ ] ] since @xmath321 is complex gaussian random variable , the inner expectation @xmath322 can be decomposed to a real and an imaginary part as @xmath323&= & \text{e}\left [ a + i b \big| a^2+b^2 > u_k - \tilde{u } \right]\\ & = & \text{e}\left [ a \big| a^2+b^2 > u_k - \tilde{u } \right]+ i \text{e}\left [ b \big| a^2+b^2 > u_k - \tilde{u } \right]\\ \ ] ] to ease notation , let us denote @xmath324 . focusing on the real part , we use the total expectation law again to obtain : @xmath325 & = & \text{e}\left[\left . \text{e}\left[a \big| a^2 > u_k -\tilde{u } - b^2 , b^2\right ] \right| a^2 + b^2 > u_k - \tilde{u } \right]\\ & = & \text{e}\left[\left . \text{e}\left[a \big| \vert a \vert > \tilde{\tilde{u } } , b^2\right ] \right| a^2 + b^2 > u_k - \tilde{u } \right]\\ & = & \text{e}\left[\left.\underbrace{\text{e}\left[sign(a ) \vert a \vert \big| \vert a \vert > \tilde{\tilde{u } } , b^2\right ] } _ { ( \ast \ast ) } \right| a^2 + b^2 > u_k - \tilde{u } \right ] . \ ] ] now , we show that the sign of @xmath326 is independent of the amplitude @xmath327 and @xmath328 . @xmath329 thus , the sign of @xmath326 , the amplitude @xmath327 and @xmath328 are independent . consequently , the inner expectation @xmath330 is equal to : @xmath331&= & \text{e}\left[sign(a)\right]\text{e}\left [ \vert a \vert \big| \vert a \vert > \tilde{\tilde{u } } , b^2\right]\\ & = & 0 . \ ] ] from symmetry , same result can be obtained for the imaginary part . thus , the vectors are still zero mean vectors . however , the variance of the vector entries increase , given that the vector norm is greater than a threshold . remember , by evt @xcite , the tail is exponentially distributed with rate parameter @xmath239 . thus , it follows that @xmath332 \\ & \stackrel{(b)}{= } & \sum_{n=1}^r \text{e}\left[\left . \vert h_{i , n } \vert^2\right| \vert { \mathbf{h}}_i \vert^2 > u_k \right ] \\ & \stackrel{(c)}{= } & r \text{e}\left[\left . \vert h_{i , n } \vert^2\right| \vert { \mathbf{h}}_i \vert^2 > u_k \right ] \label{eq : element conditional expectation}\end{aligned}\ ] ] where ( a ) is the result of computing the expected norm of an i.i.d . complex normal random vector , given that it exceeded a threshold @xmath88 . ( b ) follows from the linearity of expectation operator . ( c ) is since the elements of @xmath13 are identically distributed . we point out that choosing vectors with norm greater than a threshold increase the variance of the entries to @xmath333 . similar to the entries conditional expectation , it can be shown that the matrix entries are still uncorrelated in pairs . namely , @xmath334&\\ & = \text{e}\left [ h_{i , n}^{\ast}h_{i , m } \big| \sum_{n=1}^r \vert h_{i , n } \vert^2 > u_k\right]&\\ & = \text{e}\left[\left . \text{e}\left [ h_{i , n}^{\ast}h_{i , m } \big| \sum_{n=1}^r \vert h_{i , n } \vert^2 > u_k , \sum_{\substack{m=1 \\m\neq n}}^r \vert h_{i , m } \vert^2 = \tilde{u } \right]\right| \sum_{n=1}^r \vert h_{i , n } \vert^2 > u_k \right]&\\ & = \text{e}\left[\left . \underbrace{\text{e}\left [ h_{i , n}^{\ast}h_{i , m } \big| \vert h_{i , n } \vert^2 > u_k - \tilde{u } \right]}_{(\tilde{\ast})}\right| \sum_{n=1}^r \vert h_{i , n } \vert^2 > u_k \right ] & \ ] ] similar to the method above , we decompose @xmath335 to its real and imaginary part . thus , the inner expectation @xmath336 is @xmath337&\\ & = \text{e}\left [ a h_{i , m } - i b h_{i , m } \big| a^2 + b^2 > u_k - \tilde{u } \right]&\\ & = \text{e}\left [ a h_{i , m } \big| a^2 + b^2 > u_k - \tilde{u } \right]- i \text{e}\left [ b h_{i , m } \big| a^2 + b^2 > u_k - \tilde{u } \right ] & \ ] ] using the total expectation law on the real part of the expectation , we have @xmath338&\\ & = \text{e}\left[\left . \text{e}\left [ a h_{i , m } \big| a^2 > u_k - \tilde{u } - b^2 , b^2 \right ] \right| a^2 + b^2 > u_k - \tilde{u } \right]&\\ & = \text{e}\left[\left . \underbrace{\text{e}\left [ sign(a ) \vert a \vert h_{i , m } \big| \vert a \vert > \tilde{\tilde{u } } , b^2 \right]}_{(\tilde{\ast \ast } ) } \right| a^2 + b^2 > u_k - \tilde{u } \right ] . & \ ] ] now , similar to the above method , let us show that the sign of @xmath326 is independent of @xmath327 and @xmath339 . @xmath340 thus , the inner expectation @xmath341 can be expressed as @xmath342&= & \text{e}\left[sign(a)\right]\text{e}\left [ \vert a \vert h_{i , m } \big| \vert a \vert > \tilde{\tilde{u } } , b^2\right]\\ & = & 0 . \ ] ] from symmetry , same result can be obtained for the imaginary part . thus , given that the norm is greater than a threshold , the vector entries sill are uncorrelated . thus , claim [ claim . entries var of above thr vec ] follows . a hermitian random matrix is unitary invariant if the joint distribution of its entries does not change under unitary transformation , namely , @xmath343 for any unitary matrix @xmath344 . it is well known that a gaussian matrix is unitary invariant ( e.g. , @xcite ) . it is not hard to see that @xmath345 . that is , the wishart matrix is also unitary invariant . similarly , as @xmath346 , @xmath347 . since @xmath348 is unitary invariant , by ( * ? ? ? * lemma 2.6 ) , there is a decomposition such that @xmath349 , @xmath244 being unitary and independent of @xmath350 , hence @xmath351 and part ( i ) follows with @xmath352 . note that if we are only interested in the decomposition , and the independence between the unitary matrix and the eigenvalues is not required ( as is the case in the lower bound below ) , the proof is simpler since @xmath200 is hermitian . for part ( ii ) , note that if @xmath353 , then @xmath354 . this is since the unitary transform does not change the vector norm , hence the set of instances @xmath355 with @xmath356 is the same as those with @xmath357 . the rest follows in the same manner . first , note that , @xmath358,\ ] ] where @xmath359 . accordingly , since @xmath360 is convex in @xmath278 for the corresponding @xmath361 , we apply jensen s inequality to the r.h.s . and note that @xmath362 = 1 $ ] , proving the bound . we wish to bound the following expectation , when @xmath250 and @xmath251 , independent of @xmath252 : @xmath363 = \frac{1}{\pr\left({\mathbf{x}}+{\mathbf{y}}>u\right ) } \\ \left[\int_{x=0}^{u } \int_{y = u - x}^{\infty}\log(1+x)f_{{\mathbf{x}}}(x ) f_{{\mathbf{y}}}(y ) dy dx + \int_{x = u}^{\infty } \int_{y=0}^{\infty}\log(1+x)f_{{\mathbf{x}}}(x ) f_{{\mathbf{y}}}(y)dy dx \right].\end{gathered}\ ] ] performing integration on @xmath364 , we have the following . [ eq . mmse lower bound ] & e[(1+x)|+ > u ] & + & = & where @xmath365 is the upper incomplete gamma function . further , note that the complement cdf of the @xmath366-distribution is @xmath367 . accordingly , to evaluate ( [ eq . mmse lower bound ] ) , we need to evaluate both @xmath368 and @xmath369 . we begin with integration by parts on @xmath368 , where the anti - derivative is @xmath370 . we have @xmath371 where the last equality follows since @xmath372 . now , to bound the above expression , we restrict our attention to @xmath253 . since @xmath373 , utilizing claim [ claim . lower bound on upper gamma s geq 2 ] , we have @xmath374 next , let us address @xmath369 . again , using claim [ claim . lower bound on upper gamma s geq 2 ] , we have @xmath375 let @xmath376 . accordingly , we have @xmath377\end{aligned}\ ] ] where @xmath378 is the beta function . note that while @xmath379 as @xmath380 , the integral converges as this is the mean of @xmath381 when @xmath382 has a beta distribution . accordingly , we have @xmath383.\ ] ] to complete the proof , remember that @xmath384 . thus , we have @xmath385 > \frac{\gamma(r)}{\gamma(r , u/2)}\bigg(\frac{e^{-u/2 } u^{r-1}}{2^{r-1 } \gamma(r ) } \left[\log ( u ) + \psi(r - j+1 ) - \psi(r ) \right ] \\ + \frac{\gamma(r - j+1,u/2)}{\gamma(r - j+1 ) } \log(1+u ) + e\frac{\gamma(r - j,1+u/2)}{\gamma(r - j+1 ) } \bigg ) . \end{gathered}\ ] ] in this paper , we suggested a distributed multiuser scheduling algorithm which utilizes the multiuser diversity and achieves the optimal capacity scaling laws ( for large number of users ) . specifically , we offered threshold - based algorithms , and characterized the scaling law of the expected capacity under linear decoding , e.g. , zf , zf - sic and mmse . to support the results , we provided both tractable analysis which gave insightful results as well as simulations which showed tightness even for moderate number of users . we concluded that the distributed algorithm achieves the optimal scaling laws , i.e. , @xmath386 . li , ping and paul , debashis and narasimhan , ravi and cioffi , john , `` on the distribution of sinr for the mmse mimo receiver and performance analysis , '' in _ information theory , ieee transactions on _ , vol 52 , 2006 , pp . 271286 . j. kim , s. park , j. lee , j. lee , and h. jung , `` a scheduling algorithm combined with zero - forcing beamforming for a multiuser mimo wireless system , '' in _ ieee vehicular technology conference _ , vol . 1 , 2005 , pp . 211215 . m. airy , s. shakkattai , and r. heath jr , `` spatially greedy scheduling in multi - user mimo wireless systems , '' in _ the thirty - seventh asilomar conference on signals , systems and computers _ , vol . 1.1em plus 0.5em minus 0.4emieee , 2003 , pp . 982986 . c. swannack , e. uysal - biyikoglu , and g. wornell , `` low complexity multiuser scheduling for maximizing throughput in the mimo broadcast channel , '' in _ proc . allerton conf . communications , control and computing _ , 2004 . g. primolevo , o. simeone , and u. spagnolini , `` channel aware scheduling for broadcast mimo systems with orthogonal linear precoding and fairness constraints , '' in _ ieee international conference on communications _ , vol . 4 , 2005 , pp . 27492753 . t. yoo , n. jindal , and a. goldsmith , `` finite - rate feedback mimo broadcast channels with a large number of users , '' in _ information theory , 2006 ieee international symposium on_.1em plus 0.5em minus 0.4em ieee , 2006 , pp . 12141218 . t. yoo and a. goldsmith , `` on the optimality of multiantenna broadcast scheduling using zero - forcing beamforming , '' _ selected areas in communications , ieee journal on _ , vol . 24 , no . 3 , pp . 528541 , 2006 . h. weingarten , y. steinberg , and s. shamai , `` the capacity region of the gaussian multiple - input multiple - output broadcast channel , '' _ information theory , ieee transactions on _ , vol . 52 , no . 39363964 , 2006 . z. shen , r. chen , j. andrews , r. heath , and b. evans , `` low complexity user selection algorithms for multiuser mimo systems with block diagonalization , '' _ signal processing , ieee transactions on _ , vol . 54 , no . 9 , pp . 36583663 , 2006 . k. jagannathan , s. borst , p. whiting , and e. modiano , `` efficient scheduling of multi - user multi - antenna systems , '' in _ modeling and optimization in mobile , ad hoc and wireless networks , 2006 4th international symposium on_.1em plus 0.5em minus 0.4emieee , 2006 , pp . x. qin and r. berry , `` exploiting multiuser diversity for medium access control in wireless networks , '' in _ infocom 2003 . twenty - second annual joint conference of the ieee computer and communications . ieee societies _ , vol . 2.1em plus 0.5em minus 0.4emieee , 2003 , pp . 10841094 . k. bai and j. zhang , `` opportunistic multichannel aloha : distributed multiaccess control scheme for ofdma wireless networks , '' _ vehicular technology , ieee transactions on _ , vol . 55 , no . 3 , pp . 848855 , 2006 . j. kampeas , a. cohen , and o. gurewitz , `` capacity of distributed opportunistic scheduling in heterogeneous networks , '' in _ proceedings of the annual allerton conference on communication control and computing _ , 2012 . j. choi and f. adachi , `` user selection criteria for multiuser systems with optimal and suboptimal lr based detectors , '' _ signal processing , ieee transactions on _ , vol . 58 , no . 10 , pp . 54635468 , 2010 . j. laneman and g. wornell , `` distributed space - time - coded protocols for exploiting cooperative diversity in wireless networks , '' _ information theory , ieee transactions on _ , vol . 49 , no . 10 , pp . 24152425 , 2003 . l. wang , c. chiu , c. yeh , and c. li , `` coverage enhancement for ofdm - based spatial multiplexing systems by scheduling , '' in _ ieee wireless communications and networking conference , wcnc_.1em plus 0.5em minus 0.4emieee , 2007 , pp . 14391443 . m. pun , v. koivunen , and h. poor , `` opportunistic scheduling and beamforming for mimo - sdma downlink systems with linear combining , '' in _ ieee 18th international symposium on personal , indoor and mobile radio communications _ , 2007 , pp . 16 . a. janssen , j. van leeuwaarden , and b. zwart , `` gaussian expansions and bounds for the poisson distribution applied to the erlang b formula , '' _ advances in applied probability _ , vol . 40 , no . 1 , 122143 , 2008 .

the `` unified scheme '' for active galactic nuclei ( agn ) has enjoyed remarkable success in explaining not only the spectroscopic and polarimetric properties of various classes of objects , but also their morphologies in the light of certain emission lines , the overall radio structure and spectra of quasars , and the existence of ir - luminous , but optically - faint agn ( e.g. antonucci 1993 ; cimatti et al . 1993 ; capetti , macchetto , & lattanzi 1997 ; hines et al . 1995 ) . the idea relies on the presence of a nominally toroidal structure surrounding the central engine which obscures the compact object and broad - line region ( blr ) for high - inclination sightlines . from such vantage points , the bright nuclear emission components are seen primarily in light which has become polarized by scattering off dust and/or electrons located above and below the polar openings of the torus . for these objects ( e.g. seyfert 2 galaxies , narrow - line radio galaxies , _ iras _ qsos ) , extinction of the direct optical and near - ir continuum also increases the contrast of the emission from the more ubiquitous narrow - line region ( nlr ) , earning their type-2 spectroscopic designation and resulting in extreme optical / far - ir colors . by contrast , objects of type 1 are thought to be viewed through the toroidal hole , a perspective which provides the observer with a relatively unattenuated view of the central engine and blr . type-1 objects include the broad - line quasars , broad - line radio galaxies , and seyfert 1 galaxies . when our sightline lies very near the axis of a relativistic jet emanating from the compact object , doppler - boosted synchrotron emission from the jet dominates the spectrum and a highly variable and polarized blazar is observed . though the model does not explain the essential difference between radio - loud and -quiet agn , it successfully applies within each radio class . stockman , angel , & miley ( 1979 ) were the first to point out that the @xmath4-vector of optical polarization for quasars associated with double - lobe radio sources tends to be aligned with the radio source axis on the sky . this tendency toward `` parallel '' polarization was soon found to exist also among seyfert 1 galaxies ( antonucci 1983 ) and radio galaxies ( antonucci 1982 , 1984 ; rudy et al . 1983 ) , and to extend to radio structure at the milliarcsecond scale ( rusk & seaquist 1985 ) . though interest in the radio - optical alignment was soon eclipsed by antonucci s discovery ( 1983 ; see also antonucci & miller 1985 ) of very strong perpendicular polarization in some type-2 seyfert galaxies , the result of stockman et al . ( 1979 ) was early evidence for aspect - dependent classification of agn . there is no doubt that the dominant polarizing mechanism for most type-2 objects is small - particle scattering . spectra of polarized flux ( @xmath5 ) often show broad permitted lines that are not evident in the total flux spectrum but are characteristic of a classical emission - line qso . in addition , the polarized continuum occasionally exhibits the `` 3000 bump '' , a pseudo - continuum from and balmer free - bound emission arising within the blr ( grandi 1981 ; wills , netzer , & wills 1985 ) . the perpendicularity between the optical polarization @xmath4-vector and radio morphology is simply a result of the scatterers lying along the same general direction from the central source as the plasmons . in contrast , the polarization properties of radio - loud , non - blazar , type-1 agn are not well - studied . overall , the polarization of these objects rarely exceeds 2% ( stockman , moore , & angel 1984 ) , so the sources are difficult to study in detail . the interpretation of weak polarization is also more ambiguous , since synchrotron emission , scattering in an extended halo or accretion disk , or even partial extinction by a medium of aligned dust grains are all potentially viable mechanisms ( e.g. , stockman et al . 1979 ; antonucci 1984 ) . discrimination among these possibilities requires spectropolarimetry of sufficient quality to measure the polarization of the emission lines _ vs. _ the continuum and to compare the continuum shape in polarized flux with that in total flux . such a study is justified by the potential for new , qualitative information on the structure of the inner regions of qsos and possibly on the salient distinction between radio - loud and -quiet objects . in this paper we present the results of linear spectropolarimetry of 8 double - lobed quasars from the original stockman et al . ( 1979 ) list which show alignment between the axis of radio emission and optical polarization . the data were obtained using the steward observatory 2.3 m bok telescope and ccd spectropolarimeter ( schmidt , stockman , & smith 1992 ) . wavelength coverage generally spans a single order of the low - resolution grating , @xmath6 , however on occasion a second tilt was used to include an important emission line . the data typically consist of @xmath7 observations from several epochs over the period @xmath8 , each observation being the result of @xmath320 min of integration in a @xmath9 sequence of waveplate orientations . entrance slits were @xmath10 width @xmath11 51 length , providing a spectral resolution of @xmath12 and ample real estate for sky subtraction . basic properties of the targets and a summary of continuum polarization values measured at each epoch are listed in tables 1 and 2 , respectively . only pks 0405@xmath13123 showed such weak polarization in the initial observation that accurate spectropolarimetry was impractical . with that lone exception , the new spectropolarimetric results are broadly consistent with `` white - light '' and filtered measurements from the 1970 s and 1980 s ( e.g. , stockman , moore , & angel 1984 ; berriman et al . 1990 ; webb et al . 1993 ) . in the spectropolarimetric results , the maximum position angle difference between the optical polarization and radio axis on either arcsecond ( as ) or milliarcsecond ( mas ) scales is @xmath14 ( table 3 ) , so the new observations substantiate a long - term relationship between the two quantities . for the 7 objects observed in detail , polarimetric variability was small , when present at all ( see 4.2 ) . thus , the data for each object were coadded , weighting according to the statistical quality of each observation . the combined results are shown in figures @xmath15 , plotted as observed quantities _ vs. _ rest wavelength . each figure portrays , from bottom to top , the total flux spectrum @xmath16 , the stokes flux @xmath17 for a coordinate system aligned with the overall position angle of polarization , the rotated stokes parameter @xmath18 in percent , and the electric - vector position angle @xmath19 . for objects whose polarization position angles do not vary strongly with wavelength , such presentations are preferable to plots of the degree of polarization @xmath20 and polarized flux @xmath1 because they avoid having to apply statistical corrections for the positive bias of measurement errors on @xmath20 . the final polarimetric signal - to - noise ratio for each object is @xmath21 per 20 wide bin , which is somewhat broader than a spectral resolution element . the sample was chosen primarily to be tractable with a 2.3 m telescope . generally , sources are optically bright , @xmath22 , and modestly polarized , @xmath23 . all lie above a galactic latitude of 40 except 4c 34.47 ( @xmath24 ) . in any case , the lack of measurable polarization in the emission lines in any object ( 3 ) rules out a significant interstellar component due to aligned grains in either the milky way or quasar host galaxy . the targets have total ( core + lobes ) radio luminosities @xmath25 ( w hz@xmath26 ) and core 5 ghz - to - optical flux ratios greater than 10 ( table 1 ) . one of the most core - dominant of the radio sources , 4c 34.47 , is also a known superluminal source ( barthel et al . redshifts span the range @xmath27 , so the + [ ] complex lies in the observed spectral range for all objects ; lower-@xmath28 objects also provide , and 3c 95 exhibits @xmath29 . in general , the emission - line components are very prominent in total flux . because targets were selected from a list already known to be polarized and with an orientation related to the radio structure , our sample is not well - suited in isolation for statistical evaluation of polarization properties _ vs. _ the predictions of a particular model ( e.g. , the unified scheme ) . instead , our goal is to examine what can be learned about the polarizing mechanism in the aligned objects and possibly about the central regions of radio - loud quasars . comparisons with spectropolarimetric compilations of other agn may be informative but relevant selection criteria should be kept in mind . for example , cohen et al . ( 1999 ) presented observations of 13 low-@xmath28 radio galaxies . that sample contains 8 narrow - line objects whose nuclei are presumably highly obscured ; 6 of these reveal broad lines in polarized flux . of the 5 broad - line radio galaxies , one was chosen by cohen et al . because other observations suggested that it was heavily obscured . the remainder show no tendency toward alignment between optical polarization and radio structure , some contain large starlight components in their optical continua , and only one exhibits a true quasar - like spectrum . more recently , malkan ( 2000 ) has made a study of several low - polarization qsos of a variety of types , including 3 radio - loud objects . a few of the targets observed here have been studied with spectropolarimetric techniques by others . antonucci ( 1988 ) presented data of miller and goodrich on 4c 34.47 and ton 202 , pks 0405@xmath13123 was studied by antonucci et al . ( 1996 ) , and malkan ( 2000 ) discusses 4c 09.72 . in none of these results do the emission lines or 3000 bump appear in polarized light , though the previous data on 3c 34.47 and ton 202 are of low signal - to - noise ratio and the polarization of pks 0405@xmath13123 was very weak also when studied by antonucci ( @xmath30 ; @xmath31% ) . from inspection of figures @xmath15 , it is clear that the following characteristics are universal among the objects studied : 1 . a position angle of polarization that is essentially constant with wavelength , 2 . a degree of polarization ( % ) ] that is diminished at the locations of broad and narrow emission lines , and 3 . a reduction in polarization in the interval occupied by the 3000 bump : 2500 @xmath32 5000 . malkan ( 2000 ) notes that the combination of all 3 of the above trends is found _ only _ in the radio - loud objects he studied . the results contradict a claim of stockman et al . ( 1984 ) that the degree of polarization in a broad blue filtered bandpass generally exceeds that measured in the red . however , stockman et al . point out that definite wavelength dependence was detected in only 2 of their objects . one is the @xmath33 , radio - quiet broad absorption line qso 1246@xmath13057 ( e.g. , schmidt & hines 1998 ) , where the optical band examines a region well short of the 3000 bump . evidence among the remainder of their sample is largely statistical in nature owing to large measurement uncertainties . of the 4 objects which are in common with our spectropolarimetric sample , only ton 202 showed a blue - red polarization difference exceeding the combined 1@xmath34 error bar : in that case , @xmath35 , as we find . trends 2 ) and 3 ) above produce a spectrum of stokes flux @xmath36 which increases smoothly toward shorter wavelengths , showing no evidence of line emission . the confidence which can be placed on the claim of unpolarized lines varies from object to object according to the degree of polarization , data quality , and emission - line strength . we have used simple spectral synthesis techniques to estimate the fraction of the observed broad - line flux , @xmath37 , which could share the polarization characteristics of the continuum , gauging the appearance of a polarized emission line in the models by eye . because of the difficulty of setting a `` continuum '' level beneath the 3000 bump , we confine this test to the most prominent broad emission line in the spectrum : , , or @xmath29 , as appropriate . the results are listed in table 4 as 3@xmath34 upper limits to the polarized fraction of these features . limits range from @xmath38 for the line of 4c 34.47 to @xmath39 for the weak feature in 4c 09.72 . four of the 7 targets yield 3@xmath34 upper limits so low that only a portion of the line could be polarized like the continuum . this conclusion contrasts sharply with typical polarized type 2 sources , where broad lines generally appear _ only _ in polarized light , @xmath40 . it should be pointed out that the above - quoted confidence levels are quite conservative : several unpolarized emission features are present in most spectra , and statistical limits for the ensemble would be tighter than for the single brightest line . furthermore , since the 3000 bump is thought to arise largely as gaseous emission within the blr , the observed decline in polarization within the bump in each object is strong evidence for an unpolarized broad - line component . as an aside , we point out that many of the objects have sufficiently bright [ ] @xmath414959 , 5007 lines to rule out the possibility that a significant portion of the narrow - line emission shares the polarization of the continuum . our results demonstrate that the polarization of the optical / radio aligned quasars studied here is a property solely of the continuum which underlies 3000 bump and is not shared by the emission lines . we have characterized the shape of this smoothly rising polarized continuum by a best - fit power law in stokes flux ( @xmath42 . the resulting spectral indices @xmath43 are listed in table 4 . the results presented above lead to several general conclusions regarding the polarizing mechanism in these radio - optical aligned quasars : 1 . each object is dominated by a single polarizing mechanism . this mechanism acts entirely , or in large part , on the continuum of the object . if scattering is involved , the particles can not lie substantially outside either of the emission - line regions . electrons are possible scatterers , but considering that the blr in objects of this luminosity may be as large as 1 pc in radius ( kaspi et al . 2000 ) , dust particles may also be viable if the grains are entrained in a wind near the condensation radius ( kartje 1995 ) . 3 . the fact that the polarizing mechanism is aware of the radio axis on pc through mpc scales implicates an origin that is related to the collimation of , or radiation by , the particle beam : i.e. the accretion disk or emission from the jet itself . these two alternatives are discussed in detail below . it is popular to interpret the `` big blue bump '' which dominates the optical / uv portion of the spectral energy distribution of a typical agn as emission from an accretion disk that surrounds the compact object . though the shape of this bump is not a power - law over long spectral baselines , francis et al . ( 1991 ) measured optical - near uv continuum slopes for a large number of radio - loud and -quiet qsos , after correcting for the contributions of the balmer - continuum and emission . the resulting histogram reveals that the great bulk of the objects falls in the range @xmath44 , where @xmath43 is the spectral index @xmath45 . it can be seen from comparison with table 4 that this range also encompasses all of the slopes measured for the _ polarized flux _ spectra of the radio - optical aligned quasars studied here . the continuum shapes of these objects are therefore consistent with a spectrally - neutral polarizing mechanism applied to the light of the big blue bump . unfortunately , actual models of the polarization produced by physically - thin accretion disks are not as categoric ( koratkar & blaes 1999 ; koratkar et al . 1995 ; antonucci et al . 1996 ) . briefly , simple optically - thick disk models imply a high opacity due to electron scattering , hence a substantial polarization when viewed at sufficiently large inclination , @xmath46 . this prediction is modified in spectral regions where the absorption opacity is also high , so strong polarimetric signatures are expected near bound - free edges . spectropolarimetric observations have been made of a number of polarized qsos in the region around the lyman edge ( e.g. koratkar et al . 1995 , 1998 ; antonucci et al . 1996 ; schmidt & hines 1998 ) . the data fail to show either the strong predicted continuum polarizations or the expected discontinuities . moreover , scattering in an optically - thick , physically - thin disk results in a polarization @xmath4-vector which is _ perpendicular _ to the disk axis ; the latter is thought to approximately coincide with the radio axis . this relationship is , of course , opposite to the trend observed among the radio - optical aligned quasars . thus , simple thin disks seem to be ruled out on a number of accounts , a conclusion which seemed inevitable even in 1986 ( antonucci 1988 ) . suggested remedies to the situation include physically thick disks , winds , and additional scattering structures ( see , e.g. , kartje 1995 ) . unfortunately , such complications reduce the predictability of models and as a result the utility of polarimetry as a diagnostic of the central structure of the nucleus . a discussion of scattering disks in agn would be incomplete without mentioning the case of oi 287 . this object is an unusually strong narrow - line source whose polarization is extremely high ( @xmath47 ) , constant over time , and shows a position angle that is precisely aligned with the double - lobed radio structure ( rudy & schmidt 1988 ) . like the radio - optical aligned quasars , the narrow emission lines are unpolarized . however , in oi 287 broad and are evident in polarized light ( goodrich & miller 1988 ; see also antonucci , kinney , & hurt 1993 ) , a fact which motivated a model involving obscuration and scattering of light originating in the central source and blr by a physically ( and optically ? ) thin disk that is viewed very nearly edge - on . while oi 287 demonstrates the feasibility of thin scattering disks resulting in strong parallel polarizations , its overwhelming narrow - line spectrum and polarized broad lines indicate that the source is viewed at a high inclination and may well be the product of a polarizing mechanism unrelated to that dominating the objects studied here . synchrotron emission is a well - known source of optical continuum and polarization in certain classes of agn . it originates both from resolved , kpc - scale jet features such as those which emanate from several radio - loud agn , and from variable components in the spectral energy distributions of blazars . the latter presumably also arise in small - scale jet structures which in the unified scheme ( e.g. antonucci 1993 ) are doppler - boosted when a relativistic particle beam is viewed from near the axis of ejection . recall that in the unified scheme , broad - line quasars are seen from a sufficiently small inclination that the torus does not intercept sightlines to the nucleus ( @xmath48 ) , so a weak optical synchrotron component from the relativistic jet would not be a great surprise . with the exception of the enormous jet of 3c 273 , the resolved optical jets associated with local agn have projected lengths of a few kpc ( e.g. , scarpa et al . 1999 ) . at a distance typical of the sources in table 1 , they would subtend @xmath491 on the sky and be very difficult to discern against the bright nuclear source . both well - studied local examples m87 and 3c 273 show superluminal motion ( biretta & meisenheimer 1993 ; unwin et al . 1985 and references therein ) and have been mapped at @xmath501 resolution in spectral index and polarization ( perlman et al . 1999 ; conway & rser 1993 ) . the optical jets are not only associated with radio counterparts but are aligned with radio structure on the mas scale as well as with the giant lobes . the jets of both m87 and 3c 273 exhibit optical @xmath4-vectors of polarization that are generally perpendicular to the projected jet axes in the inner regions ( @xmath51-fields parallel to the jets ) , flipping by @xmath390 in knots that are presumed to be the sites of shocks . as a result , the polarization of each jet summed over its entire length is modest ( @xmath52% ) . finally , optical spectral indices measured for the local jet sources are : m87 , @xmath53 independent of location ( biretta & meisenheimer 1993 ) and 3c 273 , @xmath54 ( conway & rser 1993 ) . radio - optical spectral indices of jets associated with pks 0521@xmath13365 , 3c 371 , and pks 2201 + 044 measure @xmath55 ( scarpa et al . these slopes compare well with the spectral indices of polarized flux for the targets studied here ( @xmath56 ) , except for 4c 34.47 ( @xmath57 ; figure 6 and table 4 ) . the difficulty with the hypothesis that kpc - scale jets are the origin of polarization in the radio - optical aligned quasars is the low optical luminosities of known examples . monochromatic values at @xmath58 summed over the lengths of the above - mentioned jets range from @xmath59 ergs s@xmath26 @xmath26 for pks 2201 + 044 to @xmath3@xmath60 ergs s@xmath26 @xmath26 for 3c 273 km s@xmath26 mpc@xmath26 and @xmath61 are assumed . ] . these must be compared with the optical luminosity for a typical quasar from our sample ( @xmath62 , @xmath63 ) of @xmath310@xmath64 ergs s@xmath26 @xmath26 . indeed , if the jet of 3c 273 were unresolved from the central source , it would give rise to a net optical polarization for the quasar of @xmath65% . of course , it is possible that our selection of the most highly polarized quasars has also identified sources with unusually bright or polarized optical jets . however , the structures would have to be at least an order of magnitude more luminous than any local example and be uniformly polarized near the synchrotron maximum of @xmath66% to yield the levels of polarization listed in table 2 . more promising is the possibility of synchrotron emission from a misdirected blazar core component or an intrinsically weak end - on blazar . except for 4c 34.47 , the spectral indices of polarized flux of the radio - optical aligned quasars are consistent with the optical - near uv range measured for blazars ( e.g. landau et al . 1986 ) , and a synchrotron component with moderately strong polarization ( @xmath67 ) that provides @xmath5010% of the total continuum flux would account for the observed levels of polarization . the canonical benchmark for a highly - polarized quasar has been that the broadband optical polarization exceed @xmath68 ( e.g. moore & stockman 1981 ; stockman et al . 1984 ; wills et al . this criterion was based on the bimodal distribution of polarization observed among survey qsos ( stockman 1978 ) , a fact that suggested different polarizing mechanisms dominate in the two regimes . while it is true that quasars exceeding the 3% limit are very likely to show blazar - like properties , it is quite possible indeed probable that some weakly - polarized objects also contain feeble synchrotron components . remarkably , the best example of a miniblazar is found in the prototypical emission - line quasar 3c 273 ( impey , malkan , & tapia 1989 ; wills 1989 ) . highly variable in polarization but generally showing @xmath69% at optical wavelengths , 3c 273 occasionally exhibits `` flares '' when the @xmath70-band polarization reaches 2% ( courvoisier et al . 1988 ; impey et al . 1989 ; de diego et al . dilution by the 3000 and big blue bumps cause the degree of polarization to decline toward the blue and the emission lines are measured to be patently unpolarized ( smith , schmidt , & allen 1993 ; de diego et al . 1992 ) . from the same observations , it was found that the spectrum of polarized flux in the optical - uv is rather steep , @xmath71 , and variable . though the polarization position angle has been measured to vary through a range of at least 120 , values tend to cluster around a mean of @xmath72 ( impey et al . 1989 ) , close to the @xmath73orientation of the resolved jet and radio lobes . thus , 3c 273 could well be classified as a quasar that shows alignment between its optical polarization and radio structure . a lower - luminosity analog is the type-1 seyfert galaxy ngc 4151 , which contains a variable continuum component that has been studied extensively throughout the electromagnetic spectrum and whose optical polarization lies within @xmath310 of the axis of the two - sided arcsecond - scale radio jet ( e.g. , schmidt & miller 1980 ; pedlar et al . 1993 ) . the classic test for blazar activity is variability . the polarimetric data available to us ( stockman et al . 1984 ; webb et al . 1993 ) coupled with the observations summarized in table 2 provide 5@xmath1310 epochs of broadband measurements for each object in our sample . the data span a total time base of typically 20 yr and sample intervals of about a month to a decade . a simple @xmath74 analysis of the measurements uncovers a few interesting cases . foremost is the aforementioned pks 0405@xmath1312 , which was found to have @xmath75% on two occasions by stockman et al . ( 1984 ) but has always been measured at @xmath490.30% since . the position angle of polarization , which is probably a more robust indicator since it is not as sensitive to differences in waveband or to variable dilution , shows values for this quasar that range over nearly the entire compass . even in 1978@xmath1380 , when the source was comparatively strongly polarized , a range of @xmath76 was measured . unfortunately , pks 0405@xmath1312 has been so weakly polarized of late that accurate spectropolarimetry has been impossible ( see also antonucci et al . 1996 ) . three other targets from our sample show some evidence for polarimetric variability . the @xmath74 index for the position angle measurements of ton 202 , 4c 37.43 , and 4c 09.72 are 35 , 23 , 23 for 9 , 4 , 6 degrees of freedom , respectively , corresponding to probabilities of chance occurrence of @xmath77 . the range of variation is not great , amounting to @xmath78 for 4c 09.72 and 4c 37.43 and @xmath320for ton 202 . however , these cases are unlikely to be statistical flukes , since measurements for b2 1208 + 32 and 4c 34.47 , acquired to similar accuracy with the same instruments and over similar baselines , cover a total span in @xmath19 of just 12 and 8 , respectively , and yield @xmath79 for 6 degrees of freedom for each object . unfortunately , the detection of variability requires high relative precision in the measurements , and available data for these quasars are not of uniformly high quality . if 3c 273 were not virtually on our doorstep , it is likely that its miniblazar source would remain undiscovered even today . while polarimetric variability may be a sufficient condition for identifying the synchrotron mechanism , it may not be necessary . even the most active blazars often exhibit `` preferred '' polarization position angles in extensive monitoring campaigns ( e.g. angel et al . 1978 ) , suggesting a systemic axis possibly associated with that of the accretion disk . the synchrotron components proposed here would be much more stable than those of classical blazars , but this is qualitatively consistent with jets viewed at larger angles to the ejection axis . a strong link between the existence of a beamed optical component and core - dominant radio structure is well - established ( e.g. impey & tapia 1988 ; wills et al . 1992 ) , and strong optical polarization is almost always seen in superluminal sources . it is therefore interesting that the most core - dominant of our targets , 4c 34.47 , with @xmath80 , lies in the range where a majority of sources would show high polarization ( impey et al . 1989 ) . 4c 34.47 has also shown superluminal motion with the rather modest apparent speed of @xmath81 ( barthel et al . the projected expansion axis was oriented at @xmath82 , very nearly aligned with the large - scale structure and only 19 from the optical polarization position angle . a simple beaming model suggests that the jet lies within 44 of the line of sight . once again we draw a parallel to 3c 273 , which is a superluminal source showing @xmath83 and a corresponding viewing angle of @xmath4920 ( unwin et al . 1985 ) . on the other hand , the unusually flat spectral index of polarized flux measured for 4c 34.47 would require a miniblazar source of unprecedented properties one with a remarkably flat synchrotron spectrum and/or strong intrinsic wavelength dependence to its polarization the preceding sections present modest evidence that the optical polarization of radio quasars may be due largely to synchrotron emission from misdirected or intrinsically weak blazar core components . because this hypothesis implies the presence of an exposed relativistic jet , and such jets are presumably absent in radio - quiet objects , it might be expected that radio quasars would be over - represented among agn with small but significant polarization in large - scale surveys . of course , this implies equality between the classes with respect to other variables ( dust content , orientation , starlight fraction , etc . ) . in their statistical analysis of the polarization of palomar - green ( pg ) qsos , berriman et al . ( 1990 ) discriminated explicitly according to radio properties . their figure 3 indeed shows that a significantly larger fraction ( 5:17 ) of radio - loud objects ( those with @xmath84 ) are found with @xmath85% than for the radio - quiet sample ( 9:80 ) . four of the radio - luminous pg sources are spectropolarimetric targets studied here . however , berriman et al.s breakdown according to radio core dominance showed no correlation within either the radio - loud or -quiet samples , and a major conclusion of that paper was the general lack of evidence for synchrotron components in the sources studied . the techniques applied by berriman et al . ( 1990 ) were effective at discriminating against highly polarized agn of either the scattering or blazar variety , but since the pg catalog is derived from optical selection criteria , the radio - loud subsample is small . the criteria used polarimetric variability and the dependence of @xmath20(% ) on radio core dominance and optical color were inefficient at identifying weak nonthermal components when the observable itself was known to a typical accuracy of only @xmath86% ( @xmath87 ) . indeed , the prototype miniblazar , 3c 273 , fails to stand out in any of the diagnostic tests made by berriman et al . , despite the fact that it is a particularly bright and easily measured member of the pg catalog . the weak polarization that is observed at wavelengths as long as 2 @xmath88 m in 3c 273 as well as for 3 of the spectropolarimetric targets studied here ( sitko & zhu 1991 ) indicates that the putative synchrotron components would also be minor contributors to the observed flux in the infrared . unfortunately , none of the sources studied here were detected by _ iras _ , but they would make interesting _ sirtf _ targets . detailed spectral energy distributions , plus comprehensive polarimetric monitoring and additional investigations of the type carried out here , may be required to confirm the existence and prevalence of synchrotron components as significant sources to the optical continua of quasars . special thanks go to d. hines , m. malkan , and the referee , r. antonucci , for illuminating discussions and critical readings of the manuscript . support for this work was provided by nsf grants ast 91 - 14087 and ast 97 - 30792 and by the director of kpno . 1988 in supermassive black holes , ed . m. kafatos ( cambridge : cambridge univ . press ) , 26 1993 , , 31 , 473 antonucci , r. , geller , r. , goodrich , r.w . , & miller , j.s . 1996 , , 472 , 502 antonucci , r. , kinney , a.l . , & hurt , t. 1993 , , 414 , 506 antonucci , r.r.j . , & miller , j.s . 1985 , , 297 , 621 barthel , p.d . , hooimeyer , j.r . , schilizzi , r.t . , miley , g.k . , & preuss , e. 1989 , , 336 , 601 berriman , g. , schmidt , g.d . , west , s.c . , & stockman , h.s . 1990 , , 74 , 869 biretta , j.a . , & meisenheimer , k. 1993 , in jets in extragalactic radio sources , eds . rser & k. meisenheimer ( berlin : springer - verlag ) , 159 capetti , a. , macchetto , f.d . , & lattanzi , m.g . 1997 , , 476 , l67 cohen , m.h . , ogle , p.m. , tran , h.d . , goodrich , r.w . , & miller , j.s . 1999 , , 118 , 1963 conway , r.g . , & rser , h .- j . 1993 , in jets in extragalactic radio sources , eds . rser & k. meisenheimer ( berlin : springer - verlag ) , 199 courvoisier , t.j .- l . , robson , e. , hughes , d. , blecha , a. , bouchet , pl , krisciunas , k. , & schwarz , h. 1988 , , 335 , 330 de diego , j.a . , prez , e. , kidger , m.r . , & takalo , l.o . 1992 , , 396 , l19 fanti , c. , fanti , r. , formiggini , l , lari , c. , & padrielli , l. 1977 , , 28 , 351 francis , p. , hewett , p.c . , foltz , c.b . , chaffee , f.h . , weymann , r.j . , & morris , s.l . 1991 , , 373 , 465 goodrich , r.w . , & miller , j.s . 1988 , , 331 , 332 grandi , s.a . 1981 , , 251 , 451 hines , d. , schmidt , g.d . , smith , p.s . , cutri , r. , & low , f.j . 1995 , , 450 , l1 impey , c.d . , malkan , m.a . , & tapia , s. 1989 , , 347 , 96 impey , c.d . , & tapia , s. 1988 , , 333 , 666 kaspi , s. , smith , p.s . , netzer , h. , maoz , d. , jannuzi , b.t . , & giveon , u. 2000 , , in press kartje , j.f . 1995 , , 452 , 565 kellerman , k.i . , sramek , r.a . , schmidt , m. , green , r.f . , & shaffer , d.b . 1994 , , 108 , 1163 koratkar , a. , antonucci , r.r.j . , goodrich , r.w . , bushouse , h. , & kinney , a.l . 1995 , , 450 , 501 koratkar , a. , antonucci , r.r.j . , goodrich , r. , & storrs , a. 1998 , , 503 , 599 koratkar , a. , & blaes , o. 1999 , , 111 , 1 landau , r. et al . 1986 , , 308 , 78 malkan , m.a . 2000 , private communication pedlar , a. , kukula , m.j . , longley , d.p.t . , muxlow , t.w.b . , axon , d.j . , baum , s. , odea , c. , & unger , s.w . , 263 , 471 perlman , e.s . , biretta , j.a . , zhou , f. , sparks , w.b . , & macchetto , f.d . 1999 , , 117 , 2185 price , r. , gower , a.c . , hutchings , j.b . , talon , s. , duncan , d. , & ross , g. 1993 , , 86 , 365 rudy , r.j . , & schmidt , g.d . 1988 , , 331 , 325 rudy , r.j . , schmidt , g.d . , stockman , h.s . , & moore , r.l . 1983 , , 271 , 59 rusk , r. , & seaquist , e.r . 1985 , , 90 , 30 scarpa , r. , urry , c.m . , falomo , r. , & treves , a. 1999 , , 526 , 643 schmidt , g.d . , & hines , d.c . 1999 , , 512 , 125 schmidt , g.d . , & miller , j.s . 1980 , , 240 , 759 schmidt , g.d . , stockman , h.s . , & smith , 1992 , , 398 , l57 sitko , m.l . , & zhu , y. 1991 , , 369 , 106 smith , p.s . , schmidt , g.d . , & allen , r.g . 1993 , , 409 , 604 stockman , h.s . 1978 , in pittsburgh conference on bl lac objects , ed . a.m. wolfe ( pittsburgh : university of pittsburgh ) , 149 stockman , h.s . , angel , j.r.p . , & miley , g.k . 1979 , , 227 , l55 stockman , h.s . , moore , r.l . , & angel , j.r.p . 1984 , , 279 , 485 unwin , s.c . , cohen , m.h . , biretta , j. , pearson , t.j . , seielstad , g.a . , walker , r.c . , simon , r.s . , & linfield , r.p . 1985 , , 289 , 109 webb , w. , malkan , m. , schmidt , g.d . , & impey , c.d . 1993 , , 419 , 494 wills , b.j . 1989 , in bl lac objects , eds . l. maraschi , t. maccacaro , & m .- h . ulrich ( berlin : springer ) , 109 lcccccc 3c 95 & 0349@xmath13146 & 0.616 & 16.2 & 26.9 & @xmath89 & 13 + pks 0405@xmath1312 & 0405@xmath13123 & 0.574 & 15.8 & 27.2 & @xmath90 & 384 + b2 1208 + 32 & 1208 + 322 & 0.388 & 16.0 & 25.7 & @xmath91 & @xmath91 + ton 202 & 1425 + 267 & 0.362 & 15.7 & 25.6 & @xmath92 & 19 + 4c 37.43 & 1512 + 370 & 0.371 & 15.5 & 26.0 & @xmath93 & 24 + 3c 323.1 & 1545 + 210 & 0.264 & 16.7 & 26.0 & @xmath94 & 41 + 4c 34.47 & 1721 + 343 & 0.206 & 16.5 & 25.8 & @xmath95 & 353 + 4c 09.72 & 2308 + 098 & 0.432 & 16.0 & 26.1 & @xmath96 & 28 + llcccc 3c 95 & 1995 nov & 7 & @xmath97% & @xmath98 + & 1996 dec & 12 & @xmath99% & @xmath100 + & 1999 oct & 2 & @xmath101% & @xmath102 + + pks 0405@xmath1312 & 1995 nov & 1 & @xmath103% & @xmath104 + & 1999 oct & 1 & @xmath105% & @xmath106 + + b2 1208 + 32 & 1996 mar & 6 & @xmath107% & @xmath108 + & 1996 jun & 1 & @xmath109% & @xmath110 + & 1996 dec & 6 & @xmath111% & @xmath112 + & 1997 dec & 2 & @xmath113% & @xmath114 + + ton 202 & 1996 mar & 2 & @xmath115% & @xmath116 + & 1996 jun & 4 & @xmath117% & @xmath118 + & 1997 apr & 2 & @xmath119% & @xmath120 + & 1997 dec & 2 & @xmath121% & @xmath122 + & 1998 apr & 3 & @xmath123% & @xmath120 + & 1999 may & 2 & @xmath124% & @xmath125 + + 4c 37.43 & 1996 may & 7 & @xmath126% & @xmath127 + & 1999 may & 3 & @xmath128% & @xmath129 + + 3c 323.1 & 1996 jun & 5 & @xmath130% & @xmath131 + & 1998 apr & 5 & @xmath132% & @xmath133 + + 4c 34.47 & 1996 jun & 2 & @xmath134% & @xmath135 + & 1998 may & 1 & @xmath136% & @xmath137 + & 1998 sep & 3 & @xmath138% & @xmath139 + & 1998 nov & 1 & @xmath140% & @xmath141 + & 1999 may & 1 & @xmath142% & @xmath139 + + 4c 09.72 & 1996 dec & 12 & @xmath143% & @xmath144 + & 1997 dec & 3 & @xmath145% & @xmath146 + & 1998 nov & 1 & @xmath147% & @xmath148 + lcclc 3c 95 & 0.82 & 176 & 164 ; @xmath91 & 12 + pks 0405@xmath1312 & 0.27 & 124 & 12 ; 170 : & @xmath1368 + b2 1208 + 32 & 1.04 & 18 & 3 ; @xmath91 & 15 + ton 202 & 1.74 & 68 & 53 ; @xmath91 & 15 + 4c 37.43 & 1.24 & 100 & 109 ; @xmath91 & @xmath139 + 3c 323.1 & 1.47 & 22 & 20 ; @xmath91 & 2 + 4c 34.47 & 0.95 & 148 & 162 ; 167 & @xmath1314 + 4c 09.72 & 1.32 & 117 & 145 ; @xmath91 & @xmath1328 + lcc 3c 95 & @xmath131.1 & 0.8 + b2 1208 + 32 & @xmath131.0 & 1.0 + ton 202 & @xmath130.8 & 0.8 + 4c 37.43 & @xmath130.8 & 1.2 + 3c 323.1 & @xmath130.9 & 0.6 + 4c 34.47 & + 0.1 & 0.4 + 4c 09.72 & @xmath130.8 & 1.5 + ( thin line and right ordinate ) , normalized at long wavelengths . _ middle panel : _ stokes parameter @xmath18(% ) for a coordinate system rotated to the systemic position angle of polarization . _ top panel : _ position angle of polarization in degrees . note depressions in the degree of polarization in the `` 3000 bump '' and at each major emission line , and the corresponding smoothly rising polarized flux spectrum . artifacts caused by strong terrestrial absorption and emission features are marked.,title="fig : " ] 6.2truein

in order to capture some qualitative features of @xmath4 dimensional gravity , it is always useful to consider simpler toy models that are free of many of the technical difficulties of the full theory . this is particularly true of @xmath0 gravity . before 1984 , roughly speaking , the theory was considered to be `` too trivial '' to deserve any attention . it was with the work of deser , jackiw and thooft @xcite,@xcite , who considered point particle solutions , and later with the papers by achucarro and townsend @xcite and witten @xcite , that an avalanche of papers on @xmath0 gravity followed . a very wide spectrum of issues have been addressed in this years , from the `` problem of time '' to the `` loop representation '' , and more recently , black hole solutions @xcite . for recent reviews see @xcite and particularly @xcite . for a review of the @xmath0 black hole see @xcite . @xmath5dimensional gravity has been successful in providing simplified models in the study of black hole physics . in particular , the case when a negative cosmological constant is added ( anti de sitter space - times ) has been extensively studied in the last years . on the one hand it has been shown @xcite that making certain identifications in anti de sitter ( ads ) geometry , one obtains a black hole solution ( btz black hole ) . on the other hand , in string theories it has been found that there is a whole family of supersymmetric black brane solutions whose `` near horizon '' geometry is the product of spheres and the three dimensional ads spacetime @xcite . the study of the quantum properties of the btz black hole has been studied using many approaches , all of them giving a quantum mechanical derivation of the black hole entropy @xcite,@xcite,@xcite . the case of a positive cosmological constant has also received recent attention . maldacena and strominger @xcite have given a cft derivation of the entropy associated to the cosmological horizon , following earlier results by carlip @xcite . more recently , similar derivations in the euclidean continuation have given an alternative explanation of the hawking - gibbons entropy of de sitter space - time @xcite . in this note we consider three dimensional gravity with a cosmological constant . the purpose of the paper is to explore a little - known duality that exists between lorentzian @xmath0 gravity with a _ positive _ cosmological constant and euclidean gravity with a _ negative _ constant @xcite . as first shown in @xcite , @xmath0 gravity can be formulated as a theory of connections . there are two ways of writing the 3-dimensional action that involve connections . the einstein - hilbert - palatini action has as independent fields a `` spin '' connection @xmath6 and a ( non - degenerate ) co - triad field @xmath7 . the `` pure - connection '' formulation considers a chern - simons action where @xmath6 and @xmath7 are combined into a single connection @xmath8 living in a larger gauge group . this later description simplifies the structure of the theory and it is possible to completely characterize the state space . when one works in the so called `` frozen formalism '' , i.e. , in the reduced phase space of the theory , one is dealing with the `` true degrees of freedom '' of the theory . that is , roughly speaking , an equivalence class of classical solutions where one does not distinguish between configurations that are related by a `` gauge transformation '' . however , in the cases of interest for this note ( @xmath9 signature with @xmath10 and @xmath11 with @xmath12 ) , both the space of histories and the reduced phase spaces have been shown to _ coincide _ @xcite . thus , we have a mathematical identical description of the two systems . this is quite surprising and disturbing at first . but , as it is the case in many physical situations , the `` physics '' is in the interpretation of the formalism . as we shall see , there is a precise way to reconstruct two space - times , each one with a different signature , for each point of the ( reduced ) phase space . this construct also provides an explicit mapping between this space - times , or using a `` wheelerism '' , a _ wick rotation without wick rotation_. let us illustrate this duality using some heuristic arguments . take the euclidean btz solution with positive mass @xmath13 and no angular momentum . in three spacetime dimensions this means that mass is dimensionless and @xmath14 has units of length . ] , @xmath15 and make the following identifications : @xmath16 we obtain , @xmath17 here we have changed the names of the variables @xmath18 and @xmath19 . this solution looks like a cosmological `` big bang '' solution , with a singularity at @xmath20 . in the next sections we will see that this is indeed the case , and that this space - time is in a precise sense , dual to the btz solution . now take ( [ btz ] ) with @xmath21 $ ] . these are `` particle '' solutions , with no horizon and naked , conical singularities at @xmath22 . the angle deficit of this cones is @xmath23 . for @xmath24 we get the singular free global ads spacetime . proceeding with ( [ analitic ] ) in this solution we get @xmath25 this represents a spacetime with a cosmological horizon at @xmath26 and a conical singularity at @xmath22 with deficit angle @xmath27 . near the horizon , defining @xmath28 , we find , @xmath29 note that in order that the euclidean continuation of this solution is singularity free in the @xmath30 plane , we must identify the time @xmath31 with period @xmath32 . this period can be associated to the inverse temperature of the solution , which in this case is @xcite , @xmath33 it is worth noting that we could formally associate a temperature to the dual of the particle case , which had no horizon and thus no natural notion of temperature . reciprocally , the originally hot btz black hole is continued to a cosmological solution with no natural temperature , for which one could again assign some formal temperature . the structure of the paper is as follows . in sec . [ sec:2 ] we recall the gauge theory formulation of @xmath0 gravity with a cosmological constant , and the equivalence between the choices of signature and cosmological constant @xmath34 that we are interested in . in sec . [ sec:3 ] we explore in detail this duality , and apply it to the interesting case of the btz black hole ` family ' . in particular , we give a rigorous justification for the ` dualities ' presented above . we end with a discussion in sec . [ sec:4 ] . in the appendix , we consider the case of non - zero angular momentum . in this section we recall some basic features of @xmath0 gravity in the presence of a cosmological constant . the einstein - hilbert action for 3-dimensional gravity , @xmath35:=\frac{1}{2\pi } \int_m \d^3\!x\,\sqrt{|g|}(r-2\lambda)\ , , \ ] ] can be rewritten in terms of a @xmath36form connection @xmath6 and a @xmath36-form triad field @xmath7 as follows @xcite , @xmath37:= \frac{1}{2\pi } \int_m \epsilon_{ijk } \left(f(\omega)^{ij}\wedge e^k -\frac{\lambda}{3 } e^i\wedge e^j \wedge e^k \right)\ , , \label{actionx}\ ] ] where the capital indices denote internal vectors , that is , vectors in a fiducial 3 dimensional vector space @xmath38 . the field @xmath39 is the soldering form that maps internal vectors to tangent vectors to the 3 manifold @xmath13 : @xmath40 . in the internal vector space @xmath38 there is a fixed fiducial metric @xmath41 whose signature coincides with the one of the space - time metric . therefore , in the case of euclidean gravity @xmath42 and for lorentzian signature it is the 3 dimensional minkowski metric @xmath43 with signature @xmath44 . there is an lorentz invariant ( or rotation invariant in the euclidean case ) volume element @xmath45 in @xmath38 defined from @xmath41 . the @xmath46 with `` upstairs '' indices is defined from @xmath45 by `` raising the indices '' with @xmath47 . we can recover the space - time metric @xmath48 from the soldering form @xmath49 , @xmath50 finally , @xmath51 is the curvature @xmath52form of the connection @xmath53 , _ i.e. _ , @xmath54 . the variation of ( [ actionx ] ) with respect to @xmath55 yields , @xmath56 and the equation from @xmath7 is , @xmath57 the first equation states that the torsion of @xmath6 vanishes , and implies that the covariant derivative defined by @xmath58 coincides with the one compatible with the co - triad , that is , @xmath59 . we can therefore replace @xmath60 by @xmath61 in ( [ ele2 ] ) and we arrive at einstein equations : @xmath62 recall that this equation implies that the space - time has constant scalar curvature proportional to @xmath34 : @xmath63 therefore , @xmath64 . note that this result is independent of the signature of the space - time and therefore , of the gauge group we are considering . at this point it is convenient to use a fact from 3 dimensions : the dimensions of the `` lorentz group '' and the manifold coincide . that is , the lie algebra @xmath65 and the internal space @xmath38 can be identified . furthermore , the killing - cartan metric on the lie algebra @xmath66 is proportional to the internal metric @xmath41 . note that the connection is a `` lie - algebra valued '' @xmath36 form , and its components are labeled by two vectorial indices ( the natural labels in the vectorial or defining representation ) . however , in @xmath67 dimensions , the adjoint representation and the defining representation of @xmath65 coincide . it is simpler to work with @xmath36index labels ( the natural ones in the adjoint representation ) . we therefore define @xmath68 it is in this step that the euclidean and lorentzian theories have different expressions . this can be understood from the fact that we are using the internal metric @xmath69 to raise the indices of the @xmath45 in equation ( [ isom ] ) : @xmath70 . we can now rewrite the generalized covariant derivative in terms of @xmath71 as , @xmath72^i,\ ] ] where @xmath73^i={\epsilon^i}_{jk}\,\omega_\mu^j\,v^k$ ] . from ( [ isom ] ) it also follows that the generalized curvature tensor two form @xmath74 is , @xmath75 the palatini action can be rewritten as , @xmath76&= & \frac{1}{2\pi}\int_m \left(2f(\omega)_i\wedge e^i -\epsilon_{ijk}\frac{\lambda_l}{3 } e^i\wedge e^j \wedge e^k \right ) \qquad \;\;{\rm lorentzian,}\label{actionl}\\ s_e[e,\omega]&=&-\frac{1}{2\pi}\int_m \left(2f(\omega)_i\wedge e^i+\epsilon_{ijk}\frac{\lambda_e}{3 } e^i\wedge e^j \wedge e^k \right ) \qquad{\rm euclidean . } \label{actiony}\end{aligned}\ ] ] the actions ( [ actionl ] ) and ( [ actiony ] ) , in spite of the fact that are defined for different lie groups and have different signs , are in a precise sense , equivalent . this can be seen at different levels ( for details see @xcite ) . the easiest way of showing their equivalence is to generalize the observation of witten @xcite that one can define a new connection @xmath77 from the old connection @xmath78 and the co - triad @xmath7 , living on a larger lie algebra . this generalization , originally due to romano @xcite , can be applied to any action that has the structure of actions ( [ actionl ] ) and ( [ actiony ] ) . it is defined , in particular , for any semi - simple lie group @xmath79 and depends on a free parameter @xmath80 . form : @xmath81 . ] . the result is that any such action is equal to a chern - simons action based on the @xmath80-extension , @xmath82 , of the original group @xmath79 . for the cases of interest , that is , for the groups @xmath83 and @xmath84 , their @xmath80-extension coincide , being the group @xmath1 , when @xmath85 . the gravity actions can be put in the @xmath86 form if we construct the @xmath87 field from the drei - bein @xmath7 as follows : @xmath88 . then , the gravity actions ( [ actionl]),([actiony ] ) give rise to the _ same _ chern - simons action when @xmath89 and @xmath90 . for details of this equivalence , manifested also at the phase space level see @xcite . the other cases give rise to the groups @xmath91 when @xmath92 and @xmath93 , and @xmath94 when @xmath95 and @xmath96 . in this note we shall restrict our attention to the cases in which the @xmath80-extension corresponds to the _ de - sitter _ group @xmath1 . ] . solutions to the equation of the motion coming from the chern - simons action are flat connections , and two solutions related by a gauge transformation are regarded as physically indistinguishable . thus , the physical phase space @xmath97 is the moduli space of flat @xmath1 connections or , in other words , the homomorphisms from @xmath98 to @xmath1 ( modulo the adjoint action of the group ) . there are several remarks . first , a point in @xmath97 represents both ( an equivalence class of ) lorentzian space - times ( 3-geometries ) with a _ positive _ cosmological constant and euclidean space - times with a _ negative _ cosmological constant ! thus , there exists , at the fundamental level , a mathematical equivalence between the two theories . it is precisely this duality that we shall explore in this note . second , it is no coincidence that the group @xmath1 appears as the chern - simons gauge group , since it is also the global symmetry group of the de - sitter ( ds ) space in three dimensions . in fact , one knows that a space - time of the desired topology can be found by appropriate identifications and/or quotients of the ds space by a subgroup of the @xmath1 group @xcite . the relation between the elements of the group @xmath79 , associated to every non - contractible loop , and the quotient construction is what allows us to reconstruct a spacetime given a set of group elements . furthermore , one is able to reconstruct _ two _ space - times : one lorentzian ( @xmath99 ) and one euclidean ( @xmath12 ) . for , the global de - sitter solutions can be viewed as embedded in four dimensional minkowski space - time ; the lorentzian de - sitter ( @xmath100 ) is the hyperboloid defined by @xmath101 , where @xmath102 , and the euclidean de - sitter ( @xmath103 ) can be taken as one of the hyperboloids defined by @xmath104 . it is easy to see that the scalar curvature of these space - times is positive for @xmath100 and negative for @xmath103 , so they are indeed solutions to the einstein equations ( [ einst ] ) . it is important to note that this holonomy " duality is , in a sense , complementary to the well known wick rotation ( analytic continuation ) . in this later case , one in changing the signature of the embedding space so the symmetry group of the global , maximally extended , geometry is not preserved , whereas the sign of the cosmological constant , and therefore the scalar curvature , is preserved . for instance , the wick rotation of anti de sitter ( ads ) , a lorentzian space - time with @xmath12 and group @xmath105 is euclidean de - sitter @xmath103 for which @xmath12 and @xmath106 . in this section , we construct explicitly the duality mentioned in previous sections for a particular family of ( very well known ) solutions : the btz family . in the euclidean side this includes the btz black hole ( of positive mass ) , going through the zero mass solution and the so called mass gap " finishing with the global euclidean de - sitter . we shall study their lorentzian duals and comment on their thermodynamic properties . the lorentzian de sitter spacetime @xmath100 can be defined as the surface @xmath107 embedded in the four dimensional minkowski spacetime with line element , @xmath108 the topology of the space - time is @xmath109 . this spacetime is maximally symmetric , i.e. , it has six killing vectors ( three rotations and three boosts ) , which form a @xmath1 isometry group . a standard choice of coordinates @xmath110 that cover the entire space - time are @xcite , @xmath111 the induced metric then takes the form , @xmath112\label{4.4}\ ] ] where @xmath113 , @xmath114 , @xmath115 . as can be easily seen , the spacetime is a two sphere that contracts to its minimum area @xmath116 at @xmath117 and then expands again . in spite of the convenience of working on a globally defined coordinate patch , we shall introduce a popular choice of coordinates in which the metric looks first let us introduce some other embedding coordinates that will cover the region @xmath118 and @xmath119 , @xmath120 here @xmath121 , @xmath122 $ ] and @xmath123 . the metric induced in this region is , @xmath124 defining the coordinates @xmath125 and @xmath126 we get , @xmath127 where @xmath128 , @xmath114 , @xmath129 $ ] . in this schwarzschild " coordinates , @xmath22 , the origin , is the location of an observer and @xmath130 is its cosmological ( event ) horizon @xcite . using standard arguments in which one analytically continues the solution to a euclidean ( in this case , locally isometric to the three sphere @xmath131 ) spacetime , and requires regularity in the @xmath132 plane , one can assign a temperature to the horizon to be @xcite , @xmath133 this temperature is just ( @xmath134 times ) the surface gravity of the ( properly normalized ) killing vector generating the horizon @xcite . in the global coordinate system ( [ global ] ) the observer is located at the north pole " of the two - sphere ( @xmath135 ) . the antipodal point would be represented , in a penrose - carter diagram , as living in another asymptotic region " @xcite , and is not covered by the coordinate patch ( [ 4.5 ] ) . the euclidean de sitter space - time @xmath103 can be defined as the surface @xmath136 embedded in the four dimensional minkowski spacetime with line element , @xmath108 this space - time is disconnected , each connected component being a hyperboloid . it is standard to choose one of them , say , the upper one as the space - time . the topology of the space - time is @xmath137 . this spacetime is maximally symmetric , i.e. , it has six killing vectors ( three rotations and three boosts ) , which form a @xmath1 isometry group . a standard choice of coordinates @xmath138 that cover all the spacetime ( @xmath139 ) are , @xcite , @xmath140 the induced metric then takes the form , @xmath141\ ] ] where @xmath142 , @xmath114 , @xmath143 . the point @xmath144 is a regular point and is just a singular point of the coordinate system . this space - time can also be put in a static form . for that we need to introduce some new embedding ( covering the same region of spacetime ) as follows , @xmath145 the induced metric then takes the form , @xmath146\label{4.18}\ ] ] where @xmath147 , @xmath114 , @xmath148 . defining the coordinates @xmath149 and @xmath150 we get , @xmath151 where @xmath128 , @xmath114 , @xmath152 . this is precisely the wick transform " of ( a patch of ) the global anti de sitter ( ads ) spacetime through the transformation @xmath153 . in this part we construct , in an explicit fashion , a euclidean btz black hole of mass @xmath13 . we do this in order to illustrate the procedure that will lead to the ( one parameter family of ) dual lorentzian cosmological space - times . for simplicity we shall consider the non - rotating black hole . the rotating case can be found in the appendix . the starting point is euclidean de sitter @xmath103 ( [ 4.13 ] ) . next , we consider the killing vector field , @xmath154 which generates boosts in the plane @xmath155 and it is tangent to the de sitter space . note that @xmath156 . let us now define the embedding ( [ 4.17 ] ) , changing the names of the coordinates as follows , @xmath157 we have the line element , @xmath158 \label{met}\ ] ] where @xmath147 , @xmath159 , @xmath160 . note that @xmath161 , and @xmath162 is an affine parameter of the finite boost " along the integral curves of @xmath163 . the metric ( [ met ] ) induces naturally the topology @xmath137 over the region covered by the embedding ( [ btz : emb ] ) . note however , that from a canonical perspective , where we require the space - time to be of the form @xmath164 , we are forced to take out the point @xmath165 from the space - time , and as a consequence of that , the topology will change to @xmath166 . we shall construct the solution and see that in the btz case , one can add the point @xmath165 to the space ` generated ' via the hamiltonian @xmath0 evolution and still have a regular space - time with a different topology . the btz solution is constructed by periodically identifying @xmath162 with @xmath167 , where @xmath168 . here @xmath169 is simply a parameter labeling the space - time . what one does is to compactify along @xmath162 if we consider the point @xmath165 to be part of space - time we have a @xmath166 topology , and if we remove it , we end up with a space - time of topology @xmath170 ( needed from the canonical perspective ) . note that even when we pick this later choice , we are respecting the @xmath171 periodicity of the @xmath172 coordinate . this will translate into a trivial holonomy along its ( artificially constructed , topological nontrivial ) orbits . let us see that we indeed get the btz solution . writing , @xmath173 and @xmath174 , we have @xmath175 where @xmath147 , @xmath176 , @xmath177 . now , changing coordinates @xmath178 we get , @xmath179 we see that the horizon is at @xmath180 ( @xmath165 ) and the @xmath31 coordinate , the parameter along the killing field @xmath181 is naturally periodic with period @xmath182 . from here we see that the temperature that we should assign is , @xmath183 finally , the mass of the space - time is @xmath184 , where the value @xmath185 is assigned , as usual , to the solution with @xmath186 . the following questions come to mind : what is the relation between this construction and the reduced phase space in terms of @xmath187 ? in what sense can we say that the holonomies along the orbits of @xmath163 are related to the parameter @xmath188 ? the relation between the appearance of this non - trivial holonomies and the construction described above deserves more attention . the procedure involves the construction of a 3-dimensional space - time @xmath13 , out of a point of the reduced phase space . this procedure naturally implies some sort of gauge fixing " in order to select a representative form the equivalence class of flat @xmath1 connections , that will have a corresponding space - time locally isometric to @xmath189 . to see that the space - time constructed via identifications has associated to it a flat @xmath1 connection whose holonomies parameterize the reduced phase space we need to use the structure available in @xmath189 . the @xmath1 3-d connection is constructed out of the structure defined on the embedding space - time , namely , the flat @xmath1 connection compatible with the minkowski metric connection parallel transports vectors with internal indices . in minkowski space - time there is a canonical , globally defined soldering form that translates internal vectors into space - time vectors . we shall use this structure to refer to internal vectors as simply ` vectors ' ] and the action @xmath190 of the group @xmath1 acting on @xmath189 ( i.e. , @xmath191 , @xmath192 ) . the hyperboloids corresponding to de - sitter space - times are invariant subspaces of the @xmath1 action @xmath190 , and the group acts transitively . that is , given any two points @xmath193 and @xmath194 on @xmath13 , we can always find an element @xmath195 such that @xmath196 . let us now define a canonical ( auxiliary ) 3-d flat connection on @xmath13 as follows : given an open curve @xmath197 on @xmath13 , with @xmath193 as its starting point and @xmath194 as its end point , and an initial vector @xmath198 , we parallel transport the vector using the flat @xmath1 connection of the embedding space - time . thus , the connection so defined is a true @xmath1 connection on @xmath13 . furthermore , given that the action @xmath190 of @xmath1 on @xmath189 is linear , the derivative @xmath199 coincides with it , and is also given by an @xmath1 action on tangent vectors . thus , if we parallel transport a 3-dimensional vector ( using the natural connection ) along the orbits of the kvf @xmath163 from the origin @xmath200 to say , the point @xmath194 defined by @xmath201 , the original vector and the parallel transported one will be the same " ( for instance in a cartesian coordinate system ) . how can we say that the parallel transported vector will be different from the original one when we are using the flat constructed above ? since we are identifying @xmath194 with the origin , we are at that step closing the curve @xmath197 into a loop , and we have to bring back " the vector to the starting point . that is , we push " forward the vector from @xmath194 to @xmath193 using the mapping @xmath202 . thus , the element of the group @xmath1 that provides the parallel transport ( in this case , a boost ) , becomes the holonomy along the non - contractible loop @xmath197 . therefore the `` parallel transported '' vector has as its holonomy precisely the group element in @xmath1 corresponding to a boost along @xmath163 with parameter @xmath203 . this is the relation between holonomies and space - time identifications . to see more about the relation between geometry and holonomies ( known as geometric structures ) see @xcite now , the topology of the spacetime is @xmath170 and therefore we have two non - contractible generators of @xmath204 . the gauge invariant information , i.e. , the coordinates of @xmath97 are four parameters @xmath205 @xcite , where @xmath206 are rotation parameters corresponding to a holonomy along the @xmath207 $ ] generator of @xmath98 , and @xmath208 correspond to boost parameters along the dual plane in @xmath189 . concretely , representatives of the @xmath209 $ ] equivalence class are given by the closed orbits of the @xmath210 vector field and representatives of @xmath211 $ ] are given by the orbits of @xmath212 . in this language , a ( euclidean ) btz black hole of mass @xmath13 and @xmath213 has coordinates in @xmath97 equal to @xmath214 . in order to construct the holonomy - duals to the btz black hole , we will proceed analogously . first , we take the _ same _ killing vector @xmath154 and identify points in @xmath100 space along this killing vector . this procedure will induce closed time - like curves in the region where @xmath215 , therefore we will take out this region from the spacetime and take the surface @xmath216 as a `` chronological singularity '' . the region of interest its covered by the following patch of coordinates @xmath217 with @xmath218 and @xmath219 . in these coordinates the metric takes the form , @xmath220 we can now safely identify @xmath162 with @xmath221 , which makes the surfaces of constant time compact @xmath2 manifolds . notice that in order to relate it to the btz black hole we have the correspondence @xmath222 . the chronological singularity @xmath223 is now located at @xmath117 , which is a naked `` initial '' singularity . for big values of @xmath224 , the spatial slices have the topology of a 2torus whose radius are expanding exponentially in time . the parameter @xmath225 , labeling the particular space - time , gives the ratio between the two radii . for small @xmath224 one of the radius goes to @xmath226 while the other shrinks to zero ( see figure 1 ) . in order to put this solution in a familiar form set @xmath227 , @xmath228 and @xmath229 . thus , we have , @xmath230 where @xmath231 , @xmath232 , and @xmath114 . these solutions are not new . they are a subclass of the general solution found by ezawa @xcite ( see also @xcite and @xcite ) . it is important to note that in terms of the reduced phase space language , these solutions have coordinates @xmath233 , and that the holonomy - duality is relating space - times ( [ btze ] ) and ( [ ezawa ] ) when we set @xmath234 . figure 1 : the `` big bang '' spacetime ( a ) vs. the btz spacetime ( b ) . in ( b ) we represent the plane @xmath235 in the coordinates given by the metric ( [ btzrho ] ) . each point of this plane has a circle on it , which completes the spacetime ( with its @xmath236 topology . @xmath169 represents the radius of the circle at the origin . in ( a ) , the parameter @xmath225 defines the geometry of the @xmath237 constant torus . the duality identifies @xmath238 with @xmath169 . in this part we shall consider another class of solutions . just as the @xmath2-btz - cosmological solutions were found by means of an identification along a boost , the solutions in this section will be constructed using a rotation . point particle solutions for de - sitter spacetime were studied by deser and jackiw @xcite . they are a generalization of the point particle solutions for flat 3-dimensional gravity studied in @xcite . recall from sec . [ sec:3.a ] that the global de sitter has a @xmath109 topology . thus , as argued by deser and jackiw , one can not construct a one - particle space - time by removing a wedge " from the @xmath239 spatial slice since there will be , at least , two conical singularities . in our coordinate system @xmath110 ( [ 4.4 ] ) the two particles will be located at antipodal points on the sphere ( @xmath135 and @xmath240 ) . however , the schwarzschild patch ( [ 4.5 ] ) only sees one of the observers / particles and the other one is hidden behind the cosmological horizon . the ` remove a wedge ' construction of deser and jackiw can be interpreted in our language as , again , an identification along the orbits of a killing field . the main difference with the btz construction is that now we take a rotation kvf @xmath241 instead of a boost . when one uses a rotation field to construct a new spacetime , one needs due care . for , one can not simply identify , say , a point with angle coordinate @xmath190 with the point @xmath242 , for @xmath243 . this naive quotient procedure leads ( for almost all values of @xmath244 ) to a space that is not even a manifold . instead , the construction involves the arbitrary choice of a value of , say @xmath245 , such that the points @xmath245 and @xmath246 are to be identified and the wedge between @xmath244 and @xmath171 is removed from the manifold . even though the construction procedure is different from the one used in the btz case , the picture of having an element of @xmath1 , the holonomy , for each non - contractible loop remains unchanged . this is so since by cutting the wedge , one has two curvature - singular points at the poles of the sphere that have to be removed from the manifold . thus , the resulting topology of @xmath13 is @xmath166 and we have one non - trivial loop around the @xmath247 . as an end result , we have that the holonomy along this loop , the ( new ) orbit of the vector @xmath248 , is a rotation ( in so(3,1 ) ) by @xmath244 . in this case we take the region of @xmath100 space covered by the patch of coordinates described below . @xmath249 here @xmath121 , @xmath122 $ ] and @xmath123 . the metric induced in this region is , @xmath250 cutting the wedge means restricting the range of @xmath190 to @xmath251 and identifying @xmath245 with @xmath252 , where @xmath253 $ ] . thus , the space - time has a deficit angle of @xmath23 . when @xmath254 , there is no deficit and we recover global de sitter . if we now change coordinates to @xmath255 , @xmath256 and @xmath257 we get , @xmath258 where @xmath259 , @xmath260 $ ] and @xmath128 . the cosmological horizon is now located at @xmath261 and , following the argument given in the introduction , the temperature of the horizon is @xmath262 . the deficit angle is manifested by the fact that the ` area ' of the horizon is now @xmath263 . let us now consider the euclidean case . the starting point is euclidean de sitter spacetime ( [ 4.13 ] ) with topology @xmath137 . now it is possible to have a single point particle located at the ` origin ' @xmath165 . if we remove this world - line , we are left with a space - time with topology @xmath166 , with a single non - contractible loop . we shall use the same rotational kvz @xmath264 as in the case of lorentzian signature to identify along its orbits . the resulting space - time , with deficit angle @xmath23 is given by , @xmath265 which corresponds to the wick transform of the ` point particle ' anti - de sitter spacetime of mass @xmath266 . just as we did for the @xmath2-cosmological space - times , we could , formally , assign a temperature to this point particle solution to be equal to @xmath267 . this spacetime is the limit @xmath268 in both lorentzian and euclidean signatures . for the euclidean signature , we can simply take the limit of ( [ btze ] ) when @xmath268 , @xmath269 now , @xmath31 is no longer a periodic coordinate ( recall that its range , for the case @xmath270 was @xmath176 ) , so we have a space - time with topology @xmath271 and zero temperature . note that we would have arrived at the same space - time in the limit @xmath272 in ( [ 3.32 ] ) . in this case , @xmath273 had the interpretation of the range of @xmath190 after one has removed the wedge from the spacetime , so the limit @xmath274 might be interpreted as the limit in which the conical singularity opens up to form a cylinder , thus recovering the @xmath271 topology . this space - time is referred as the vacuum since it is the wick rotation of the btz black hole of zero mass and this is considered to be the zero energy solution @xcite . the @xmath268 limit of out lorentzian solutions has also some similar features . that is , starting from the @xmath2-cosmological solutions ( [ ezawa ] ) , and taking the limit @xmath275 we get , @xmath276 where @xmath172 is not periodic and it takes values in @xmath277 , @xmath231 and @xmath114 . it is again a space - time with topology @xmath166 . starting from the point - particle family ( [ 3.32 ] ) , it corresponds to taking the limit @xmath278 , where the cosmological horizon shrinks to zero , and it can be interpreted as the 2-sphere opening up to a cylinder . again . the temperature we would assign to this vacuum state would be zero . the main results of this note can be summarized in fig.2 . here , we have the four possibilities of signatures and signs of the cosmological constant . in the left column we have lorentzian space - times and in the right their corresponding analytic continuations ( wick transforms ) , as shown by the ` w ' arrows ( for wick ) . we have listed the space - times considered in this note . the ` holonomy ' transform ( h - duality ) corresponds to the diagonal ` h ' arrows going from the lower - left corner to the upper right corner and relating the space - times studied in this note . there are several remarks that need to be made . 1 . _ relation to wick transform_. as already discussed , the standard wick transform and the holonomy duality are in a sense , complementary . first , the wick - duality for space - times is restricted to complex space - times that admit both an euclidean and a lorentzian real section . normally this is possible only in the presence of a time - like kvf . furthermore , the mapping in not strictly between globally defined space - times , since , for instance , it is only the exterior region of the lorentzian btz black hole that gets mapped " to the euclidean sector . on the other hand , the holonomy transform , being a duality at the phase space level , relates globally defined space - times . furthermore , the existence of a kvf is not essential for the definition of the duality . the wick rotation has been extensively used in gravitational physics . not only is it used for qft calculations , but also in connection with gravitational thermodynamics . standard arguments of hawking and gibbons @xcite assign a temperature to the lorentzian space - time by looking at the period of the wick - rotated time in the dual space - time . this physical equivalence between space - times is used regularly in the study of thermodynamics of black holes . one may wonder if the holonomy dual space - times can be regarded as having more than just the phase - space equivalence . is it justified to assign thermodynamical properties to the dual space - time to define , say , a temperature and entropy for space - times that do not even have a horizon ? can we perform a calculation in one domain and conclude something about the other domain just as we do with the wick transform ? + affirmative answers to this intriguing questions would indeed be of some interest since we could , for instance , pass from the physics in @xmath3 to @xmath279 composing a @xmath38 and a @xmath280-transform . what would the relevance be for an hypothetical @xmath281 conjecture remains an open question . thermodynamics_. let us elaborate on the possible assignment of a temperature to the space - times studied in previous section . let us first recall the standard argument to define a physical temperature for a euclidean space - time . first , we start we a lorentzian stationary black - hole space - time , with an @xmath282-dimensional ( null ) horizon ( with @xmath283 being the dimension of space - time ) . in the euclidean sector , the horizon is now @xmath284 dimensional and ` time ' is a periodic coordinate . one asks for regularity of the space - time at the horizon in order to fix the periodicity of the euclidean time . even if we start with a euclidean solution with non - trivial topology , the fact that it is regular is not enough to define a temperature . we need , as in the schwarzschild black hole case , the lorentzian time that _ measures proper time at infinity _ in order to have the ` right ' temperature . the same is true for the de - sitter space - time where one needs to choose the time - parameter as the proper time of the observer whose cosmological horizon we want to consider . the periodicity of the wick - transformed time is then related to the temperature . + let us now consider the space - times of interest for this note . the ( non - rotating ) euclidean btz solutions have a naturally defined temperature @xmath285 . on the other hand , the @xmath280-dual space - time is a spatially closed space - time , with no time - like kvf and no horizon . one may wonder what , if any , would be the meaning of formally defining thermodynamic parameters and , in particular , a _ constant _ temperature to the whole space - time . we shall return to this question shortly . + the lorentzian de - sitter space - time together with the point particle solutions ( [ 3.32 ] ) have horizons with a temperature ( assigned via the wick rotation ) given by @xmath286 . the dual space - times ( [ 3.33 ] ) do not have a periodic ` time ' coordinate @xmath31 and therefore , no natural temperature . however , if we declare the temperature of this space - times to be @xmath287 , ( i.e. , the temperature of its @xmath280-dual ) then the natural period for @xmath31 is @xmath288 . now , the space - times ( [ 3.33 ] ) are the wick transform of the family of point particle @xmath3 space - time ( with @xmath289 ) . therefore , using the standard wick transform argument , we can assign a temperature to the family of point - particle-@xmath3 space - times to be , @xmath290 note that with this definition , the @xmath3 space - time ( corresponding to @xmath291 and @xmath254 ) is a ` thermal state ' with temperature @xmath292 , recovering the suggestion of strominger and others @xcite . + in recent years there have been several attempts to calculate the entropy of the btz black hole . one of those calculations , originally due to carlip @xcite ( see also @xcite ) , starts by considering the @xmath293 chern - simons phase space for a @xmath170 space - time . then , one imposes boundary conditions , motivated by the lorentz black hole horizon , and computes the bekenstein - hawking entropy using the wzw theory induced on the horizon . one intriguing fact about this calculation is that , in the euclidean domain , one does not have the geometrical interpretation of the lorentzian horizon , and without the knowledge of its origin , we would only be faced with some boundary conditions on phase space inducing a wzw theory on the boundary . + since the @xmath280-duality is precisely based on the fact that the two gravity theories share the same phase space , then we could perfectly well interpret the result of the calculation as giving us the entropy of the lorentzian cosmological space - time . the correspondence is given as follows . the horizon @xmath165 of the euclidean black hole corresponds in the lorentzian side to the initial @xmath20 singularity ( this is a striking property , because @xmath20 is a _ chronological _ singularity and the horizon @xmath165 , is just a coordinate singularity ) . this singularity is a mild one , since it is manifested by the fact that the spatial slice is a degenerate torus . if would be of interest to geometrically interpret the boundary conditions in the lorentzian theory and to identify the calculated entropy with some geometrical invariant , the analogous of the horizon area in the bh case . + another interesting calculation refers to the entropy associated to the cosmological horizon in de - sitter . the first approach , very much in the style of carlip is given in @xcite . maldacena and strominger again impose some boundary conditions in the @xmath293 chern - simons phase space and compute the bekenstein - hawking entropy using the cft on the boundary of the region , in this case , the horizon . the computed entropy is , as expected , one fourth of the cosmological horizon radius , @xmath294 . the other calculation @xcite computes the gibbons - hawking entropy by considering the partition function in the wick - rotated theory ( corresponding to a chern - simons theory on @xmath295 ) . the maldacena - strominger results can be re - interpreted in the same way that carlip s results were ( see also @xcite ) . that is , one can argue that the calculation very well applies to the @xmath280-dual theory , namely , to a euclidean de - sitter space - time with topology @xmath296 ( [ 3.33 ] ) , and assigns to it an entropy @xmath297 . now , recall that these space - times are the wick - duals of the point particle @xmath3 space - times , and therefore one is naturally lead to assign to these space - times an entropy as well . for the global @xmath3 the associated entropy would be @xmath298 . just as for the cosmological space - times , we would like to interpret this entropy in geometrical terms . 3 . _ path integrals , wick rotations , etc .. _ the fact that three dimensional gravity has a dual description , both as a theory of triads and connections @xmath299 and as a theory of space - time metrics , poses a puzzle . on the one hand , there is the not so unpopular wisdom that there is a deep relation between the lorentzian and euclidean path integral formulation of gravity . independently of the viewpoint one adopts , be it that the euclidean path integral is useful for calculating lorentzian transition amplitudes , or that the euclidean regime is fundamental , there is some belief that euclidean methods are relevant for quantum gravity and bh - thermodynamics . in fact , this is the standard justification for the extended use of the wick - rotation . on the other hand , when the theory is formulated as a diffeomorphism invariant theory of connections , one looses even the notion of a space - time metric ; it now becomes a _ derived _ notion and , therefore , the role of the wick rotation turns out to be not so obvious ( see @xcite ) . for , in the completely solvable models like @xmath86 theories and @xmath0 gravity with no cosmological constant , the rigorously defined measures on the space of histories have heuristic analogs involving a factor of the form @xmath300 for both gauge groups ( @xmath83 and @xmath84 ) , so the signature of the space - time metric seems to play no role . if one takes the viewpoint that the heuristic path integrals based on the actions ( [ actionl ] , [ actiony ] ) fully represent quantum gravity in three dimensions , then one is naturally lead to conclude that the @xmath280-duality is , in a sense , more fundamental and natural than the wick transform . for , the @xmath280-dual theories not only share the space of histories , in their chern - simons version , but also have the property that the actions ( [ actionl],[actiony ] ) , when evaluated on the shared history , differ only by a sign @xcite . thus , _ the path integral for these two actions must be equivalent _ ( since the factor involved is of the form @xmath300 and the actions are real ) . implications of the previous discussion are intriguing . first , one can justify the formal assignments of thermodynamical properties like entropy and temperature to cases where , classically , there is no time - like kvf and no horizon . this can be done by applying the standard arguments of gibbons - hawking together with the @xmath280-duality@xcite . second , there seems to be some tension between the relevance of the wick transform for quantum gravity in its metrical / geometrodynamical formulation and the holonomy duality for its connection - dynamics formulation . it would be of interest to fully understand the role of these dualities in the context of path integrals for diffeomorphism invariant theories . finally , we could speculate about some implications for the @xmath280-dual of the @xmath301 correspondence and what it could tell us about quantum gravity in higher dimensions . the authors would like to thank the participants of the 15th pacific coast gravity meeting , february 1999 , and in particular abhay ashtekar , mximo baados and steven carlip for discussions . this research was supported in part by the national science foundation under grant no . phy94 - 07194 . ac would like to thank the itp for its hospitality during which part of this work was completed . ac was also supported by dgapa - unam proy . in121298 and by conacyt ( mxico ) ref . i25655-e . the work of ag was also supported in part by national science foundation grant phy97 - 22362 , and by funds from syracuse university . let us start by considering the euclidean btz black hole with angular momentum , @xmath302 where @xmath303\right\}^{1/2}\ ] ] @xmath304 ( recall that @xmath305 ) . in order to construct this space - time , we can again start with the metric ( [ met ] ) , @xmath158 \label{a3.18}\ ] ] where @xmath147 , @xmath159 , @xmath160 . note that @xmath161 , and @xmath162 is an affine parameter of the finite boost " along the integral curves of @xmath163 . the region that the embedding ( [ btz : emb ] ) covers has the topology @xmath166 . the rotating btz solution is constructed by specifying the points that need to be identified in order to have a two torus . just as in the non - rotating case , one of the generators is given by the ( closed ) orbits of the @xmath306 kvf , with period @xmath171 . thus , one only needs to specify the other generator . this is done by periodically identifying the point @xmath307 with the point given by a boost along @xmath163 with parameter @xmath308 , together with a rotation along the kvf @xmath309 equal to @xmath310 . here @xmath311 are the parameters labeling the space - time . what one does is to compactify along the line , in the plane @xmath312 , connecting , say @xmath313 and @xmath314 and one ends up with a space - time of topology @xmath170 . we can define the new periodic coordinates @xmath315 along the generators of the torus , @xmath316 where @xmath317 and @xmath318 . then , the line element has the form , @xmath319\ ] ] in order to put it in the standard form , let us define new coordinates @xmath320 where @xmath114 and @xmath321 . then , the metric takes the proper radial coordinates " form @xcite , @xmath322 finally , if we define @xmath323 , we recover the standard form ( [ jnot0 ] ) . in terms of the gauge invariant information , i.e. , the coordinates of @xmath97 given by the four parameters @xmath205 @xcite , a ( euclidean ) btz black hole of mass @xmath13 and angular momentum @xmath324 has coordinates in @xmath97 equal to @xmath325 . the corresponding lorentzian dual space - times are constructed using the same holonomy parameters . starting from the metric ( [ bang ] ) , @xmath326 along the @xmath172 direction we keep the @xmath171 periodicity , and the other generator is constructed via a boost in the @xmath327 direction with parameter @xmath328 together with a rotation along @xmath306 with parameter @xmath329 . thus , the spacetime has has coordinates in @xmath97 equal to @xmath330 . it is the holonomy dual to the btz black hole when we set @xmath331 and @xmath332 . we can define the new periodic coordinates @xmath333 along the generators of the torus , @xmath334 where @xmath317 and @xmath318 . then , the line element has the form , @xmath335 \\ & = & -l^2 \d t^2 + l^2 \cosh^2 t | \d\chi + z(t ) \d \phi|^2 \ , \end{aligned}\ ] ] where @xmath336 is the so called modulus of the torus , which completely determines the conformal geometry of it . for @xmath337 , this parameter goes to the constant @xmath338 , and the torus will evolve blowing up with the conformal factor @xmath339 . for @xmath117 the torus degenerates to a circle of radius @xmath340 . finally , the line element can be put in the form , @xmath341 that can be heuristically obtained from ( [ jnot0 ] ) by the substitution @xmath342 , @xmath343 , @xmath344 and by the relabeling @xmath345 and @xmath346 . the reduced phase space for a @xmath170 topology space - time is four dimensional and parameterized by @xmath347 . however , the btz black hole family spans only a two dimensional surface ( given by @xmath169 and @xmath348 , with @xmath349 ) , while the lorentzian cosmological solutions are well behaved for any point on @xmath97 . what are then , the euclidean duals to the cosmological solutions that are off the btz torus ? these solutions have been studied by carlip and teitelboim @xcite who refer to them as off - shell " black holes . they are ( euclidean ) @xmath170 solutions , with a conical singularity and a twist at the @xmath165 horizon . thus , one can not include this point in the ( maximally evolved ) space - time and obtain a regular space - time with the @xmath166 topology of the black holes . if one did not ask for regularity at the horizon and only asked for a space - time defined in the region @xmath350 ( as in the cosmological case for @xmath351 ) , the off - shell " black holes would be perfectly valid , as solutions to the einstein equations . let us first construct the euclidean space - times corresponding to a ` rotating particle ' at the origin . these spaces are the wick transforms of the ( naked singularity ) lorentzian space - times with a rotating particle at the origin . the topology of these space - times , both lorentzian and euclidean is @xmath271 . thus , we only need to specify the holonomy along the unique non - trivial loop . the explicit procedure for the construction of the spacetime includes some cutting and gluing together , starting from the euclidean de - sitter space . recall that in the case of zero angular momentum , the construction involved cutting a wedge , removing the points inside the wedge , and gluing together the lines that defined the original wedge ( for a constant time ) . now , when angular momentum is present , we still cut a wedge , but now we shall identify points that are not only related by a rotation but also a boost . note that the particular rotation and boost that we have chosen in the construction of this appendix are such that they leave invariant orthogonal planes , which implies that , as elements of the group @xmath1 , they commute . the starting point is again the region of @xmath189 parameterized by ( [ 4.17 ] ) and with line element , @xmath352\label{4.18bis}\ ] ] where @xmath147 , @xmath114 , @xmath148 . here , the kvf along which the identification will be made are @xmath353 and @xmath354 . we again cut a wedge with angle @xmath355 and identify points that differ by a rotation by @xmath355 along @xmath264 and a boost with parameter @xmath329 along @xmath163 . in order to implement these identifications , let us define the new coordinates @xmath356 then , @xmath357\label{4.18b}\ ] ] where @xmath128 , @xmath259 and @xmath358 . now , we define the new coordinates @xmath359 and @xmath360 , we get , @xmath361 note that this is the wick transform of the ads - point particle solution given in @xcite . the construction of the lorentzian solution follows the same steps . the starting point are the embeddings ( [ 3.26 ] ) and the induced metric in de - sitter , @xmath362 again , we cut a wedge of size @xmath355 and identify using a rotation by @xmath355 along @xmath264 together with a boost along @xmath163 of parameter @xmath329 . with the new coordinates @xmath363 and @xmath364 we get , @xmath365 which becomes , after defining @xmath366 , @xmath367 , @xmath368 m. baados , _ trends in theoretical physics ii _ h. falomir , r.e . gamboa saravi and f. schapopsnik , ( aip conf . proc . 484 , 1999 ) . ( m. baados , three dimensional quantum geometry and black holes " , preprint hep - th/9901148 . ) a. ashtekar and j. romano , _ phys . _ * b229 * ( 1989 ) 56 ; a. ashtekar , v. husain , c. rovelli , j. samuel and l. smolin , _ class . quantum grav . _ * 6 * ( 1989 ) l185 ; l. smolin , in the _ proceedings of the john s hopkins conference on knots , topology and quantum field theory _ , ed . l. lusanna ( world scientific , singapore , 1989 ) . s. carlip , _ knots and quantum gravity _ , ed . j. baez ( oxford , oxford university press , 1994 ) . ( s. carlip , geometric structures and loop variables in ( 2 + 1)-dimensional gravity " , preprint gr - qc/9309020 . )

since the prediction of keldysh and kopaev , @xcite there has been a long and intense pursuit to realise a condensed phase in solid state excitonic , and related , systems . in particular , polaritons in semiconductor microcavities , the coupled eigenstates of an exciton with a cavity photon , @xcite represent ideal candidates for observing condensation phenomena . the very light mass of these composite bosonic particles promises relatively high transition temperatures . in the past decade , improvements in the growth technology of semiconductor heterostructures have made the study of high - quality strongly - coupled planar microcavities almost routine for iii - v and ii - vi semiconductors . the high degree of external control of these systems , and the possibility of direct detection for them has opened the route towards a new generation of fast optical matter - wave lasers and amplifiers . @xcite more recently , concerted experimental efforts have been devoted to the realisation of a bose - einstein condensate of microcavity polaritons . @xcite on the experimental side , a challenge to the realisation of a condensed polariton phase might be represented by the finite quality of the cavity mirrors and the resultant short polariton lifetime , of the order of picoseconds . in addition , due to the ` bottleneck effect ' , @xcite the relaxation of polaritons to the zero momentum state can be delayed , hindering the creation of a thermal population in the lowest energy state . it has however been recently shown @xcite that thermalization processes due to particle - particle scattering can be dramatically magnified by increasing the value of the ( non - resonant ) pump power , and by positively detuning the cavity energy above the excitonic energy . under these conditions , the progress towards a zero momentum quasi - equilibrium condensate has been significant , @xcite including a non - linear threshold behaviour in the emission intensity at zero momentum , @xcite the investigation of the second order coherence function , @xcite a characteristic change in the momentum space distribution above threshold , @xcite and evidence that the equilibration time is shorter than the polariton lifetime has been seen . @xcite finally , very recently , a clear demonstration of condensation of cavity polaritons has been demonstrated in cdte . @xcite kasprzak and collaborators have shown that condensation of equilibrated polaritons can be achieved for effective temperatures around @xmath0k and evidence for condensation has been seen in the occupation function , the first order coherence ( both in time and in space ) and in the spontaneous appearance of linear polarisation of the condensate emission . alongside the experimental effort , a significant theoretical effort has been invested in analysing properties and predicting signatures of polariton condensation . @xcite much of this work focuses on modelling the conditions under which condensation can occur , both in equilibrium , @xcite and considering the effects of pumping and decay . @xcite possible signatures include the nonlinear relation of emission at zero momentum to pumping power , @xcite changes to the linewidth , @xcite the pl spectrum and the angular distribution of radiation , @xcite and spontaneous polarisation of emitted radiation . @xcite in this paper , we discuss the optical properties of condensed polaritons , focusing our interests on absorption , photoluminescence ( pl ) and resonant rayleigh scattering ( rrs ) . from our study we conclude that rrs , the coherent scattering by disorder of polaritons into directions other than that of the original probe , represents a powerful tool for investigating the condensed phase . we will show that signatures of condensation are visible in the rrs spectrum , allowing a direct probe of the collective excitation properties of the polariton condensate . in particular , we will show that , above the threshold for condensation , strong emission from the collective sound mode both above and below the chemical potential can be seen in the rrs spectra ( see fig . [ fig : rrsca ] ) . in contrast such features are expected to be much harder to observe in usual pl emission spectra , where the spectrum is dominated by the very strong condensate emission at the chemical potential , which is likely to mask these more subtle features . in addition , we will show that the rrs spectra directly reflects the disorder averaged density of states of excitonic particle - hole like excitations , i.e. bound excitons coupled to the coherent photon field . for @xmath1 as a function of the dimensionless momentum @xmath2 and rescaled energy @xmath3 , for zero detuning , rabi splitting @xmath4mev , temperature @xmath5k , and a disorder strength characterize by an inverse scattering time @xmath6mev . ( a ) non - condensed regime ( dimensionless density @xmath7 ) ; ( b ) condensed regime ( @xmath8 ) ; ( c ) condensed regime ( @xmath9 ) . ( the parameters chosen for these plots are the same as those used later for spectral weight and photoluminescence . ) the value of the chemical potential is explicitly marked [ horizontal green ( gray ) line ] . while in the non - condensed regime , rrs emission is always above the chemical potential , in the condensed phase , emission from the collective sound mode is seen both above and below the chemical potential . ] resonant rayleigh scattering depends on disorder to scatter polaritons between momentum states . in order to carry on our analysis , we introduce a realistic description of disorder and analyse its effects on rrs and more generally on other optical probes , like absorption and pl . to do this , we make use of a quantitatively accurate model for exciton disorder , @xcite and numerically evaluate the distribution of excitonic energies and oscillator strengths associated with a given disorder potential . an accurate treatment of exciton disorder for quantum wells in microcavities is important because the large ratio of exciton mass to photon mass means that those exciton states contributing most to the thermally populated polaritons are strongly localised , i.e. can not be treated within the first born approximation . such a treatment shows that , for a given exciton energy and momentum , there is still a distribution of possible oscillator strengths . for the exciton states relevant to polariton formation , the distribution of oscillator strengths varies from a narrow gaussian at low energies ( below the band edge ) to a porter - thomas distribution at high energies , as is known from previous works . @xcite here , we study the effect of this distribution on the many - body physics , and in particular , on rrs and pl spectra . in this paper , we consider the case relevant to the existing cdte or gaas microcavities of high - quality quantum wells , where the typical excitonic disorder amplitude is smaller than the rabi splitting . in this regime , we can make use of the _ coupled oscillator model _ or wave vector conserving approximation , as explained in ref . . as will be explained in more detail later on , this corresponds to approximating the full exciton green s function by its momentum - diagonal part ( or equivalently considering its disorder average ) , and treating perturbatively the off - diagonal terms . in this way , the translational invariance broken by the presence of the disorder is restored , polaritons have a defined wave vector and , at high enough densities , condense in the lowest momentum state . such a treatment implies that excitonic disorder , being on short length scales , does not lead to spatial inhomogeneity on the length scale associated with the polariton . extended polaritons are formed from a superposition of many localised excitons , therefore recovering translational invariance at the level of polaritons . at low densities , in the non - condensed state , many observable properties can be adequately described by the coupled oscillator model and so can be found from the mean squared oscillator strength at a given energy . in this limit our method recovers the well known results for the inhomogeneous broadening of the polariton pl , @xcite and by considering in addition the mean fourth power of the oscillator strength the averaged rrs response . @xcite however , when condensed , there are observable effects associated with the full distribution of oscillator strengths . in particular , the non - vanishing probability of excitons to have arbitrarily small oscillator strengths has direct consequences for optical probes , including both rrs and pl . in ref . , we considered specific aspects of resonant rayleigh scattering arising from the model discussed in this paper . here , we provide and compare further experimental probes of condensation , and discuss the underlying physical mechanisms involved . all these optical probes , including rrs , are in addition to a non - resonant pumping of the microcavity polaritons . a related problem is studied in ref . . there , the rayleigh scattering of a strong resonant pump is considered ( treating disorder perturbatively ) , thus the probe is the pump . in this paper , in contrast , one has to distinguish two types of coherence . the first type is the internal coherence of the condensed system , which arises spontaneously following non - resonant pumping . the second type is the coherent scattering of an external laser probe on disorder , which can be strongly modified by the presence of the condensate . the form of the density of states , optical density , and distribution of oscillator strengths found from the numerical calculations are also important in other thermodynamic properties and probes of polaritons . one such example is the calculation of the linear blue - shift vs. density in the low density regime . the inclusion of exciton states that couple weakly to photons extends the validity of the current model to densities beyond those in similar treatments , @xcite and provides a stronger basis for the use of those models in the regimes where they are valid . the paper is arranged in the following sections : we describe the effect of quantum well disorder on the exciton oscillator strength in sec . [ sec : disor ] . there , we also give details of the numerical analysis ( sec . [ sec : numer ] ) , which will be used in the many - body model introduced in sec . [ sec : model ] to describe the thermodynamics of polariton condensation ( sec . [ sec : phase ] ) . section [ sec : photo ] describes the optical probes such as spectral weight and photoluminescence ( sec . [ sec : spect ] ) and resonant rayleigh scattering ( sec . [ sec : rrsca ] ) . conclusions are collected in sec . [ sec : concl ] , while appendix [ sec : phase - sens - detect ] explains how one can detect rrs using phase sensitive measurements . the problem of an exciton in a disordered quantum well has been studied at some length in the last two decades . @xcite quantum well disorder can arise due to interface and alloy fluctuations and affects the properties of the excitonic linewidth and absorption spectrum . similarly to refs . , we will assume the external disorder potential to be correlated on a length scale larger than the exciton bohr radius . accordingly , we factorise the excitonic in - plane relative and centre of mass coordinates @xmath10 and assume that the disorder affects only the excitonic centre of mass motion @xmath11 , while the internal degrees of freedom can be restricted to the ground state hydrogenic state @xmath12 . here , @xmath13 is the exciton bohr radius and @xmath14 the reduced mass ( henceforth we will set @xmath15 ) . neglecting the transverse degrees of freedom related to the confinement of the excitons in the qw , the energy associated to the wavefunction @xmath16 is given by the sum of the relative motion energy @xmath17 [ i.e. , the band gap minus the exciton binding energy , @xmath18 and the energy related to the centre of mass motion @xmath19 : @xmath20 \phi_{\alpha } ( { { \mathbf r } } ) = \varepsilon_{\alpha } \phi_{\alpha } ( { { \mathbf r}})\ ; . \label{eq : centr}\ ] ] here , the effective disorder potential @xmath21 represents the microscopic structural disorder averaged over the electron - hole motion . @xcite this can be approximated , e.g. , with a gaussian noise correlated on a length scale @xmath22 with variance equal to @xmath23 : @xmath24 where @xmath25 is the quantisation area . it will be convenient , later on , to introduce the scattering time @xmath26 where @xmath27 is the two - dimensional density of states in the clean limit . in two dimensional non - interacting systems , all states are localised by the disorder potential . however , the localisation length and the character of the excitonic wavefunction change significantly from below to above the band edge @xmath17 . well below the band edge , low - energy lifshitz tail states , @xcite are well localised in deep potential minima , with a nodeless ( roughly gaussian ) shape . these states are rare , because they occur due to large , rare , fluctuations of the disorder potential . in contrast , higher energy states above the band edge have a fractal - like shape with many nodes and can be approximated by a random superposition of plane waves with the same momentum @xmath28 . here , the localisation mechanism is closely related to quantum mechanical interference effects . accordingly , the change of the shape of the centre of mass wavefunctions across the band edge is accompanied by a dramatic change in the excitonic oscillator strength . @xcite the oscillator strength describing coupling of a quantum well exciton to light , @xmath29 , is given by the probability amplitude of finding an electron and a hole at the same position and with centre of mass momentum equal to the photon momentum @xmath30 and is therefore proportional to the fourier transform of the centre of mass wave function , @xmath31 : @xmath32 where , @xmath33 is the dipole matrix element . the dispersion for photons in a microcavity of width @xmath34 is given by @xmath35 , where @xmath36 , and can be approximated , for small momenta , by a parabolic dispersion @xmath37 , with the photon mass given by @xmath38 . because there may be many different exciton wavefunctions corresponding to similar exciton energies , the oscillator strength @xmath39 is a random quantity , which varies both in phase and magnitude . considering many disorder realisations , we find a distribution of squared oscillator strengths , @xmath40 as a function of the energy @xmath19 and momentum @xmath30 . this distribution reflects the statistical properties of the centre of mass excitonic wave functions . to compare to experimentally relevant observables , we introduce the density of states ( dos ) , @xmath41 and the mean squared oscillator strength : @xmath42 where @xmath43 is the average over different disorder realisations . these quantities are related to the excitonic optical density by the relation : @xmath44 experimentally , the excitonic optical density can be measured by dividing the pl emission by the excitonic occupation . @xcite focusing for the moment on average properties , rather than on the entire distribution , a simple expression exists for the mean squared oscillator strength in the high energy limit . in this limit , where the dos is flat and energy independent , @xmath45 , making use of the born approximation , the squared average oscillator strength @xmath46 does not depend separately on the disorder potential correlation length @xmath47 and variance @xmath48 , but instead only on the scattering time @xmath49 ( see fig . [ fig : gepsp ] ) : @xmath50 where @xmath51 is the free particle dispersion . from this form , and the effectively constant density of states , one can see that both the mean squared oscillator strength and the excitonic optical density are symmetrical . the comparison of this approximate form to the numerical simulation is shown in fig . [ fig : gepsp ] , in contrast , for energies much below the band edge , the specific asymptotic expression of @xmath46 depends whether the correlation length @xmath47 is smaller or larger than the localisation length @xmath52 . in the white noise limit,@xcite @xmath53 , one can show that the centre of mass wavefunction @xmath11 can be approximated by a gaussian centred at a randomly distributed site @xmath54 , @xmath55 and therefore giving a squared oscillator strength proportional to : @xmath56 thus , here the distribution of squared oscillator strengths is very narrow , with a mean square value given by the above form . similarly , much theoretical ( and numerical ) work has been done to establish the energy dependence of the density of states . in the low energy tail , as before , the specific asymptotic form of the dos depends on the value of the correlation length @xmath47 . in the white noise limit , one can show that @xmath57 , while , in the opposite ( classical ) limit , @xmath58 , instead one has @xmath59 ( see , e.g. , ref . ) . in general , for the finite values of the disorder correlation length corresponding to typical experiments , the regions in energy where one of these two analytical regimes apply are very restricted , and therefore a numerical analysis is required . numerical analysis is also essential in order to account for the distribution of squared oscillator strengths @xmath60 near the band edge . this distribution changes substantially from low energy states to high energy ones . as we discuss more in detail in the next section , for high energy states the oscillator strength distribution is governed by a porter - thomas law , while for lifshitz tail states the distribution follows a narrow gaussian - like distribution . neither of these distributions apply for energies around the band edge . however , it is states near the band edge that have the largest optical density , and so the distribution of oscillator strengths in this region has a significant impact on derived quantities . it thus becomes essential to use numerical analysis to find the entire distribution of oscillator strengths . to photons of momentum @xmath61 ( a ) ( from ref . , for comparison ) and @xmath62 ( b ) , where all exciton states found numerically . results are taken from 160 different realizations of disorder potential , and for ( b ) , coupling strengths from eight different photon momenta @xmath63 with the same value of @xmath64 are combined . the mean squared averaged oscillator strength @xmath65 for the same value of momentum [ lower red ( gray ) points ] and the density of states @xmath66 [ upper green ( gray ) points ] are also explicitly plotted . inset : fit of @xmath67 to the expression with a renormalized energy @xmath68 . ] to photons of momentum @xmath61 ( a ) ( from ref . , for comparison ) and @xmath62 ( b ) , where all exciton states found numerically . results are taken from 160 different realizations of disorder potential , and for ( b ) , coupling strengths from eight different photon momenta @xmath63 with the same value of @xmath64 are combined . the mean squared averaged oscillator strength @xmath65 for the same value of momentum [ lower red ( gray ) points ] and the density of states @xmath66 [ upper green ( gray ) points ] are also explicitly plotted . inset : fit of @xmath67 to the expression with a renormalized energy @xmath68 . ] versus energy and momentum [ or equivalently angle @xmath69 . note that the scale in angle is not linear . the free particle dispersion @xmath70 [ solid green ( gray ) line ] , and the a trace of the mean squared oscillator strength for two representative values of momenta , @xmath71 ( @xmath72 ) and @xmath73 ( @xmath74 ) [ red ( gray ) plus symbols ] are explicitly plotted ( cf fig . [ fig : gepsp ] ) . the figure is adapted from ref . . ] to solve eq . , we exactly diagonalise this problem within a finite basis set , using conjugate - gradient minimisation with pre - conditioning of the steepest descent wavevector ( for a detailed discussion of this method , see ref . ) . we find the exact eigenvalues and eigenfunctions in a finite basis on a grid of @xmath75 points for a system of size @xmath76 m , @xmath77mev , @xmath78 and excitonic mass @xmath79 . these parameters give an inverse scattering time of @xmath80mev . for this choice of the grid , one can show that convergence is reached . from the evaluated eigenvalues @xmath81 and eigenstates @xmath82 over 160 realisations of the disorder potential , we can derive the excitonic density of states , the oscillator coupling strength and its squared average , eq . , which we plot in figs . [ fig : gepsp ] , and [ fig : geps2 ] , while the corresponding optical density is plotted in fig . [ fig : optic ] . the lower panel of fig . [ fig : gepsp ] shows the squared coupling strength @xmath60 versus energy for a fixed value of momentum , @xmath83 , corresponding , for a cavity of @xmath84ev , to an angle of @xmath85 . note that , because of the presence of disorder in the quantum well , one photon with a given momentum couples with many exciton states with different energies . we will see later on therefore that a polariton with a given momentum is formed by the superposition of one photon state @xmath86 and many exciton states @xmath87 . these states are more or less strongly coupled depending on the distribution of oscillator strength for that given momentum . as fig . [ fig : gepsp ] shows , by probing a quantum well at a large angle , i.e. with high momenta photons , the excited excitons with larger oscillator strength are the ones which are almost - delocalised in nature and with a many - node fractal - like shape . by taking the average over many ( 160 ) disorder realisations , the squared average oscillator strength @xmath65 shown in the inset in fig . [ fig : gepsp ] is well described by the lorentzian shape predicted by the born approximation of eq . , with a fitted width of @xmath88mev in good agreement with the theoretical value @xmath80mev . however , the peak of the lorentzian does not coincide with the energy of the clean limit , @xmath68 , but is renormalised down in value , as can be shown by employing a self - consistent born approximation . in contrast , as shown in the upper panel of fig . [ fig : gepsp ] , for photons with zero momentum , the maximum value of the oscillator strength characterises excitonic states below the band edge which are more localised in nature . this can be easily understood by the following qualitative argument : at very low energies , the excitonic state is strongly localised in a deep potential minimum and has no nodes . increasing the energy , at first the localisation length increases [ e.g. , @xmath89 in the white noise limit ] and thus increases the oscillator strength . however , eventually , the wavefunction starts developing nodes and consequently the squared average oscillator strength decreases . when @xmath90 , only the high energy side of @xmath91 can be described by the born limit . for the chosen values of the disorder potential correlation length @xmath47 and variance @xmath48 , an analytical expression for the low energy ( lifshitz ) tail is not known . the crossover from localised to more plane - wave - like excitonic states obtained by increasing the value of the photon momentum is plotted in fig . [ fig : geps2 ] . at high momenta , the maximum of the squared average oscillator strength @xmath65 shown in fig . [ fig : geps2 ] follows the free particle dispersion , plus a renormalisation down in energy , which decreases for higher photon momenta , i.e. these states are described well by free particles , including disorder in the first born approximation . at low momenta , the states which have the stronger oscillator strength are effectively localised . the crossover , as seen in the contour plot , happens at relatively large angle , @xmath92 , because of the large ratio of exciton to photon mass : the crossover momentum is set by exciton mass , but its conversion to an angle depends on effective photon mass . for a connected reason , the thermally populated polariton states are formed out of strongly localised ( i.e. beyond first born approximation ) excitonic states : for values of temperature and photon mass relevant for experiments , thermal population of polaritons extends up to around @xmath93 in momentum , which as seen in fig . [ fig : geps2 ] corresponds to exciton states not accurately described by the born approximation . for this reason , in the following we will concentrate on the oscillator strength corresponding to @xmath94 , @xmath95 . mev and the fwhm @xmath96mev . ] finally we plot in fig . [ fig : optic ] the optical density . for the chosen values of @xmath47 and @xmath48 , the optical density shows a maximum around @xmath97mev below the band edge , a full width at half maximum ( fwhm ) of approximatively @xmath96mev and a clear asymmetry of the line - shape . it is well known @xcite that , by indicating with @xmath98 the confinement energy of the lowest state in a typical potential minimum , the excitonic line shape is determined by the ratio @xmath99 and , for a finite value of this , the optical density develops an asymmetry towards higher energies . for our choices of parameters , @xmath100 . asymmetry of the optical density of quantum well excitons has been also measured experimentally ( see , e.g. , ref . ) , by dividing the measured pl by the excitonic occupation . as mentioned earlier , the full distribution of oscillator strengths , and not just its mean squared value will be important . it is useful to discuss here some technical details of how this is extracted . the numerical analysis provides the excitonic eigenvalues and eigenstates only within a finite interval in energy . ( the lowest energy states arise from rare potential fluctuations , which would require a larger region of space to be sampled ; the highest energy states have spatial variation on lengthscales finer than our grid . ) within the interval of energies found , averages are performed making use of the raw data coming from 160 realisations of the disorder potential ; outside this interval , averages are taken by extrapolating the numerics . in particular , in the low energy lifshitz tail , as the oscillator strengths have a narrow gaussian distribution around its squared averaged value , we approximate the distribution of @xmath95 with a delta function at its extrapolated value @xmath101 . in the very high energy region , instead , we make use of the porter - thomas distribution , @xmath102}{\sqrt{2 \pi x \bar{x } } } } \ ; , \ ] ] where @xmath103 , again extrapolating the fitted value for the squared averaged oscillator strength . in addition , we fix the overall scale of @xmath95 to match integrated optical intensity with half the experimentally measured rabi splitting @xmath104 squared : @xmath105 this normalisation accounts for factors other than the wavefunction in eq . ( [ eq : oscil ] ) . having described the effect of quantum well disorder on the energies of excitons and on their dipole coupling to the cavity photons , we will next introduce the many - body hamiltonian that will allow us to describe a system of thermalized interacting polaritons which , for a given density and temperature , can condense to a superfluid phase . the hamiltonian describing a model of excitons with random energies @xmath19 , corresponding to the centre of mass eigenstate from eq . , dipole coupled to a cavity mode can be written as @xmath106 the description of exciton - exciton interaction in this model is somewhat simplified : it imposes the constraint that two excitons can not occupy the same energy level @xmath19 , because of their composite fermionic nature . this constraint is implemented by representing each energy level [ where energy levels are distributed according to the density of states , @xmath66 ] by two possible states , respectively @xmath108 and @xmath109 . these two states are then represented by occupation of two fermionic states , so that the ground state @xmath110 , and an excitonic state @xmath111 , eliminates the unphysical states @xmath112 and @xmath113 . in this way saturation effects are introduced to all orders . note that these fermionic operators @xmath114 can not be written as a linear combination of free electron and hole operators this is a result of having already included the intra - exciton coulomb interaction . the transition matrix element between a photon state @xmath115 and an exciton state , @xmath116 , is given by @xmath29 from eq . and evaluated numerically , as explained in the previous section . it is convenient to rescale this coupling according to @xmath117 where @xmath118 is the inverse level spacing @xmath119 , measured in units of the excitonic rydberg energy . this corresponds to measuring the density of particles per bohr radius squared . using these units , we may write the total number of excitations @xmath120 n a dimensionless form by introducing a dimensionless density of particles @xmath121 , or equivalently @xmath122 , where @xmath123 is the physical areal density of particles . and thermal occupation factor [ red ( gray ) line starting at top left ] for @xmath124mev and @xmath125k . ] in the limit of vanishing density , our occupancy constraint has no effect , and so this description is equivalent to one treating excitons as bosons . as will be shown later , in sec . [ sec : blues ] , for low densities , our occupancy constraint is also equivalent to a model of excitons as bosons , but with some effective saturation interaction . thus , as in the bosonic case , saturation effects are linear @xcite and the blue - shift increases linearly in density . there is however an important difference to those previous works . in contrast to the clean systems studied there , our effective saturation interaction has a coefficient related to disorder and temperature rather than to the rabi splitting and bohr radius only . however , increasing the density , saturation effects , and thus the underlying fermionic degrees of freedom become more and more important , and the conventional linear treatment becomes inadequate . in our treatment saturation effects are introduced exactly and taken into account to all orders . our model is accurate as far as the occupancy is restricted to the strongly localised lifshitz states in the tail below the band edge ( see fig . [ fig : occup ] ) . all the results described here respect this limit . one can moreover show that it is this density regime which is relevant for on - going experiments in cdte . @xcite higher energy states , beyond the lifshitz tail , contribute to the optical response , however they are only weakly occupied , and so the above limit remains valid . what the model eq . does not include is the coulomb interaction between excitons with different energies @xmath19 , i.e. , at low densities , on different localisation sites @xmath54 . this contribution is expected to be small in the low density regime . this low density regime will be observed all the way through this paper . at the same time we are not including double - occupancy of a single exciton energy level , which could be important at higher densities . in the following , we will make extensive use of the numerical analysis of oscillator strength distributions explained in the previous section . let us note here again that in the non - condensed regime , the full distribution is unnecessary the excitonic optical density is sufficient to derive the polariton dispersion and its inhomogeneous broadening . however , in the condensed phase , and in order to correctly describe resonant rayleigh scattering mechanisms , we will see that it is of fundamental importance to consider the entire distribution of oscillator strength . we consider a thermal equilibrium system of polaritons , as describing the situation with non - resonant pumping and strong thermalization processes . this limit has been recently demonstrated to be accessible in on - going experiments . @xcite the total density of polaritons is fixed by introducing a chemical potential @xmath126 . making use of standard path integral techniques and of the grand canonical ensemble , @xmath127 , we integrate over the fermionic fields , and thus write the partition function @xmath128 $ ] , with @xmath129 , in terms of the following imaginary time action : @xmath130 = \int_0^\beta d\tau \sum_{{{\mathbf p } } } \psi_{{{\mathbf p}}}^ * \left(\partial_\tau + \tilde{\omega}_{{{\mathbf p}}}\right ) \psi_{{{\mathbf p } } } - \operatorname{tr}\ln g^{-1 } \ ; . \label{eq : actio}\ ] ] here , @xmath131 is the , energy level diagonal , inverse single - particle green s function : @xmath132 where @xmath133 and @xmath134 are respectively the excitonic and photonic energies measured with respect to the chemical potential . within this path integral formulation , it is possible to show @xcite that the constraint on the site occupation can be taken into account elegantly by shifting the fermionic matsubara frequencies according to @xmath135 which we will assume from here on in . we will see later in section [ sec : photo ] how the formation of polaritons , the result of strong coupling between photons and excitons , influences the optical response and in particular how the optical response is affected when the system enters a condensed phase . this section will discuss the mean - field equations , and resultant phase diagram ; however first we wish to stress an important point . as observed previously , the relevant polariton states form from the mixing of a photon with a given momentum to many localised excitonic states with different energies . the strength of this mixing is determined by the distribution of oscillator strengths . although the quantum well excitons may be disorder - localised , for a weak disorder potential , with @xmath136 , and in the absence of strong photon disorder ( which in some cases can also be relevant @xcite ) the resultant polaritons are delocalised . as such , polaritons will be described by a momentum quantum number , and condense , as in the usual picture , in the lowest momentum state . the static and uniform minimum @xmath137 of the action can be found from the equation of motion , or saddle - point equation , which has to be solved together with the mean - field equation for the total number of excitations : @xmath138\ ; . \label{eq : homo2}\end{gathered}\ ] ] here , the spectrum of particle - hole excitations , i.e. the quasi - particle spectrum discussed later , is controlled by the energy of an exciton in presence of a coherent photon field @xmath139 , and is given by : @xmath140 in the following , we will solve simultaneously these coupled equations , both above ( @xmath141 ) and below ( @xmath142 ) the critical temperature for condensation . it is interesting to compare these with similar equations which have been already considered , e.g. , in the bcs theory of superconductivity , @xcite or in models describing gases of fermionic atoms close to a feshbach resonance.@xcite considering particularly the case of feshbach resonance , the density of particles is typically kept fixed , while the interaction strength can be varied externally to span across a bec condensate of tightly bound fermionic molecules and a bcs state of loosely bound fermionic pairs . in our case , two main important differences arise . firstly , density is the parameter which is changed . secondly , the coupling strength of excitons with cavity light depends on the excitonic energy . there is a wide variety of behaviour that can come from these coupled equations and , controlled by the distribution of oscillator strengths and by changing the density of states . by solving the ` gap equation ' eq . in the limit @xmath143 , let us concentrate for the moment on the critical temperature @xmath144 as a function of the chemical potential @xmath126 , forgetting about the corresponding excitation density . at the phase boundary , where @xmath145 , in contrast to inside the condensed phase , the squared oscillator strength appears only linearly in eq . , this expression therefore involving only the optical density @xmath146 . it is clear that the standard bcs result can be thought of as a limiting case where the effective optical density is much wider than the temperature . @xcite in the opposite limit , of a narrow delta - like optical density , one can solve the gap equation analytically . @xcite by taking into account the density equation , the critical temperature can be expressed as a function of the excitation density rather than as a function of the chemical potential . in the standard bcs theory there is no modulation of the interaction strength and therefore no difference between optical density @xmath146 and density of states @xmath66 . in our system instead one has to introduce separately both quantities , and further a realistic description of the dos is required in order to measure the density in physical units . in previous treatments , @xcite where the dos has been attributed an arbitrary gaussian - like shape ( and moreover where the coupling strength was kept fixed ) , the density was measured in arbitrary units . note however that , the fact that the density of states asymptotes to a constant at large densities , and the presence of those exciton states that couple only weakly to light , is responsible only for the asymptotic form of the phase boundary at large densities ( well beyond the range shown in fig . [ fig : criti ] and beyond the validity of our model ) . at smaller densities , such states have a little effect on the phase boundary . in order to calculate the phase boundary using our numerically calculated distributions , we will solve eq . ( [ eq : homo1 ] ) and eq . ( [ eq : homo2 ] ) numerically . how the disorder averaged sums in these equations are to be evaluated has been discussed sec . [ sec : numer ] . versus the dimensionless density @xmath147 for effective zero detuning @xmath148mev ( @xmath149 is the energy at which the excitonic optical density has its maximum , and so as seen in fig . [ fig : optic ] , @xmath150 implies @xmath151mev ) ( solid black ) and for positive detuning @xmath152mev ( @xmath153mev ) ( dotted - dashed black ) and @xmath4mev . the mean - field boundaries for the two different detunings are cut - off by the expected linear dependence of the critical temperature , as indicated at extremely low densities [ orange ( gray ) solid and dotted - dashed lines ] . the horizontal dashed line marks the temperature of @xmath125k which will be used in later figures . ] the critical temperature @xmath144 at which the system , for a given density , goes from a non - condensed phase to a condensed phase , is plotted in fig . [ fig : criti ] for two different values of the detuning @xmath154 , where @xmath149 is the energy at which the excitonic optical density has its maximum ( see fig . [ fig : optic ] ) . it is clear that increasing the detuning of the cavity mode above the excitonic one decreases the overall photonic fraction and therefore decreases the value of the critical temperature at a given density of excitations . conversely , if one fixes temperature , the critical density for condensation is higher for larger detuning , as fig . [ fig : criti ] clearly shows . this dependence is easy to understand in the extremely low density limit , where the critical temperature varies linearly with density , and is inversely proportional to the polariton mass : by increasing the detuning , the polariton mass also increases . having discussed the thermodynamics via the mean - field equations , we now turn in the rest of this paper to the analysis of the optical responses and probes of microcavities , both below and above threshold for condensation , with particular emphasis on the resonant rayleigh scattering emission . such optical responses are described by considering fluctuations about the mean - field theory outlined in this section . the spectrum of these fluctuations , and the various optical response functions that these fluctuations describe is the subject of sec . [ sec : photo ] . as well as describing the optical response , fluctuations about the mean field theory can also be important in describing , in the extremely low density limit , corrections to the mean field thermodynamics . since the subject of this paper is optical probes , we will not discuss the subject of fluctuation corrections to the critical temperature here , as they have been discussed elsewhere @xcite . for low densities , a bosonic form can be recovered from our hamiltonian by using the holstein - primakoff transformation to express the two - level systems in terms of bosonic operators , and so this low density limit can of course be treated within our model . a smooth crossover between the low density ( fluctuation dominated ) , and the higher density ( long - range interaction dominated ) regimes can be found by considering fluctuation corrections to the density , as was studied in ref . the optical absorption and emission spectra of a microcavity can be derived by first finding the green s function describing photon propagation . the non - condensate part of these response functions ( i.e. away from zero momentum ) may be found in practice by considering the green s function for fluctuations about the mean - field solution , @xmath155 , where @xmath156 is a bosonic matsubara frequencies , and expanding the action up to quadratic terms : @xmath157 this gives the matsubara inverse photonic green s function : @xmath158 physical response functions can be found by analytic continuation of the imaginary time matsubara green s function to real times . @xcite the matrix elements of @xmath159 can be expressed in terms of the bare photon energy @xmath160 and the excitonic quasi - particle energy @xmath161 as follows , @xmath162 where @xmath163 it is useful to decompose the photonic green s function into its momentum diagonal and off - diagonal contributions : @xmath164 by treating the off - diagonal terms perturbatively , the translational invariance can be recovered and polariton eigenstates can be labelled by momentum vectors . in the next section , we will focus on the diagonal terms , which characterise both the spectral weight and the photoluminescence emission . section [ sec : rrsca ] is then dedicated to rrs , for which we will see that the off - diagonal terms are necessary in order to describe the scattering of an incident photon ( via the excitonic component of the microcavity polariton mode ) into directions other than its original direction . @xcite secondary emission from a semiconductor microcavity after optical excitation is the source of both incoherent pl and coherent rrs . at short times , this emission is dominated by rrs , the coherent scattering from disorder , and so is at the energy of , and coherent with , the incident radiation . at longer times , phonon and particle - particle scattering destroy coherence , and redistribute the energy , leading to a quasi - equilibrium distribution of energies , and thus the incoherent pl emission intensity , @xmath165 is given by the bose occupation factor @xmath166 times the spectral weight : @xmath167 the spectral weight can be interpreted as an absorption coefficient @xcite ( the probability to absorb a photon minus the probability to emit a photon ) , where negative values of @xmath168 represent gain . in contrast , the pl @xmath169 is always positive . in calculating the pl , it is convenient to make an approximation by neglecting multiple polariton scattering , while still including the effects of exciton - disorder scattering to all orders . this is discussed in ref . , where comparison between this approximation and exact numerical calculations show this approximation to be remarkably good . physically , this is a good limit to consider because the typical exciton - disorder scattering times are very short compared to the inverse frequencies considered in pl . as pl depends on the momentum diagonal part of the photon green s function , neglect of multiple scattering means in practice averaging over disorder realisations at the level of the inverse photon green s function , @xmath159 . since off - diagonal terms in the inverse green s function , eq . , break translational invariance , they average to zero , and so scattering between different photon momentum states can thus be neglected . the off - diagonal terms neglected here will however play a crucial role in the case of rrs response , as discussed below . the spectral weight calculated from this formula , along with the quasi - particle density of states is shown in fig . [ fig : spwei ] . as a function of the dimensionless momentum @xmath2 and rescaled energy @xmath3 , for zero detuning @xmath170 = 0 ( @xmath171mev ) , @xmath4mev , @xmath5k and photon mass @xmath172 . right column : plot of the quasi - particle dos for the same choice of parameters as the left column , respectively : ( a ) non - condensed ( @xmath7 , @xmath173mev ) ( the bare exciton and photon dispersions [ blue ( dark gray ) dotted line ] and the upper and lower polariton dispersions [ red ( gray ) solid line ] obtained from the effective coupled oscillator model are shown for comparison ) ; ( b ) condensed ( @xmath174 , @xmath175mev ) ; ( c ) condensed ( @xmath9 , @xmath176mev ) . ] we first discuss this in the non - condensed case , where @xmath177 , and so @xmath178 , and the photon s green function becomes diagonal in particle - hole space , and simplifies to : @xmath179 this expression describes the coupling of one harmonic oscillator ( the photon mode ) to many harmonic oscillators ( the exciton modes ) . in the limit of small density , the chemical potential is far below all exciton modes , and so @xmath180 . in this limit , the result is identical to a bosonic description of excitons sometimes also called the _ linear dispersion model_. the underlying fermionic structure appears as a reduction of the effective exciton - photon coupling , due to saturation effects described by the @xmath181 term , and discussed in more detail below . from this _ linear dispersion model _ , describing one oscillator coupled to many , there are in general two broadened modes at high and low energies here these are the lower polariton ( lp ) and upper polariton ( up ) and a continuum of modes associated with the exciton optical density . however , when considering the corresponding spectral weight , modes are weighted by their photonic component , so one primarily sees the lp and up modes , and only weak emission near the bare excitonic states the excitonic ` dark ' states . these three features are clearly visible in fig . [ fig : spwei ] , and have been previously predicted by houdr _ et al . _ @xcite making use of a simplified and exactly solvable model . note that in the limit of an infinitely narrow spectral width , these dark exciton states become entirely dark , having vanishing photon component . the presence of ` dark ' exciton states , coexisting with strong coupling polaritons can explain the simultaneous observation of large rabi splittings and long decay times seen in some experiments . @xcite when the rabi splitting is substantially larger than the exciton inhomogeneous broadening , ( in our case we have @xmath182mev and the fwhm of the optical density @xmath183mev ) , there is a substantial difference between broadening of lower and upper polaritons . this is because the high energy tail of the optical density decays as a power law , while the low energy lifshitz tail decays exponentially . thus , the optical density has a larger value at the up mode than at the lp mode , giving a larger broadening compared to the almost vanishing width of the lp . this description of the polariton linewidth due to the excitonic inhomogeneous broadening coincides with that of whittaker @xcite , and it has been well tested experimentally . @xcite in addition , the location of the lp and up can be found by making use of an effective two - oscillator model , i.e. assuming a narrow delta - like optical density , @xmath184 where @xmath149 is the location of the maximum optical density and effectively , the exciton energy . in this case the system reduces to two coupled oscillators , giving unbroadened lp and up poles at @xmath185 where , @xmath186 is the reduced rabi splitting , due to saturation effects at higher densities . the reduction of @xmath187 splitting thus translates directly into a reduction of the lp up splitting , and thus a blue - shift of the lp , which at small densities can be shown to be linear . for comparison , the results of this formula are shown by the solid ( red ) lines in panel ( a ) of fig . [ fig : spwei ] . in sec . [ sec : blues ] we will discuss the calculation of this linear blue - shift in the low density regime , by evaluating @xmath187 as a function of density , making use of the full density of states . let us now turn to the signatures of condensation as seen in the spectral weight , and thus in the pl emission ( but which are most probably masked in the pl emission by strong emission from the condensate mode ) . when condensed , the polariton modes are replaced by new collective modes : @xcite the lower polariton becomes the linear goldstone mode , and two branches appear below the chemical potential . the appearance of new excitation branches below the chemical potential is generic to condensation , @xcite however the experiments required to probe these modes are not easy in other bose condensed systems , such as atomic gases . for this reason , let us briefly discuss the physical origin of these new branches , and the reason they may be observed in optical response of polariton systems . the bogoliubov spectrum arises because of the possibility of processes that spontaneously either create or destroy two non - condensed particles . ( such processes arise due to scattering from or to the condensate . ) as a result of these processes , there is mixing between the propagation of an extra particle , or propagation of a missing particle ( i.e. a hole ) . ( such a language of particle and hole refers to the normal state quasi - particles , in the current case the normal state polaritons . ) in the normal state , one can separately calculate the spectral weight of particle excitations , which have weight only above the chemical potential , and hole excitations which have weight only below the chemical potential . when condensed , as the `` particle '' and `` hole '' spectral weights become mixed , this mixing can lead to spectral weight below the chemical potential that is associated with particle propagation . to observe this weight it is however necessary to have the ability to inject a particle which is not a quasi - particle of the condensed system , i.e. not a bogoliubov quasiparticle . in atomic experiments this is hard to achieve , but for polaritons can be naturally achieved by injecting a photon . these new branches below the chemical potential are seen as optical gain in the spectral weight ( see fig . [ fig : spwei ] ) . note that the presence of pumping and decay will modify the linear dispersion of the goldstone mode at low momentum , making it diffusive , @xcite and that quantisation by disorder may also have some effect . @xcite the spectral weight also contains information about the excitonic quasi - particle dos of the system . as discussed above for the non - condensed case , this is visible via the appearance of ` dark ' exciton states [ see figs . [ fig : vsdet ] and [ fig : spwei](a ) ] . when condensed , there is a coherent field that modifies the energies of these excitonic quasi - particle states , as given in eq . ( [ eq : quasi ] ) , and thus modifies their density of states . the density of states given by taking the energy of such modes is shown by the right column of fig . [ fig : spwei ] , and can be compared to the corresponding faint features seen in the photon spectral weight in the left column . for @xmath90 as a function of the detuning @xmath170 and energy @xmath188 , for a fixed and very low value of density ( @xmath189 , @xmath190mev ) and @xmath191k ( @xmath192 ) . the bare exciton and photon dispersions [ blue ( dark gray ) dotted line ] and the upper and lower polariton dispersions [ red ( gray ) solid line ] obtained from the effective coupled oscillator model are explicitly shown . ] the change to the quasi - particle density of states that occurs on condensation requires some explanation . as mentioned earlier , the change of this spectrum is similar to that seen in the bcs theory of superconductivity , as such it is surprising that there is no gap in the density of states of fig . [ fig : spwei ] . in fact , there would be a gap if the distribution of oscillator strengths were replaced with the mean square oscillator strength . in that case , there would be a branch cut , and a @xmath193 singularity in the density of states . however , with the full distribution of oscillator strengths , one finds that there is always a non - vanishing probability of arbitrarily small oscillator strengths . ( note that if the oscillator strength were zero , the corresponding state would not contribute to any photon response , but for an arbitrarily small coupling , it has some contribution . ) thus , the contribution of these weakly coupled exciton states in effect smoothes out the gap . thus , this system can be vaguely described as ` gapless fermion condensation ' , analogous to ` gapless superconductivity ' , but through a mechanism very different to the standard abrikosov - gorkov mechanism considered in superconductors . @xcite as one moves away from the chemical potential , the mixing of ` particle ' and ` hole ' modes described by the bogoliubov spectrum reduces . as a result , far above the chemical potential , the modes are much as the uncondensed case , and the far below the chemical potential , the new modes disappear . to explain this quantitatively , it is clearer to discuss the case of weakly interacting bose gas . @xcite writing the bogoliubov mode energy as @xmath194 , where @xmath195 is the mean - field value , the spectral weight is given by : @xmath196 it is thus clear that at large momenta , where @xmath197 , the coefficient of the first delta function ( modes above the chemical potential ) will go to one , and the the coefficient of the second ( modes below ) will be suppressed to zero ( roughly quadratically in energy ) . in contrast to the power law suppression of the spectral weight of modes far below the chemical potential , the pl signal from modes far above zero is suppressed exponentially by the thermal occupation of these modes [ see eq . ] , while below the chemical potential , there is no such decay . however , this discussion neglects emission from the condensate mode , which should included at zero momentum , as defined in eq . . to see a noticeable change of the normal modes requires a relatively large condensate density . due to instrumental broadening , the presence of emission from this large condensate density might obscure emission from the normal modes , so they may only be weakly visible , as shown in fig . [ fig : plumi ] . as discussed below , rrs may provide a means of escape from this problem . as a function of the dimensionless momentum @xmath2 and rescaled energy @xmath198 , for detuning @xmath199mev , @xmath4mev , @xmath5k and @xmath200 : ( a ) non - condensed ( @xmath7 , @xmath173mev ) ; ( b ) condensed ( @xmath201 , @xmath202mev ) ; ( c ) condensed ( @xmath9 , @xmath203mev ) . ] before discussing the rrs response , let us briefly discuss a point mentioned above the calculation of the lp blue - shift as a function of density , making use of the full density of states , and our model of saturation effects . this can be observed in pl experiments by the change of blue - shift as a function of intensity of the non - resonant pumping . in the non condensed regime , the blue - shift of the lp in our model is a consequence of the saturation of the ( disordered ) energy levels . following the narrow band - width limit described in eq . ( [ eq : lpeup ] ) , this saturation blue - shift can be found if one has an expression for the chemical potential at a given temperature and density . in the low density limit , the expression for density @xmath147 as a function of @xmath126 [ eq . ( [ eq : homo2 ] ) with @xmath177 ] can be inverted in terms of elementary functions . as illustrated by fig . [ fig : occup ] , in the low density limit , the chemical potential is far below the band edge . thus , there are two significant contributions to density : one from low energies , in the tail of the dos , but at large occupation , and one from high energies , in the tail of the occupation , but at large dos , thus : @xmath205 here @xmath206mev is the energy which characterises the exponential decay of dos in the tail of the lifshitz states this coefficient is extracted by an exponential fit to our numerical dos . at a temperature of @xmath207mev , the dominant contribution to the density is given by the second integral and the chemical potential increases logarithmically with the density : @xmath208 from this expression , and from the coupled oscillator expressions of the lp eq . , we can explicitly derive the reduced rabi splitting due to saturation effects , @xmath209 and the lp blue - shift , which is thus linear in this low - density regime : @xmath210 where we have reintroduced here the density @xmath211 per unit volume . remaining at small densities , but now considering the low temperature limit , @xmath212 , then the dominant term in eq . ( [ eq : density - assympt ] ) is the first , and so eq . ( [ eq : appbl ] ) should be modified by replacing the thermal length @xmath213 , by a characteristic disorder length @xmath214 . it is instructive to compare this result with that for a clean system ; in this case blue - shift of the lp has been attributed either due to coulomb interaction @xcite or to saturation effects @xcite ( where the expression given here is valid only in the dilute limit ) : @xmath215 because the excitonic rydberg @xmath216 can be of the same order of magnitude as the rabi splitting @xmath104 , in a clean system the two shifts can be expected to be of the same order of magnitude . considering just saturation effects , the difference between clean and dirty systems is that , at low temperature , in a clean system blue - shift depends on exciton number per square bohr radius , while in a disordered system the relevant length is that characteristic of the disorder potential , which is in general larger than the bohr radius . finally , we wish to observe that there is a distinction between coulomb and saturation effects : coulomb interactions result in a blue - shift of both the lp and up , while saturation leads to a blue - shift of the lp and a red shift of the up , i.e. a collapse of the rabi splitting . thus , these effects can be experimentally distinguished , and their relative magnitudes determined . in the condensed regime , the equivalent of the lp is the goldstone linear mode , which by definition starts at the chemical potential . thus , the observed blue - shift is a direct observation of chemical potential vs. density . figure [ fig : vsden ] shows the variation of the spectral weight at zero momentum as a function of density , from which the energy of the zero momentum lp and up modes can be extracted . also shown , for comparison , is the chemical potential vs.density ( dashed green line ) . the locking of the lp mode to the chemical potential beyond the critical density ( red solid line ) is clearly visible , and the behaviour of the lp mode vs density allows the extraction of density from an experimental measurement of lp blue - shift . for @xmath90 as a function of the dimensionless density @xmath147 and the rescaled energy @xmath217 , for zero detuning ( @xmath218mev ) and @xmath191k ( @xmath192 ) . the rescaled chemical potential @xmath219 [ green ( light gray ) dashed ] , the non - condensed lower and upper polariton modes from the two coupled oscillator model [ blue ( dark gray ) dotted ] , and the critical density for condensation @xmath220 [ vertical solid red ( dark gray ) line ] are shown for comparison . ] as described in refs . , the rrs intensity describes the probability to inject a photon into the cavity with momentum @xmath30 , and detect a coherent photon at @xmath221 with @xmath222 . the intensity of such a signal may be written as : @xmath223 since this involves the off - diagonal momentum space components of the green s function , it is necessary to include the off - diagonal momentum components of the inverse green s function eq . . however , as discussed in sec . [ sec : photo ] , we will consider only a single polariton scattering and again neglect multiple polariton scattering ( but include all orders of exciton - disorder scattering ) . such an approximation is reasonable for the same reasons discussed at the start of sec . [ sec : photo ] . in the following , we propose using the rrs signal to probe the excitation spectrum in the presence of a polariton condensate . it is therefore necessary to be able to separate the rrs signal from the strong photoluminescence that would arise from the equilibrium state with large polariton density . further , since a condensed polariton system may have a strong nonlinear response to an applied probe ( i.e. stimulated scattering if the probe significantly affects the population of polariton modes ) , it is necessary to use a weak rrs probe . fortunately , the coherent nature of rrs allows exactly this : one can detect a weak rrs signal by phase sensitive measurement , as is discussed in detail in appendix [ sec : phase - sens - detect ] . hence , the limit on intensity of the probe is provided by the sensitivity of the ccd camera , and not by the incoherent photoluminescence background . this therefore allows a probe sufficiently weak that only the linear rrs response is seen , and nonlinear effects can be avoided . when non - condensed , the single polariton scattering expansion of the photon green s function in the off - diagonal terms gives @xmath224 the factors @xmath225 appearing here can be interpreted as a filter , allowing a response only when the outgoing ( or equivalently incoming for @xmath226 ) momentum has an energy close to the polariton mode at the given energy . this means that @xmath227 and so is responsible for the ring - shaped rrs signal observed in experiments . @xcite in contrast , the term @xmath228 describes scattering between momentum states via polariton - exciton - polariton scattering . this term has a large variation from disorder realization to disorder realization , and is the reason for the speckle seen in rrs experiments . this speckle , and the disorder averaged rrs intensity are shown for comparison in fig . [ fig : excsp ] . as the precise speckle pattern depends on the precise disorder realization , the most we can reasonably do is to describe the statistical properties of this speckle , thus the disorder averaged rrs signal shown in fig . [ fig : rrsca ] would in experiment represent the envelope of the rrs speckle pattern . for a single disorder realization ( gray dotted line ) and its disorder average ( black cross symbols ) versus energy for @xmath229 and a @xmath230 azimuthal angle , @xmath231 and @xmath232 . for comparison , we also plot the excitonic optical density [ green ( light gray ) plus symbols ] . ] when condensed , the expression for the rrs intensity becomes more complicated , but can still be written in the form of a filter and scattering part : @xmath233 where the filter and scattering terms are respectively : @xmath234 as in the non - condensed case , the filter function @xmath235 restricts the allowed incoming and outgoing momenta to those for which the normal modes have the injected energy . however , as discussed earlier , in the presence of a condensate , the normal modes supported are no longer the lp and up , but instead the new bogoliubov - like quasi - modes . the decay of spectral weight associated with modes far below the chemical potential , as discussed in sec . [ sec : spect ] still applies ( this can clearly be seen in fig . [ fig : rrsca ] , fig . [ fig : spwei ] and fig . [ fig : plumi ] ) . however , since the energy of the rrs signal is controlled by the energy of the incident photon , it is possible to study modes far above the chemical potential , which would have negligible equilibrium occupation , and thus negligible weight in the pl signal . both for this reason , and because of the ability to discriminate between the coherent rrs signal and inchoherent photoluminescence , rrs provides a very powerful tool to study the interesting properties of condensed polariton systems . in this paper we have analysed and compared different optical responses of a condensate of microcavity polaritons . one of the conclusions of this work is that , compared to photoluminescence studies , the energy and momentum resolved resonant rayleigh scattering spectra provides many advantages in probing a condensed phase and in studying the associated coherent excitation spectrum . in particular , resonant rayleigh scattering , by collecting the signal at a different angle from that at which the system is probed , allows one to look directly at characteristic features of the condensed phase excitation spectrum . such features , like the goldstone modes both above and below the chemical potential , are weak in photoluminescence emission and are likely to be masked by the strong emission from the condensate in the lowest momentum state . in contrast , rayleigh scattering , being coherent with the probe , can be distinguished from condensate emission . one can ask why the accurate description of excitonic disorder considered in this paper is necessary , when the microcavity photon introduces a long length scale which averages over disorder . one might also compare this system with other excitonic systems , such as double quantum wells , @xcite where high density of dipole - dipole interacting excitons quickly screens the disorder . @xcite the answer to this question is that the large ratio of exciton to photon mass means firstly that exciton density required for condensation in microcavities is far less than in double quantum wells ; and secondly that those exciton states involved in forming the thermally occupied polariton modes are strongly localised exciton states ( i.e. influenced by disorder beyond born approximation ) . one of the consequences of an accurate treatment of disorder in microcavities is that the blue - shift of the lower polaritons due to saturation effects depends on a length scale characteristic of disorder and temperature , which can be much larger than the bohr radius , this length playing an equivalent role in the clean system . this effect can be important in determining the polariton densities from the measured blue - shift in current experiments . an important consequence of disorder is that the exciton - light coupling strength is characterised by a full distribution . in the non - condensed phase , the optical density alone determines the photoluminescence response , while for resonant rayleigh scattering the averaged fourth power of the oscillator strength is required . however , in the condensed phase one has to consider the entire distribution for each exciton energy . for energies close to and above the band edge , there is always a non - vanishing probability of an arbitrarily small oscillator strength . for this reason , in contrast to examples like bcs superconductivity , one can show that the quasi - particle spectrum does not have a hard gap . in conclusion , we have considered how an accurate treatment of disorder on the single - particle excitonic level , when elevated to the many body problem of interacting microcavity polaritons , leads to a variety of interesting features in various optical responses and probes of the condensed phase . we are grateful to roland zimmermann , wolfgang langbein and ben simons for suggestions and useful discussions . f.m.m . and m.h.s would like to acknowledge financial support from epsrc . j.k . would like to acknowledge financial support from the lindemann trust . this work is supported by the eu network `` photon mediated phenomena in semiconductor nanostructures '' hprn - ct-2002 - 00298 . in this appendix we show how the coherent nature of the resonant rayleigh signal allows a weak coherent probe to be detected in the presence of strong incoherent photoluminescence . the rayleigh scattering probe is a perturbation , @xmath236 , which in the rotating wave approximation may be written as : @xmath237 this describes a probe at wavevector @xmath30 , frequency @xmath238 , of strength @xmath239 . the rrs signal is the coherent scattering of this probe to other wavevectors . to isolate the part of the emission that is coherent with the probe , one may use a homodyne measurement , interfering the emission with part of the probe signal . this corresponds to measuring the spectrally resolved emission intensity : @xmath240 where @xmath241 \left[\hat{\psi}^{}_{{\mathbf q}}(0 ) + a_1e^{i\phi } \right ] |n\rangle \ ; , \end{gathered}\ ] ] and where @xmath242 is the strength of the homodyne mixing , and @xmath243 is a phase delay introduced between the probe and the homodyne signal . the states @xmath244 are the eigenstates of the system without the probe , given by @xmath245 ; @xmath246 is the free energy , for normalization . the principle of phase sensitive detection is to vary the phase delay , and to extract the part of @xmath247 which depends on this phase delay . since the background pl does not depend on the phase delay , this allows one to separate a small but phase dependent signal from a strong but phase independent background . the part of the signal that depends on phase is : @xmath248 \ ; . \label{eq:4}\end{gathered}\ ] ] standard first order time - dependent perturbation theory , using the perturbation in eq . yields : @xmath249e^{-i \omega t } + [ \ldots ] e^{i\omega t } \ ; . \label{eq:5}\end{gathered}\ ] ] the term written @xmath250 $ ] is similar to the first term , but with @xmath251 and @xmath252 . inserting this signal into eq . and then into eq . ( [ eq:1a ] ) , it can be seen that such a term gives a signal at frequency @xmath253 , and so can be clearly separated and ignored . inserting the first term into eq . , one can write : @xmath254 \frac { \langle n | \hat{\psi}^{\dagger}_{{\mathbf q } } | m \rangle \langle m | \hat{\psi}^{}_{{\mathbf p } } | n \rangle } { e_n - e_m - \omega - i \eta } \ ; , \label{eq:6 } \end{gathered}\ ] ] which is the definition of the retarded green s function , @xcite @xmath255 . by repeating the same analysis for the second term in eq . ( [ eq:4 ] ) , one can then write : @xmath256 where @xmath257 is the phase of the retarded green s function . hence , the phase - dependent part of the luminescence is given by the off - diagonal in momentum space part of the retarded green s function , i.e. the rrs signal . h. deng , d. press , s. gotzinger , g. s. solomon , r. hey , k. h. ploog , y. yamamoto , phys . * 97 * , 146402 ( 2006 ) . j. kasprzak , m. richard , s. kundermann , a. baas , p. jeanbrun , j. keeling , f. m. marchetti , m. h. szymanska , r. andr , j. l. staehli , v. savona , p. b. littlewood , b. devaud , and le si dang , nature , * 443 * , 409 ( 2006 ) . m. h. szymanska and p. b. littlewood , solid sate commun . * 124 * , 103 ( 2002 ) ; m. h. szymanska , p. b. littlewood , and b. d. simons , phys . a * 68 * , 013818 ( 2003 ) . a. kavokin and g. malpuech , _ cavity polaritons _ , in thin films and nanostructures vol . 32 ( elsevier , ny , 2003 ) . j. keeling , p. r. eastham , m. h. szymanska , p. b. littlewood , phys . lett . * 93 * , 226403 ( 2004 ) ; phys . b * 72 * , 115320 ( 2005 ) . f. m. marchetti , b. d. simons , and p. b. littlewood , phys . b * 70 * , 155327 ( 2004 ) ; f. m. marchetti , m. h. szymanska , p. r. eastham , b. d. simons , and p. b. littlewood , solid state commun . * 134 * , 111 ( 2005 ) . d. sarchi and v.savona , aip conf . proc . * 893 * , 401 ( 2007 ) . p. schwendimann and a. quattropani , phys . b * 74 * , 045324 ( 2006 ) . r. zimmermann , f. grosse , and e. runge , pure and appl . chem . * 69 * , 1179 ( 1997 ) ; r. zimmermann and e. runge , phys . ( a ) * 164 * , 511 ( 1997 ) . i. carusotto and c. ciuti , phys . lett . * 93 * , 166401 ( 2004 ) ; c. ciuti and i. carusotto , phys . status solidi b * 242 * , 2224 ( 2005 ) . s. d. baranovskii and a. l. fros , fiz . . poluprovodn . * 12 * , 2233 ( 1978 ) [ sov . . semicond . * 12 * , 1328 ( 1979 ) ] . r. zimmermann , e. runge , and v. savona , `` theory of resonant secondary emission : rayleigh scattering versus luminescence '' , in _ quantum coherence , correlation and decoherence in semiconductor nanostructures _ ( ed . t. takagahar ) , elsevier science ( 2003 ) . m. holland , s.j.j.m.f . kokkelmans , m.l . chiofalo , and r. walser , phys . * 87 * , 120406 ( 2001 ) ; e. timmermans , k. furuya , p.w . milonni , and a.k . kerman , phys . a * 285 * , 228 ( 2001 ) ; y. ohashi , a. griffin , phys . lett . * 89 * , 130402 ( 2002 ) .

clusters of galaxies are the last and most massive class of bound structures to form out of the big bang ( @xmath0 ) . cluster assembly combines irregular accretion from surrounding diffuse matter , especially from filaments , and violent mergers with other clusters . as in the universe at large , most cluster matter is non - baryonic , `` dark matter '' . the baryonic matter within clusters is predominantly hot , diffuse plasma ( @xmath1 ) ; the intracluster medium ( icm ) . the icms are shocked and stirred throughout their formation @xcite . the resulting icm flow structures provide vital information about the large scale dynamics of cluster formation , as well as about the physics of the icm plasma@xcite . cluster - scale shocks , in particular , are tell - tale consequences of merger events @xcite . tracing and deciphering these icm flow structures are essential to our understanding of cluster formation . while thermal x - rays have successfully revealed some icm shocks and other flow features @xcite , distinct x - ray signatures are often subtle and hard to isolate and measure cleanly . observations of the sunyaev - zeldovich effect add some statistical information about the icm pressure and velocity distributions@xcite . but , a full , clear picture requires additional , complementary tools . thankfully , non - thermal synchrotron emissions from @xmath2 gev electrons ( cre ) in the icm and in embedded objects may help reach this goal . so far , the principal non - x - ray tool in efforts to characterize icm dynamics has been cluster - scale , diffuse radio emissions ( radio `` halos '' , `` mini - halos '' and `` relics '' ) by electrons within the icm that are energized or re - energized by cluster shocks and/or turbulence @xcite . there are multiple plausible origins for those diffuse cre , which in various models may be recent or ancient @xcite . their links to local , current icm dynamics result from the fact that they have rather short lifetimes against synchrotron losses and inverse compton scattering of cmb photons ( @xmath3 100 myr ) . observed cre must be `` energetically younger '' than this . since icm flow velocities are characteristically @xmath4 kpc / myr , such embedded cre must have been energized relatively close by on cluster scales . although these emissions place essential constraints on icm physics , their translation into distinct and firm icm dynamical properties has so far proven to be quite challenging @xcite . at the same time , radio - bright , bipolar plasma outflows from active galactic nuclei ( radio agn , henceforth , simply agn ) are common in clusters @xcite . interactions between those outflows and icm dynamical structures on scales of 10s to 100s of kpc offer an additional and potentially very useful diagnostic probe of the icm dynamics and physics . that is quite separate from the increasing evidence that agn - icm feedback loops on those scales in the cores of relatively undisturbed clusters are central to the properties of those icms @xcite . the focus of our discussion here , instead , is the consequences of agn outflow encounters with cluster scale icm flow structures ( especially shocks ) , and , in particular , impact on the consequent dynamical form of the agn structures that develop as the agn plasma penetrates and generate cavities in the icm . specifically , we explore major distortions to the agn flows and the potential for using such distortions to identify and characterize icm flow structures , including shocks . in subsequent publications we will present analyses of high resolution 3d mhd simulations of agn outflows in related dynamical icm situations . there we will examine details such as coherence and disruption of the agn flows , development of turbulence , magnetic field evolution , as well as relativistic electron propagation , acceleration and emissions . our purpose now is more restricted ; namely it is to outline a simple , but coherent formalism that we have found to be useful in predicting behaviors of the above simulation experiments and that hopefully captures essential physics needed to interpret observed agn interactions with dynamical icms . we note , in the context of the hedla workshop that inspired this work , that many of the dynamical components laid out below should be testable in laboratory experiments . the best known and most widely discussed agn distortion produced by asymmetric icm interactions are those that have formed tails as a result of relative motions between the agn host galaxy and the surrounding icm . so - called `` narrow angle tailed '' ( nat ) agns , in which the originally collimated agn outflows appear deflected into a pair of quasi - parallel trails in the wake of the galaxy are especially common in disturbed clusters , where relatively high velocity icm motions are likely @xcite . if there are apparent , but less dramatic deflections , the agns are typically called `` wide angle tailed '' ( wat ) agns @xcite . both are obvious candidates as probes of icm dynamics @xcite . the idea that icm shocks encountering plasma bubbles or cavities left behind by expired agn might rejuvenate their cre populations has also been widely discussed @xcite . to the best of our knowledge , however , expected behaviors of active agn outflows colliding with icm shocks has not been previously outlined , although there have been several recent suggestions that observed nat agn may , in fact , be associated with cluster radio relics @xcite . here we do not address that issue directly , but do outline the essentials of dynamical encounters of that kind , in order to facilitate future discussions . basics of the formalism we outline have been in the literature for some time . our objective is to set them down as a package , extending them as needed for application to the issues at hand , while looking for parameterizations that may be useful for diagnostic purposes . we specifically target agn outflow / shock encounters , although the necessary formalism has broader applications . we also include along the way some preliminary results from simulations in order to illustrate the kinds of behaviors that result . we ignore here a number of complications that will ultimately be important to detailed models , but that do not seem essential to a `` first order '' picture . although our goal in this paper is to set out a simple set of analytic relationships , we do show illustrations of behaviors from high resolution numerical simulations that will be discussed in detail in subsequent publications . those simulations were carried out using the wombat mhd code employing a second order mhdtvd riemann solver @xcite . an adiabatic equation of state with @xmath5 was employed . steady bi - polar jets were formed within a small cylindrical volume , where the jet pressure , @xmath6 , was kept approximately equal to the ambient pressure , and the jet plasma was accelerated to a velocity , @xmath7 . all the simulations shown in this paper kept the density within the jet - source cylinder , @xmath8 , at 1% of the undisturbed ambient density . we have , however , varied these parameters to confirm relations described below . we also allowed jet power to cycle off in some cases . the simulated jets were weakly magnetized with a toroidal geometry and nominal @xmath9 within the jet - source cylinder . magnetic stresses are too small to influence dynamics at the levels relevant to this paper . the simulations did include cre ; in follow - up discussions we will discuss in detail synchrotron emission properties . that is beyond the scope of the present paper , however . the paper outline is this . section [ sec : shock - cavity ] reviews the interactions between a plane shock and a low density cavity . in section [ sec : jet - wind ] we outline in simple terms the propagation of a jet into a head or tail wind ( section [ subsec : term ] ) and into a cross wind ( section [ subsec : cross ] ) that would , for instance , represent post - shock ambient conditions . for completeness section [ sec : cocoon ] sets out a simple model to estimate the form of the plasma cocoon inflated by jets nominally prior to the above interactions . typically the initial encounter between an icm shock and an agn will involve the shock penetrating the agn cocoon / cavity inflated by the agn outflow . the basic dynamics of such an encounter are qualitatively simple : the core of the cavity is crushed relatively quickly , while the perimeters of the cavity roll into a vortex ring structure @xcite . here we simply outline of the physics . consider then a planar icm shock of mach number , @xmath10 , where @xmath11 is the speed of the shock in the icm and @xmath12 is the icm sound speed . @xmath13 and @xmath14 are the pre - shock icm pressure and mass density , while @xmath15 ( @xmath16 ) is the icm adiabatic index . typical icm shocks will have modest mach numbers , @xmath17 , @xcite but to allow a wider application of this discussion we do not restrict @xmath18 . we then assume the shock encounters a low density cavity , which we take initially to be in pressure equilibrium with the icm ( @xmath19 ) . let the undisturbed density inside the cavity be @xmath20 , with @xmath21 . the initial cavity sound speed is , @xmath22 , where , acknowledging that the cavity includes relativistic plasma , we allow the cavity potentially to have an adiabatic index , @xmath23 , distinct from the icm . as the shock encounters the cavity , it penetrates at a speed @xmath24 pulling icm gas with it . pressure balance between the original cavity and un - shocked icm leads to the condition , @xmath25 . the original icm - cavity contact discontinuity ( cd ) follows the shock at a speed , @xmath26 , intermediate between the icm shock and the cavity shock . at the same time a rarefaction propagates back into the post - shock icm at the post - shock icm sound speed . pfrommer and jones @xcite presented the full , nonlinear , 1d riemann solution to this problem . here we need only a few , approximate behaviors in order to understand what happens during the encounter . generally , even though the shock speed is greater inside the cavity , the shock strength is less than the incident shock ; that is , @xmath27 . there is no general , analytic formula for @xmath28 , although it can be found numerically from the riemann solution . in the strong shock limit , when @xmath29 , pfrommer and jones found @xmath30 ; that is @xmath31 . in the weak shock limit ( as @xmath32 ) , of course , @xmath33 . in either regime the internal shock speed considerably exceeds the incident , icm shock speed whenever @xmath21 . consequently , the shock passes through the cavity much more quickly than it propagates around it . the velocity of the cd within the cavity , @xmath34 , obtained in the riemann solution is , @xmath35 this measures the cavity collapse rate . it will generally be less than @xmath36 . however , pfrommer and jones found , so long as @xmath37 and @xmath21 , that @xmath38 . under those circumstances , the cavity is crushed ( the initial cd pushes all the way through the cavity ) _ before _ the icm shock propagates around it . we note that if the cavity has previously developed a thick boundary layer within which @xmath39 is not very small ( e.g. , through turbulent mixing ; see section [ sec : cocoon ] ) , the internal shock speed within this layer will be slower than above , but @xmath28 can remain comparable to @xmath18 through that boundary layer . this can , for instance , significantly influence such things as particle acceleration during shock passage and shock amplification of turbulence . if the initial cavity is `` round '' ( e.g. , a sphere ) , the shock intrusion into the cavity begins sooner and is usually stronger at the normal , `` point of first contact . '' on the other hand , the oblique penetration towards the extremes of the cavity boundary also generates vorticity . an initially spherical cavity will , thus , evolve into a torus , or vortex ring that mixes icm and cavity plasma @xcite . analogously , even if the cavity boundary is planar , but the shock normal is oblique to the cavity boundary , refraction of the shock and penetrating cd will produce vorticity and mixing @xcite . examples of shocks impacting icm cavities have been presented previously @xcite . outcomes are also evident in figures [ fig : aljet ] and [ fig : orjet ] . if an agn jet remains active following shock impact , so that it drives into the post - shock icm , its propagation will be modified because the post - shock icm is denser than the pre - shock icm and also because the post - shock icm plasma is put in motion by the shock . in this section we outline a toy model to address these effects . the basic geometry is illustrated in figure [ fig : cartoon ] . cartoon illustrating the basic geometry of bipolar agn jets propagating into an ambient wind.,scaledwidth=48.0%,scaledwidth=41.0% ] we approximate the dynamics as that of a jet within an ambient wind , decomposing the wind velocity into components parallel ( head wind or tail wind ) and orthogonal ( cross wind ) to the jet velocity vector , @xmath40 , near its terminus . that is , we set @xmath41 where @xmath42 . when @xmath43 ; that is , whenever there is a tail wind , @xmath44 , or a head wind , @xmath45 , the rate of advance of the jet terminus will be modified ; when @xmath46 the jet propagation will be deflected laterally . although we present the model in the context of a post - shock flow , the formalism applies to any relative motion between the agn and its ambient medium . indeed the model is essentially a merger and modest extension of classical cartoons of agn jet propagation @xcite . simply stated , jet advance and deflection are influenced by parallel and transverse ram pressures created by the wind . we begin by considering the influence of an ambient wind motion aligned with the jet velocity ; that is , consequences of @xmath43 . to keep the discussion simple we consider here only steady , collimated jets ( zero opening angle ) and homogeneous ambient media . we adopt the standard picture that the advance of the jet terminus into the icm can be expressed in terms of the propagation of a contact discontinuity that forms at the head of the jet @xcite . _ relative to the agn _ , that boundary ( the jet `` head '' ) propagates at velocity , @xmath47 , where @xmath48 represents the length of the jet from the agn . there will be a bow shock propagating into the wind , ahead of the head , as well as a reverse , `` terminal '' shock in the jet . the simple , 1d cartoon model ignores these details , drawing a box around all this and assumes the jet thrust ; that is , the total jet momentum flux coming in through an area , @xmath49 , is balanced by the total wind momentum flux coming in on the opposite side of the box , again , through an area , @xmath50 . in reality , because both jet and wind plasma will , as `` back flow '' , exit the face of the box through which the jet enters ( and box sides ) , the effective sizes of the box faces need to be larger than @xmath50 . we can crudely account for this asymmetry by setting the effective area allocated to the wind , @xmath51 , to be larger than the nominal cross section of the jet . we comment , as well , that even for a collimated jet the radius , @xmath52 , will not generally be a constant along the length of the jet , especially when it becomes over or under pressured with respect to its surroundings . in that case , for instance , the propagating jet plasma will execute a sequence of expansions and rarefactions around an equilibrium radius . consequently , the jet velocity , @xmath7 , density , @xmath8 , and pressure , @xmath6 , will all vary along a real jet . our analysis here does not try to account for dynamics at that level of detail , so we will assume below that the jet has some suitably chosen characteristic radius , @xmath52 , density , @xmath8 and pressure , @xmath6 . we allow both the jet velocity and equation of state ( eos ) to be either non - relativistic or relativistic , although we will assume the propagation of the wind and the head through the wind in the agn reference frame are non - relativistic . then , the jet 4-velocity is @xmath53 , with @xmath54 the lorentz factor of the jet velocity , @xmath7 , while the enthalpy density of the jet plasma in the jet plasma frame is @xmath55 , with @xmath56 the internal energy density of the jet plasma , including rest mass energy , @xmath57 . the jet momentum flux density in the agn frame can be written @xcite @xmath58 the jet thrust is @xmath59 for a non - relativistic eos , where @xmath60 , @xmath61 , while for a relativistic eos , where @xmath62 , @xmath63 . it will be useful later on also to have expressions for the jet energy flux density in the agn frame . subtracting off the rest mass energy this is @xmath64,\end{aligned}\ ] ] so that the `` luminosity '' of the jet is @xmath65 we define the _ internal _ jet mach number as , @xmath66,@xcite where @xmath67 is the jet sound speed , with @xmath68 . for a relativistic eos @xmath69 , while @xmath70 , so @xmath71 . note in this limit that the jet mach number , @xmath72 , which , if @xmath73 gives @xmath74 . this makes clear that the fact that a jet velocity is close to the speed of light does not , by itself , imply a high mach number for the jet dynamics . a non - relativistic eos with @xmath75 recovers the familiar @xmath76 . with these definitions the momentum flux density is @xmath77.\ ] ] setting @xmath78 , this becomes simply @xmath79 . \label{eq : momflux}\ ] ] for a relativistic eos @xmath80 , whereas for a non - relativistic eos , @xmath81 , the usual gas adiabatic index . written in this form the expression for the jet momentum flux is virtually the same for relativistic or non - relativistic jets . if , as is often the case , @xmath82 we can neglect the second term in square brackets . then , the jet thrust for either relativistic or non - relativistic flow is @xmath83 in either regime the thrust depends simply on the internal jet pressure and the internal jet mach number . the jet luminosity is not quite so tidy , but still simple to express . for a non - relativistic eos the luminosity would be @xmath84 a_j,\ ] ] which , with @xmath85 , but @xmath86 and @xmath87 takes the simple familiar form , @xmath88 at the other extreme , with a relativistic eos , @xmath89 , and @xmath90 , the luminosity is also simple to express ; namely , @xmath91 since , now @xmath92 . again , in this form the relativistic and non - relativistic expressions are almost the same . the jet luminosity can , like the thrust , be described in terms of the jet pressure and mach number , but now also in terms of a jet speed , @xmath7 , ( which may @xmath93 ) . we return now to estimating the rate at which the jet terminus propagates through its surrounding plasma . the momentum flux balance condition determining the propagation velocity , @xmath94 , is actually measured in the frame of the head , so we should transform the momentum flux relations to that frame . however , provided @xmath95 , it is easy to show that the fractional change in @xmath96 is small . we will neglect that correction below . the momentum balance condition becomes , @xmath97.\ ] ] using equation [ eq : momflux - ap ] this leads to @xmath98 \approx \mathcal{m}_j^2 \frac{\gamma'}{\gamma_w}\frac{a_j p_j}{a_h p_w } \label{eq : headv - vel}\ ] ] or , @xmath99 \approx \mathcal{m}_j^2 \frac{\gamma'}{\gamma_w}\frac{a_j}{a_h}\frac{p_j}{p_w } \label{eq : headv - mach}\ ] ] where @xmath100 and @xmath101 are the mach numbers of the jet head advance and the head / tail wind with respect to the agn . then , of course , @xmath102 is the wind sound speed . in equations [ eq : headv - vel ] and [ eq : headv - mach ] the upper ( lower ) sign corresponds to a tail ( head ) wind . if @xmath103 we have the simple result @xmath104 that is , the mach number of the advance of the head relative to its ambient medium is similar to the internal mach number of the jet modified by a factor that depends mostly on the ratio of the integrated jet pressure , @xmath105 , to the integrated wind pressure ( isotropic pressure , not ram pressure ) `` across the head '' , @xmath106 . in our simulation experiments with steady , fixed axis , non - relativistic jets @xmath107 provides reasonable matches between simulated jet propagation and equations [ eq : headv - vel ] , [ eq : headv - mach ] and [ eq : headv - app ] . some insights into appropriate @xmath108 are useful going forward . our discussions relate to jets that are pressure confined as they propagate . so , we expect @xmath109 , where @xmath110 represents the pressure of whatever medium is immediately surrounding the jet . if a jet enters a region where it is out of pressure balance , it will generally expand or converge to compensate . as noted above , the pressures within simulated jets are actually quite non - uniform as a result . on average , however , we find , independent of the pressure assignd to a simulated jet at its origin , the average pressure along the propagating jet becomes roughly comparable to the ambient pressure . we also find in simulations and argue in section [ sec : cocoon ] that the pressure inside jet cocoons are commonly roughly similar to those in the undisturbed surroundings , even though the cocoon formation drives a shock into its surroundings . consequently , we expect within a factor of a few that @xmath111 . nonetheless , we leave such ratios as undetermined in our analyses , in order to reveal their roles . we note also for non - relativistic jets with non - relativistic eos that , since for both the jet and the ambient medium , @xmath76 , equation [ eq : headv - app ] can be written @xmath112 independent of @xmath108 . this recovers the commonly applied assumption that , modulo `` an efficiency factor '' ( @xmath113 ) the advance speed of the jet head is roughly the jet speed reduced by a factor of the square root of the density ratio between the ambient medium and the jet when the jet and its advance are both highly supersonic @xcite . it is further evident from these relations , as we would expect intuitively , that the advance speed of the jet with respect to the agn , @xmath114 , is greater when it propagates downstream with a wind ( a `` tail wind '' ) than if it propagates upstream into a wind ( a `` head wind '' ) . in fact a jet propagating into a sufficiently strong head wind can be stopped , or even reversed . using equation [ eq : headv - app ] the approximate condition for the head wind to stop forward progress of the head is @xmath115 we have verified this condition in our simulations ( see figure [ fig : aljet ] ) . although the above results would be applicable for any agn relative motion through its ambient medium when the agn jets are aligned with the relative motion , we introduced the issue in the context of post - shock icm flows . in that case it is useful to express these relations in terms of wind properties resulting from icm shocks . to keep it simple , we assume that the pre - shock icm plasma was at rest with respect to the agn , although that is easily modified . again expressing the icm shock mach number as @xmath18 we have ( assuming @xmath116 ) , @xmath117 where , as above , the subscript index ` i ' identifies the unshocked icm , and ` w ' indicates post - shock icm ( the `` wind '' ) . applying equations [ eq : jump - d ] - [ eq : jump - a ] to equation [ eq : headstop ] we obtain an approximate relation for the strength , @xmath118 , of an icm shock that can stop the advance of an approaching jet of internal mach number , @xmath119 in a `` head on collision '' when the agn is at rest in the undisturbed icm ; namely @xmath120 where @xmath121 . as noted above , we expect for steady jets with fixed axes , @xmath122 . the function @xmath123 , for @xmath124 , corresponding to expected `` merger - related '' icm shock strengths . as a rough rule of thumb , then , an agn jet running head on into a shock will be stopped or reversed if the mach number of the shock is comparable to or even a bit less than the mach number of the jet . the above behaviors are illustrated in figure [ fig : aljet ] for two , simulated non - relativistic shock - jet interactions with different relative shock strengths . the views are 3d volume renderings of jet mass fraction ; that is , only plasma that originated from the agn is shown . in both cases , @xmath125 , and the shock normal was aligned with the ( bi - polar ) jet axis . icm shock propagation was left - to - right , although in this view the shock plane was rotated 40 degrees from the line of sight , in order to reveal the structures more clearly . the location of the agn is marked by an x in each case . the color map of the mass fraction tracer runs from `` white '' ( 100% ) through yellow , red , green and blue ( @xmath126 30% ) . the figure upper panel represents the outcome for an icm shock with @xmath127 , whereas the lower panel involves a @xmath128 shock . both are shown at approximately the same time interval since initial contact between the shock and the left - facing ( upwind ) jet . in the bottom panel the stronger shock has left the computational domain to the right after reversing the left - facing jet and crushing the two original jet plasma cocoons and stripping them from the jets . the `` smoke ring '' to the right is the resulting vortex ring structure , whose formation out of the pre - shock cavities was outlined in the previous section . the flaring seen at the right end of the remaining ( downwind ) jet represents the head of that jet . the shocked jet plasma is unable to propagate back to the agn through the strong right - facing wind behind this shock . it is not able to refill a cocoon ( see section [ sec : cocoon ] ) . the jet remains `` naked '' . several agns in disturbed clusters have been seen that are candidates for this interaction . probably the best known example is the radio source ` c ' in a2256 , @xcite which has a roughly 1 mpc , very thin ( unresolved ) `` tail '' extending to the west of an agn , but nothing evident to the east . indeed , that agn is seen projected near a strong radio relic in the cluster , suggesting that it could have passed through a moderately strong cluster merger shock . the weaker and slower icm shock involved in the upper figure [ fig : aljet ] panel dynamics is still in the volume illustrated , although not directly visible in this rendering . in the undisturbed icm its position along the jet axis is about 2/3 of the distance from the agn x to the right end of the right - facing jet head . inside the right - side cocoon , the shock has just reached the head at this time . we found experimentally in this case that an incident shock with @xmath129 would just stop the approaching ( left facing in the figure ) jet . illustration from simulations of icm shocks colliding head on with @xmath125 agn jet pairs . both panels are volume renderings of a passive jet mass - fraction tracer . shock propagation was left to right . x marks agn location . the jet axes are rotated 40 degrees out of the plane of the sky with the left jets approaching the observer . * top * : @xmath130 . * bottom * : @xmath131 . see text for details.,scaledwidth=48.0%,scaledwidth=41.0% ] thus , evidence for relative agn foreshortening to one side has the potential to find and even measure the strengths of icm shocks or winds . in the presence of a cross wind , @xmath46 ( see figure [ fig : cartoon ] ) , the jet is subjected to an unbalanced transverse ram pressure force , @xmath132 . particularly , if that cross wind results from a crossing shock , the jet cocoon ( cavity ) will be crushed and stripped away from the propagating jet ( see figure [ fig : orjet ] ) . from that point on the jet interacts directly with the wind , and we can estimate the induced transverse pressure gradient within the jet as @xmath133 then the transverse acceleration of a steady jet is determined by the relation @xmath134 we can use @xmath135 as a way to estimate the length @xmath136 over which the transverse ram pressure from the wind will deflect the jet by 90 degrees ; that is , the `` jet bending length '' . neglecting the ( initially small ) second term inside the parentheses in equation [ eq : ftrans ] this gives , @xmath137 where the final expression has used equation [ eq : momflux - ap ] . analogous to the results of the previous subsection we can also write this final form in terms of the jet mach number , @xmath119 and the cross - wind mach number , @xmath138 , @xmath139 if the jet and wind pressures are comparable , the ratio of the jet bending length , @xmath136 , to the jet diameter , @xmath140 , is roughly @xmath141 . a little bit of algebra shows that equation [ eq : ellbend ] matches the bending radius of curvature derived in @xcite for an agn moving supersonically through an ambient medium at right angles to the jet axis . of course , only relative motion matters , and here we emphasize that it is not necessary that the cross wind is supersonic for the bending to develop . the bending radius will , however , scale inversely with the square of the mach number of relative motion . so , once again , obvious distortion in the agn points to relatively large mach number of the relative motion between the agn and its immediate surroundings . assuming the cross wind under discussion comes entirely from an icm shock propagating transverse to the jet axis , we can use equations [ eq : jump - d ] - [ eq : jump - a ] to give a relation between the jet bending length , the jet radius , the jet mach number and the icm shock mach number , @xmath18 , @xmath142 by convention , when the jet length , @xmath48 satisfies @xmath143 , but @xmath144 the resulting agn morphology would be describe as a `` wide angle tail '' radio galaxy , or a wat " . as the ratio @xmath145 becomes smaller , the morphological label would shift to `` narrow angle tail '' or `` nat '' . figure [ fig : orjet ] illustrates results from a simulation of a @xmath131 shock that crossed from the left and collided at right angles with a pair of @xmath146 jets oriented vertically in the image . equation [ eq : lbend - mi ] predicts @xmath147 , which is actually very close to the empirical result of the simulation . as in figure [ fig : aljet ] this image shows a volume rendering of the passive jet mass fraction tracer . the shock plane has again been rotated 40 degrees from the line of sight to make 3d structures more distinct . we comment on the close resemblance between the structures visible in figure [ fig : orjet ] and the `` classic '' nat source , ngc1265 @xcite , whose morphology has long been modeled in terms of a strong cross wind in the rest frame of the host galaxy . here the wind is a feature of the post - shock environment . that has also been suggested for ngc1265 @xcite . simulation of an @xmath131 shock that crossed from the left and collided at right angles with an @xmath146 agn jet pair ( vertical jet axis in the image ) . shown is a volume rendering of a passive jet mass - fraction tracer after the shock has passed through the volume shown . x marks agn location . view orientation is the same as figure [ fig : aljet ] . see text for details.,scaledwidth=48.0%,scaledwidth=41.0% ] note that the two jets in figure [ fig : orjet ] remain stable long after they are deflected by @xmath148 degrees into `` tails '' , even though turbulent mixing regions develop around them and enlarge downstream . the tails merge at the far right into a pair of merging vortex rings that formed along the lines outlined in section [ sec : shock - cavity ] as the icm shock crushed the two jet cocoons produced before shock impact . the jets can be traced as coherent structures almost to the vortex rings . the shock itself has left the observed volume to the right at the time shown . finally , we comment on the more complex case where an incident shock collides at an arbitrary angle with respect to the agn jet axis , or more generally when the agn jets encounter an oblique wind . since @xmath149 is initially the same for both jets , the rate of jet deflection indicated by equation [ eq : ftrans ] would , before deformation , be the same for both jets ( see figure [ fig : cartoon ] ) . the aligned wind velocity , @xmath150 , has the same magnitude , but opposite sign on the two jets . the downwind jet , where @xmath151 , thus advances more rapidly . as the jets begin to be deflected , however , the upwind jet , where @xmath45 , is deflected to be more nearly transverse to the wind ( @xmath149 increases along the jet on that side ) , while the downwind jet is deflected to be more nearly aligned with the wind ( @xmath150 increases along this jet ) . consequently , the upwind jet becomes more sharply bent than the downwind jet . we have confirmed in simulations a regular transitioun along these lines between the aligned wind interactions described in section [ subsec : term ] . and cross wind interactions outlined at the beginning of this section . evidently , such asymmetries can provide information about both the relative mach numbers of the jets and a wind and the relative orientations of the wind and the jets . the shape evolution of the two jets depends distinctively on the ratio of two mach numbers as well as the orientation between the agn jet axis and the wind velocity vector , @xmath152 . thus , the resulting shape provides a means to determine both @xmath153 and @xmath154 . also , it is obvious that if the `` weather conditions '' encountered by jets vary as they extend through the icm , additional morphological features are likely to develop that can be used to identify and characterize these icm structures . we began this discussion with an outline of shock propagation through a pre - existing cavity formed by agn outflows ; that is a jet cocoon . then we outlined propagation of the jets themselves in the post - shock flow . the shocks both crush the cocoon and may , if the wind `` stripping '' is faster than replacement from the jet terminus , remove the cocoon ( see figures [ fig : aljet ] and [ fig : orjet ] ) . we did not , however , address what might be learned from the cocoons themselves , in order to tie the pieces together , here we outline briefly some of the related physics connecting the jet propagation to the formation of the jet cocoon before shock impact . since our purpose is only to lay out a rough picture of cocoon inflation , we ignore in this section complications such as relative motions between the agn and its ambient icm , the presence of large scale pressure or density gradients , or icm turbulence . all of these will modify cocoon morphology and quantitative measures , although they should not fundamentally change the basic picture presented here . in simple terms the cocoon represents the reservoir of plasma that has previously passed through the jet . the cocoon plasma will generally be of substantially lower density than the surrounding icm , with the two media nominally separated by a contact discontinuity . because it is also a `` slip surface '' , that boundary is likely to be unstable to kelvin - helmholtz instabilities ( khi ) , so that some degree of mixing will take place ( unless , for instance , magnetic fields in or around the cocoon are strong enough and coherent enough to stabilize the contact discontinuity @xcite ) . figure [ fig : noshock - color ] shows volume renderings of cocoons formed from two simulated steady agn jet pairs . in each case the agn ( marked by an ` x ' ) expelled identical , but oppositely directed jets . the rendered quantity is the same passive mass fraction tracer shown in figures [ fig : aljet ] and [ fig : orjet ] . in this case the jet axis is viewed in the plane of the sky . the upper image corresponds to a @xmath125 jet pair , while the lower image represents the cocoons of a @xmath146 jet pair . the structures are viewed when @xmath155 . the jets , themselves , are faintly recognizable inside the cocoons and along the cocoon axes . the cocoon boundaries are clearly influenced by khi ; some mixing has occurred . indeed the small differences between the left and right cocoons come from detailed differences in the khi on the two sides that are seeded by the mismatch between a mathematically circular jet and a cartesian numerical grid . khi details depend on the exact placement of the agn on the numerical grid . the jet mass fraction dominating the images @xmath156 . despite such complications , there is a relatively clear cocoon boundary , so we assume below that the jet cocoon and the surrounding icm are cleanly separated . jet `` backflow '' cocoons formed by @xmath125 ( * top * ) and @xmath146 ( * bottom * ) steady jets . a passive , jet mass - fraction tracer is volumed rendered in the images . the jets are `` in the plane of the sky '' and shown when they have approximately the same lengths , @xmath48 . the location of the agn is marked by an ` x ' . see text for details.,scaledwidth=48.0%,scaledwidth=41.0% ] the energy deposition from the jet into the cocoon will generally drive a shock laterally outward into the icm . we will call this the `` inflation shock '' . provided the jet terminus propagates supersonically into the icm ( @xmath157 , equation [ eq : headv - mach ] ) there will be a bow shock attached to the jet head , which will generally merge with the inflation shock . in figure [ fig : noshock - color ] the tips of the two cocoons are confined by bow shocks . the tighter `` mach cone '' of the higher mach number jet leads to `` sharper points '' in the cocoons . although the bow shock may be quite strong , both observations and simulations suggest that the inflation shock is typically rather weak . for example the @xmath146 jets shown in the bottom of figure [ fig : noshock - color ] and in figure [ fig : noshock - mach ] produce bow shocks with @xmath158 ( @xmath159 ) on the nose of the jet in agreement with expectations from equation [ eq : headv - app ] . on the other hand , over much of their surfaces the accompanying both inflation shocks have mach numbers , @xmath160 for the duration of the simulation ( see figure [ fig : noshock - mach ] ) . there is no fixed value , however . the relatively small mach numbers of most of the inflation shock surfaces also point to the fact that the pressure inside the shocks and inside the cocoons is not much greater than the ambient pressure , @xmath13 ( see equation [ eq : jump - p ] ) . because these properties do not inherently lead to scale - free structures , we do not try to model cocoon formation in terms of self - similar behavior @xcite . volume rendering of the shocks associated with the @xmath146 simulation shown in the bottom image of figure [ fig : noshock - color ] . the * top * image matches the orientation of the bottom image in figure [ fig : noshock - color ] , while the * bottom * image has the jet axis rotated 40 degrees , with the left jet approaching ( so similar to figures [ fig : aljet ] and [ fig : orjet ] . shocks are color coded , with the strongest shocks , @xmath161 shown in white , with @xmath162 represented by purple . x marks agn location . see text for details.,scaledwidth=48.0%,scaledwidth=41.0% ] still , it makes sense to estimate the volume of the cocoon simply by comparing the work required to inflate the cocoon to the energy that has passed down the jet into the cocoon . very simply , we set @xmath163 , where @xmath164 is a numerical factor that accounts for energy lost by the cocoon plasma as it inflated . accounting for adiabatic work done on the icm would lead to @xmath165 . for a non - relativistic jet with @xmath87 , this would give @xmath166 , while for a jet with a relativistic eos and @xmath167 , the equivalent consequences would give @xmath168 . empirical estimates from non - relativistic simulations do , indeed , suggest @xmath169.@xcite the exact value for @xmath170 is not essential to our purposes here . as a primitive model for cocoon geometry we assume the cocoon length is set by @xmath171 , giving @xmath172 where @xmath173 represents a characteristic cocoon cross sectional area . we do not imply in this that the cocoon cross section is constant , and , indeed , it generally will not be ( see figure [ fig : noshock - color ] ) . in practice we treat it as a geometric average cross section . to simplify our treatment further , we express the jet luminosity in the form applicable to both relativistic and non - relativistic jets suggested by equations [ eq : ljet - nonrel ] and [ eq : ljet - rel ] ; that is , @xmath174 where @xmath175 . for a relativistic jet @xmath73 in this expression . from equation [ eq : headv - app ] we write @xmath176 combining equations [ eq : cvol - simp ] , [ eq : ljet - simp ] and [ eq : headv - simp ] gives a simple estimate for @xmath173 , @xmath177 we expect @xmath178 , and , indeed using equation [ eq : ac - nonrel ] , below , and the simulation parameters for the cocoons shown in figure [ fig : noshock - color ] we get consistent estimates for @xmath173 using @xmath179 . note that , if these various jet parameters are time independent , the effective cocoon cross section , @xmath173 is steady in time . that is , the ratio @xmath180 . we briefly address alternative possibilities below . for a relativistic jet with a relativistic eos , @xmath181 , while @xmath182 , and equation [ eq : ac - gen ] becomes , @xmath183 while for a non - relativistic jet , @xmath184 the only difference between equations [ eq : ac - rel ] and [ eq : ac - nonrel ] is a relative factor @xmath185 in equation [ eq : ac - rel ] . we can see from these relations that the cocoon cross section scales with the geometric mean of the jet cross section , @xmath50 and the jet head cross section , @xmath51 . it is strongly boosted if the jet mach number is large ( especially for non - relativistic jets ) , and , with less sensitivity , to a large density contrast between the icm and the jet , @xmath186 . that is , cocoons will be somewhat fatter for jets with larger density contrast , @xmath187 , but especially fatter for larger mach numbers . we are assuming in these comments that , as argued previously , we should expect @xmath188 . the mach number sensitivity of @xmath173 is evident in a comparison of the non - relativistic jet cocoons shown in figure [ fig : noshock - color ] , where the only difference in the two simulations was jet mach number ; above , @xmath125 , while below , @xmath146 . the @xmath146 cocoon , in particular , is , as noted above , strongly tapered , representing the fact that the mach cone at its nose has strongly confined it . we associate @xmath173 with the mean cross sectional area , which here would be roughly midway between the agn and the jet terminus . the reader may have noticed at the same time that the cocoons from the jets shown in figure [ fig : noshock - color ] are relatively skinny compared to typical , observed radio galaxy cocoons . that is characteristic of the properties of moderate mach number , non - relativistic agn jet simulations that are steady and maintain a precise axis for the jets . on the other hand , if either the jet power cycles substantially or the axis of the jet wanders or precesses , the relative rate at which the jet advances is reduced @xcite . in order to maintain the same cocoon volume indicated in equation [ eq : cvol - simp ] the cross section must increase . this effect can , be roughly accounted for in equations [ eq : headv - simp ] - [ eq : ac - nonrel ] by adjusting the effective head area , @xmath51 . for instance , a precessing jet will balance its thrust during a precession period over an area , @xmath189 , where @xmath190 is the instantaneous length of the jet . applied into equation [ eq : headv - simp ] the head advance rate , @xmath191 . if we maintain constant jet luminosity , @xmath192 , so @xmath193 . now , @xmath194 const , rather than @xmath180 . simultaneously , the volume swept out by the precessing jet itself will be @xmath195 , where @xmath196 is the opening angle of the precession cone . this sets a scale for the effective @xmath197 . these effects both fatten the cocoons relative to what we derived above . finally , we mention that equation [ eq : ac - rel ] predicts rather fat cocoons for high mach number relativistic jets propagating in an environment where @xmath198 . that is consistent with results of numerical simulations of relativistic hydrodynamical jets @xcite , and also simulations of jets dominated by their poynting flux ( so , internally highly relativistic ) @xcite . the plasma media in galaxy clusters , icms , are dynamical . they are stirred by many processes associated with cluster assembly . the dynamical state of an icm provides unique information about how this takes place , so it is important to find and evaluate icm dynamical conditions , especially those far away from equilibrium . x - ray observations can reveal critical information about icm thermodynamical properties , some statistical characteristics of icm dynamical states and find relatively strong shocks in higher density icm regions . but , much of the story is not visible in the x - ray window . on the other hand , agns in so - called `` radio mode '' are common in clusters , and especially in disturbed clusters . those agns expel fast plasma jets that plow through the icm . those interactions will influence the icm and its evolution . more to the point of this paper , however , is the considerable impact of icm dynamics on the trajectories and of the agn jets and the distributions of their debris . in particular , impacts between icm shocks and winds will distort and even disrupt the agn structures that form . through an understanding of how those distortions develop and how they depend on the agn and icm properties we hope to open a clearer window to revealing the dynamical structures of icms . by relating those structures to other information about the dynamical state of the cluster , and , for instance , evidence for merging , strong accretion events or gravitational disturbance by dark matter halos , these insights can provide vital probes of cluster formation processes and their relative roles . we have laid out in this paper a simple summary of some of the essential dynamical relationships involved in agn / icm interactions , with a special focus on interactions involving icm shocks . the nominal target application is icm shocks colliding with agn outflows . we pointed out , however , that the relations we developed have application to any relative motion between the agn and its immediate environment . the formalism allows agn jets that are either non - relativistic or relativistic . we set down basic relations to evaluate shock interactions with the low density cavities created by agn jets , as well as to follow the propagation of the jets within post - shock flows . for completeness , we used the same formalism to provide a basic context for the formation of the cavities themselves . one notable aspect of the relationships we derive is the essential roles of the internal mach number of the jet flow and the mach number of the icm shock . more directly , the ratio of these two mach numbers seems central to evaluating the interactions between the agn and a dynamical ambient environment . this provides a potentially useful link that can help develop quantitative understandings of icm dynamical states that otherwise are likely to remain obscure for some time . twj and bo were supported in this work at the university of minnesota by nsf grant ast1211595 . cn was supported by an nsf graduate fellowship under grant 000039202 . twj , cn and bo gratefully acknowledge support and hospitality of the minnesota supercomputing insitute . we thank an anonymous referee for constructive comments that improved the manuscript . 2ifxundefined [ 1 ] ifx#1 ifnum [ 1 ] # 1firstoftwo secondoftwo ifx [ 1 ] # 1firstoftwo secondoftwo `` `` # 1''''@noop [ 0]secondoftwosanitize@url [ 0 ] + 12$12 & 12#1212_12%12@startlink[1]@endlink[0]@bib@innerbibempty a. kravtsov & s. borgani , `` formation of galaxy clusters '' , araa * 50 * , 353 - 409 ( 2012 ) a. schekochihin & s. cowley , `` turbulence , magnetic fields and plasma physics in clusters of galaxies '' , phys . plasmas * 13 * , 56501 - 56508 ( 2006 ) m. markevitch & a. vikhlinin , `` shocks and cold fronts in galaxy clusters '' , phys * 443 * , 1 - 53 ( 2007 ) r. adam , b. comis , i. bartalucci , a. adane , p. ade , p. andre , a. beelen , a. benoit , a. bideaud , n. billot , o. bourrion , m. calvo , a. catalano , g. coiffard , a. daddabbo , f. desert , s. doyle , j. goupy , b. hasnoun , i. hermelo , c. kramer , g. lagache , s. leclercq , j. macias - perez , j. martino , p. mauskopf , f. mayet , a. monfardini , f. pajot , e. pascale , l. perotto , e. pointecouteau , n. pnthieu , g. pratt , v. reveret , a. ritacco , l. rodrigues , g. savini , k. schuster , a. sievers , s. triqueneaux , c. tucker & r. zylka , `` high angular resolution sunyaev - zeldovich observations of macs j1423.8 + 2404 with nika : multi - wavelength analysis '' , a. & a. * 586 * , 122 - 138 ( 2016 ) a. bonafede , h. intema , m. bruggen , m. girardi , m. nonino , n. kantharia , r. van weeren & h. rottgering , `` evidence for particle re - acceleration in the radio relic in the galaxy cluster plckg287.0 + 32.9 '' , ap.j . * 785 * , 1 - 14 ( 2014 )

in recent years , there has been intense research work in the topic of highly inhomogeneous light beams , and in particular , in light carrying orbital angular momentum ( oam)also called twisted light ( tl ) @xcite . the research in tl spans several areas , such as the generation of beams @xcite , the interaction of tl with atoms and molecules @xcite or with condensed matter @xcite . tl has already proved to be useful in applications . the most notable example is perhaps the optical trapping and manipulation of microscopic particles @xcite . applications in other fields are also sought , for example in quantum information technology , where the oam adds a new degree of freedom encoding more information @xcite . in addition , theoretical studies in solid state physics predict , for instance , that tl can induce electric currents in quantum rings @xcite , and new electronic transitions ( forbidden for plane waves ) in quantum dots @xcite . this all suggests that tl can be a new powerful tool to control quantum states in nanotechnological applications . two features of tl are particularly striking . first , tl exhibits a vortex or phase singularity at the beam axis . second , polarization and oam are so intermixed that two beams having the same oam but opposite circular polarization behave in a completely different way . this is in contrast to what happens to plane waves , where the polarization alone does not determine other important properties . these two features can also be found in other inhomogeneous beams , namely the so - called azimuthally polarized @xcite fields . the interaction between tl and matter is particularly interesting due to the inhomogeneous nature of the tl , and it is worth revisiting its mathematical formulation . the most general form to describe the light - matter interaction is the minimal coupling hamiltonian , where the electromagnetic ( em ) fields enter through their potentials . in many cases of interest , the hamiltonian can be rewritten in terms of em fields using gauge transformations , i.e. , transformations among potentials that preserve the em fields @xcite . usually , the transformations are accompanied by approximations . one of the best - known among these hamiltonians is the dipole - moment approximation ( dma ) . it can be derived under the assumption that the em fields vary little in the region where the matter excitation takes place , and effectively the electric field @xmath0 is treated as spatially homogeneous . the dma hamiltonian then takes the form @xmath1 , where @xmath2 is the dipole moment of the material system . while gauge invariance is a symmetry property of the electromagnetic interaction and therefore all observable quantities have to be independent of the special choice of a gauge , this independence is usually lost when approximations are performed @xcite . a standard example is the @xmath3-@xmath4 two - photon transition in a hydrogen atom where it has been explicitly shown that in the case of the @xmath5 coupling already very few intermediate states are sufficient to obtain a very accurate results while in the case of the @xmath6 coupling a very large number of intermediate states is required @xcite . a similar behavior has been found in calculations of the level width of microwave transitions for the measurement of the lamb shift @xcite . also for single photon transitions in a h@xmath7 ion large differences between the two gauges have been found when variational wave functions for the molecular orbitals are used with , e.g. for the @xmath8-@xmath9 transition , a strong preference of the @xmath10 coupling @xcite . the dma form of light - matter coupling is advantageous for several additional reasons : because the dma only contains the electric field , it is manifestly gauge invariant . the momentum operator has a clear physical meaning and can be used directly for the calculation of quantities like current densities . in contrast , in the case of the minimal coupling there is an additional , gauge - dependent contribution , usually called the diamagnetic current @xcite . the difference between the canonical and the mechanical ( or kinetic ) momentum may also lead to an apparent ambiguity in the definition of the photon momentum , as has been discussed in detail in a recent review by barnett et al . finally , since the dma interaction hamiltonian is linear in the field it can be easily treated perturbatively while the minimal coupling hamiltonian contains terms linear and quadratic in the potential which therefore have to be combined in a proper way when calculating optical nonlinearities . all these arguments show that a coupling in terms of the electric field has clear advantages . in fact , when the light field is sufficiently homogeneous over the size of the matter system the dma is perfectly applicable . this holds for example in atomic and molecular physics and also in the case of nanoscale systems such as quantum dots , where the matter states are highly localized . when the inhomogeneous nature of the field becomes important the dma can not be used anymore . one could perform calculations with the minimal coupling hamiltonian , which contains the vector potential . still , it is appealing to work with a hamiltonian which contains the electric and magnetic fields only , because then the theory is manifestly gauge invariant . of course , using the so - called poincar gauge @xcite one can formally rewrite the hamiltonian in terms of fields , however , it is not always possible to express the fields explicitly . a desirable expression would be one resembling the electric dipole - moment hamiltonian , but retaining the spatial dependence : @xmath11 . we will call this _ electric field coupling _ ( efc ) hamiltonian . in addition to the dipole - moment coupling , the efc hamiltonian also contains higher order couplings in the electric field , e.g. , quadrupole moments . in some situations the spatial inhomogeneity of the field can be kept in an efc hamiltonian in a parametric way , while the transition matrix elements are determined by the coupling via the electric dipole term only @xcite , which has been used to describe for instance four - wave - mixing phenomena @xcite . because for tl new transitions are induced by its oam @xcite , such an approach would not describe the main feature of tl and , thus , it is crucial to include the spatial dependence also in the transition matrix elements . under certain assumptions , e.g. , for the interaction with localized structures placed at the beam maximum , it is possible to cast the hamiltonian in a efc form , and thus , describe the modified selection rules @xcite for example using a power - zienau - woolley transformation @xcite . however , in this paper we show that for tl - matter interaction in the vicinity of the beam axis an efc hamiltonian can not in general be used . the tl - matter interaction at the phase singularity for highly focused beams has been analyzed using a multipolar expansion for electric and magnetic fields @xcite , which already revealed that higher order electric and magnetic terms can be of significant importance . nevertheless , for a subgroup of tl beams we will show that it is possible to derive an electric multipolar hamiltonian for the tl - matter interaction close to the phase singularity , which offers the advantages of a dma hamiltonian . we organize the article as follows . as a next step we briefly revisit the concepts of gauge transformation and dma hamiltonian necessary to understand the discussion ahead . in sec . [ sec : tl ] we introduce the mathematical representations of tl . in sec . [ heur ] , using a heuristic derivation much alike the one found in the literature for the dma , we arrive at the new expression for the tl - matter hamiltonian . section [ sec : bfields ] shows that the atypical behavior of the electric and magnetic fields of tl is in part responsible for the need to modify the hamiltonian . section [ lag ] is devoted to a careful derivation of the new hamiltonian . we wrap up with the conclusions in sec . [ conc ] . in the appendix we discuss why the efc gauge , which seems to be the natural extension of the dma , is not useful for tl . the starting point for a mathematical description of the effect of light on matter is the minimal coupling hamiltonian , that expresses the external em fields in terms of a scalar @xmath12 and a vector @xmath13 potential . for a single particle of mass @xmath14 and charge @xmath15 under a static potential @xmath16 , the hamiltonian reads @xmath17 ^ 2 + v(\mathbf r ) + { q } \ , u(\mathbf r , t ) \,.\ ] ] it is obtained from the lagrangian @xmath18 via the canonical momentum @xmath19 and the legendre transformation @xmath20 . the relationship between potentials and the electric @xmath21 and magnetic @xmath22 fields are @xmath23 gauge transformations are defined such that they preserve the electric and magnetic fields [ eq : gauge - transf ] @xmath24 where @xmath25 is the scalar gauge transformation function . since the canonical momentum ( [ eq : canonical ] ) depends on the vector potential it is obviously gauge dependent . in cases where the em fields vary little on the scale of the system , taken to be centered around @xmath26 , a gauge transformation is sought that would render @xmath27 in the region around @xmath26 . assuming that for external radiation @xmath28 , this is achieved by the gppert - mayer gauge transformation @xmath29 @xcite leading to the new potentials [ eq : gm - gauge ] @xmath30 by neglecting the derivatives of the old vector potential in the relevant region of space we can obtain @xmath31 . this leads to the well - known dma hamiltonian @xmath32 which is evidently gauge - invariant . it should be understood that the requirement @xmath33 in an extended region of space is a very stringent one , for it demands the magnetic field to be zero in that region , in violation to maxwell s equations for a propagating field . a striking feature of the dma is that operators retain their physical meaning . as an important example , we look at the momentum . due to the fact that @xmath27 , the canonical momentum in the new gauge is equal to the mechanical momentum @xmath34 . the mechanical momentum is indeed important , since it is a form - invariant operator @xcite . as such , its eigenvalues are independent of the gauge and are therefore representing physical quantities . the canonical momentum , on the other hand , is not form - invariant . this is a drawback , because the quantized version of the canonical momentum , i.e. , the operator @xmath35 , is typically used to perform calculations ( see sect . iv.a.2b of ref . @xcite ) . in order to obtain measurable quantities such as current densities , the correction due to the vector potential have to be taken into account , e.g. , in terms of a diamagnetic current @xcite . the physical ( i.e. , gauge - independent ) current is then given by the sum of two gauge - dependent contributions . it is obvious that the dma can not be used to describe the interaction of tl with objects placed close to the beam center because there the electric field and thus the whole light - matter coupling vanishes . in analogy with the dma , a transformation function @xmath36 could be used @xcite which keeps the spatial dependence of the vector potential , yielding new potentials [ eq : local - dipole ] @xmath37 as expected , the new scalar potential has the dipole - like form as that resulting from the gppert - mayer transformation , but now with a position - dependent electric field . like in the case of the dma the new vector potential does not vanish , but it contains spatial derivatives of the old one . again , for sufficiently localized charges and a sufficiently smooth vector potential it may be permissible to disregard these terms resulting in new potentials @xmath38 and @xmath39 , with the concomitant benefits of equality of momenta . we call this the electric field coupling ( efc ) approximation . however , we will show below that for tl in the region close the the beam axis this gauge is not useful . let us now come to the case of tl . a tl beam can have different radial profiles such as laguerre - gaussian or bessel type beams . here we will consider the case of a bessel beam , which has the advantage of being an exact solution of maxwell s equations @xcite . in mathematical terms , the vector potential of a monochromatic tl beam in cylindrical coordinates @xmath40 , can be described by @xmath41 with components @xcite [ eq : a_bessel ] @xmath42\ , , \\ \label{eq : a_bessel_phi } a_\varphi(\mathbf r , t ) & = & \sigma f_{q_r\ell}(r ) \sin[(\omega t - q_z z ) - ( \ell + \sigma)\varphi]\ , , \\ \label{eq : a_bessel_z } a_z(\mathbf r , t ) & = & -\sigma \frac{q_r}{q_z } f_{q_r\ell+\sigma}(r ) \sin[(\omega t - q_z z ) - ( \ell + \sigma)\varphi ] \ , , \nonumber \\\end{aligned}\ ] ] with frequency @xmath43 , wave vectors @xmath44 and @xmath45 , related by @xmath46 , @xmath47 being the index of refraction of the medium , and @xmath48 , @xmath49 , @xmath50 denoting unit vectors in cylindrical coordinates . the integer @xmath51 is related to the oam of the beam , as will be discussed in more detail below . the circular polarization of the field , given by polarization vectors @xmath52 , is singled out with the variable @xmath53 , which yields left ( right)-handed circular polarization for the values @xmath54 . sometimes , in particular in the quantum theory of light , @xmath53 is referred to as the spin angular momentum of the photon with @xmath55 being the spin per photon @xcite . the radial profile of the beam @xmath56 is a bessel function : @xmath57 , with @xmath58 being the amplitude of the potential . note that @xmath59 is a measure of the beam waist . the vector potential of eq . ( [ eq : a_bessel ] ) satisfies the coulomb gauge condition @xmath60 and the vectorial helmholtz equation @xcite . of the in - plane components of the vector potential ( and thus of the in - plane components of the electric field ) for a non - vortex beam ( @xmath61 ) and tl beams ( @xmath62 ) . ] figure [ fig : profile ] shows the beam profile @xmath63 of the in - plane components @xmath64 and @xmath65 and thus also of the in - plane components of the electric field for three different values of the oam : @xmath66 . in the region close to @xmath67 , we observe a main difference that exists between non - vortex beams and tl . while for non - vortex beams ( @xmath61 ) the amplitude has a maximum at @xmath67 , for tl ( @xmath68 ) the amplitude of the in - plane components is zero there . close to the origin , the profile of the in - plane components can be approximated by @xmath69 . since for bessel functions the relation @xmath70 holds , it can be seen from eq . ( [ eq : a_bessel ] ) that the structure of the beam is unchanged if simultaneously the parameters @xmath71 are replaced by @xmath72 . therefore in the following we will restrict ourselves to tl beams with @xmath73 . in the paraxial approximation , when @xmath74 , the @xmath75-component of the vector potential is disregarded . this case has been extensively used in the literature @xcite . the vector potential in the paraxial approximation @xmath76 then reads @xmath77 in this approximation the positive and negative frequency components of @xmath78 are eigenstates of the angular momentum operator @xmath79 with eigenvalue @xmath80 , which can thus be identified with the oam per photon @xcite . although this does not strictly hold in the non - paraxial case , for the sake of brevity , in the following whenever we refer to the oam of the beam , we are implicitly referring to the oam of its paraxial version . the integer @xmath51 is also sometimes called the topological charge @xcite . in this section , following the spirit of the dma , we derive a gauge transformation that captures the essential features of tl , and at the same time retains the advantages of the dma . the derivation is intended to be intuitive and self - evident , and is only done for the paraxial vector potential eq . ( [ eq : ainitial ] ) . a formal analysis leading to the same results will be given in sec . [ lag ] , where the more general form of the vector potential eq . ( [ eq : a_bessel ] ) will be used and also the limitations of the paraxial approximation will be discussed . we are interested in the interaction of tl with a planar , localized structure , such as a quantum disk or a quantum dot , whose lateral dimensions are smaller than the the characteristic radial length scale @xmath81 of the beam ( i.e. , @xmath82 ) . if such a structure is placed at a position with non - vanishing electric field , in particular close to the beam maximum , the conditions of the dma are satisfied and thus the dma is well applicable . however , this is different in the case of such a structure centered at @xmath83 . at the beam center the radial profile of the vector potential can be approximated by @xmath84 , with @xmath85 . note that the vector potential eq.([eq : ainitial ] ) and consequently also the electric field at @xmath83 are zero and thus , within the dma there would be no interaction whatsoever . motivated by the efc hamiltonian , we try a gauge transformation function of the form @xmath86 where we have defined a two - dimensional in - plane position vector @xmath87 out of the 3d vector @xmath88 , and added a constant prefactor @xmath89 to be determined later . for @xmath90 the transformation obviously reduces to the efc case . according to eqs . ( [ eq : gauge - transf ] ) the potentials in the new gauge are calculated to @xmath91 \ , \hat{\mathbf z } \ , , \nonumber \\ & & u^{pa}\,'(\mathbf r , t ) = - \frac{1}{\beta } \mathbf r \cdot \mathbf e^{pa}(\mathbf r , t ) \,.\end{aligned}\ ] ] while the radial dependence of the scalar potential as well as of the in - plane components of the new vector potential is @xmath92 like in the case of the old vector potential , the remaining component @xmath93 is @xmath94 and can thus be neglected in the region close to the singularity . this is consistent with keeping terms up to order @xmath95 . in the efc gauge ( @xmath90 ) we then have @xmath96 and @xmath97 . thus , for @xmath98 the new vector potential is not smaller than the old one clearly demonstrating that this gauge does not help to reduce the difference between mechanical and canonical momentum . on the other hand , when @xmath99 the in - plane components @xmath100 and @xmath101 of the new vector potential vanish for @xmath102 . as a result , @xmath103 . the hamiltonian then reads @xmath104 we achieve a hamiltonian which contains an efc - like term , but with a different prefactor . furthermore , since the new vector potential vanishes , the canonical and mechanical momenta are equal . we will refer to the transformation according to eq . ( [ eq : chi_beta ] ) with @xmath105 as the _ tl gauge_. the very reason for the new prefactor @xmath106 is the existence of a vortex , that causes the first term of an expansion of the vector potential near @xmath26 to be proportional to @xmath107 . for the case @xmath108 , the gauge transformation with @xmath109 is not advantageous . this is because @xmath110 , as seen by inspecting eq . ( [ eq : heur_anew ] ) , and it can not be neglected . conversely , choosing @xmath111 ( for @xmath112 ) would remove the @xmath113-component but keep the @xmath114-component . it is legitimate to wonder why there is such an asymmetry between tl fields having the same @xmath51 but differing in their circular polarization state , while an asymmetry of this type is not present in plane waves . we will further explore this in the next section . the aforementioned results suggest that there are two topologically distinct classes of tl fields , depending on the combination of oam and circular polarization , which we will study now in detail . we calculate the electric and magnetic fields using the full form of the vector potential [ eq . ( [ eq : a_bessel ] ) ] . a plot of two representative cases of electric and magnetic fields for @xmath115 and @xmath116 , at @xmath117 and @xmath118 is presented in fig . [ fig : efields ] . for comparison , the fields of a non - vortex beam with @xmath119 and @xmath120 are also shown . the vectorial character of the non - vortex beam is similar to a plane wave with perpendicular @xmath121 and @xmath122 fields . the amplitude is radially modulated according to the bessel function @xmath123 . in contrast , in the case of tl with @xmath115 the field profiles are much more complex . when @xmath124 , the electric field is oriented azimuthally around the beam axis , and the magnetic field in the central region points inwards . for other values of @xmath125 ( or @xmath75 ) , the patterns change , but eventually both magnetic and electric fields cycle through the radially - like and azimuthally - like polarization patterns . in contrast , when @xmath99 , the fields look entirely different , and never evolve into azimuthal or radial patterns . we refer to these two as the anti - parallel [ sign@xmath126 sign@xmath127 and the parallel [ sign@xmath128 sign@xmath127 beam classes . the field patterns shown in fig . [ fig : efields ] gives an indication why a gauge , in which the scalar potential provides the dominant contribution to the coupling could be found in the parallel class but not in the anti - parallel class . in the central region the field lines of the electric field in the parallel class are similar to a vector field close to a saddle point . such a vector field can indeed be written as the gradient of a scalar field . in contrast , in the anti - parallel class the field lines of the electric field are obviously closed indicating that this is dominantly a vortex - type field which has a non - vanishing curl and therefore can not be obtained as a gradient of a scalar potential . hence , in any gauge the interaction will mainly originate from the vector potential . & & + & @xmath129 & @xmath130 $ ] & @xmath131 $ ] + & @xmath132 & @xmath133 $ ] & @xmath134 $ ] + & @xmath135 & @xmath136 $ ] & @xmath137 $ ] + & @xmath129 & [ cols= " < " , ] + & @xmath135 & 0 & 0 + it is our interest to study the region close to the phase singularity @xmath83 . thus , we provide analytical results for the field amplitudes in this region expanded in powers of @xmath138 . table [ table : fields ] presents the lowest non - vanishing orders in @xmath138 of the electric and magnetic fields in the plane @xmath118 and at @xmath117 obtained from the full vector potential in eq . ( [ eq : a_bessel ] ) . in table [ table : fields_parax ] the same fields but obtained from the potential in the paraxial approximation [ eq . ( [ eq : ainitial ] ) ] are given . note that the in - plane vector potential of eq . ( [ eq : ainitial ] ) gives rise to a @xmath75-component of the magnetic field ; this component , however , has a prefactor @xmath139 and therefore , in order to be consistent with the paraxial approximation , it has been omitted in table [ table : fields_parax ] . indeed , it is clearly seen that if in table [ table : fields ] all terms containing a factor @xmath139 or @xmath140 are neglected the fields of table [ table : fields_parax ] are obtained . let us first compare the full form and the paraxial case for the parallel class [ sign(@xmath51)@xmath141sign(@xmath53 ) , in our case @xmath120 ] . in the paraxial approximation we obtain pure in - plane fields with electric and magnetic fields having the same dependence on @xmath114 . when calculated from the full vector potential , the magnetic field is slightly rescaled , the correction being of second order in the small parameter @xmath139 . both fields acquire a small @xmath75-component which is of first order in @xmath139 . additionally it is proportional to @xmath143 and thus decreases faster for @xmath144 than the in - plane components . thus , these corrections are negligible in the region close to the phase singularity and the assumptions of the paraxial approximation are well satisfied in this region . in the anti - parallel class [ sign(@xmath51 ) @xmath145 sign(@xmath53 ) , here @xmath108 ] , on the other hand , the @xmath75-components of the electric and magnetic fields still contain the small parameter @xmath139 , however the radial dependence is now proportional to @xmath147 . the in - plane component of the electric field is still proportional to @xmath95 , thus at sufficiently small @xmath114 the electric field is always dominated by the @xmath75-component . this clearly demonstrates that the paraxial approximation , which neglects the @xmath75-component , is not applicable in the region close to the phase singularity . indeed , a careful look at the @xmath75-component of the vector potential in eq . ( [ eq : a_bessel ] ) reveals that already there the small factor @xmath139 is counterbalanced by a @xmath114-dependence which is one order lower than for the in - plane components and therefore dominates close to @xmath83 . the dominance of the @xmath75-component of the fields is closely related to the field profiles in the anti - parallel case shown in fig . [ fig : efields ] . as already mentioned , the electric field profile has a non - vanishing curl which is oriented in the @xmath75-direction . according to maxwell s equation this curl is associated with the time - derivative of a magnetic field , which therefore necessarily has to have a strong @xmath75-component close to @xmath83 . half an oscillation period later , the roles of electric and magnetic fields in the central column of fig . [ fig : efields ] are interchanged . then the magnetic field lines are closed circles being associated with a strong @xmath75-component of the electric field close to the center . for angular momenta @xmath148 the magnetic field in the anti - parallel class has an additional correction which is of second order in @xmath139 but which has a @xmath114-dependence proportional to @xmath149 . for sufficiently small radii this is the dominant contribution to the fields . thus , in this case close to the center the beam is dominated by the magnetic field . this holds in particular for the case @xmath150 , in which there is a non - vanishing in - plane magnetic field at the beam center while the electric field vanishes at this point . this is again an indication that an efc - like hamiltonian is not applicable since with such a hamiltonian the interaction with matter is described only in terms of the electric field . some research articles in the topics of highly focused tl and azimuthally / radially - polarized fields report similar findings to ours . paraxial beams of tl can be focused using , e.g. , high - na lenses @xcite or a nanoantenna @xcite . the theoretical analysis of focusing , based entirely on electric and magnetic fields , can be done using the formalism by wolf @xcite , and the results @xcite show important similarities with the field patterns presented in fig . [ fig : efields ] . azimuthally- and radially - polarized fields are a special class of tl fields @xcite . the field patterns of azimuthally / radially - polarized _ non_-paraxial bessel beams presented by ornigotti et al . @xcite are also in agreement with our findings . regarding the magnitude of the fields near @xmath83 zurita - snchez et al . @xcite have shown that , for the strongly focused azimuthally - polarized beam they studied , the magnetic interaction overcomes the electric interaction near the phase singularity ; recently , their findings have been corroborated by the theoretical study of klimov et al . @xcite in the case of focused laguerre - gaussian beams . finally , in their research on highly - focused tl beams , monteiro et al . @xcite , iketaki et al . @xcite and klimov et al . @xcite report that interesting effects only occur when @xmath151 and @xmath108 . the overall similarities are no coincidence , for the vector potential eq . ( [ eq : a_bessel ] ) in contrast to eq . ( [ eq : ainitial]) shares with the aforementioned _ non_-paraxial beams the important feature of having a non - negligible @xmath75-component , which we have shown to give rise to the described features . we are now in a position to clarify the findings in the heuristic derivation of the tl - matter coupling shown in sec . there it was assumed that there is no @xmath75-component in the vector potential . from table [ table : fields ] we see that the @xmath75-component of the fields are negligible only in the parallel class . in the anti - parallel class they are proportional to @xmath152 . because for @xmath153 the magnetic field can not be neglected compared to the electric field , we were not able to derive an efc - like hamiltonian . in other words a hamiltonian representation given solely in terms of the electric multipoles , such as @xmath154 , is insufficient to describe the tl - matter interaction for the anti - parallel class . the use of the gauge transformation function @xmath155 found in sec . [ heur ] can be motivated using formal arguments . in the following we use the more general form of eq . ( [ eq : a_bessel ] ) for the vector potential in the coulomb gauge . for charged particles localized around the same center , a power - zienau - woolley ( pzw ) transformation can be done using the gauge function @xmath156 where @xmath157 is given in the coulomb gauge . this is the generalization to inhomogeneous fields of the gppert - mayer transformation ( dma ) , and leads to the so - called poincar gauge @xcite . in our work we focus on the interaction of tl with planar systems . therefore , if we consider a charge distribution mainly extended in the @xmath158 plane for a fixed @xmath75 , the quantity @xmath159 scales only in the in - plane component with @xmath160 ( see , e.g. , ref . @xcite ) . defining @xmath161 , the gauge function reads @xmath162 for small systems ( @xmath163 ) the radial dependence can be approximated by @xmath164 , which leads to @xmath165 . with these simplifications , we evaluate the integrals eq . ( [ eq : chi_poincare ] ) , and obtain @xmath166 the in - plane part of the transformation function @xmath155 is exactly the same as we got in sec . in addition , there is a new term arising from the non - vanishing @xmath75-component of the vector potential . note that we have neither required @xmath167 in the new gauge , nor have we neglected @xmath168 . additionally , for non - vortex fields ( @xmath61 ) with negligible @xmath75-component our result coincides with that of the efc hamiltonian . here we have motivated the use of eq . ( [ eq : chipzw ] ) by showing that the tl gauge function can formally be derived by a pzw transformation for charged particles localized in a planar structure ( constant @xmath75 ) . since any gauge transformation function can be postulated and used to cast the potential in suitable forms , the tl gauge can also be applied to other more general structures for variable @xmath75 ( with varying degrees of accuracy or usefulness ) . we see that the natural extension of the dma to the case of tl beams is slightly different from the plain efc hamiltonian . because of the generalized use of efc hamiltonians @xcite , it is worth exploring further its connection to our result . to this end , let us simply postulate a general gauge transformation of the form @xmath169 where @xmath170 is any number . clearly , we can recover the tl gauge by @xmath171 and @xmath172 . in contrast , when setting @xmath173 , eq . ( [ eq : chigeneral ] ) reduces to the efc gauge . in the following , we will again only consider the case of @xmath174 and polarization @xmath116 , since , as already discussed , there are no essential differences in the case with negative @xmath51 and opposite sign of @xmath53 . according to eqs . ( [ eq : gauge - transf_u ] ) and ( [ eq : chigeneral ] ) , the scalar potential in the new gauge reads @xmath175 obviously in the scalar potential we recover an efc - type structure of the hamiltonian , however with in general different prefactors for the in - plane and out - of - plane components . the new vector potential in the region close to the phase singularity is given by [ eq : new_a ] @xmath176 \sin \left [ { \left ( { \omega t - { q_z}z } \right ) - \left ( { \ell + \sigma } \right)\varphi } \right]\ , , \label{eq : new_a_r } \\ { { a'}_\varphi } & = & \frac{{{\beta _ r } - \left ( { 1 + \sigma \ell } \right)}}{{{\beta _ r } } } { a_\varphi } - \frac{\sigma}{{{2\beta _ z } } } \frac{{q_r^2}}{{q_z^2}}{a_0 } \frac{{\left ( { \ell + \sigma } \right)}}{{\left ( { \ell + \sigma } \right ) ! } } { { \left ( { \frac{{{q_r}r}}{2 } } \right)}^{\ell + \sigma - 1 } } \left({q_z}z\right ) \nonumber \\ & & \mbox{\qquad}\times \left [ 1 - \frac{1}{{\left ( { \ell + \sigma + 1 } \right ) } } { { \left ( { \frac{{{q_r}r}}{2 } } \right)}^{2 } } \right ] \cos \left [ { \left ( { \omega t - { q_z}z } \right ) - \left ( { \ell + \sigma } \right)\varphi } \right]\ , , \label{eq : new_a_vp}\\ { { a'}_z } & = & \frac{{{\beta _ z } - 1}}{{{\beta _ z}}}{a_z } - \frac{\sigma}{{{\beta _ z } } } \frac{{{q_r}}}{{{q_z}}}{a_0 } \frac{1}{{\left ( { \ell + \sigma } \right ) ! } } { { \left ( { \frac{{{q_r}r}}{2 } } \right)}^{\ell + \sigma } } \left({q_z}z\right ) \nonumber \\ & & \mbox{\qquad}\times \left [ 1 - \frac{1}{{\left ( { \ell + \sigma + 1 } \right)}}{{\left ( { \frac{{{q_r}r}}{2 } } \right)}^{2 } } \right ] \cos \left [ { \left ( { \omega t - { q_z}z } \right ) - \left ( { \ell + \sigma } \right)\varphi } \right ] \nonumber \\ & & - \frac{{2}}{{{\beta _ r}}}\frac{{{q_z}}}{{{q_r}}}{a_0}\frac{1}{{\ell ! } } { \left ( { \frac{{{q_r}r}}{2 } } \right)^{\ell + 1 } } \sin \left [ { \left ( { \omega t - { q_z}z } \right ) - \left ( { \ell + \sigma } \right)\varphi } \right]\,,\label{eq : new_a_z}\end{aligned}\ ] ] where we have used the expansion in powers of @xmath138 and kept all terms up to the order @xmath177 . as discussed in sec . [ heur ] for the paraxial approximation , also here it is obvious that the efc gauge with @xmath178 is not useful for tl because the transformed in - plane components of the vector potential are not smaller than the original ones . in fact , for @xmath179 they are in general even larger . a more detailed discussion of the efc gauge can be found in the appendix . in the following we will restrict ourselves to the tl gauge @xmath180 and @xmath181 and discuss the vector potential for the different cases . we remind that @xmath182 and @xmath183 are proportional to @xmath95 while @xmath184 . we first examine the new vector potential in the parallel class , i.e. , sign(@xmath51 ) @xmath141 sign(@xmath53 ) or , more explicitly , @xmath99 . the results are a direct extension to those found by the heuristic derivation in sec . [ heur ] . in the parallel class the radial dependence of the transformed potential is the same as for the original one . moreover , each component of the vector potential contains a term proportional to the small quantities @xmath185 . since we assume a planar structure these terms can be neglected . then the expressions simplify to @xmath186 a_{z}(\mathbf r , t ) \,.\end{aligned}\ ] ] the first thing to notice is that the components @xmath187 and @xmath188 of the new vector potential are zero , as we have already found in sec . therefore , in the hamiltonian @xmath189 the in - plane tl - matter interaction can be expressed solely by a dipole - like term @xmath190 with a prefactor different to the efc hamiltonian . for the @xmath75-component we have both a dipole - like term , but also a term @xmath191 which still contains the vector potential . we point out that @xmath192 a'_z\,^2(\mathbf r , t ) \propto ( q_r r)^{2\ell+2}$ ] , and may be safely disregarded . it is interesting to also compare again the canonical and mechanical momenta . their difference is given by @xmath193 also here the canonical and mechanical momenta are the same for the in - plane components and only in the @xmath75-component a difference in the momenta arises , which is however of the order @xmath143 and therefore one order higher than the correction to the in - plane momenta in the original gauge . let us now consider what happens in situations of experimental and application interest . we first address the situation when the interaction with the system only occurs through the in - plane components of the field , for example in the selective excitation of heavy holes in a quantum dot . then , the tl - matter interaction reads @xmath194 and is modeled by electric multipoles only with all the benefits of a dma . effectively we end up in the desirable situation where the vector potential is eliminated , as also shown in sec . nevertheless the description is beyond the dma because it keeps the full spatial dependence of the electric field and thus can give rise to transitions which are forbidden in the case of excitation by plane waves , for example transitions from envelope function with @xmath195-type symmetry in the valence band to those with @xmath196-type symmetry in the conduction band or vice versa . next , we consider the case where the system interacts with the @xmath75-component of the field , for example in intersubband transitions in quantum wells @xcite or the excitation of light holes @xcite . here , the electric multipoles are accompanied by a magnetic term arising from the non - vanishing @xmath75-component of the vector potential . however , since no atypical behavior of the fields near the phase singularity occurs , it is expected that the electric interaction is larger than the magnetic one as usually happens . one could then safely only retain the electric multipolar term , and possibly neglect the difference between momenta . therefore for the parallel class a hamiltonian with only electric dipole - moment terms having the correct prefactors can describe the tl - matter interaction at the phase singularity . for the anti - parallel class , we already found that a description with electric field only is not sufficient . still , we can gain valuable insights from studying the anti - parallel case with sign(@xmath51 ) @xmath145 sign(@xmath53 ) , i.e. , @xmath124 . here , the vector potential reads : @xmath197 a_\varphi(\mathbf r , t ) \ , , \\ \label{eq : apbeta4_varphi } a_\varphi'(\mathbf r , t ) & = & \frac{2\ell}{\ell+1 } a_\varphi(\mathbf r , t ) - \frac{q_r^2}{2q_z^2 } \frac{q_z z}{\ell } \left [ ( \ell-1 ) - \frac{4\ell(\ell-1)}{(q_r r)^2}\right ] a_r(\mathbf r , t ) \ \label{eq : apbeta4_z } a_{z}'(\mathbf r , t ) & = & \left [ \frac{\ell-1}{\ell } - \frac{2q_z}{q_r}\right ] a_z(\mathbf r , t ) - \frac{q_r}{q_z } \frac{q_z z}{\ell } \left [ \frac{q_r r}{2 } - \frac{2\ell}{q_r r}\right ] a_r(\mathbf r , t ) \,.\end{aligned}\ ] ] here we have kept the terms @xmath198 since , in contrast to the parallel class , they are now accompanied by radial dependencies proportional to @xmath199 and @xmath200 times the original vector potential . thus , the transformed vector potential becomes even stronger close to the phase singularity . the magnetic interaction resulting from these terms may be comparable or even surpass the electric interaction . this is in agreement with previous results for highly focused beams , where a magnetic field contribution stronger than the electric field contribution at the phase singularity was found @xcite . it is also interesting , that even far from the phase singularity , the in - plane term @xmath188 does not vanish . let us study this in more detail using as an example the excitation of a quantum dot placed at the beam axis by a tl beam and energy close to the qd band - gap . considering again the case of optical transitions with in - plane matrix elements such as the heavy hole - to - conduction band transitions , we neglect the @xmath75-component of the interaction , and also the terms proportional to @xmath201 . then , the hamiltonian reduces to @xmath202 \,.\end{aligned}\ ] ] [ we note that the angular component of the momentum vector reads @xmath203 , where the canonical momentum @xmath204 is in fact an angular momentum @xcite . ] though there is an efc - type hamiltonian , clearly the in - plane vector potential remains in the hamiltonian . we wonder how electric and magnetic contributions compare to each other . let us specifically consider the case @xmath150 . then , the electric multipolar term is proportional to @xmath205 . on the other hand , the magnetic term in eq . ( [ eq : h_anti - parallel ] ) is proportional to @xmath206 . if we assume that momentum and position vector are proportional to each other , as it is so in the dma ( since @xmath207 $ ] ) , it becomes clear that one should not a priori neglect the magnetic interaction , for it may be comparable or even larger the electric interaction , in particular at the phase singularity . when the @xmath75-component of the fields become also important , it is clear that also here the vector potential remains in the hamiltonian . thus , for the anti - parallel class the tl - gauge transformation , though being mathematically correct , is in general not advantageous . we have studied the tl - matter interaction close to the beam axis . in contrast to conventional light beams , twisted light has a phase singularity at the point @xmath83 , and a strong intermixing between polarization and oam . we distinguished the tl beams into two topologically different classes , namely the parallel class where handedness of circular polarization ( i.e. , the photon spin ) and oam have the same sign and the anti - parallel class where the signs of circular polarization and oam differ . to obtain a hamiltonian which includes the em fields instead of the potentials , we suggested to use a new gauge , the tl gauge . for the parallel class , the tl gauge leads to a hamiltonian which has a dipole - type structure , but a different prefactor . for in - plane problems it takes the simple form @xmath208 . the prefactor is mandatory to describe the correct interaction and to achieve the identity of canonical and mechanical momentum . the origin of the prefactor in the tl gauge is the vortex , which exists at the phase singularity . for the anti - parallel class we showed that the tl gauge , which casts the hamiltonian at least partly into electric fields , is not in general advantageous as the vector potential can not be eliminated nor neglected . because in the anti - parallel class magnetic effects can not be neglected compared to the electric ones , the hamiltonian should include magnetic as well as electric terms , and their relative strength must be analyzed in the particular problem at hand . we compared the tl gauge to the more common dma and efc hamiltonians . while for structures located close to the beam maximum the dma is applicable , for structures located close to the beam center it can not be used since the electric field at the phase singularity vanishes . we have also pointed out that the use of the paraxial approximation close to the phase singularity may be misleading and should be avoided at least in the anti - parallel beam class . in contrast to other gauges the tl gauge depends explicitly on beam parameters , in particular on the oam @xmath51 . on the one hand this is clearly a restriction , but on the other hand , when tl is used to excite structures in the region of the beam center , this is usually done with the aim to address specific transitions which are driven by a light field with a given value of @xmath51 . in this case a beam with a well - defined @xmath51 is used and thus , at least for beams within the parallel class , the tl gauge for this experimental set - up is well defined and can be used to write the coupling completely in terms of the electric field . in comparison to other gauges , like the poincar - gauge or the multipolar gauge , the tl gauge offers the same advantages for tl beams in the parallel class as the dma offers for slowly varying light beams : in contrast to the poincar - gauge , the tl gauge can be simply evaluated and leads to explicit formulas . for in - plane problems @xmath209 contains only the electric field , which makes it manifestly gauge invariant and secures the physical meaning of the momentum operator . furthermore , it contains all the higher order electric field couplings like coupling to quadrupole terms in a compact , appealing form . g. f. quinteiro would like to thank p. i. tamborenea for fruitful discussions in the topic of gauge - invariance in general and applied to tl . he also thanks the argentine research agency conicet for financial support through the `` programa de becas externas '' . in this appendix we will discuss in some more detail why the efc gauge is not useful for tl . we obtain the efc gauge from our more general gauge function eq . ( [ eq : chigeneral ] ) by setting @xmath210 . from the general formulas ( [ eq : new_a ] ) we can then calculate the new potentials . in the parallel class , when restricting ourselves to the lowest non - vanishing order in @xmath138 ( which is the order @xmath51 for the in - plane components and @xmath177 for the @xmath75-component ) and , as discussed in sec . [ sec : parallel ] , neglecting terms involving the small quantities @xmath198 , the expressions simplify to we see that the vector potential in the efc gauge grows with @xmath51 . at first glance the eqs . ( [ eq : apbeta1 ] ) might look surprising since they seem to violate the uniqueness of the em fields : if one wanted to calculate the z - component of the magnetic field using the in - plane components of the new vector potential , one would find that @xmath212 which would violate the independence of the em field on gauge transformations . such a contradiction is only apparent for the following reason . we approximated the in - plane components of the vector potential to lowest order in @xmath213 , i.e. , @xmath214 . under this approximation @xmath215 and thus there is no contradiction . the transformed vector potential also reveals a difference between canonical and mechanical momentum @xmath216 , which also grows with @xmath51 . therefore , when the efc gauge is applied to high-@xmath51 tl beams at the phase singularity and the canonical momentum instead of the mechanical momentum is used in calculations , a significant error may be introduced . @xmath217 a_\varphi(\mathbf r , t)\ , , \mbox{\qquad } \\ \label{eq : apbeta3_varphi } & & a_\varphi'(\mathbf r , t ) = \ell a_\varphi(\mathbf r , t ) \nonumber \\ & & \mbox{\quad } - \frac{q_r^2}{2q_z^2 } ( q_z z ) \left [ ( \ell-1 ) - \frac{4\ell(\ell-1)}{(q_r r)^2}\right ] a_r(\mathbf r , t)\ , , \\ \label{eq : apbeta3_z } & & a_{z}'(\mathbf r , t ) = - \frac{2q_z}{q_r } ( \ell+1 ) a_z(\mathbf r , t ) \nonumber \\ & & \mbox{\quad } - \frac{q_r}{q_z } ( q_z z)\left [ \frac{q_r r}{2 } -\frac{2\ell}{q_r r}\right ] a_r(\mathbf r , t ) \,.\end{aligned}\ ] ] again , the in - plane components grow with increasing @xmath51 . furthermore , like in the case of the tl gauge in the anti - parallel class ( sec . [ sec : antiparallel ] ) , the vector potential exhibits new terms containing @xmath218 multiplying the original vector potential . thus , also in the efc gauge the transformed vector potential becomes even stronger close to the phase singularity . for both reasons the efc gauge is not useful in the anti - parallel class . 61ifxundefined [ 1 ] ifx#1 ifnum [ 1 ] # 1firstoftwo secondoftwo ifx [ 1 ] # 1firstoftwo secondoftwo `` `` # 1''''@noop [ 0]secondoftwosanitize@url [ 0 ] + 12$12 & 12#1212_12%12@startlink[1]@endlink[0]@bib@innerbibempty @noop * * , ( ) @noop * * , ( ) @noop _ _ ( , ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop _ _ ( , ) @noop _ _ ( , ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop _ _ ( , , ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop _ _ ( , ) @noop * * , ( ) s. j. van enk and g. nienhuis 1994 europhys . lett . * @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop _ _ ( , )

there are two meanings ordinarily associated with the noun _ paradox_. either it is an apparently absurd or counter - intuitive statement that further explanation nevertheless proves to be well - founded , or it is a statement that is taken to be self - contradictory or intrinsically unreasonable ( adapted from @xcite ) . often , in the former meaning , a paradox is also said to be _ apparent_. in brief , the intent of this paper is is to show that a number of paradoxes in chain dynamics are indeed apparent paradoxes , as they can be explained within a unified theory . to this end , in sections [ sec : dissipation ] and [ sec : shocks ] , building on earlier work of oreilly and varadi @xcite and others before them , we propose a theory for dissipative singularities in a one - dimensional inextensible continuum , which we conventionally call a _ string_. the chain problems to which this theory is to be applied are presented in section [ sec : chain_problems ] along with a brief history of the ( supposedly ) paradoxical solutions that have been proposed so far . in due order , these very problems are then solved within the present theory in sections [ sec : falling ] , [ sec : folded ] , and [ sec : sliding ] . although attained in a different perspective , the conclusions reached here are in tune with those of grewal , johnson , and ruina @xcite , and of hamm and gminard @xcite before them . there is a common ground between these approaches which perhaps deserves further study and experimental validation . here we derive the balance equations that govern the dynamic evolution of inextensible strings with singularities in the velocity field from an appropriate extension of rayleigh s dissipation principle . the development illustrated below largely rests on sect . 2.2 of @xcite , whose major essential features will be briefly recalled whenever the need for self - consistency is more urgently felt . rayleigh s dissipation principle finds its complete formulation within thermodynamics , where it is also known as the principle of _ minimum entropy production _ , see ( * ? ? ? 2.2.2 ) . here we shall only need its purely mechanical variant , which can be easily obtained by specializing ( and partly extending ) the treatment in sect . 2.2.3 of @xcite to the purely isothermal case . a similar development suited for classical mechanical systems with a finite number of degrees of freedom has recently been proposed in @xcite in response to a different attitude towards non - conservative systems @xcite . here we shall extend the approach suggested in @xcite to one - dimensional continua with singularities , showing that rayleigh s formalism is indeed more far - reaching than commonly believed . for a material body , we denote by @xmath0 the power expended by all active agencies directly applied to it from the exterior ( and which are given explicitly in terms of the body s motion ) . letting @xmath1 denote the kinetic energy ( in a given inertial frame ) , we call @xmath2 the _ net working _ , as it represents the amount of external power still available to the body once that going into the motion has been accounted for . in general , @xmath3 can be expressed as a _ linear _ functional in the _ generalized _ velocities of the body ( including the classical velocity field @xmath4 and any extra field describing the evolution of additional degrees of freedom ) . by duality , the specific form of @xmath3 also identifies the _ generalized forces _ corresponding to the generalized velocities . in the presence of internal constraints , the _ total working _ @xmath5 is defined as @xmath6 where @xmath7 is the power expended by the ( generalized ) reactions enforced by the constraints . the version of rayleigh s principle that we posit here states that the true evolution of a body is identified by requiring that a functional @xmath8 , representing a _ dissipation potential _ , be such that @xmath9 is stationary with respect to all virtual motions for which the generalized forces are held fixed . formally , this amounts to requiring that @xmath10 provided that @xmath11 is linear in the ( generalized ) velocity variations . classically , @xmath8 is a _ , positive semi - definite functional , related to the _ dissipation _ @xmath12 in the body through @xmath13 were this always the case , rayleigh s principle would be a genuine _ minimum _ principle , and @xmath12 is actually minimized by the true motion compared to all virtual ones subject to the same generalized forces . here , for the class of problems at hand , we shall afford a generalization of . consider an inextensible string moving in the three - dimensional space @xmath14 . we denote by @xmath15 the _ motion _ of the string relative to a reference configuration , described by the convected co - ordinate @xmath16 , which by the constraint of inextensibility can be taken to represent the arc - length in both the reference configuration and the present one , the latter being the image @xmath17,t)$ ] at time @xmath18 of the former . the velocity field @xmath4 is defined as @xmath19 likewise , @xmath20 is the acceleration field . by requiring that the length of any arc @xmath21 be preserved in time , we easily obtain that @xmath4 must obey the constraint @xmath22 where @xmath23 denotes the unit tangent field in the present configuration . we further assume that a _ shock _ traverses the reference configuration , represented by a propagating front @xmath24 upon which the velocity field @xmath4 is discontinuous , while @xmath25 itself is continuous . the point @xmath26 designates the position in space of the shock . we call it a _ singular point _ , for brevity ; the arcs of the string that do not contain singular points will be called _ regular_. letting for any function @xmath27 denote by @xmath28 and @xmath29 the right and left limits of @xmath27 , that is , the limits for @xmath30 and @xmath31 , respectively , we set , as usual , @xmath32 we denote by @xmath33 the velocity of the singular point @xmath34 . it is not a material velocity and it must be continuous across the shock : @xmath35 hence @xmath36 figure [ fig : shock ] illustrates the general situation envisaged here . , where the unit tangent @xmath37 jumps from @xmath38 to @xmath39 as the arc - length parameter @xmath40 traverses @xmath41 , is acted upon by the force @xmath42 . the arc under consideration is described by the interval @xmath21 . ] equations and represent constraints that must be enforced at all times along the string and at the singular point @xmath34 , respectively . to free the velocity field @xmath4 from its constraint we account for the power @xmath7 expended by the associated reactive force . we write @xmath7 in the form @xmath43 where @xmath44 and @xmath45 delimit an interval which comprises the propagating front , @xmath46 , and @xmath47 represents an unknown _ tension _ along the string , which may suffer a jump at the singular point @xmath34 and will eventually be determined by requiring to be satisfied . in a different fashion , we shall enforce by requiring @xmath33 as defined by to be continuous across @xmath34 . an integration by parts performed in both integrals obtained by splitting the integral in over the intervals @xmath48 $ ] and @xmath49 $ ] leads us to the following equivalent expression for @xmath7 , @xmath50 letting @xmath51 denote the force acting per unit length of the string and @xmath52 its ( uniform ) linear mass density , if an external concentrated force @xmath53 is applied to @xmath34 ( as for the interaction with an obstacle ) , the net working @xmath3 can be written as @xmath54 which is easily derived by splitting as above the integral of the kinetic energy over the intervals @xmath48 $ ] and @xmath49 $ ] . for simplicity , we shall only consider a portion of string with a single singular point @xmath34 , which is also identified with the only source of dissipation . the rayleigh potential @xmath8 , which is to be related to the dissipation @xmath12 in the body not necessarily through , reduces to a function of objective dissipation measures . we take these latter to be the shock s speed @xmath55 in the reference configuration and the jump @xmath56 of the the velocity across the singular point @xmath34 , @xmath57 . two identities , which follow from , are instrumental to expressing the jump terms in both and in a way that makes only @xmath33 appear there as the natural velocity against which power is expended at the singular point @xmath34 . these are as follows [ eq : v_0_identities ] @xmath58 @xmath59 and are supposed to provide linear forms in @xmath33 and @xmath55 , so as to identify by duality the appropriate generalized forces expending power against these kinematic measures in and . we heed for later use that also allows one to rewrite as @xmath60 by using both and in and , and performing arbitrary variations @xmath61 and @xmath62 of both @xmath55 and @xmath4 , respectively , while leaving all generalized forces frozen , we arrive at @xmath63\cdot{\delta{\bm{v}}}ds , \ ] ] where use has also been made of the conditions @xmath64 which limit the variation of the motion to the portion of string under consideration . correspondingly , @xmath65 which follows from remarking that @xmath66 and by is further changed into @xmath67 requiring the variational principle in to be identically valid for arbitrary variations @xmath62 and @xmath61 , we arrive at the equations [ eq : evolution_equations ] @xmath68,\ ] ] @xmath69 @xmath70 where has also been used . these are the evolution equations of the theory ; they are valid along any regular arc and at any singular point of an inextensible string . while is clearly the expected balance equation for linear momentum on regular arcs of the string , with @xmath47 the internal tension , and are still to be related to the classical balances of linear momentum and energy at a singular point . we shall see shortly below how this can be achieved . i wish first employ equations to relate the rayleigh potential @xmath8 to the dissipation @xmath12 in the system . by use of and in and , with the aid of the identities and , we can rewrite @xmath5 as @xmath71 which by and becomes @xmath72 where @xmath73 clearly acquires the meaning of the energy dissipated per unit time in the string , since the singular point @xmath34 is admittedly the only point where energy may get lost . by its very physical interpretation , we shall consider @xmath12 as the primary constitutive function of the theory ; the dissipation potential @xmath8 , which enters the evolution equation is to be derived from @xmath12 through . solving this equation by the method of characteristics , we arrive at the following explicit expression for @xmath8 , to within an arbitrary additive constant , @xmath74 by comparing and , we easily see that whenever @xmath75 equation can equivalently be written as @xmath76 which has the advantage of linking directly the jump in tension to the dissipated energy . it follows from that whenever @xmath12 is homogeneous of degree @xmath77 in all its variables , @xmath78 of which is just a special case . in the present context , there is no guarantee that holds . more generally , we shall assume that @xmath12 is positive semi - definite and vanishing when _ either _ @xmath55 _ or _ @xmath79 vanish . for such a @xmath12 , @xmath8 as delivered by identifies through the evolution equation for @xmath41 required by the theory . this equation together with and will determine the motion of a singular string . equation is not expected to be valid in general . nonetheless , the proposal for @xmath12 made in @xcite , whereas does not comply with , still makes satisfied . it was assumed in @xcite that @xmath80 with @xmath81 a phenomenological , dimensionless coefficient . it readily follows from that the rayleigh function @xmath8 associated with this @xmath12 is given by with @xmath82 . the expression for @xmath12 in will also be adopted in the rest of the paper , for it is amenable to a physical interpretation based on the classical laws of impact dynamics ( see section 4[sec : dissipation_chains ] below ) . we close this section by showing how equation and are related to the classical balances of linear momentum and energy at the singular point @xmath34 . that expresses precisely the former follows from writing as in @xcite the balance of linear momentum for the whole arc @xmath83 $ ] in the form @xmath84 splitting the integral on the left side of this equation over the intervals @xmath48 $ ] and @xmath49 $ ] , both of which depend on time , and taking the limits as both @xmath85 and @xmath86 , which isolates the singular point @xmath34 from the regular arcs adjacent to it , we reobtain . in a similar way , granted that the dissipation @xmath12 is concentrated at @xmath34 , the balance of energy over the whole arc @xmath83 $ ] reads as @xmath87 whence by the same reasoning as above we arrive at the following localized energy balance , @xmath88 for @xmath89 , this balance equation coincides with ( 5 ) of @xcite , provided that we identify the power supply @xmath90 introduced there at @xmath34 with @xmath91 . likewise , it reduces to the analogous equation ( 2.7)@xmath92 of @xcite , provided that there @xmath93 incorporates both the energy dissipated at the singular point and the power expended by the external force . equation also agrees with ( 13 ) of @xcite , provided that the power supply @xmath94 introduced there is written as @xmath95 . now , is nothing but , as shown by using both and and combining the result with . this also shows that and together imply . thus , our interpretation of the evolution equations and as balance laws is completely justified . often in the following we shall turn to as an equivalent formulation for in the presence of a propagating shock , and so our theory will be based on , , and . the shocks considered in the preceding section , which manifest themselves at singular points in the middle of a moving string , will be called _ internal _ to distinguish them from the _ external _ shocks considered here . an external shock manifests itself whenever a moving string gets in contact with a still , shapeless reservoir from where it is ejected and set in motion or where it comes to an abrupt halt . this will be our way of treating continuous systems with a variable mass . essentially , this idea goes back to cayley @xcite , who introduced the notion of _ continuously _ imparted impacts , equally able of setting resting matter in motion and of putting moving matter at rest . though our derivation of the balance equations governing an external shock will be similar to the one presented in section 2[sec : classical ] for internal shocks , the reader should not be deceived into thinking that an external shock is nothing but an internal shock with zero velocity on one side . in an external shock the string is supposed to be quiescent on one side of the shock , but its shape is completely undetermined and it only plays the role of providing sources for both linear momentum and energy . we distinguish two types of external shocks , as to whether the quiescent string follows or precedes the moving string . both situations are depicted in fig . [ fig : external_shocks ] . we denote by @xmath96 and @xmath97 the sources of linear momentum and energy , respectively , for the former case , and by @xmath98 and @xmath99 the corresponding sources for the latter case . we note in passing that @xmath100 and @xmath101 are the forces transferred to the quiescent string in the two case , while @xmath102 and @xmath103 are the corresponding energy sources for the quiescent string in the two cases . , these may equally well represent energy gains or energy losses . ] the external shock represented in fig . [ fig : external_shocks](a ) is characterized by a front @xmath104 evolving in time in both the reference and present configurations , so that the string s velocity @xmath4 jumps from @xmath105 to nought at the singular point @xmath106 . by requiring the velocity @xmath107 of the geometric point that instantaneously coincides with @xmath108 to vanish identically in time , as commanded by the kinematic compatibility with the quiescent string , we readily obtain that @xmath109 which here replaces . must be tangent to the moving string , and so this must move along its length at least just before coming to an abrupt stop . ] the balance of linear momentum for the arc of string in fig . [ fig : external_shocks](a ) , comprised between @xmath110 and @xmath111 , can be written in the form @xmath112 where as in section [ sec : dissipation ] @xmath51 denotes the force acting per unit length of the string . accounting in for the dependence of @xmath113 on time and taking the limit as @xmath44 tends to @xmath113 , we arrive at @xmath114 where @xmath115 is the tension in the moving string at the singular point @xmath108 . by we give the following form @xmath116{\bm{t}^\ast_+}.\ ] ] if the balance of linear momentum in echoes , the balance of energy in here becomes @xmath117 where account has also be taken of the fact that the force @xmath96 expends no power since @xmath118 . by performing the time differentiation in and then taking the limit as @xmath119 , also with the aid of , we obtain the following balance law for the energy at the singular point @xmath108 , @xmath120{\dot{s}^\ast_+}.\ ] ] reasoning in exactly the same way , we derive the analogs of the kinematic condition and balance equations and at the singular point @xmath121 depicted in fig . [ fig : external_shocks](b ) where a quiescent string is abruptly set in motion . letting @xmath122 denote the velocity of the moving sting at @xmath121 and @xmath123 the inward unit tangent , we arrive at the following equations @xmath124 @xmath125{\bm{t}^\ast_-},\ ] ] @xmath126{\dot{s}^\ast_-},\ ] ] where @xmath127 is the tension in the moving string at @xmath121 . in particular , kinematic compatibility requires the string to be injected at @xmath121 along its tangent and the balance of linear momentum requires the force @xmath98 to be directed along the string s tangent as well . although equations and are valid in general , precisely as are and , they all are compatible with a special class of motions for a string which will be our object in the following sections , the one in which a string slides everywhere along its length . such motions are characterized by the kinematic requirement @xmath128 enforced along the entire moving string . it is compatible with the inextensibility constraint only is @xmath129 is assumed to be a function of time only , @xmath130 . under this assumption , both and reduce to [ eq : external_shock_plus_minus ] @xmath131 and , , , and collectively simplify into @xmath132 @xmath133 these equations will often be called upon in the following . in the following sections we shall apply the general theory presented so far to three specific problems in the dynamics of chains . below in this section the reader will be introduced to these problems collectively , also indulging in some historical notes on the supposedly paradoxical aspects that they have shown since the work of buquoy nearly two centuries ago .. if attention is restricted to an appropriate subsystem , all the problems treated here could also be seen as problems for systems with a variable mass . ] before doing so , however , we need supplement the theory by showing that the choice for the shock dissipation @xmath12 in is the most appropriate for chains . resuming the notation introduced in @xcite , for a shock in a chain we shall denote by @xmath90 the energy supply at a singular point , and so hereafter we shall set @xmath134 the kink formed in a chain where a shock is propagating through it is likely to involve an impact between the chain s links that come there in contact . when two smooth rigid bodies collide in a single point of their boundaries , the variation @xmath135 in the total kinetic energy during the impact can be expressed as @xmath136 where @xmath137 is newton s restitution coefficient and @xmath138 is the kinetic energy of the velocity differences , that is , the kinetic of the imaginary motion that would result from attributing to all the particles in the colliding bodies the velocities @xmath139 they lose in the impact . in other words , letting @xmath140 be the velocity of the @xmath141th particle before impact and @xmath142 that after impact , @xmath143 and @xmath144 where @xmath145 is the mass of the @xmath141th particle and the sum is extended to any discretization of of both colliding bodies . when a chain is described as a one - dimensional , inextensible string , as we did above , and a shock propagating though it is envisaged as a localized mass transfer with rate @xmath146 associated with the energy supply @xmath90 , directly suggests taking the rate of change of @xmath138 as @xmath147 so that the justifies , after setting @xmath148 in the following we shall not insist in representing the coefficient @xmath149 through in terms of the restitution coefficient for the impact of two chain s links . rather , we shall consider @xmath149 as a phenomenological coefficient pertaining only to the shape of the links in a chain and the material that constitutes them . of we shall only retain the bounds it imposes on @xmath149 , @xmath81 , . ] and we shall say that @xmath150 is the completely elastic limit , as there would be no dissipation at the shock , while @xmath151 is the completely inelastic limit , for which the dissipation at the shock would be maximum , for a given velocity jump @xmath79 . as we shall see , @xmath149 is a measurable parameter , and indeed it has been measured already , though perhaps inadvertently . how chains have come to be associated with paradoxes is quite an interesting story , which here i only recount briefly . these _ paradoxes _ are indeed counter - intuitive predictions which some theories made and others did not support , which were sometimes sustained by experimental observations and sometimes were not . often it was considered paradoxical in this context relaxing an assumption mostly taken for granted and regarded as natural , but indeed unnecessary . the examples treated in the following are concerned with chains falling in one fashion or another in the terrestrial gravitational field . for them , for example , it might have been regarded as paradoxical a theory that predicted falling with an acceleration greater than the gravitational acceleration @xmath152 , not to mention the possibility that in the continuum limit either the velocity or the acceleration became unbounded during the motion . specifically , these are the problems we shall treat here ( and certainly , not for the first time ) . problem ( a ) , or the problem of a _ falling chain _ : a chain initially kept at equilibrium over a smooth , rigid plane , is let go under the action of its own weight ; describe the motion of the chain while the links coming in contact with the supporting plane accumulate over it , and compute the total force impressed on it , that is , the _ dynamical weight _ of the chain , see fig . [ fig : chain_sketches](a ) . problem ( b ) , or the problem of a _ folded chain _ : a chain is initially held in a u - shaped configuration , with one arm ending in a fixed point and the other free to fall under its own weight ; describe the motion until when the chain is fully stretched vertically and compute the dynamical weight of the chain , see fig . [ fig : chain_sketches](b ) . problem ( c ) , or the problem of a _ sliding chain _ : a chain is supported by a smooth , rigid plane and leaves it though a hole under the action of gravity ; describe the motion of the chain hanging vertically below the plane and compute the dynamical weight of the chain , see fig . [ fig : chain_sketches](c ) . it is hard to say which of these problems is the oldest . for example , a variant of problem ( a ) is treated in tait and steele s treatise @xcite . here one end of a uniformly heavy chain hangs over a small , smooth pulley and the other end is coiled up on a table ; if the free arm of the chain preponderates the antagonistic arm is pulled upward and keeps removing links from the table ; if , on the contrary , the free arm is outweighed by the other arm , this problem becomes very similar to problem ( a ) . problem ( b ) is treated in love s book @xcite ; here the free chain s arm is supposed to fall with acceleration @xmath152 and the jump in tension at the tight chain s bight is derived from the localized balance of linear momentum . problem ( c ) featured in both jean s @xcite and love s @xcite treatises , which arrive at the same equation of motion obtained by cayley @xcite from his general variational principle for the motion of systems whose mass grows by continuously taking into `` connextion '' ( in cayley s own language ) the contents of a quiescent reservoir . the solution method suggested by those textbooks was based on the balance of linear momentum , and no mention was made there of cayley s principle , though both methods produced one and the same equation of motion when applied to problem ( c ) . in solving both problems ( b ) and ( c ) , love s approach , unlike others , requires computing the tension within the chain . moreover , also in view of our further development , i found very illuminating the following sentence from love s book @xcite : `` it is important to observe that discontinuous motions such as are considered here in general involve dissipation of energy . '' before discussing the details of the solutions to problems ( a)-(c ) afforded by the theory proposed in this paper , also to place them in a clearer perspective , i summarize below some other , newer approaches to these problems , some perhaps more convincing than others . one popular solution to problem ( a ) assumes that the chain s links fall with acceleration @xmath152 and computes the force that they exert on the supporting plane by balancing the linear momentum when they come abruptly at rest @xcite . the conclusion is that at each instant in time the dynamical weight of the chain is three times the static weight of the chain collected on the plane up to then . this solution was already common knowledge by the early 1950s @xcite and also received some experimental confirmation , mixed with the voicing of some concerns as to the validity of the assumption that the chain falls with acceleration @xmath152 , or _ freely _ , as one might also say @xcite . it was indeed shown experimentally by hamm and gminard @xcite that a chain does _ not _ fall freely on a supporting plane , but with an acceleration _ greater _ than @xmath152 . this naturally led then to envision the existence of a dynamical force pulling the chain down , which can only be exerted by the solid surface on which the chain is accumulating , and for which an explicit representation was also proposed , based on a model for the growth of the accumulating coil @xcite . as remarked in @xcite , this force is for all intents and purposes a _ tension _ exerted by the supporting surface , pulling downward the falling chain . as a result , the dynamical weight of the chain now exceeds three times the static weight of the accumulated chain ( even dramatically so , as the experimental figures on @xcite suggest that it is nearly 6 times the total static weight of the chain at the time immediately preceding its complete deposition ) . in section [ sec : falling ] below , we shall make a more definite contact with both this study and that of grewal , johnson and ruina @xcite , which are equally central to our development . problem ( b ) is perhaps the most controversial of the three considered here . if , as we learned above , love @xcite assumed that the chain s arm with a free end falls freely , hamel @xcite , as before him had done kucharski @xcite , rejected this assumption and based instead the solution of the problem of the folded chain on the conservation of the total mechanical energy . this latter implies that the velocity of the falling arm increases indefinitely as the chain approaches the fully stretched configuration . this can be easily understood by realizing that conservation of energy requires a finite amount of potential energy to be transferred to a moving chain fragment whose mass decreases to nought as the chain straightens to its full length . direct inspection of the equation of motion derived from the energy conservation shows that also the acceleration diverges in the same limit , and so the chain is bound to fall faster than in a free fall . again a positive tension , @xmath153 , applied to the falling arm at the chain s bight must be responsible for such a faster fall . it was realized in @xcite that conservation of energy actually requires that @xmath153 equals the tension , @xmath154 , acting at the bight on the resting arm . as also pointed out in @xcite , different _ constitutive _ choices for these tensions would grant a full host of evolutions for the folded chain , including the one where the free arm falls freely @xcite . the problem , however , is that we are not at liberty of choosing constitutive laws for @xmath153 and @xmath154 , at least as long as we model chains as _ inextensible _ strings , for which the tension in an internal _ reactive _ force to be determined . we shall show in section [ sec : folded ] how this task can be accomplished within our theory . the chain problems treated here have been selected with the purpose of proving that , treated in the appropriate continuum theory , they are free from any paradox . studies that treat the problem of a folded chain with the methods and language appropriate to a discrete systems have also been pursued ; among these , two in particular deserve notice @xcite . in the classical solution of problem ( c ) , when the chain starts sliding downward from rest and zero dangling length , the motion is predicted to take place with uniform acceleration @xmath155 . although nothing would of course be paradoxical with such a prediction , echoing love @xcite , sommerfeld @xcite remarked that in the corresponding motion the total mechanical energy is not conserved and made carnot s energy loss mechanism responsible for this @xcite . such a conclusion was rejected in @xcite and @xcite in favour of a lagrangean theory that enforces energy conservation . thus , if not paradoxical , problem ( c ) has become at least controversial , even more so on account of the fact that a convincing explanation for the energy loss can be given in terms of the ( inelastic ) impacts suffered by the links continuously set in motion by the sliding chain . in the following three sections we shall study the chain problems described in this section within the general theory proposed here . the only constitutive parameter will be @xmath149 in . whenever @xmath156 , the singular point which will be appropriately identified for the three problems , is the site of a dissipative internal shock . further sources of dissipation may arise from external shocks , if present ; for them expresses in this theory the analogue of what in sommerfeld s language @xcite is carnot s energy loss . in this section and in the following two we solve problems ( a ) , ( b ) , and ( c ) , respectively , by applying the theory put forward in sects . [ sec : dissipation ] and [ sec : shocks ] . in particular , we shall make use of equation to describe the motion of a regular arc in a chain , of to balance the linear momentum at a singular point , and of , which combined with embodies the balance of energy at a singular point . by , here acquires the form @xmath157 all the above equations , which are dynamical in nature , must be combined with the condition for kinematic compatibility in . and @xmath108 , the former being an internal shock and the latter an external one . the height of the falling chain at time @xmath18 is @xmath158 ; the length being deposited on the supporting plane at the same time is @xmath159 , where @xmath160 is the total length of the chain . @xmath161 is the reactive , impulsive force continuously imparted by the plane to the chain at @xmath34 . ( b ) the time of fall @xmath162 scaled to @xmath163 in as a function of the parameter @xmath149 . , title="fig : " ] and @xmath108 , the former being an internal shock and the latter an external one . the height of the falling chain at time @xmath18 is @xmath158 ; the length being deposited on the supporting plane at the same time is @xmath159 , where @xmath160 is the total length of the chain . @xmath161 is the reactive , impulsive force continuously imparted by the plane to the chain at @xmath34 . ( b ) the time of fall @xmath162 scaled to @xmath163 in as a function of the parameter @xmath149 . , title="fig : " ] figure [ fig : falling_chain](a ) , which is a decorated version of the sketch in fig . [ fig : chain_sketches](a ) , describes our idealized model for a falling chain . here @xmath34 denotes the point where the chain comes in contact with a smooth , rigid plane ; we think of it as a singular point for the chain dynamics , where the unit tangent @xmath37 jumps from @xmath164 to @xmath165 in the cartesian frame @xmath166 at the value @xmath167 of the arc - length parameter that describes the material points in the chain with origin in its free end . an internal shock in propagating at @xmath34 with speed @xmath55 . at the point @xmath108 , which we shall imagine to be arbitrarily close to @xmath34 , the chain comes in contact with the coil of links already deposited on the supporting plane . this is the site of an external shock , governed by ( the plus side of ) equations . here we study the dynamics of this chain in the special class of motions that preserve its shape , characterized by . by the inextensibility constraint , at instant @xmath18 in time @xmath168 is the velocity of all links in the moving chain . the compatibility condition indeed readily gives that @xmath169 and so the internal shock is standing in space , as also is the external shock at @xmath108 , by . the theory we adopt does not require any specific assumption on the shape of the quiescent coil deposited on the plane . we only need to know the force @xmath100 it suffers from the chain . since we assume @xmath108 to be arbitrarily close to @xmath34 , we identify the tension @xmath115 featuring in with the right limit @xmath153 of the tension in the chain at @xmath34 and we write as @xmath170 at @xmath34 the chain is also subject to the linear momentum supply @xmath161 , which here plays the role of @xmath53 in and is taken to be orthogonal to the supporting plane ( since the latter is smooth ) , @xmath171 , with @xmath172 . having a reactive nature , @xmath173 is to be determined along with the motion . writing in the frame @xmath166 , we readily conclude that [ eq : falling_chain_singular_momentum_balance ] @xmath174 where @xmath154 is the left limit of the tension at @xmath34 , responsible when positive for a faster fall of the chain . making use of in , we see that @xmath175 , and so the coil of accumulating links is free from any horizontal forces . to determine @xmath173 , we need to know @xmath154 , which since @xmath172 is subject to the bound @xmath176 if , as in fig . [ fig : falling_chain](a ) , @xmath158 denotes the hight of the falling chain at time @xmath18 and @xmath160 is the total length of the chain , then @xmath159 is the length of the chain accumulated on the plane . correspondingly , @xmath177 , where a superimposed dot denotes differentiation with respect to time , and becomes @xmath178 since along the falling chain @xmath179 and @xmath180 , equation yields @xmath181 being the left side of equation independent of @xmath40 and @xmath182 at the free end @xmath183 , it readily follows that @xmath184 since @xmath185 at @xmath34 , combining , , , and , we obtain the equation @xmath186 which actually describes the motion of the chain only in terms of the constitutive parameter @xmath149 . shortly below we shall give an analytic representation of certain solutions of . some of their qualitative features can be easily obtained directly from . first , if @xmath187 , then @xmath188 , and so the chain falls with an acceleration greater than @xmath152 . second , if @xmath81 and @xmath189 , which is when the chain is released with initial zero velocity , then @xmath190 all along the motion , and so becomes @xmath191 which coincides with ( 5 ) of @xcite , provided that @xmath149 is set equal to the parameter @xmath192 introduced there . accordingly , by , , and , we can write the tension @xmath154 exerted on the falling chain at @xmath34 as @xmath193 which obeys and vanishes for @xmath151 , that is , when the chain is in free fall . by inserting into , we arrive at @xmath194 which agrees with ( 6 ) of @xcite . moreover , the tension @xmath154 was denoted @xmath195 in @xcite and we may identify with @xmath173 the force denoted @xmath196 there . by comparing and we can easily express our @xmath149 in terms of @xmath195 and @xmath196 , in terms of @xmath195 and @xmath196 in @xcite . that formula would be reconciled with by exchanging @xmath195 and @xmath196 . ] @xmath197 an identification which is further confirmed by comparing with ( 10 ) of @xcite . the advantage of and over is that it expresses both @xmath195 and @xmath196 in terms of a single constitutive parameter , @xmath149 . thus , for the record , the special chains ingeniously constructed in @xcite would correspond to @xmath198 and @xmath199 , see equations ( 18 ) and ( 19 ) of @xcite , respectively . @xmath173 is also the dynamic weight of the chain that is yet to fall ; to obtain the total dynamic weight @xmath200 of the falling chain we need to add to @xmath173 the static weight of the chain that has already fallen : @xmath201 once is solved for @xmath158 , shall deliver @xmath200 as a function of time . hamm and gminard @xcite used @xmath149 ( which they called @xmath192 ) , as a fitting parameter to explain their measurements indicating that a chain was indeed falling with acceleration greater than @xmath152 . they determined @xmath202 and , more importantly , they found that this value was independent of the material properties of the floor on which the chain was falling , supporting our view that @xmath149 is indeed a constitutive parameter of the chain only . the ( implicit ) solution of can be found by quadratures . to free the solution from inessential physical constants , we rescale both lengths and time . lengths will be rescaled to the total length @xmath160 of the chain . time will be rescaled to @xmath203 which is the time of _ free fall _ from the height @xmath160 . thus , letting @xmath204 equation is transformed into @xmath205 where a prime @xmath206 denotes differentiation with respect to @xmath207 . correspondingly , letting @xmath208 be the total static weight of the chain , we obtain from that @xmath209 the solution of under the initial conditions @xmath210 and @xmath211 reads as @xmath212 whence it follows that the _ time of fall _ @xmath162 is given by @xmath213 which is function of @xmath149 only . figure [ fig : falling_chain](b ) shows the graph of this function ; it is apparent that @xmath214 , with equality holding only for @xmath151 . it also makes us appreciate quantitatively how faster than freely a chain may actually fall . figure [ fig : falling_chain_y_p_plots](a ) illustrates the graphs of @xmath215 against @xmath207 according to for three values of @xmath149 . they indicate that , apart from the case where @xmath151 , @xmath216 diverges to @xmath217 as @xmath218 , which indeed follows directly from , as also does the similar divergence of @xmath219 in the same limit . as a consequence , by , @xmath200 is predicted to diverge to @xmath220 as @xmath218 . by , equation can be given an easier expression , @xmath221 which shows how @xmath200 diverges as @xmath222 , unless @xmath151 or @xmath150 . the former is the case of free fall , when reduces to @xmath223 . the latter is the case of no dissipation at the internal shock , when reduces to @xmath224 . for @xmath150 appears a bit peculiar , we shall offer below an independent argument that would suggest that @xmath149 should indeed be greater than or equal to @xmath199 . ] the unboundedness of @xmath216 , @xmath219 , and @xmath200 could be regarded as paradoxical . it is indeed a consequence of the continuum model being used here , which can hardly be considered valid up to when @xmath225 vanishes .. ] following @xcite and @xcite , we introduce a cutoff length , @xmath226 , which could be estimated as consisting of up to 3 chain s links @xcite , so that @xmath215 would be subject to the bound @xmath227 . such a bound alters only marginally the time of fall @xmath162 , especially if @xmath228 is sufficiently small , but it makes the limiting value of @xmath200 finite . figure [ fig : falling_chain_y_p_plots](b ) illustrates the graphs of @xmath200 against time for several values of @xmath149 ; they are drawn up to the time when @xmath229 . we close this section by computing the total dissipation associated with our solution of the falling chain problem . there are two independent sources of dissipation in this system , one for each shock . at the internal shock , by and , the dissipative power @xmath90 is given by @xmath230 at the external shock , by and , the dissipative power @xmath97 is given by @xmath231 that both these expressions are compatible with the total balance of energy is seen by computing the time rates of the total kinetic energy @xmath232 and the total potential energy @xmath233 making use of it is a simple matter to show that @xmath234 which coincides with ( 13 ) of @xcite . since @xmath235 , requires that for the total dissipation to be non - negative ( which is the same as requiring that @xmath236 ) , @xmath149 must obey the inequality @xmath237 . precisely the same strategy adopted in section [ sec : falling ] to solve the problem of a falling chain is followed here to solve the problem of a folded chain . to avoid unnecessary repetitions , i shall only outline the major steps of the solution and comment the results . figure [ fig : folded_chain](a ) illustrates the problem in the notation introduced in our theory . the length of the fallen chain is denoted by @xmath225 ; @xmath160 is the total length of the chain , so that the distance between fixed and free ends is @xmath238 , and @xmath34 is the singular point where the chain is tightly folded . as in @xcite and @xcite , we further assume that a vertical force @xmath239 , with @xmath240 , is applied to the free end of the chain . the motion is assumed to take place under gravity and along the vertical line , so that only @xmath158 suffices to describe it . taking the origin of the arc - length parameter @xmath40 in the fixed point of the chain , we denote by @xmath241 the value of @xmath40 designating the front of the shock . there , as @xmath40 increases , @xmath37 jumps from @xmath164 to @xmath242 in the cartesian frame @xmath166 shown in fig . [ fig : folded_chain](a ) . the kinematic compatibility condition here gives @xmath243 which replaces . since the free end moves with speed @xmath244 , inexensibility and require that holds also here , and implies that @xmath245 setting @xmath246 equal to the tension at the free end and integrating , we arrive at @xmath247 by , the balance of linear momentum at @xmath34 reduces to the equation @xmath248 which combined with and delivers @xmath249 and @xmath250 which is the equation of motion for @xmath158 . for @xmath151 , equation reduces to equations ( 15 ) and ( 16 ) of @xcite . moreover , setting @xmath251 in equation ( 7.8 ) of @xcite , we recover our equation , provided that @xmath149 is identified with the parameter @xmath137 of @xcite . since the arm of the chain ending in the fixed point is at rest , the dynamical weight of the whole chain is simply @xmath252 , which by becomes @xmath253 for the folded chain the only source of dissipation is at the internal shock . it is an immediate consequence of and that @xmath254 which coincides with ( 7.7 ) of @xcite and agrees with ( 15 ) of @xcite for @xmath151 . moreover , the power @xmath255 expended by the force @xmath256 is easily computed to be @xmath257 the balance of total energy requires that @xmath258 starting from the following expressions for @xmath259 and @xmath260 , @xmath261 @xmath262 it is an easy exercise to show that is indeed valid along all solutions of . a special case of is worth considering in more details . if @xmath263 , becomes @xmath264 it follows from that @xmath265 whenever @xmath266 . if , in addition , @xmath151 then @xmath267 , which means that the free end of the chain falls freely . however , if @xmath268 then @xmath269 and the chain falls faster . introducing the scaled variables defined in , we write as @xmath270 valid whenever @xmath271 . equation can be solved by quadratures . setting @xmath272 and letting @xmath211 , we arrive at the following ( implicit ) representation for the solution , @xmath273 and correspondingly @xmath200 in becomes @xmath274+\eta.\ ] ] to illustrate this solution we consider the case when @xmath275 , so that the chain is fully extended upward before being set free . the time of fall @xmath162 , which is delivered by the formula @xmath276 is plotted in fig . [ fig : folded_chain](b ) as a function of @xmath149 . again @xmath162 is equal to the time of free fall @xmath163 only for @xmath151 ; otherwise , @xmath277 . figure [ fig : folded_chain](c ) shows how @xmath225 grows in time for three values of @xmath149 . here , unless @xmath151 , @xmath216 diverges to @xmath220 as @xmath218 . in fig . [ fig : folded_chain](d ) are plotted in a logarithm scale the graphs of @xmath200 as delivered by . for @xmath151 , @xmath278 , so that the dynamic weight of the completely fallen chain is @xmath279 . for @xmath268 , @xmath200 diverges to @xmath220 as @xmath280 and each graph in fig . [ fig : folded_chain](d ) is plotted up to the time when @xmath281 , where @xmath228 is the same cutoff parameter introduced in section [ sec : falling ] . an idealized sliding chain is represented in in fig . [ fig : sliding_chain](a ) . here a heap of links , whose specific shape and structure is ignored , is sitting on a smooth , rigid plane and act as a reservoir for the chain sliding away under gravity through a hole in the plane . an internal shock travels through the singular point @xmath34 as does an external shock through @xmath121 , which we take to be extremely close to @xmath34 . the origin of the arch - length co - ordinate @xmath40 is such that the falling free end of the chain has @xmath282 , where @xmath160 is the total length of the chain . as above , we denote by @xmath283 and @xmath241 the propagating shocks fronts . with this choice for the orientation of @xmath40 , at @xmath34 the unit tangent @xmath37 jumps from @xmath284 to @xmath285 . kinematic compatibility acquires again the form , both shocks are standing in space , and is replaced by @xmath286 where @xmath158 denotes the length of the chain fallen at time @xmath18 . accordingly , becomes @xmath287 since the tension vanishes at the free end of the chain , integrating we readily arrive at @xmath288 as shown in fig . [ fig : sliding_chain](a ) , the plane exerts an impulsive reaction , @xmath171 , with @xmath172 . at @xmath34 the balance equation for linear momentum then requires that [ eq : sliding_chain_tau_m_and_phi_0 ] @xmath289 combining together , , and , we obtain the equation of motion for @xmath225 , @xmath290 and @xmath291 which delivers an admissible reaction only is @xmath292 , which is indeed a consequence of , if @xmath266 and @xmath293 , both taken as valid here . for @xmath150 equation coincides with both cayley s equation @xcite and the equation proposed in @xcite . from , we also derive the dynamical weight @xmath200 of the chain , @xmath294 identifying @xmath154 with the tension @xmath127 at @xmath121 , by and , we also conclude that @xmath295 , and so the heap of links suffers no horizontal force as a consequence of the chain s motion . by and , we easily see that the power @xmath99 expended at the external shock is @xmath296 while by the power expended at the internal shock is @xmath297 since the total kinetic and potential energies are here expressed as @xmath298 respectively , a simple computation shows that @xmath299 along every solution of , as expected . a special solution of deserves notice by its simplicity . if @xmath300 and @xmath189 , then either @xmath301 or @xmath302 with @xmath303 which corresponds to a uniformly accelerated fall with acceleration @xmath304 . correspondingly , the time @xmath162 for the complete fall of the chain is @xmath305 , where @xmath163 is the same as in , and @xmath306 where @xmath307 is again the total static weight of the chain . more generally , introducing the same scaled variables as in , we give and the following forms @xmath308 @xmath309 both valid for @xmath310 . letting @xmath311 and @xmath211 , the implicit solution of reads as @xmath312 valid for @xmath313 . the corresponding time of fall @xmath162 is plotted in fig . [ fig : sliding_chain](b ) for different values of @xmath314 . i am indebted to peter palffy - muhoray for having sparkled my recent interest in chain dynamics and to james a. hanna for having encouraged me to pursue further the method of _ external shocks _ presented in this paper ( of course , all possible mistakes remain my own . ) whittaker et . 1937 _ a treatise on the analytical dynamics of particles and rigid bodies_. cambridge : cambridge university press , 4th edition . reissued with forward in the cambridge mathematical library series 1988 . 1921 _ theoretical mechanics : an introductory treatise on principles of dynamics_. london : cambridge university press , 3rd edition . available at https://archive.org/details/theoreticalmecha00loveiala . feynman rp , leighton rb , sands m. 2011 _ the feynman lectures on physics _ , volume 1 . new york : basic books . available at http://www.feynmanlectures.info/. see , in particular , http://www.feynmanlectures.info/exercises/falling_chain.html .

recently , there has been a great deal of activity aimed at studying the synchronization of coupled oscillators of diverse nature @xcite . the theory of synchronization implies that even in uncoupled state the individual elementary units exhibit self - sustained oscillations . however no less interesting are the systems where local coupling is essential for the very generation of oscillations and not only for their modulation or phase adjustment . as far back as in early 1970s , smale @xcite constructed a counterintuitive mathematical example of a biological cell modeled by the chemical kinetics of four metabolites , @xmath0 , such that the reaction equations @xmath1 for the set of metabolites , @xmath2 , had a globally stable equilibrium . the cell is `` dead '' , in that the concentrations of its metabolites always tend to the same fixed levels . when two such cells are coupled by linear diffusion terms of the form @xmath3 , where @xmath4 is a diagonal matrix with the elements @xmath5 , however , the resulting equations are shown to have a globally stable limit cycle . the concentrations of the metabolites begin to oscillate , and the system becomes `` alive '' . in smale s words : _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ there is a paradoxical aspect to the example . one has two dead ( mathematically dead ) cells interacting by a diffusion process which has a tendency in itself to equalize the concentrations . yet in interaction , a state continues to pulse indefinitely . _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ the reaction equations involved in smale s model were too general to appeal to any specific process . since then , inspired by his pioneer work , a number of models have been proposed containing biologically plausible mechanisms by which coupling of identical nonoscillating cells could generate synchronous oscillations . among them are a model of electrically coupled cells , characterized by an excitable membrane and calcium dynamics @xcite , a model in which coupling a passive diffusive cytoplasmic bulk with an excitable membrane ( having an activator - inhibitor dynamics ) produces a self - sustained oscillatory behavior @xcite , an analog operational - amplifier implementation of neural cells connected by passive coupling ( where conductance of the resistive connection simulates the diffusion coefficient ) @xcite , etc . the authors of the last model suggested a term awakening dynamics for the phenomenon . the subject of the present paper is an emergence of collective oscillations in a system of coupled nonoscillatory consumer - resource ( cr ) pairs . owing to simplicity , this model in a sense may be considered minimal . our choice of coupled cr equations as a matter of enquiry is dictated primarily by the ubiquity and importance of cr interactions . cr models are the fundamental building blocks used in mathematical description and simulation of ecosystems . depending on a specific nature of the involved cr interactions , they can take the forms of predator - prey , plant - herbivore , parasite - host , and victim - exploiter systems @xcite . however applications of the cr models extend far beyond the ecology and are found wherever one can speak of win - loss interactions . in its broad meaning , resource is any substance which can lead to increased growth rate of the consumer as its availability in the environment is increased . as this takes place , the resource is certainly consumed . consuming the resource means tending to reduce its availability . when carefully examined , cr models are identified in the following fields : epidemiology ( susceptible and infected ( * ? ? ? * ch . 10 ) ) , laser dynamics ( photons and electrons ( * ? ? ? 6 ) ) , labor economics ( share of labor and employment rate @xcite ) , theoretical immunology ( antigens and b lymphocytes @xcite ) , kinetics of chain chemical reactions ( lipid molecules and free radicals @xcite ) , and in numerous other studies from diverse disciplines . we consider the situation when each of two consumer species exploits one respective resource only . as explained in appendix , terms `` consumer '' and `` resource '' in our model may bear not only their literal ecological meaning , but the physical meaning of photon density and population inversion in a laser cavity as well . both resources are being supplied with constant rates like in a chemostat and consumed according to a simple mass - action kinetics . the resources are thought to be noninteractive . when uncoupled , self - inhibition of the individual consumer population is due to intraspecific interference . the coupling is assumed to originate solely from the interspecific interference competition between the consumers and quantitatively expressed by a bilinear term combining the competitor densities . the per individual loss rate of either consumer is proportional to the density of its counterpart . representation of competition between species in terms of loss - coupling dates back to the classical model of lotka volterra gause ( lvg ) @xcite . the lvg model operates with carrying capacities of the species , rather than referring explicitly to any essential resources . as shown by macarthur @xcite , lvg equations may be considered as a quasi - steady - state approximation to the cr equations accentuating resource - mediated nature of competition , under the assumption of relatively rapid dynamics of the involved resources . thereafter trophic competition have developed into a major descriptor of competition in the ecological literature generally , and in conceptualizing ecosystems as systems of coupled cr oscillators specifically @xcite . in contrast to the prevailing models of competition we consider the case of pure interspecific interference competition between the consumers with no consumption - induced contribution . actually , our model is nothing more nor less than lvg equations augmented with the rate equations for the resources . another key assumption of the model is that the dynamics of the consumers is much faster than that of the resources . from a physical perspective , by and large similar equations with the like time hierarchy ( fast consumer and slow resource ) have been in use for coupled longitudinal modes in a semiconductor laser with an intracavity - doubling crystal since the work of baer @xcite . these equations have been treated mostly numerically . the notable analytical result belongs to erneux and mandel @xcite who succeeded to show that the system admits antiphase periodic solutions by reducing it to the equations for quasi - conservative oscillator . however this result has to do with the onset of low - amplitude quasi - harmonic oscillations . unlike their study , our approach deals with well - developed high - amplitude essentially nonlinear oscillations . besides , we propose the model to be valid not only for competing laser modes , but for loss - coupled lasers as well . we analyze the model using geometric singular perturbation technique according to which the full system of equations is decomposed into fast and slow subsystems . as we shall see below , the model reveals qualitatively different behavior at intense and weak competition between the consumer species . if coupling is strong , one of the consumers wins and completely dominates . when coupling is weak , the model exhibits low - frequency antiphase relaxation oscillations with each species alternatively taking the dominant role . the two - consumer , two - resource model we consider is the following nondimensional system of four ordinary differential equations : [ coupled - uv ] @xmath6 here dots indicate differentiation with respect to the nondimensional time variable @xmath7 , @xmath8 and @xmath9 ( @xmath10 ) are quantities measuring the respective population sizes of @xmath11th resource and @xmath11th consumer , @xmath12 ( @xmath10 ) is the inflow rate of @xmath11th resource , @xmath13 is a parameter representing consumer self - limitation , @xmath14 ( @xmath15 and @xmath16 ) quantifies the inhibitory effect of @xmath17th consumer on the growth of @xmath11th consumer due to coupling , and @xmath18 is a singular perturbation parameter indicating that the dynamics of the consumers is much faster than that of the resources . it should be mentioned that being proportional to its dimensional prototype , @xmath9 directly represents population density of consumer species and is always nonnegative . quantity @xmath8 , however , is not a population size in the true sense of the word . it is rather an affine transformation of a population size of the form @xmath19 , where @xmath20 and @xmath21 are scaling constants . this is done for reasons of mathematical convenience . unlike a purely linear transformation , an affine map does not preserve the zero point , so in @xmath22 corresponds to zero population size in reality . nevertheless , from here on we shall apply the term `` resource '' to @xmath8 for brevity . for more details and discussion on the derivation of the model from different perspectives the reader is referred to appendix . when @xmath23 , the communities are uncoupled and completely independent . an isolated cr pair is governed by equations [ isolated - uv ] @xmath24 there exist two nonnegative steady states : [ steady - uv ] @xmath25 the linearization of takes the form @xmath26 where ( @xmath27 , @xmath28 ) is one of the above steady states . at , one eigenvalue is negative and one is positive : @xmath29 , @xmath30 . thus is a saddle point . at , @xmath31 and @xmath32 , so the steady state is a stable node / focus . specifically , focus is the case for @xmath33 whence one obtains an asymptotic estimate @xmath34 fig . [ fig01]c illustrates this focus - node bifurcation numerically . damped oscillations is a well - known inherent feature of the photon - carrier dynamics in class - b lasers . an isolated cr system admits therefore , only solution which tends asymptotically towards the unique steady state . periodic solutions are excluded . however the temporal dynamics of approaching this steady state essentially depends on the interplay between @xmath35 and @xmath36 , and is worth another look . there are in fact three timescales , @xmath37 , @xmath38 and @xmath39 , involved in the cr system when it is overdamped , and two@xmath37 and @xmath38when underdamped . system is singularly perturbed because the derivative of one of its state variables , @xmath40 , is multiplied by a small positive parameter @xmath35 . singular perturbation cause two - time - scale behavior of the system characterized by the presence of slow and fast transients in the system s response to external stimuli . replacing @xmath7 in with a fast time variable @xmath41 and setting @xmath42 we obtain the fast subsystem [ isolated - fast ] @xmath43 where prime means differentiation with respect to @xmath44 . in the stretched timescale @xmath44 the slow resource variable @xmath45 , according to , is replaced by its initial value and reckoned as constant parameter . equation is of a logistic type which has the solution @xmath46 valid for @xmath47 . after a lapse of considerable time ( in the fast scale ) @xmath40 converges to either of two fixed points depending on a sign of @xmath45 : @xmath48 it means that every single trajectory starting within the positive quadrant of @xmath49 plane far enough from the stable steady state will run almost parallel to the vertical axis and hit the line @xmath50 practically in a finite time of order @xmath37 . this is shown in fig . [ fig01]a b . now set @xmath42 in to get the slow subsystem [ isolated - slow ] @xmath51 the equation describes a slow manifold consisting of two lines in the @xmath49 plane : @xmath52 and @xmath53 . by , the former attracts all the trajectories in the first quadrant , while the latter all those in the second . inasmuch as the quasi - steady state of , @xmath54 , is an isolated root of and @xmath28 is a stable solution of for any @xmath55 , the assumptions of tikhonov s theorem @xcite are satisfied , and one may proceed to approximate @xmath45 and @xmath40 in terms of the solution of the reduced system in the slow timescale . it means that after arriving at the slow manifold @xmath52 , the representing point of the full system will move along the manifold toward the equilibrium point with a characteristic velocity of order @xmath38 . in the immediate proximity to equilibrium the behavior of the trajectory is determined by the type of the fixed point , whether it is a node or a focus . eventually this depends on the value of @xmath36 . namely , close to a stable node , the system has two distinct real negative eigenvalues , one fast ( @xmath56 ) , and one slow ( @xmath57 ) : @xmath58 since @xmath59 , trajectories starting off the associated eigenvector 2 ( which is tangent to the slow manifold @xmath52 ) converge to that vector along lines almost parallel to eigenvector 1 ( which is parallel to the vertical axis @xmath40 ) . as they approach vector 2 they become tangent to it and move along it up to the very nodal point ( fig . [ fig01]a ) . the characteristic time constant of this final stage is of order @xmath39 . in the vicinity of a stable focus the motion is qualitatively different : trajectories still keep converging to the equilibrium , but no longer follow the slow manifold @xmath52 ( fig . [ fig01]b ) . the reason is that for a focus eigenvectors are complex . recalling that condition must be true for a focus , calculate the eigenvalues in a limit case of @xmath60 : @xmath61 loosely speaking , one may think that near a focal point separation of state variables into slow and fast ones ceases to have its conventional meaning . there is no point to talk about motion along the slow manifold since there are no reduced one - dimensional systems corresponding to the neighborhood of a focus . the perturbation parameter @xmath35 affects only the frequency of damped oscillation , but not the damping rate . the radius of the focal spiral uniformly shrinks with a time constant of order @xmath38 . to @xmath38 for small @xmath36 there is a way to recast the equations near the focal equilibrium @xmath62 in a convenient form where the perturbation parameter @xmath35 does not multiply any right hand side . namely , performing the scaling @xmath63 , @xmath64 and @xmath65 , where @xmath66 , one obtains @xmath67 here dots stand for differentiation with respect to the time variable @xmath68 . equations represent a weakly perturbed hamiltonian system . for @xmath69 the system is pure hamiltonian and admits the first integral @xmath70 which is a conserved quantity ( @xmath71 ) . the periodic solutions of the hamiltonian system form a one - parameter family with the equilibrium @xmath72 as center point . the condition @xmath73 makes equations quasi - conservative with phase trajectories slowly spiralling to the equilibrium ( fig . [ fig02 ] ) . system can be rewritten as a second order differential equation for @xmath74 only : @xmath75 this is the equation for a nonlinear quasi - conservative oscillator . as is seen from , for @xmath76 and @xmath77 the oscillator is not only quasi - conservative , but also a quasi - linear with the frequency @xmath78 . these conditions motivate introducing the new variable @xmath79 . as a result one obtains @xmath80 this equation of quasi - conservative , quasi - linear oscillator will be used later on when analyzing the onset of synchronous periodicity . in the following , we will distinguish two main types of coupling : strong , @xmath81 , and weak , @xmath82 . physically feasible equilibria , or fixed points , @xmath83 , of are those for which @xmath84 . we denote the interior fixed point by @xmath85 , where the subscripts stand for the consumers . lack of a certain index at a boundary fixed point means that the consumer concerned is not present ( extinct ) . thus @xmath86 and @xmath87 designate either of one - consumer equilibria corresponding to dominance , while @xmath88 means both consumers having been washed out . model has four feasible steady states . to @xmath38 for small @xmath36 [ ss ] @xmath89 where @xmath90 . boundary equilibria @xmath91 , @xmath92 and @xmath93 always exist . @xmath94 exists if [ existf12 ] @xmath95 in case the expression with `` @xmath96 '' in is realized , while in case the one with `` @xmath97 '' is feasible . @xmath91 is always unstable , because two of the four associated eigenvalues are positive : @xmath98 correct to @xmath38 in @xmath35 , the eigenvalues for @xmath92 and @xmath93 are [ lambdaf1f2 ] @xmath99 whence it follows that and are stable when [ stabilityf1f2 ] @xmath100 respectively . the necessary and sufficient conditions for all the eigenvalues of the jacobian matrix , @xmath101 evaluated at @xmath94 , to have negative real parts are , from the routh hurwitz criterion , [ routh ] @xmath102 where @xmath103 , @xmath104 , @xmath105 and @xmath106 are the coefficients of the characteristic polynomial @xmath107 of : @xmath108 to analyze the validity of we put @xmath109 , where @xmath110 . then for to be true , @xmath111 should be met , which is incompatible with strong coupling , yet possible for weak coupling . conditions and are always valid . as is known , guarantees a simple complex conjugate pair of eigenvalues corresponding to a linearization about steady state @xmath94 to have negative real part . for @xmath111 it boils down to @xmath112 whence it follows that @xmath113 with regard to a fairly small value of @xmath35 , may be thought to be broken under most physically meaningful conditions unless coupling is infinitesimally weak . hence , normally , condition of the routh hurwitz criterion is never fulfilled and the interior fixed point @xmath94if it exists is always unstable by growing oscillations . existence and stability conditions of possible nonnegative equilibrium points are summarized in table [ tab1 ] . .existence and stability conditions of nonnegative equilibria of system . [ cols="<,<,<",options="header " , ] [ tab1 ] strong coupling in makes possible bistability of boundary equilibria , as evident from table [ tab1 ] . when both @xmath92 and @xmath93 are stable with an unstable coexistence steady state @xmath94 , the system being studied is able to exhibit a hysteresis effect . given the strong coupling , suppose , that the inflow of resource 2 is kept at some constant level , @xmath114 , while the inflow of resource 1 , @xmath115 , steadily increases from a value less than @xmath116 along the line @xmath114 in the @xmath117 parameter plane , as shown in fig . [ fig03]a . referring to and , one sees that initially @xmath92 is unstable , @xmath93 is stable and @xmath94 does not exist with the complete dominance of consumer 2 . within the interval @xmath118 steady state @xmath92 becomes stable yet empty and @xmath94 becomes existent ( according to ) yet unstable , so the situation remains unchanged until point @xmath119 in fig . [ fig03]a has been reached from the left . for a larger @xmath115 , state @xmath93 gives up its stability and the system jumps to @xmath92 . consumer 2 gets washed out , while consumer 1 takes over . if now we start reducing @xmath115 , the system remains in @xmath92 until @xmath115 drops to the lower critical value @xmath116 , beyond which @xmath92 is no longer stable and there is a reverse jump to @xmath93 . in other words , as @xmath115 progresses along @xmath114 there is a discontinuous switch from consumer 2 to consumer 1 at @xmath120 , while as @xmath115 retraces its steps , there is a discontinuous switch from consumer 1 to consumer 2 at @xmath116 . [ fig03]b illustrates how the steady state level of consumer 1 responds to infinitesimally slow changes of the inflow rate of its own resource . the hysteresis is made possible thanks to the concurrent stability of both boundary equilibria and instability of the interior fixed point for @xmath121 . in terms of electronics , such a situation would describe a flip - flop circuit bistable multivibrator having two stable conditions , each corresponding to one of two alternative input signals . as seen from table [ tab1 ] , the very existence of interior equilibrium @xmath94 in a case of weak coupling ( condition ) implies instability of both boundary fixed points , @xmath92 and @xmath93 . system happens to possess four nonnegative steady states , none of them being stable . as we have found , @xmath94 is unstable through growing oscillations . in such a case , the model would thus be expected to have a limit cycle in its four - dimensional phase space corresponding to self - sustained oscillations . as is known , hopf bifurcations come in both super- and subcritical types . if a small , attracting limit cycle appears immediately after the fixed point goes unstable , and if its amplitude shrinks back to zero as the parameter is reversed , the bifurcation is supercritical ; otherwise , it s probably subcritical , in which case the nearest attractor might be far from the fixed point , and the system may exhibit hysteresis as the parameter is reversed . to check whether the bifurcation is supercritical or subcritical , we employ the averaging method @xcite . for negligible @xmath36 the problem reduces to weakly coupled quasi - conservative oscillators . introduce the new time @xmath122 , where @xmath123 . the new dynamic variables @xmath124 and @xmath125 are defined by formulas @xmath126 and @xmath127 . thus the variables @xmath124 and @xmath125 are the respective deviations of @xmath128 and @xmath129 from their standalone steady state values . with these new variables equations become [ hamilton4 ] @xmath130 where dots mean differentiation with respect to @xmath68 and @xmath131 are the scaled coupling strengths . for brevity , we restrict our consideration to the case of identical cr pairs and symmetric coupling setting @xmath132 and @xmath133 . with these assumptions system can be transformed to two coupled second - order equations for quasi - conservative , quasi - linear oscillators . this can be done as follows . first , and are solved for @xmath134 and @xmath135 . secondly , and are differentiated with respect to time to get @xmath136 and @xmath137 . thirdly , @xmath134 , @xmath135 , @xmath138 and @xmath139 are plugged into the equations for @xmath136 and @xmath137 . and finally , @xmath125 are replaced by the new variables @xmath140 : @xmath141 . as a consequence , we arrive at the following coupled equations : @xmath142 where @xmath143 we seek the solution of in the quasi - harmonic form @xmath144 where @xmath145 and @xmath146 are slowly varying amplitudes and phases , and @xmath147 is the frequency of the synchronous oscillations . differentiating the assumed form of @xmath140 and equating the result to the assumed form of @xmath148 yields the first pair of relationships between @xmath145 and @xmath146 : @xmath149 then , differentiating @xmath148 and substituting the resulting expression for @xmath150 as well as the assumed forms for @xmath140 and @xmath148 into yields the second pair of equations relating @xmath145 and @xmath146 . separating into equations for the rate of change of @xmath145 and @xmath146 one obtains @xmath151 where @xmath152 to this point no approximations have been made except of the expansion in powers of @xmath153 in . averaging equations over the period @xmath154 and considering @xmath145 , @xmath146 , @xmath155 and @xmath156 to be constants while performing the averaging , one obtains the following equations describing the slow variations of @xmath145 and @xmath146 : [ aver1 ] @xmath157 where @xmath158 . any stable fixed points of could mean that the phase difference between the coupled oscillators do not change in time ( @xmath159 ) , and the oscillations are periodic with constant amplitudes @xmath145 . thus , finding the conditions when these fixed points are stable would mean finding the conditions at which the synchronization occurs . equations show that the first approximation of the averaging method does not predict any nonzero steady - state amplitudes . the reason for this seems to lie in the fact that the even terms in do not contribute to the value of integral . to take account of them we have to employ in what follows the second approximation of the averaging method . as to the phase difference , @xmath160 , we see that stable is only antiphase steady - state regime with @xmath161 . let @xmath162 be the vector of the right - hand sides of . also denote the operation of averaging by the angle brackets @xmath163 . then @xmath164 will mean the right - hand sides of . following the conventional procedure we retain only vibrational terms in @xmath162 , @xmath165 and integrate @xmath166 up to an arbitrary function of the amplitudes chosen for simplicity to be equal to zero : @xmath167 now , upon calculation of the jacobian matrix @xmath168 , we can write down the second approximation symbolically as @xmath169 ( the terms now neglected by averaging are of a higher order of magnitude with respect to the small parameters than the terms neglected in the first approximation . ) due to the symmetry of the case we may set @xmath170 enabling us to present the resulting equations of the second approximation as compact as possible : [ aver2 ] @xmath171 the stable steady - state solution for the phase difference is @xmath161 . inserting this value into and putting @xmath172 we get three different steady - state solutions for @xmath20 ( including the trivial one ) , of which only one , @xmath173 being stable . formula indicates a soft transition to sustained oscillations as the coupling parameter @xmath174 is progressively increased from a critical value . thus we expect that the hopf bifurcation is supercritical . [ fig04 ] plots @xmath175 phase planes for coupling strength below and above the hopf bifurcation . when @xmath176 ( consumers are decoupled ) the internal equilibrium is a stable focus ( fig . [ fig04]a ) . for @xmath177 ( very weak coupling ; just below the bifurcation ) @xmath94 is still a stable focus , though a very gently winding one : the decay is slow ( fig . [ fig04]b ) . for @xmath178 ( very weak coupling ; just above the bifurcation ) there is an unstable focus at @xmath94 and a stable oval limit cycle of small size representing low - amplitude quasi - harmonic oscillations ( fig . [ fig04]c ) . on further increasing of coupling strength the limit cycle continuously grows in size and takes an irregular shape indicating nonlinearity of the synchronous oscillations ( fig . [ fig04]d f ) . as a practical matter , the range of very weak coupling not too far away from the hopf bifurcation , where oscillations are quasi - linear and quasi - harmonic , is of less concern to us than is the range of far more feasible not - too - weak coupling , corresponding to well - developed substantially nonlinear oscillations . we are going to demonstrate that given conditions , system exhibits relaxation oscillatory behavior , with the two coupled cr pairs being antiphase locked . by the assumption , @xmath179 , meaning that system is singularly perturbed . the slow variables are resources , @xmath180 and @xmath181 , and the fast variables are consumers , @xmath182 and @xmath183 . the standard practice of reducing such systems is the adiabatic elimination of the fast variables , when the left - hand side in the fast equation is replaced by zero , thus turning this differential equation into an algebraic equation . it is assumed that the fast variables quickly relax to their momentary equilibrium , quasi - steady - state , values obtained from the algebraic equations , in which the slow variables are treated as parameters . `` frozen '' slow variables do not move substantially in this short adaptation time of the fast variables . quasi - steady - state values of the fast variables can then be expressed by values of the slow variables . the fast variables hastily adapt to the motion of the slow variables . the former are entrained by the latter . the utility of quasi - steady - state approximation is that it allows us to reduce the dimension of the system by retaining only slow variables in the model . one has to establish the validity of the adiabatic elimination in each specific case by using the recommendations of the singular perturbation theory @xcite . in particular , tikhonov s theorem @xcite requires the quasi - steady state of the fast equations to be stable . to decompose the full system into fast and slow subsystems , introduce fast time variable @xmath41 . now rescale by replacing @xmath7 with @xmath184 and , after taking @xmath42 , it becomes [ fast ] @xmath185 where prime means differentiation with respect to @xmath44 . this is the fast subsystem , where @xmath180 and @xmath181 are replaced by their initial values and treated as parameters . it yields inner solution , valid for @xmath47 . setting @xmath42 in leads to the slow subsystem [ slow ] @xmath186 which produces outer solution , valid for @xmath187 . in this singular limit as @xmath188 , the subsystem defines a slow flow on the surface ( slow manifold ) given by . outer solution is valid for those @xmath180 and @xmath181 , for which the quasi - steady states of the fast subsystem are stable . we anticipate the dynamics of the full system in its four - dimensional phase space @xmath189 to consist of two typical motions : quickly approaching the slow manifold , and slowly sliding over it until a leave point ( where the solution disappears ) is reached . after that , the representing point may possibly jump to another local solution of . thus , we ought to find all quasi - steady states of the fast subsystem , map the domains of their stability onto the slow phase plane @xmath190 , and then investigate the dynamics of the slow subsystem with piecewise continuous functions . fast subsystem , which is nothing but the conventional lvg model , has four quasi - steady states three boundary and one interior denoted by @xmath191 ( the slow variables are deemed to be frozen ) : [ qss ] @xmath192 existence and stability of these quasi - steady states is determined by @xmath190position of a representing point in the phase plane of the slow subsystem , shown in fig . [ fig05]a . @xmath191 , @xmath193 and @xmath194 always exist for all @xmath180 and @xmath181 from the positive quadrant of the slow phase plane . for not - too - weak coupling , such that @xmath195 , @xmath196 exists for all @xmath180 and @xmath181 satisfying the condition @xmath197 , i.e. within the opening of the angle formed by lines @xmath198 and @xmath199 in fig . [ fig05]a . the opening shrinks as the coupling strengths get weaker . jacobian matrix of the fast subsystem , @xmath200 has the following sets of eigenvalues at : [ eigqss ] @xmath201 based on one concludes that @xmath191 ( the origin ) is always an unstable node for all @xmath180 and @xmath181 from the positive quadrant of the slow phase plane . @xmath193 is a stable node for @xmath202 , i.e. below the line @xmath203 in fig . [ fig05]a , otherwise it is a saddle . similarly , @xmath194 is a stable node for @xmath204 , i.e. above the line @xmath205 in the plane of slow variables , otherwise it is a saddle . the interior quasi - steady state , @xmath196 , is always a saddle . the performed typology of fixed points of the fast subsystem leads to three qualitatively different phase portraits depicted by fig . [ fig05]b d . suppose initially @xmath193 is stable and @xmath194 is not . consumer 1 completely dominates . this corresponds to slow variables @xmath180 and @xmath181 being somewhere below the line @xmath205 of fig . [ fig05]a . fast subsystem has phase portrait of a type shown in fig . [ fig05]b . while @xmath180 remains much greater than @xmath206 , the dynamics of the resources ( treated as bifurcation parameters in reference to the consumers ) is described by a system of two independent equations [ slow - piecewise1 ] @xmath207 which is a piecewise version of the slow subsystem for @xmath208 and @xmath209 . system has stable steady state @xmath210 where @xmath211 . this equilibrium lies in the upper left corner of fig . [ fig05]a . while heading to , the trajectory crosses the line @xmath205 and enters the domain of bistability of both @xmath193 and @xmath194 . fast subsystem takes new phase portrait of a type presented by fig . [ fig05]c . however the dominance of consumer 1 persists . upon introducing the deviations @xmath134 and @xmath212 from the respective steady states , the system becomes [ slow - piecewise1-xi - eta ] @xmath213 whence one finds @xmath214 it follows from , that the dynamics of variable @xmath180 is faster than that of @xmath181 due to small @xmath36 . clearly , the representing point must have relaxed to the vertical line @xmath215 well before approaching the horizontal line @xmath216 . further developments depend on whether the involved individual cr systems are overdamped or underdamped . if @xmath36 is great enough to damp intrinsic oscillations in the constituent cr pairs , the representing point will slide along the nullcline @xmath217 ( slow manifold of the system ) , steadily tending to point ( fig . [ fig06]a ) . if @xmath36 is small in terms of the condition , then the individual cr pairs are underdamped . in the immediate vicinity of @xmath218 the division of variables on slow and fast ones loses its meaning and the reduced equation is no longer valid . instead of one should write down two equations : [ constituent - u1v1 ] @xmath219 however this system is identical to equations for an uncoupled cr system and , in essence , describes convergence to a focal point via dying oscillations . in the plane of slow variables @xmath190 these oscillations manifest themselves in damped transverse fluctuations superimposed on the independent vertical motion along the nullcline @xmath217 ( @xmath220 ) toward point ( fig . [ fig06]b ) . by virtue of the condition , the trajectory has to cross the line @xmath199 on its way toward the neighborhood of steady state . as soon as this has happened , node @xmath193 in the plane @xmath221 will be absorbed by saddle @xmath196 . a new phase portrait of the fast subsystem takes on the appearance of fig . [ fig05]d . consumer 1 rapidly washes out , and the alternative boundary quasi - steady state @xmath194 becomes stable , with consumer 2 dominating . in terms of the four - dimensional phase space of full system , the representing point is now in the other stable branch of the slow manifold . the motion over this alternative branch obeys the piecewise subsystem [ slow - piecewise2 ] @xmath222 with the initial conditions @xmath223 and @xmath224 . the dynamics of is basically similar to that of analyzed above . system has a stable steady state @xmath225 where @xmath226 . this equilibrium lies in the lower right corner of fig . [ fig05]a . variable @xmath181 , being more rapid in comparison to @xmath180 , quickly enters the neighborhood of the nullcline @xmath227 given by @xmath228 , and then depending on the value of @xmath36finally approaches the nullcline either monotonically ( fig . [ fig06]a ) or via damped oscillations according to equations [ constituent - u2v2 ] @xmath229 ( fig . [ fig06]b ) . system describes underdamped intrinsic oscillations of uncoupled consumer 2 for small @xmath36 . at the same time , @xmath180 steadily and independently tends to @xmath230 . again , because point is located below the line @xmath231 ( on the strengths of the condition ) , the trajectory would certainly cross that line at a point @xmath232 , whereupon node @xmath194 would be absorbed by saddle @xmath196 . the system returns to the first branch of the slow manifold , and thereby the oscillatory cycle gets closed . fig . [ fig07 ] shows the results of a numerical integration of system for the case of underdamped individual consumer - research pairs . the two coupled communities execute self - sustained relaxation oscillations which are antiphase - locked . the resources @xmath180 and @xmath181 demonstrate sawtooth periodic pulses . the oscillation range for the resource levels remains finite and , what is important , it does not depend on the intrinsic second order loss @xmath36 ( measuring intraspecific interference ) . the times of motion over either branch of the slow manifold and add up to give a predominant contribution to the period of oscillations , @xmath233 . these times are determined mainly by the dynamics of the slow resource variables @xmath180 and @xmath181 and , to a zeroth approximation in @xmath35 and @xmath36 , can be found as solutions to the equations of motion and with respective boundary conditions @xmath234 and @xmath235 . in this way one obtains a quite simple estimate for the period : @xmath236 it is interesting that , according to , the period depends on the ratio of the two resource inflows , @xmath115 and @xmath237 , rather than on each of them individually , and does not depend completely on concrete value of @xmath36 . the consumers @xmath182 and @xmath183 change periodically between extinction and respective constant levels @xmath115 and @xmath237 . very brief transient from zero to flat nonzero level within each cycle is accompanied by a highly pronounced spiky overshoot . the magnitude of the spike tends to infinity as @xmath60 , in view of and . depending on the intensity of intraspecific interference , the overshoot may or may not be followed by a tail of fading high - frequency oscillations , when a consumer variable falls below its steady - state value and then bounce back above , taking some time to settle close to its steady - state value . in signal processing , such a kind of transient oscillations is known as `` ringing '' . there is no ringing if the involved cr pair does not oscillate due to significant intraspecific interference . ringing takes place if intraspecific interference is negligible and therefore the involved cr pair is characterized by underdamped intrinsic oscillations . the `` pitch '' of ringing is just the frequency of these intrinsic oscillations . one notices that when one consumer is very scanty , the coupled system behaves like an isolated cr pair . another essential feature of the dynamics is the role of the resource variables in determining when the consumers emerge and wash out . when @xmath180 , for example , rises above a threshold value ( determined by the amount of losses experienced by @xmath182 ) then @xmath182 comes into dominance causing @xmath183 in turn to disappear . so except for their transient spiking and ringing , the consumer levels , either flat nonzero or essentially zero , are determined by the hysteretic cycling of the respective resources . scrutinizing a cycle of consumer oscillations one may distinguish four parts within it : @xmath182 is essentially zero , while @xmath183 is approximately equal to its uncoupled steady - state value , @xmath237 . @xmath180 increases due to its inflow until it overcomes losses for consumer 1 ; with a sufficient resource stock , @xmath182 now emerges . the population exhibits a spike due to the fast time scale of the consumer equations . the sharp increase in population saturates the available resource level , so @xmath180 drops . cross - losses cause @xmath183 to wash out ; @xmath182 and @xmath180 relax to quasi - steady - state values , as if there were only one isolated cr pair . @xmath183 is essentially zero . @xmath181 is increasing , like @xmath180 did in part 1 ; @xmath181 surpasses the losses , @xmath183 emerges and the subsequent cross - losses cause @xmath182 to wash out . the spiking @xmath183 also causes a substantial decrease in the available stock of the associated resource . the sequence begins again . presenting his famous model smale remarked that `` it is more difficult to reduce the number of chemicals to two or even three '' @xcite . as distinct from smale s example , coupling in our case makes self - sustained synchronous oscillations possible for just two variables . as we have seen , phase trajectory of the system constantly moves from the neighborhood of unstable boundary equilibrium @xmath92 where only consumer 1 is present , to the neighborhood of @xmath93 where consumer 2 completely dominates , back to @xmath92 , and so on in cyclic alternation . this kind of trajectory was termed `` heteroclinic cycle '' by kirlinger @xcite . a heteroclinic cycle occurs when the outflow ( unstable manifold ) from one saddle point is directly connected to the inflow ( stable manifold ) of another saddle point , and vice versa . it is closely related to another notion of the nonlinear dynamics , a homoclinic cycle , which emerges when the unstable and the stable manifolds of the same saddle coincide and form a closed loop . homo- and heteroclinic cycles are not robust structures in the sense that infinitesimally small change of system parameters destroy them . however in the practical sense , any limit cycle passing in close proximity to saddle points will be indistinguishable from a heteroclinic cycle ( fig . [ fig08 ] ) . the only difference is strict periodicity , although the period of the limit cycle in a neighborhood of the heteroclinic cycle may be long . besides , at the threshold of homo-/heteroclinic bifurcation the period is susceptible to external noise . in the context of our model , as coupling becomes stronger , the stable limit cycle swells and passes closer and closer to boundary fixed points which are node - saddles or focus - saddles ( fig . [ fig04]d f ) . depending on the interplay between the parameters , eventually it may bang into one or both of these equilibria creating either a homoclinic or heteroclinic cycle , respectively . this corresponds to @xmath238 and @xmath239 . on further increasing the coupling , the saddle connection breaks and the loop is destroyed . it is worth noting that heteroclinic cycles were first found by may and leonard @xcite in a classical lvg system with competing three species . however in their model the cycle is not truly periodic : as time goes on , the system tends to stay in the neighborhood of any one boundary equilibrium ever longer , so that the `` total time spent in completing one cycle is likewise proportional to the length of time the system has been running . '' moreover , may and leonard state that `` the phenomenon clearly requires at least three competitors , which is why it can not occur in models with two competitors . '' this statement is echoed by vandermeer @xcite who extended their theory on higher dimensions : `` it appears to be the case that all cases of an odd number of species follow this basic pattern , whereas all cases of even number of species result in extinction of half of the components , leaving the other half living independently at their carrying capacities . '' in view of our results , the above conclusion is by far and away true providing one stays within the framework of classical lvg equations , which in fact imply a high rapidity of the resource dynamics . in our model of just two competitors the slowness of the resource relative to the consumer is essential for the oscillations to occur , because it provides the necessary inertia to the system . physically , our model is most likely feasible because it is based on the well - established rate equations for semiconductor lasers , and therefore should be considered as a model of anti - phase synchronization of two lasers via their loss - coupling . ecologically , the feasibility of the model is tightly bound to justification of the adopted time hierarchy in system . time scales are usually inverted in ecosystems , as opposed to lasers , the most common case being rapid consumption of food by species . however it seems reasonable to propose that our model may describe the first level of an ecosystem , at which the consumers are autotrophs and the resources are mineral nutrients . the ability to exploit different substrates leads to a possibility of stable coexistence of different organisms descending from the common ancestor . divergent evolution is just the formation of new species : due to mutations two populations emerge with the same genetic code but having proteins able to process different substrates . providing the environmental conditions are quite stable on the evolutionary timescale , the inflows of inorganic substrates from the surroundings may be considered constant and the washout time of a substrate may occur much longer than the life expectancy of a species ( recall the definition of @xmath35 from ) . we proposed a model of two cr pairs linked by interspecific interference competition . when uncoupled , an individual cr pair has a unique stable steady state and does not admit periodic solutions . if intraspecific interference within the species is strong enough , the equilibrium is nonoscillatory ( stable node ) , otherwise the steadying occurs by decaying oscillations ( stable focus ) . when coupled , the model behaves differently at strong and weak competitive interaction between the consumers . when coupling is strong , one of the consumers wins . which consumer wins or loses depends critically on the relative intensities of the resource inflows and coupling strengths . in the case of bistability , when the system acts like a bistable multivibrator ( flip - flop circuit ) , the winner may be determined by the initial conditions . any static coexistence of competing consumers is not possible . when coupling is moderately weak , the model reveals low - frequency antiphase relaxation oscillations represented by a continuous flow of rectangular pulses . the system works as an astable multivibrator continually switching between its two states , neither of which is stable . the consumers can not coexist even dynamically : in each of two alternating states one consumer completely dominates and the other is on the verge of extinction . the most intriguing feature of the model is that each of the participating cr pairs taken separately does not oscillate ; both communities are completely quiescent , however , in interaction , when coupled in a nonlinear way , the resulting system turns into a relaxation oscillator . one way or the other , it is believed that the proposed model for coupling - induced oscillations in nonoscillatory cr pairs can be considered as a minimal in that class of population - dynamical systems and its mechanism can be applied to networks with large numbers of nonoscillatory elements and complex architecture . the author declares that there is no conflict of interests regarding the publication of this paper . of all types of interactions between individuals of the same population ( intraspecific interactions ) or individuals of different populations ( interspecific interactions ) of the same trophic level competition is most commonly encountered . in a broad sense , competition takes place when each species ( individual ) has an inhibiting effect on the growth of the other species ( individual ) . an inhibiting effect should be understood to mean either an increase in the death rate or a decrease in the birth rate . here dots indicate differentiation with respect to time @xmath7 , @xmath241 represents the total biomass of @xmath17th resource ( prey ) , @xmath242 stands for the total biomass of @xmath11th consumer ( predator ) species , the constant @xmath243 defines the growth rate of @xmath17th resource , @xmath244 is the carrying capacity of @xmath17th resource , @xmath245 is the rate of uptake of a unit of @xmath17th resource by each individual of @xmath11th consumer population , @xmath246 is the conversion efficiency parameter representing an amount of @xmath17th resource an individual of @xmath11th consumer population must consume in order to produce a single new individual of that species , @xmath247 is the loss rate of @xmath11th consumer due to either natural death or emigration . all parameters in are nonnegative . macarthur assumed population dynamics of the resources to be much faster than that of the consumers which enabled him to approximate @xmath241 in by its quasi - steady - state value derived by setting the right - hand side of to zero . as a result , he succeeded to reduce slow - scale equation to the well - known lvg model @xcite more recently , such an asymptotic reduction has also been carried out for a model of competition where species ( with continuous trait ) consume the common resource that is constantly supplied , under the assumption of a very fast dynamics for the supply of the resource and a fast dynamics for death and uptake rates @xcite . cr model assumes that competition within and between consumer species is purely exploitative : individuals and populations interact through utilizing ( or occupying ) common resource that is in short supply . quite on the contrary , lvg model describes competition strictly phenomenologically , as direct interference where consumers experience harm attributed to their mutual presence in a habitat ( e.g. through aggressive behavior ) . however we have to stress that neither macarthur s reduction claims that interference competition entirely results from `` more fundamental '' trophic competition , nor it urges us to hastily consider direct competition as some derived concept . what it actually states is that at slow - time scale associated with dynamics of the consumers , the effects of exploitation competition are indistinguishable from those of interference competition . and at slow - time scale , coefficients @xmath249 of merely add to interference coefficients @xmath250 which are to be present primordially in . most mathematical models dealing with coupled cr pairs or multilevel trophic chains ignore contributions of intraspecific and interspecific interference . indeed , the empirical data like @xcite do indicate that @xmath251 may be negligible in comparison with @xmath249 . still , works advocating the explicit account for direct interference show that incorporation of self - limitation and cross - limitation terms in the equations at the consumers level can provide for the stable coexistence of many species on few resources @xcite , @xcite . instead of the logistic mode of resource supply , as is the case in macarthur s model , our model is based on so - called `` equable '' mode of resource exploitation @xcite , by which the quantities of available resources are held constant by a continuous - flow system . according to and , a constant concentration of @xmath17th resource ( @xmath15 ) flows into a defined volume with the rate @xmath253 while unused resource flows out with the per capita rate @xmath254 , in much the same manner as in a chemostat @xcite . in more exact terms , the true chemostat model for one substrate and one species looks as follows : @xmath255 where the rate of substrate uptake is expressed by the monod formula @xmath256 , @xmath257 is a saturation constant numerically equal to the substrate concentration at which the uptake rate is half the maximum , @xmath258 is the dilution rate defined as the rate of flow of medium over the volume of the bioreactor , and @xmath259 is an input concentration of the substrate . besides , in and intraspecific competition strength @xmath265 ( @xmath10 ) measures direct interference of individuals within @xmath11th consumer population with each other resulting in an additional per capita loss rate @xmath266 ; interspecific competition strength @xmath267 ( @xmath268 ; @xmath269 ) quantifies direct interference effect from @xmath68th consumer on @xmath11th consumer resulting in an additional per capita loss rate , @xmath270 , of the latter . equations contain two important assumptions . first , they assume that the resources are noninteractive . on higher trophic levels , however , resources may interact and the possibility of competition among the resources was originally pointed out by lynch @xcite . since then , a whole series of theoretical papers have been published on two - predator , two - prey systems with interference competition between two self - reproducting prey species based on the lotka volterra equations . specifically , kirlinger @xcite describes the model in which each predator specializes on one prey only , while xiang and song @xcite treat a similar model in which each predator is allowed to feed on both prey . a second assumption of our equations is that the consumers interact only directly , through interference competition , and can not compete through their use of resources , as each consumer specializes on one resource only . the theory of pure trophic competition in equable models has been developed in the works @xcite . intraspecific interference competition is allowed within the consumers as well . even though the available amount of any resource happened to be of a constant level , the population size of the associated consumer would remain finite due to self - limitation caused by direct intraspecific interference . the novelty of model is that it considers time hierarchy of macarthur s cr equations to be reversed by assuming dynamics of the consumers to be much faster than that of the involved resources and articulates the importance of direct competition mechanisms within the framework of this assumption . the parameters @xmath274 , @xmath275 and @xmath276 reflect the rapidity of the dynamics of @xmath181 , @xmath182 and @xmath183 with reference to that of @xmath180 it is assumed that @xmath277 and @xmath278 . in studying the effect of coupling , the parameters of interest are obviously the coupling strengths , @xmath279 and @xmath280 . the parameters of interest are also those which characterize the difference between the states of the uncoupled systems . the resource income rates @xmath115 and @xmath237 are used as the control parameters that distinguish the relative base states of the two systems . laser rate equations originally proposed by statz and demars @xcite are differential equations that relate two quantities : injected carrier density ( @xmath284 ) and photon density ( @xmath285 ) . for a single - mode semiconductor laser , these equations take the form ( * ? ? ? 6 ) where dots mean differentiation with respect to time @xmath7 , @xmath287 is the injection current density ( pump parameter ) , @xmath288 is the magnitude of the electron charge , @xmath289 is the active - layer thickness , @xmath290 is the confinement factor accounting for the fraction of the light power contained in the active region , @xmath291 is the group velocity of light that can be expressed through the speed of light in vacuum ( @xmath292 ) and the group refractive index of the dispersive semiconductor material ( @xmath293 ) as @xmath294 , @xmath20 is the gain coefficient , @xmath295 is the carrier density at transparency corresponding to the onset of population inversion , @xmath296 is the lifetime of the electrons in the conduction band before being lost by escape from the active region , @xmath297 is the lifetime of photons inside the cavity before going out of the cavity or being absorbed inside the cavity . in the contribution of spontaneous emission is neglected . typical parameter values for a semiconductor laser ( mostly borrowed from @xcite ) are given in table [ taba1 ] . these numerical values are used in the calculations throughout the present paper , unless otherwise noted . lll + ( r)1 - 2 notation & meaning & value + @xmath287 ( @xmath298 ) & pump current density & @xmath299 ( 1.19375 ) + @xmath289 & active layer thickness & @xmath300 + @xmath290 & confinement factor & @xmath301 + @xmath293 & group refraction index & @xmath302 + @xmath20 & differential gain coefficient & @xmath303 + @xmath295 & carrier density at transparency & @xmath304 + @xmath296 & carrier lifetime & @xmath305 + @xmath297 & photon loss time & @xmath306 + @xmath35 & @xmath307 & @xmath308 + @xmath258 ( @xmath36 ) & intrinsic second order loss & @xmath309 ( 0.01 ) + consider two ( not necessarily identical ) lasers of type and introduce additional intensity - dependent losses such that each laser , @xmath11 , of the two has a total loss represented by the sum of the constant loss , @xmath310 , plus the loss proportional to its own intensity , @xmath311 , plus the loss proportional to the intensity of the other laser , @xmath312 : where @xmath314 ( @xmath10 ) is the second order loss constant of the isolated @xmath11th laser and @xmath315 is the coupling strength measuring the cross - loss effect of @xmath11th laser on @xmath17th laser . thus according to , two lasers happen to be cross - coupled through their resonators , so that each of them can modulate the cavity loss of the other . technically , intensity - dependent intrinsic and cross - losses may be implemented , for example , using an intracavity electro - optic modulator fed by a current proportional to the output power @xcite . we will consider the pump currents @xmath316 and @xmath317 , and the coupling strengths @xmath318 and @xmath319 , as free parameters of the model . for the present , we can not judge with any confidence the numerical value of @xmath320 , however , as it is demonstrated in the main body of the paper , the exact value of the intrinsic second order loss is not all that critical and does not affect principal results of our analysis , providing that this parameter is small in a sense . for the purposes of model calculations , we adopt @xmath320 to be somewhat less than @xmath321 . presumably , hofelich - abate and hofelich @xcite were first to introduce ( in general form ) an intensity - dependent self - limitation term in the photon - density rate equation . later , that has been done in an explicit form of the second order loss @xcite . those and subsequent studies @xcite showed efficient damping of relaxation oscillations in the presence of intensity - dependent losses . as to the formal analysis of laser coupling via intensity - loss cross - modulation , very few attempts have been done so far to consider such a mechanism e.g . @xcite and @xcite . the former work , however , does not treat intrinsic second order losses . the works which address loss - coupled modes of a single laser rather than loss - coupled single - mode lasers are much more plentiful and varied , and we refer the reader for the reviews in ( * ? ? ? * ch . 8) and 12 ) . in his seminal paper @xcite , baer studied multimode regime of nd : yag ( neodymium - doped yttrium aluminum garnet ) laser with the intracavity - doubling ktp ( potassium titynal phosphate ) crystal both experimentally and numerically , and proposed the following rate equations for two coupled longitudinal modes : where @xmath323 and @xmath324 ( @xmath10 ) are the respective gain and intensity of @xmath11th mode , @xmath325 is the fluorescence lifetime , @xmath326 is the small - signal gain ( pump parameter ) for @xmath11th mode , @xmath327 is the saturation parameter which determines how strongly the intensity depletes the available gain , @xmath328 is the cross - saturation parameter for modes 1 and 2 , @xmath296 is the cavity round - trip time , @xmath329 is the loss of @xmath11th mode , @xmath35 is the nonlinear coupling coefficient , which models the intracavity - doubling crystal as an intensity - dependent loss in the laser resonator . the last two terms in and represent second order losses that are due to intracavity second - harmonic generation and sum - frequency generation , respectively . numerical calculations revealed antiphase synchronous oscillations in . in antiphase state , either mode has precisely the same time profile being shifted by @xmath330 of a period from its counterpart . this type of dynamics was later observed in a multimode @xmath331:yag laser with intracavity doubling crystal @xcite . for @xmath284 antiphase oscillators , the phase shift between any nearest neighbors is @xmath332 of a period . the scaling @xmath334 turns into nondimensional form . since the lasers are made of the same material , we can put @xmath335 , @xmath336 , @xmath337 , @xmath338 , @xmath339 , @xmath340 , @xmath341 , and @xmath342 , whence @xmath281 , @xmath282 and @xmath283 , and we arrive , dropping the prime at @xmath7 , at the set of equations . s. smale , `` a mathematical model of two cells via turing s equation , '' in _ some mathematical questions in biology v _ ( j. d. cowan , ed . ) , vol . 6 of _ lectures on mathematics in the life sciences _ , pp . 1526 , providence , ri : american mathematical society , 1974 . y. loewenstein , y. yarom , and h. sompolinsky , `` the generation of oscillations in networks of electrically coupled cells , '' _ proceedings of the national academy of sciences of the united states of america _ , vol . 98 , no . 14 , pp . 80958100 , 2001 . i. szatmri and l. o. chua , `` awakening dynamics via passive coupling and synchronization mechanism in oscillatory cellular neural / nonlinear networks , '' _ international journal of circuit theory and applications _ , vol . 36 , no . 5 - 6 , pp . 525553 , 2008 . d. s. chernavskii , e. k. palamarchuk , a. a. polezhaev , g. i. solyanik , and e. b. burlakova , `` a mathematical model of periodic processes in membranes ( with application to cell cycle regulation ) , '' _ biosystems _ , vol . 9 , no . 4 , pp . 187193 , 1977 . t. baer , `` large - amplitude fluctuations due to longitudinal mode coupling in diode - pumped intracavity - doubled nd : yag lasers , '' _ journal of the optical society of america b : optical physics _ , vol . 3 , no . 9 , pp . 11751180 , 1986 . m. devetter and j. sed , `` the relative role of interference competition in regulation of a rotifer community during spring development in a eutrophic reservoir , '' _ international review of hydrobiology _ , vol . 93 , no . 1 , pp . 3143 , 2008 . f. m. stewart and b. r. levin , `` partitioning of resources and the outcome of interspecific competition : a model and some general considerations , '' _ american naturalist _ 107 , no . 954 , pp . 171198 , 1973 . a. ghosh , b. k. goswami , and r. vijaya , `` nonlinear resonance phenomena of a doped fibre laser under cavity - loss modulation : experimental demonstrations , '' _ pramana - journal of physics _ , vol . 75 , no . 5 , pp . 915921 , 2010 . j. tian , y. yao , j. j. xiao , x. yu , and d. chen , `` tunable multiwavelength erbium - doped fiber laser based on intensity - dependent loss and intra - cavity loss modulation , '' _ optics communications _ , 285 , no . 9 , pp . 24262429 , 2012 . a. kellou , k. at - ameur , and f. sanchez , `` suppression of undamped oscillations in erbium - doped fibre lasers with intensity - dependent losses , '' _ journal of modern optics _ , 45 , no . 9 , pp . 19511956 , 1998 . f. li , j. n. kutz , and p. k. a. wai , `` characterizing bifurcations and chaos in multiwavelength lasers with intensity - dependent loss and saturable homogeneous gain , '' _ optics communications _ , vol . 285 , no . 8 , pp . 21442153 , 2012 . a. el amili , g. kervella , and m. alouini , `` experimental evidence and theoretical modeling of two - photon absorption dynamics in the reduction of intensity noise of solid - state er : yb lasers , '' _ optics express _ , vol . 21 , no . 7 , pp . 87738780 , 2013 .

increasingly the collective behavior displayed by groups of interacting dynamical components , each of which may be capable of a full range of complex dynamics , is essential to understanding and ultimately designing systems . examples in biology , physics , and engineering are diverse , ranging from understanding sensory perception to the design of antennas capable of simultaneously sending and receiving signals at the same frequency . while in general the dynamical components may be functionally distinct and heterogeneous , in many examples one is interested in the collective behavior displayed by a group of similar coupled elements . one commonly studied example is the synchronization of oscillating subsystems that interact . the ability of a group of coupled oscillators to synchronize despite a distribution in individual frequencies is a broadly applicable phenomenon . in physical systems coherent oscillations can be used to enhance detector sensitivity or increase the intensity of a power source . synchronization is also important in biological systems where the collective behaviors in populations of animals , such as the synchronized flashing of fireflies or the coherent oscillations observed in the brain , serve as examples . including a distribution in individual frequencies in an otherwise homogeneous ensemble captures some of the inevitable differences that group members will possess . although synchronization is often put forward as an example of the importance of understanding nonlinear phenomena , the intuition for it , and indeed the subsequent mathematical discussion , often reduces to simple linear ideas . for example , the famous example of huygens clocks @xcite can be understood in terms of a linear coupling of the two pendulums through the common mounting support . then the larger damping of the symmetric mode ( coming from the larger , dissipative motion of the common support ) compared with the antisymmetric mode leads at long times to a synchronized state of the two pendulums oscillating in antiphase . the nonlinearity in the system is simply present in the individual motion of each pendulum ; specifically in the mechanism to sustain the oscillations . without the drive , the oscillators would still become synchronized through the faster decay of the even mode , albeit in a slowly decaying state . rather than this mode - dependent dissipation mechanism , one might expect synchronization to arise from the intrinsically nonlinear effect of the frequency pulling of one oscillator by another . furthermore , the model describing the two pendulums , as well as most other models used to show synchronization , has dissipative coupling between the oscillators . in contrast , many physical situations have mainly reactive coupling . nanoscale mechanical oscillator arrays are one example that offer significant potential in a range of technologies @xcite,@xcite . due to the scales and numbers in these arrays , active control of individual oscillators poses a number of challenging problems . synchronization offers a potentially appealing alternative in some applications . one notable characteristic of these arrays is the predominantly reactive coupling due to elastic or electrostatic interactions . this context is the motivation for our work . in particular , we study the model@xmath0 where @xmath1 is a complex number representing the amplitude @xmath2 and phase @xmath3 of the mth oscillator @xmath4 . the natural frequency @xmath5 of each oscillator is chosen from a distribution @xmath6 . we call the width of the distribution @xmath7 . coefficients of the terms for nonlinear frequency pulling ( @xmath8 ) , dissipative coupling ( @xmath9 ) , and reactive coupling ( @xmath10 ) serve as model parameters . the coupling is taken to be infinite range or all - to - all . in this paper we study the roles of nonlinear frequency pulling and reactive coupling in the absence of the dissipative coupling using eq . ( [ eq_cross ] ) with @xmath11 . in addition the distribution @xmath6 of the @xmath5 must be specified . we study the case of positive @xmath8 and @xmath10 ; for a symmetric distribution @xmath6 the results are the same changing the sign of both @xmath8 and @xmath10 . alternatively , when only nonlinear saturation and dissipative coupling are present ( @xmath12 ) eq . ( [ eq_cross ] ) reduces to @xmath13 the behavior for general @xmath7 and @xmath9 of eq . ( [ amplitudephase ] ) has been analyzed by matthews et al . if the width @xmath7 of the distribution @xmath6 is narrow , so that the time evolution of the magnitudes @xmath14 is fast compared with that of the phase dispersion , and @xmath9 is small , @xmath14 rapidly relaxes to a value close to unity , and the only remaining variable for each array member is the phase @xmath3 . equation ( [ amplitudephase ] ) can then be reduced to a simple form @xcite,@xcite , often known as the kuramoto model @xmath15 that has been the subject of numerous studies @xcite . in the absence of coupling each oscillator in this model would simply advance at a rate that is constant in time , but with some dispersion of frequencies over the different elements . equation ( [ kuramoto ] ) is an abstraction from the equations describing most real oscillator systems , leaving out many important physical features . the complex notation of eq . ( [ eq_cross ] ) suggests the introduction of a complex order parameter @xmath16 to measure the coherence of the oscillations@xmath17 with @xmath18 for the kuramoto model , and then the equation of motion can be written [ take the imaginary part of @xmath19 in eq . ( [ orderparameter ] ) to evaluate the sum appearing in eq . ( [ kuramoto ] ) ] as@xmath20 thus the behavior of each oscillator is given by its tendency to lock to the phase of the order parameter . the term @xmath21 acts as a locking force , and locking occurs for all oscillators with frequencies satisfying @xmath22 , with the locked oscillator phase given by @xmath23 . the magnitude @xmath24 of the order parameter must then be determined self - consistently via eq . ( [ orderparameter ] ) . the generalization of the locking force to apply to the model eq . ( [ eq_cross ] ) will be a conspicuous feature of our work . equation ( [ kuramoto ] ) is known to show rich behavior , including , in the large @xmath25 limit , a sharp synchronization transition at some value of the coupling constant @xmath26 @xcite , which depends on the frequency distribution of the uncoupled oscillators @xmath6 . the transition is from an unsynchronized state with @xmath27 in which the oscillators run at their individual frequencies , to a synchronized state with @xmath28 in which a finite fraction of the oscillators lock to a single frequency . the transition at @xmath29 has many of the features of a second order phase transition , with universal power laws and critical slowing down @xcite , as well as a diverging response to an applied force @xcite . we present a detailed analysis of the model eq . ( [ eq_cross ] ) with @xmath30 describing the synchronization of an oscillator ensemble involving reactive coupling between the elements , which then leads to synchronization through nonlinear frequency pulling . we begin in sec . [ sec_derive ] by deriving eq . ( [ eq_cross ] ) as a description of arrays of nanoelectromechanical oscillators . we then discuss the solutions to eq . ( [ eq_cross ] ) for a variety of symmetric distributions in intrinsic frequencies @xmath6 . common solution types exist for the three distinct types of frequency distributions we studied . in sec . [ sec_examples ] we introduce these solutions and the measures we use to describe them . we begin the analysis in sec . [ sec_form ] by moving to a continuum description to derive synchronization conditions . we are able to analyze the existence of two behavior types in closed form : the unsynchronized solution in sec . [ sec_onset ] and the fully locked synchronized behavior in sec . [ sec_fulllocking ] . results from our analytical arguments are combined with those from simulations to present the solutions and associated phase diagram for each frequency distribution in sec . [ sec_results ] . we begin these by considering an unbounded symmetric frequency distribution in the form of a lorentzian in sec . [ sec_lort ] . continuing in this section we then approximately double the distribution width to discuss some interesting characteristics of the well - ordered synchronized solutions as well as changes to the bifurcations . in contrast , sec . [ sec_top ] presents the results for a bounded distribution when all frequencies over some range are equally likely ( a top - hat distribution ) . this is followed in sec . [ sec_tri ] by the case of a symmetric unimodal frequency spread , namely a triangular distribution . finally , conclusions are made in sec . [ sec_conclude ] . a brief account of some results for the model eq . ( [ eq_cross ] ) has been reported previously @xcite . although the main focus of this paper is analyzing the behavior of eq . ( [ eq_cross ] ) we first show how such an equation might arise from the equations of motion of physical oscillators . as an example consider the system defined by@xmath31=0 . \label{eq_dv}\ ] ] the first two terms describe uncoupled harmonic oscillators . we suppose the uncoupled oscillators have a linear frequency that is near unity by an appropriate scaling of the time variable@xmath32 the third term is a _ negative _ linear damping , which represents an energy source to sustain the oscillations , and positive nonlinear damping , so that the oscillation amplitude saturates at a finite value . again this saturation value is chosen to be of order unity by an appropriate scaling of the displacements @xmath33 . for an example of an effective negative linear damping term in a micromechanical oscillator see references @xcite . one could also imagine implementing such an effect with an electronic feedback loop sensing each oscillator velocity and driving the oscillator with an appropriate phase . the first three terms of eq . ( [ eq_dv ] ) comprise a set of uncoupled van der pohl oscillators . the term @xmath34 describes a stiffening of the spring constant ( for @xmath35 ) and is a reactive nonlinear term that leads to an amplitude dependent shift of the resonant frequency . with @xmath36 this would give us _ duffing _ oscillators . the final term is a nearest neighbor coupling between the oscillators , depending on the difference of the displacements . this is a reactive term , that conserves the energy of the system . others @xcite have considered nonlinear oscillators coupled through their velocities ; this is a dissipative coupling that would lead to @xmath37 in the amplitude - phase reduction . the complex amplitude equation ( [ eq_cross ] ) holds if the parameters @xmath38 are sufficiently small . in this case the equations of motion are dominated by the terms describing simple harmonic oscillators at frequency one , and the time dependence remains close to @xmath39 . to formalize the smallness of @xmath38 we introduce a small parameter @xmath40 and write @xmath41 , @xmath42 , @xmath43 , @xmath44 . the oscillating displacement is written as a slow modulation of oscillations at frequency one , plus corrections @xcite@xmath45 + \epsilon x_{n}^{(1)}(t)+\ldots\label{ansatz1}\ ] ] with @xmath46 a slow time scale . , as in ref . however in the present case inspection of the van der pohl term shows that the oscillations will grow until @xmath47 . the nonlinear effects are small because the multiplying coefficients @xmath48 and @xmath49 are small . of course , this difference results from a choice of scaling of the @xmath33 variables ; for example choosing @xmath50 would result in the more familiar procedure . ] the variation of @xmath51 gives us the extra freedom to eliminate secular terms and ensure that the perturbative correction @xmath52 , as well as all higher - order corrections to the linear response , do not diverge ( as they do if one uses naive perturbation theory ) . using the relation @xmath53 ( denoting a time derivative with respect to the slow time @xmath54 by a prime ) we calculate the time derivatives of the trial solution ( [ ansatz1 ] ) [ eq : derivs]@xmath55e^{it}+c.c.\right ) + \epsilon\dot{x}_{n}^{(1)}(t)+\ldots\\ \ddot{x}_{n } & = \left ( [ -a_{n}+2i\epsilon a_{n}^{\prime}+\epsilon^{2}a_{n}^{\prime\prime}]e^{it}+c.c.\right ) + \epsilon\ddot{x}_{n}^{(1)}(t)+\ldots\end{aligned}\ ] ] substituting these expressions back into the equation of motion ( [ eq_dv ] ) , and picking out all terms of order @xmath56 , we get the following equation for the first perturbative correction @xmath57 . \label{eq_xone}\ ] ] the collection of terms proportional to @xmath58 on the right - hand side of eq . ( [ eq_xone ] ) , called the secular terms , act like a force driving the simple harmonic oscillator on the left - hand side at its resonance frequency . the sum of all the secular terms must vanish so that the perturbative correction @xmath52 will not diverge . ( terms varying as @xmath59 contribute a finite response to @xmath60 . ) this gives us an equation for determining the slowly varying amplitudes @xmath51 @xmath61=0 . \label{duffingampa}\ ] ] more informally , we might write @xmath62 so that @xmath63 and eq . ( [ duffingampa ] ) can be written in terms of the original small parameters without the formal scaling@xmath64=0 \label{duffingamp}\ ] ] with a rescaling of time @xmath65 eq . ( [ duffingamp ] ) reduces to the form eq . ( [ eq_cross ] ) except that in eq . ( [ eq_cross ] ) the nearest neighbor coupling is replaced by the all - to - all coupling convenient for theoretical analysis . in this section we introduce the types of dynamical states encountered for the model eq . ( [ eq_cross ] ) as well as diagnostic tools to characterize these states . the types of states we find for the different distributions investigated are largely the same , and so we use a particular example a lorentzian distribution of oscillator frequencies with some convenient choice of parameters @xmath8 and @xmath10and discuss how the behavior depends systematically on the distribution and other parameters later in the paper . we are mainly interested in the behavior for large numbers of oscillators @xmath66 . for the numerical simulations we are of course restricted to finite @xmath25 ( we typically use @xmath67 , but have gone up to @xmath68 to investigate some subtle effects ) . in our discussion we focus on those results that we expect to be largely independent of @xmath25 for large @xmath25 . a key diagnostic for synchronization is the complex order parameter @xmath69 defined by eq . ( [ orderparameter ] ) , introduced by kuramoto @xcite , with magnitude @xmath24 and phase @xmath70 where @xmath71 . in the large @xmath25 limit , we could use a nonzero value of the order parameter at each time as the basic criterion for a synchronized state . a synchronized periodic state with frequency @xmath72 would then have @xmath73 with @xmath74 constants . more complicated dynamical states might also occur . for a finite number of oscillators @xmath25 a precise diagnostic is harder , since there are fluctuations of order @xmath75 that make the instantaneous @xmath76 nonzero even in an unsynchronized state . for a synchronized periodic state we could require that the time average @xmath77 scales as @xmath25 as the number of oscillators changes for some frequency @xmath78 that becomes constant for large enough @xmath25 , and use this as a measure of the synchronization . in practice we use the simpler criterion that the time averaged magnitude @xmath79 is nonzero and does not appear to decrease to zero as @xmath25 increases . ( a @xmath76 fluctuating about zero with an amplitude of order @xmath75 will of course lead to a nonzero @xmath80 of this same order . ) this definition can also be applied to aperiodic synchronized states . another useful diagnostic uses the actual frequency of each oscillator defined by@xmath81 , \label{frequencylocking}\ ] ] where @xmath82 is some long averaging time and @xmath83 is a starting time sufficient to eliminate transients . a _ frequency locked _ state has a fraction of oscillators with the same frequency @xmath84 . ( the fraction should be @xmath85 and a value not decreasing to zero as @xmath25 increases ) . if not all the oscillators have the same frequency ( i.e. the fraction is not unity ) we call the state _ partially locked_. if all the oscillators have the same average frequency , we call the state _ fully locked_. to make contact with the analytic results we actually use a stricter criterion , and also require the magnitudes @xmath86 to be time independent in the fully locked state . usually we find that a nonzero @xmath80 is associated with frequency locking , but this is not always the case . _ phase locking _ is a stricter requirement than eq . ( [ frequencylocking ] ) . for phase locking we would require that @xmath87 does not diverge as @xmath88 for @xmath89 taken from some finite fraction of the oscillators . ( frequency locking is , for example , consistent with phase differences that grow diffusively proportional to @xmath90 . ) we do not investigate this stricter locking criterion . to investigate whether each locked oscillator in a frequency locked state is tightly locked in phase to the phase of the order parameter , or is fluctuating about this value , we define a _ phase synchronization index _ for each oscillator@xmath91 with @xmath92 the phase of the @xmath93th oscillator relative to the order parameter phase@xmath94 when the phase of oscillator @xmath93 is locked to the order parameter one , @xmath95 is constant and the phase synchronization index is unity , whereas as @xmath95 tends towards a uniform distribution from @xmath96 to @xmath97 the index approaches zero . we define the number of oscillators with @xmath98 very close to one as the _ tightly locked cluster_. [ tbh ] of frequencies for cutoff lorentzian with @xmath99 and cutoff @xmath100 for @xmath67 oscillators . the main graph shows the individual frequencies , and the inset a histogram.,width=316,height=282 ] we now show results of numerical simulations for the lorentzian distribution of frequencies . we have chosen to cut off the distribution at some large @xmath101 otherwise there would be a few very fast oscillators restricting the time stepping of the numerics . thus we use@xmath102{ll}$g(0)\frac{w^{2}}{\omega^{2}+w^{2}}$ & for $ \left\vert \omega\right\vert < \omega_{c}$\\ $ 0 $ & for $ \left\vert \omega\right\vert > \omega_{c}$\end{tabular } \ \ \ \ \ \ \ \ \ \ \right . \label{eq_para_lort}\ ] ] for a given choice of @xmath103 , with @xmath7 then fixed by the normalization condition@xmath104 ( it is useful to parameterize the distribution in terms of @xmath103 , since this quantity determines the values of @xmath8 and @xmath10 at which synchronization occurs in the large @xmath105 limit where the model reduces to the kuramoto phase model eq . ( [ kuramoto ] ) . ) in presenting the results we choose a distribution width such that @xmath99 . for a cutoff @xmath100 this gives @xmath106 , with @xmath107 . with no cutoff , we would have @xmath108 . the distribution of frequencies is generated from a uniform distribution of @xmath25 values @xmath109 on the interval @xmath110 using the transformation@xmath111 we present results for a deterministic set of @xmath112@xmath113 /(n-1),\ ] ] but have also used @xmath112 generated randomly on the interval . the distribution of frequencies for @xmath67 oscillators and a cutoff @xmath114 is shown in fig . [ fig_lor_dist ] . [ tbh ] oscillators with the cutoff lorentzian distribution of intrinsic frequencies of fig . [ fig_lor_dist ] . shown is the time averaged order parameter magnitude @xmath115 over a range of @xmath10 values at constant @xmath116 . solid lines are the observed solutions . the overlapping symbols are representative results from simulations following different solution branches with increasing @xmath10 ( @xmath117 ) and decreasing @xmath10 ( @xmath118 and @xmath119 ) . time traces associated with the solid symbols are shown in fig . [ fig_orderparameters ] . the dot - dashed line is at a value of @xmath120 , the order of magnitude expected for random fluctuations of the order parameter about zero.,width=316 ] a plot of the dependence of the mean magnitude of the order parameter @xmath121 as a function of @xmath10 for fixed value of @xmath116 and @xmath67 is shown in fig . [ fig_lorentzian_cut ] . this plot shows three types of states : an unsynchronized state with @xmath122 essentially zero ; and two synchronized states , one with small @xmath121 , growing from @xmath96 to about @xmath123 as @xmath10 decreases below about @xmath124 , and one with large @xmath122 that exists for all @xmath125 . ( we will discuss later such issues as whether the growth of the large @xmath126 is continuous or discontinuous near @xmath127 . ) [ tbh ] as a function of time observed at the solid symbols in fig . the three traces are for the unsynchronized state at @xmath128 ( lower trace ) , and the small amplitude synchronized state ( middle trace ) and large amplitude synchronized state ( upper trace ) both at @xmath129.,width=316 ] the variation of @xmath130 for representative examples of these three states is shown in fig . [ fig_orderparameters ] . the lower trace shows the unsynchronized state at @xmath128 in fig . [ fig_lorentzian_cut ] . the fluctuations in @xmath130 are consistent with fluctuations of the order parameter about zero with magnitude of order @xmath75 as expected for a collection of @xmath25 oscillators with different frequencies and random phases . the average frequency distribution @xmath84 ( not shown ) is unchanged from the bare distribution @xmath131 shifted by @xmath132 . this shift can be understood as arising from the nonlinear frequency shift with @xmath133 , and the coupling to a distribution of oscillators with @xmath134 and random phases . : ( a ) - actual frequency distribution ( solid line ) and bare frequency distribution shifted by @xmath132 ( dotted line ) ; ( b ) - synchronization index @xmath98.,width=288 ] the state corresponding to the middle trace in fig . [ fig_orderparameters ] is the low amplitude state at @xmath129 in fig . [ fig_lorentzian_cut ] . the mean order parameter magnitude @xmath135 is much larger than @xmath136 , suggesting this is a synchronized state . the corresponding frequency distribution and synchronization index are shown in fig . [ fig_smr_plateau ] . the frequency distribution @xmath84 ( a ) shows a small plateau of constant frequency over about 60 oscillators towards the high frequency end of the distribution ; these are the frequency locked oscillators . the synchronization index ( b ) shows that @xmath98 approaches unity for most of the locked oscillators , which means that @xmath137 for these oscillators is essentially time independent once the rotation of the phase at the order parameter frequency is subtracted out . a careful scrutiny of the two panels reveals that the plateau in @xmath84 is sharper and more extended than the one in @xmath98 , so that not all the frequency locked oscillators are time independent in the rotating frame . the dynamical state is actually quite complicated , as a review of @xmath138 shows . and order parameter @xmath76 in the complex plane at a time after transients have decayed . the + towards the center of the plot gives the value of the order parameter . the squares are oscillators locked to the order parameter frequency ( i.e. those with @xmath139 ) : solid squares are stationary when the @xmath140 dependence is removed , open squares show an additional dynamics rotating around the tail of the fixed distribution in the rotating system . the @xmath141 are oscillators that are not locked to the order parameter.,width=288 ] figure [ l_matlab_fig2 ] shows a plot of a snapshot of the complex amplitude @xmath137 of each oscillator . the + is the order parameter . the other points correspond to @xmath137 for each oscillator . it is useful to study the dynamics of this plot after rotation at the mean order parameter frequency is eliminated . as time evolves , the square symbols remain fixed in such a plot ( except for very small fluctuations ) : these represent oscillators that are locked to the order parameter . for the solid squares the complex amplitudes are essentially time independent once the phase of the order parameter is extracted . the oscillators represented by an @xmath142 on the other hand rotate clockwise or anticlockwise about the origin : these correspond to unlocked or running oscillators . the open squares exhibit a more complicated dynamics undergoing small amplitude orbits around the tail of the locked oscillator distribution . these oscillators are locked to the order parameter , since the difference of their phases from the order parameter phase does not drift over arbitrary long times . the values of @xmath84 for these oscillators are on the locked plateau in fig . [ fig_smr_plateau ] . however the amplitudes are not constant in the rotating plot , the values of @xmath98 are less than unity , and the oscillators contribute to fluctuations of the order parameter . thus the fluctuations of @xmath130 shown in fig . [ fig_orderparameters ] are not just due to finite @xmath25 effects , and we believe they would persist in the @xmath66 limit . we have not explored larger @xmath25 to pin down whether these intrinsic fluctuations are periodic or aperiodic in the large @xmath25 limit . as we discuss in more detail later , for a bounded distribution of frequencies it is possible to find a low amplitude synchronized state with @xmath143 , but with a smooth frequency distribution showing that there is no frequency locking . in fact for this state the distribution of actual oscillator frequencies does not overlap the order parameter frequency . the nonzero order parameter is caused by a systematic slowing of the phase rotation of the oscillators in the vicinity of the order parameter phase , rather than by a fraction of the phases becoming locked to the order parameter phase . for an unbounded distribution such as the lorentzian , this same effect occurs but is supplemented by the small fraction of locked oscillators . but for the large amplitude synchronized state.,width=288 ] for the same parameter values as for the previously described low amplitude synchronized state for the lorentzian distribution there is a second dynamical state that may be reached depending on the initial conditions of the simulation . this state has a much larger value of the order parameter @xmath144 and the plot of the distribution of frequencies , fig . [ fig_lgr_plateau](a ) , shows a correspondingly larger fraction of locked oscillators . ( this should be compared with panel ( a ) of fig . [ fig_smr_plateau ] . ) for the cutoff lorentzian distribution used , all the oscillators at the high frequency end of the distribution are locked , leaving only a small fraction of free running oscillators at the low end of the frequency distribution . for the @xmath66 limit and a lorentzian distribution without a cutoff , unlocked oscillators remain at both the high and low frequency ends of the distribution . for a bounded distribution of bare frequencies , such as triangular or top - hat , a fully locked state in which all the oscillators rotate with the same frequency and fixed magnitude may be found . we now turn to the analysis of the synchronization of oscillators described by eq . ( [ eq_cross ] ) . since we are interested in the behavior for a large number of oscillators , it is convenient to go to a continuum description , where we label the oscillators by their uncoupled linear frequency @xmath145 rather than the index @xmath93 , @xmath146 . introducing the order parameter eq . ( [ orderparameter ] ) , the oscillator equations can be written in magnitude - phase form as [ theta_r]@xmath147 where @xmath148 is the oscillator phase relative to that of the order parameter as before , and @xmath149 is the bare oscillator frequency shifted by @xmath132 and measured relative to the order parameter frequency @xmath150 @xmath151 note that if the order parameter is zero @xmath152 , the magnitude @xmath153 will relax to @xmath154 , and the @xmath93th oscillator will evolve at the frequency @xmath155 . the first component of this frequency shift is just the nonlinear shift at @xmath156 , and the second is from the interaction of each oscillator with the incoherent motion of the other oscillators . at each time @xmath82 the oscillators are specified by the distribution @xmath157 where @xmath158 is the fraction of the oscillators with shifted frequency @xmath149 that at time @xmath82 have magnitude between @xmath153 and @xmath159 and shifted phase between @xmath160 and @xmath161 . the order parameter is given by the self - consistency condition@xmath162 where @xmath163 is the distribution of oscillator frequencies expressed in terms of the shifted frequency @xmath149 . it is useful to split this expression into real and imaginary parts . the imaginary part is@xmath164 because the phases @xmath165 are measured relative to the orientation of the order parameter , this expresses the fact that the components of the complex amplitudes @xmath166 normal to the order parameter must average to zero . note that unlike the cases of the kuramoto model [ eq . ( [ kuramoto ] ) ] and the dissipatively coupled complex amplitude model [ eq . ( [ amplitudephase ] ) ] studied by matthews et al . , this condition is not trivially satisfied even for the case of a symmetric distribution @xmath6 , and in fact serves to determine the frequency @xmath72 of the order parameter . the frequency @xmath72 is also not trivially related to the mean frequency of the oscillator distribution . the real part of eq . ( [ selfconsistency ] ) is@xmath167 this is the condition that the components @xmath168 along the direction of the order parameter must average to the magnitude @xmath24 . this condition serves to self - consistently fix the value of @xmath24 . the expectation that eqs . ( [ theta_r ] ) might lead to synchronization follows from the behavior for narrow frequency distributions and large @xmath8 . if the width of the distribution of frequencies is small compared to the relaxation rate of the magnitude , which is of order one in the time units used in eq . ( [ eq_cross ] ) , the magnitude relaxes rapidly to the value given by the instantaneous value of @xmath160 , i.e. to the solution of@xmath169 if @xmath153 is close to one , which we will see applies at the onset of synchronization for large @xmath8 , this gives@xmath170 in this case eq . ( [ theta ] ) becomes , ignoring @xmath171 compared with @xmath172,@xmath173 this equation is the same as the one derived from the kuramoto model in eq . ( [ kuramotofield ] ) with @xmath105 playing the role of the coupling constant @xmath9 , and therefore predicts an onset of synchronization at @xmath174 @xcite . to uncover more fully the behavior of eq . ( [ eq_cross ] ) we consider two issues : the onset of synchronization , detected as the linear instability of the unsynchronized @xmath152 state , as a function of @xmath8 and @xmath10 for some given frequency distribution @xmath6 ; and the existence of a fully locked state for large values of @xmath105 . we first consider the initial onset of partial synchronization from the unsynchronized state in which each oscillator runs at its own frequency @xmath131 . we identify synchronization through a nonzero value of the order parameter . this often arises because a finite fraction of the oscillators become locked to the same frequency , which we call a partially ( fully if the fraction is unity ) locked state . however we also find situations where @xmath143 , but there is no frequency locking . the onset of synchronization can be determined by a linear instability analysis of the unsynchronized state . this is calculated by linearizing the distribution @xmath175 around the unsynchronized distribution , which is a uniform phase distribution at @xmath156 , and seeking the parameter values at which deviations from the uniform phase distribution begin to grow exponentially . this follows the method of matthews et al . @xcite , although care is needed in the analysis due to the more important role the magnitude perturbations @xmath153 play in the present case . introducing the small expansion parameter @xmath40 characterizing the small deviations from the unsynchronized state , we write@xmath176[1+\varepsilon f_{1}(\bar{\theta},\bar{\omega},t ) ] . \label{rhoexpansion}\ ] ] note that for @xmath177 this does indeed give the appropriately normalized distribution for the unsynchronized state , in which all the oscillators have unit magnitude , @xmath156 , and the phase distribution is constant . also , for nonzero @xmath40 , @xmath175 remains normalized to linear order@xmath178 providing the average of @xmath179 over @xmath160 is zero . the equation for the evolution of the radial perturbation of the distribution @xmath180 is given by noting that@xmath181 the left hand side is evaluated by expanding eq . ( [ r ] ) to first order in @xmath40 and also expanding the magnitude of the order parameter @xmath182 . , and the right hand side by the replacement @xmath183 and assuming an exponential growth or decay of the perturbation @xmath184 . the result is@xmath185 equation ( [ r1 ] ) is solved by@xmath186 with [ ab]@xmath187 to extract the equation for @xmath179 integrate the equation for the conservation of probability @xmath188 over radius . here @xmath189 is the velocity in complex amplitude space , which in polar coordinates is @xmath190 given by eqs . ( [ theta_r ] ) . again replacing @xmath191 by @xmath192 , and evaluating @xmath180 from eqs . ( [ r1ab],[ab ] ) , this gives at @xmath193@xmath194 this is solved by@xmath195 with [ cd]@xmath196}{\left ( \bar{\omega}^{2}+\lambda ^{2}\right ) \left [ \bar{\omega}^{2}+(\lambda+2)^{2}\right ] } , \label{c}\\ d & = \beta\frac{4\alpha\bar{\omega}(\lambda+1)+\lambda\lbrack\bar{\omega } ^{2}+(\lambda+2)^{2}]}{\left ( \bar{\omega}^{2}+\lambda^{2}\right ) \left [ \bar{\omega}^{2}+(\lambda+2)^{2}\right ] } . \label{d}\ ] ] we now evaluate the self - consistency condition eq . ( [ selfconsistency ] ) to first order in @xmath40 . the imaginary part is @xmath197 which to first order in @xmath40 gives@xmath198 = 0 . \label{perpint}\ ] ] similarly the real part of the self - consistency condition is@xmath199 = 2 . \label{parint}\ ] ] we want to evaluate eqs . ( [ perpint ] ) and ( [ parint ] ) at the onset of instability where the growth rate @xmath200 . we can set @xmath201 in eqs . ( [ perpint],[parint ] ) with eqs . ( [ ab ] ) and ( [ cd ] ) except in terms with @xmath202 in the denominator , since such terms may give large contributions to the integral from the region of small @xmath149 . a term involving just @xmath203 gives a finite integral , but if this is multiplied by powers of @xmath149 or @xmath204 the integral goes to zero in the @xmath200 limit . similarly for a term involving @xmath205 we must take the limit @xmath200 after doing the integral ( this is equivalent to the principal value integral ) , whereas if this term is multiplied by powers of @xmath149 we can put @xmath201 immediately . the needed integrals are [ integrals]@xmath206 the imaginary part of the self consistency condition eq . ( [ perpint ] ) becomes @xmath207 and the real part reduces to the condition for @xmath208@xmath209 we have explicitly evaluated the integrals for top - hat , triangular , and lorentzian distributions of bare frequencies . these results will be presented after we discuss full locking . we define the fully locked state as one in which all the phases are rotating at the same frequency as the order parameter , and the magnitudes are constant in time . these solutions are defined by eq . ( [ theta ] ) with @xmath210 , which with eq . ( [ rconstant ] ) can be written@xmath211 where the solution to the cubic equation ( [ rconstant ] ) for @xmath212 is to be used to form the function of phase alone @xmath213 . the function @xmath213 acts as the force pinning the locked oscillators to the order parameter , generalizing the notion introduced below eq . ( [ kuramotofield ] ) , and plays a central role in our discussion of locking . a particular oscillator , identified by its shifted frequency @xmath149 , will be locked to the order parameter if eq . ( [ locked ] ) has a solution @xmath214 [ and then @xmath153 is given by solving eq . ( [ rconstant ] ) ] and if this solution is stable . the stability is tested by linearizing eqs . ( [ theta_r ] ) about the solution . the fully locked solution will only exist if stable , locked solutions to eq . ( [ locked ] ) exist for _ all _ the oscillators in the distribution . we are thus led to investigate the properties of the function @xmath215 . in addition the self consistency condition eq . ( [ selfconsistency ] ) must be satisfied . for small @xmath171 , the magnitude @xmath216 given by eq . ( [ rconstant ] ) remains bounded away from zero for all @xmath160 , and the function @xmath213 varies continuously between minimum and maximum values @xmath217 . in this case , we immediately see that the fully locked solution only occurs for bounded distributions , @xmath218 nonzero only between finite @xmath219 and @xmath220 . in such cases we define @xmath221 , which is equal to the range of unshifted frequencies @xmath222 , as @xmath7 the width of the distribution . more generally , although @xmath213 can vary over an infinite range ( because @xmath153 may become zero ) we find that only a finite range yields stable solutions , so that again complete locking only occurs for a bounded distribution of oscillator frequencies . we first look at the fully locked solution for large values of @xmath105 . in this case the phases of the locked oscillators cover a narrow range of angles , since the range of the pinning force @xmath223 becomes large for large @xmath105 . the imaginary part of the self consistency condition eq . ( [ imselfconsistency ] ) shows that the range of phases must be around @xmath224 . equation ( [ locked ] ) can now be approximated by expanding around @xmath224 ( note @xmath225 here ) and becomes for large @xmath8@xmath226 the imaginary part of the self - consistency condition reduces to @xmath227 ( the average is over the distribution of frequencies ) , and the real part to simply @xmath228 . finally , averaging eq . ( [ alphalarge ] ) over the distribution of frequencies fixes the order parameter frequency ( the common frequency of all the oscillators)@xmath229 thus the order parameter , and all the oscillators , evolve at a frequency given by the mean of the distribution @xmath6 shifted by the nonlinear effect for @xmath156 . we now investigate the limit to the fully locked regime as we lower @xmath8 . we first summarize the argument , and then present the details . the fully locked solutions are determined by a rather complicated set of interconnected equations . they can be found by the following algorithm . for fixed values of @xmath230 solve eq . ( [ rconstant ] ) for real positive @xmath216 and hence calculate @xmath213 . using eqs . ( [ theta_r ] ) the stability of each solution is tested : the eigenvalues of this analysis are ( see eq . ( [ eigenvalues_2 ] ) below)@xmath231 with @xmath232 . from this we can identify a range @xmath233 corresponding to the range of existence and stability of locked oscillators . a fully locked solution must then satisfy the constraints@xmath234 together with the condition given by the imaginary part of the self consistency condition@xmath235\,d\bar{\omega}=0 , \label{transverse}\ ] ] where @xmath236 , and @xmath237 is then the solution to eq . ( [ rconstant ] ) . the boundary of the fully locked region occurs either when @xmath238 or when @xmath239 . this equation can be interpreted as fixing the order parameter frequency @xmath72 in terms of @xmath8 and @xmath240 . notice that eq . ( [ eigenvalues ] ) shows that for @xmath241 one eigenvalue certainly has a positive real part indicating instability , so that for stable solutions @xmath242 and @xmath243 is finite . since @xmath244 or @xmath220 is now determined , and @xmath245 , eq . ( [ transverse ] ) is an implicit equation relating the values of @xmath8 , @xmath240 and @xmath7 at the locking transition . to complete the solution , the real part of the self consistency equation@xmath246\,d\bar{\omega}=r \label{longitudingal}\ ] ] then serves to fix @xmath24 at locking , from which the value of @xmath247 at the transition to full locking can be found . we first consider the _ existence _ of a locked solution for an individual oscillator , i.e. a _ stationary _ solution of eqs . ( [ theta_r ] ) . equation ( [ rconstant ] ) gives the cubic equation for @xmath216 for each @xmath248@xmath249 of course , for the physical solution @xmath153 must be real and positive . thus we need to analyze the properties of the real positive solutions to@xmath250 as @xmath251 varies . [ tbh ] ) as @xmath251 varies . the dashed lines are @xmath252.,width=268 ] for @xmath253 the solutions to the cubic are @xmath254 . as @xmath255 increases , the solutions remain real until two of the roots collide and become complex . since the sum of the roots to eq . ( [ cubic ] ) is zero , at the collision the equation takes the form@xmath256 matching coefficients shows that collision occurs at @xmath257 when @xmath258 . the form of eq . ( [ cubic ] ) is actually already in what is known as the depressed form of a cubic equation , for which the solution is relatively simple . inspecting the form of these solutions shows that for @xmath259 there is one real solution and a complex pair in the form @xmath260 . the product of the roots is determined by the constant in the cubic , giving@xmath261 and so for @xmath262 the real root is negative , and for @xmath263 the real root is positive . these results are confirmed by numerical solution as shown in fig . [ cubic_fig ] . [ tbh ] for @xmath264 . the dashed line is @xmath265 and the dotted line @xmath266 . the condition @xmath267 is necessary ( but not sufficient ) for the solution to be stable . the lower branch always satisfies @xmath268 , and so only the upper branch may have stable regions.,width=268 ] [ tbh ] for @xmath269 . solutions exist for @xmath270 with @xmath271 . the dashed line is @xmath265 and the dotted line @xmath266.,width=268 ] thus we find the following behavior for the real positive solutions to eq . ( [ r_theta ] ) . for @xmath264 there is a root that is positive with @xmath272 for all @xmath160 , a root that varies between positive and negative values with magnitude less than @xmath273 , and a root that is negative for all @xmath160 . we have seen that the stability analysis shows that any solution with @xmath241 is unstable@xmath274 thus only the first root is relevant ( see fig . [ r_theta_bs ] ) . for @xmath269 real positive roots exist for @xmath270 with @xmath275 . in the range where there are two real positive roots , @xmath276 and @xmath277 , only the larger may be in the stable range @xmath278 , ( see fig . [ r_theta_bl ] ) . having found the stationary solutions @xmath216 , the condition for stationary phase @xmath279 can be written@xmath280 with@xmath281 the function @xmath213 acts as the force acting on each oscillator tending to pin its frequency to the order parameter frequency . the range of possible @xmath149 for locked oscillators is limited by the range of @xmath223 corresponding to stable stationary solutions . the stability of the locked solution for an individual oscillator mentioned in the previous section is given by linearizing eqs . ( [ theta_r ] ) about the stationary solutions . the linearized equations are@xmath282 the eigenvalues are@xmath283 with @xmath284 this immediately shows us @xmath278 is a necessary condition for stability , since @xmath285 ( with the equality if @xmath286 is negative so that @xmath287 are complex ) . let us first seek the condition for a root @xmath287 to become zero , signaling a stationary bifurcation . it is convenient to go back to the original equations eqs . ( [ theta_r ] ) in the form [ linear ] @xmath288 with@xmath289 then the determinant of the linear matrix derived from eqs . ( [ linear ] ) is@xmath290{cc}-\frac{\partial f}{\partial\bar{\theta } } & -\frac{\partial f}{\partial r}\\ \beta r\cos\bar{\theta } & 1 - 3r^{2}\end{array } \right\vert = -(1 - 3r^{2})\frac{\partial f}{\partial\bar{\theta}}+\beta r\cos\bar{\theta}\frac{\partial f}{\partial r}. \label{det}\ ] ] the stationary solution @xmath216 satisfies eq . ( [ r_theta ] ) and so eq . ( [ det ] ) can be written@xmath291 since@xmath292 thus a zero eigenvalue occurs at and only at stationary points of @xmath213 or at @xmath265 . the latter is where @xmath216 has a vertical tangent ( cf . [ r_theta_bl ] ) but always occurs outside the range of stable solutions , for which we know @xmath278 . the only other possibility for an instability is @xmath293 , @xmath294 . this can occur only at @xmath295 , and if @xmath296 . another result can be derived : if @xmath297 and @xmath272 then the determinant @xmath298 . this implies that the eigenvalues are real ( since the product of a complex conjugate pair is always positive ) , one positive and one negative . thus a negative slope of @xmath213 implies instability . also the hopf bifurcation can only occur at values of @xmath160 where @xmath299 . to satisfy the imaginary part of the self - consistency condition eq . ( [ transverse ] ) , the range of phases of locked oscillator phases must straddle @xmath224 . the oscillator solution here , @xmath300 is always stable for positive @xmath301 , since here @xmath302 . thus the range of possible stable stationary solutions for locked oscillators is given by the range of @xmath160 bounded on either side of @xmath224 by the closest stationary bifurcation point or by a hopf bifurcation occurring at @xmath266 , whichever is closest . we now derive the properties of the locking force , @xmath213 . for small @xmath248 , ( [ rconstant ] ) yields a stationary solution with @xmath303 and so@xmath304 with @xmath305 . thus @xmath213 is a sinusoidal function of angle in this limit . the stable solutions are the positive slope region for @xmath160 between @xmath306 and @xmath307 . [ tbh ] as @xmath240 increases , the behavior becomes quite complicated , and we have not succeeded in proving any results about the full range of possible behavior of @xmath213 . some examples are shown in fig . [ f_lambda ] . for @xmath264 we have not uncovered parameters leading to an @xmath223-curve qualitatively different from that in panel ( a ) , i.e. a single maximum and minimum . note however , that for @xmath308 values of @xmath241 are encountered , and so the hopf bifurcation may limit the range of stable solutions moving away from @xmath224 , before the maximum or minimum of @xmath223 is reached . for @xmath269 the physical solutions are limited to the range @xmath270 , and the slope of @xmath216 diverges at the end points : @xmath309 for @xmath310 and @xmath311 for @xmath312 . since@xmath313 we see that when @xmath314@xmath315 so that @xmath316 with the sign given by the sign of @xmath223 for @xmath317 ( where @xmath318 ) and opposite to the sign of @xmath223 for @xmath319 ( where @xmath320 ) . this shows , for example , that for the parameters such as those in panel ( c ) of fig . [ f_lambda ] where @xmath321 , the slope @xmath322 approaches @xmath323 as @xmath317 . this implies that either there is an additional minimum between the maximum of @xmath223 and @xmath324 as in this panel , or the maximum disappears and @xmath213 becomes monotonically increasing in this region , as for the parameters in panel ( b ) . on the other hand for @xmath325 , parameters between those of panels ( a ) and ( c ) , it turns out that @xmath326 , so that @xmath327 as @xmath328 , and @xmath213 may decrease monotonically between the maximum and @xmath317 . [ tbh ] ) for the first instability of the locked oscillator solution either side of @xmath329 for @xmath330 the value of @xmath331 is nonzero above the dashed curve , and for @xmath332 it is nonzero above the solid curve.,width=268 ] in fig . [ hopf_stationary ] the regions where the first instability on either side of @xmath224 is a hopf bifurcation ( @xmath333 when @xmath334 , with @xmath335 the eigenvalue with larger real value for @xmath330 , and @xmath336 for @xmath332 ) are plotted as a function of @xmath8 and @xmath240 . note that there may be discontinuous jumps from @xmath337 to a finite nonzero value of @xmath331 : for example on the negative @xmath160 side , the minimum in @xmath223 , giving a stationary bifurcation , may disappear by colliding with the maximum , and then the first instability jumps to the hopf bifurcation that was previously not the closest bifurcation to @xmath338 . figure [ hopf_stationary ] does not tell the whole story about whether the boundary of the locked state is a stationary or hopf bifurcation , because in general only the instability on one side of @xmath224 determines the boundary , and which side this is depends on the distribution of oscillator frequencies via the transverse self - consistency condition . this is considered further below . [ tbh ] acting on each oscillator as a function of the phase relative to the order parameter @xmath339 . the dashed portions of the curve correspond to _ stable _ solutions , and the dotted to _ unstable_. the thick solid portion denotes the distribution of oscillator frequencies that must be placed on the stable portion of @xmath223 and must also satisfy the transverse consistency condition.,width=268 ] we have found a range of @xmath160 straddling @xmath224 giving stable locked solutions for individual oscillators . for a fully locked solution , all the oscillators must have solutions to eq . ( [ omega_f ] ) in this range . in addition the imaginary part of the self consistency condition eq . ( [ imselfconsistency ] ) must be satisfied . this can be written in the form@xmath340\,d\bar{\omega}=0 , \label{perp_int_omega}\ ] ] since now there is a unique @xmath160 and @xmath153 for each oscillator frequency . changing the integration variable in eq . ( [ perp_int_omega ] ) to the angle @xmath160 yields@xmath341 the degree of freedom , for fixed @xmath301 and width of frequency distribution @xmath7 , to be determined by this condition is the order parameter frequency @xmath72 . the scheme of imposing the condition is sketched in fig . [ upper_lower ] . the thick solid line , of fixed length @xmath7 along the @xmath223 axis , is to be slid along the curve @xmath213 ( corresponding to varying @xmath72 ) until the integral eq . ( [ perp_int ] ) is zero . all quantities , except @xmath342 are positive , and so this line must straddle the origin . this must be done with all of the solid line lying within the stable ( dashed ) range of @xmath213 . for a bounded frequency distribution we can always find a fully locked solution for large enough @xmath105 . the argument is as follows . first , there is always a range of stable locked oscillator solutions straddling @xmath224 . if we define @xmath343 as the center of the distribution @xmath6 then the order parameter frequency is given by@xmath344 where @xmath345 is @xmath213 evaluated at the center ( with respect to the ordinate ) of the solid portion of the curve in fig . [ upper_lower ] , once eq . ( [ perp_int ] ) is satisfied . since it follows from eqs . ( [ f_bar_theta ] ) and ( [ rconstant ] ) that the slope of @xmath213 at @xmath224 is@xmath346 the range of phase angles of the locked oscillators of order @xmath347 is small for large enough @xmath240 or @xmath348 , and there is always a fully locked solution for a bounded frequency distribution in this limit . the center of the band can then be evaluated as @xmath343 , and so@xmath349 also , in this limit all the oscillators have essentially the same phase , so that @xmath350 and @xmath351 . this gives the complete solution for the fully locked state for very large @xmath10 . it is easiest to understand the limit of the fully locked solution by decreasing @xmath240 at fixed @xmath352 . as @xmath240 decreases , the range of the locking force @xmath223 decreases . we can continue to construct the solution as in fig . [ upper_lower ] , with the portion of the @xmath213 covered by the locked oscillators determined by the transverse self consistency condition , until the first value @xmath353 is reached for which _ either _ the lower end of the locked band would pass below @xmath243 , _ or _ the upper end of the locked band would pass beyond @xmath354 . this signals the onset of instability of the corresponding locked oscillator , either by a stationary or hopf bifurcation depending on @xmath331 at the appropriate @xmath355 or @xmath356 . @xmath355 or @xmath356 may jump discontinuously as a function of @xmath240 if a new bifurcation occurs when @xmath357 , and a new minimum and maximum of @xmath213 are created at a previously interior point of the stable band . in such a case there may be no value of @xmath240 for which one end of the band coincides with @xmath355 or @xmath356 signaling the onset of instability . presumably this corresponds to a first order transition out of the fully locked state . ] [ tbh ] plane for which the instability from the fully locked state is hopf ( shaded ) or stationary ( unshaded ) for a top - hat distribution . the dashed line shows the range of @xmath301 for which @xmath358 for a top - hat distribution . ( note the change of range of @xmath301 plotted compared with fig . [ hopf_stationary].),width=316 ] we could also imagine increasing the width of the distribution @xmath7 from the fully locked solution which occurs at small @xmath7 for fixed @xmath8 and @xmath240 , until one end of the growing band of locked oscillator solutions reaches @xmath355 or @xmath356 . for a top - hat distribution of oscillator frequencies the integral@xmath359 where the integration extends over the whole stable range of @xmath160 , provides an indicator of whether the limit is reached at the lower or upper bound of the stable band : if @xmath360 is _ positive _ , then to satisfy eq . ( [ perp_int ] ) the integral must be _ reduced _ by _ lowering _ the _ upper _ integration bound , and so the range of integration extends from the _ lower _ stability bound . this gives the condition for the maximum @xmath7@xmath361 with @xmath243 the lowest @xmath223 for stable solutions and @xmath345 the value of @xmath223 at the center of the band for @xmath362 . on the other hand if @xmath360 is negative , then the integral eq . ( [ perp_int ] ) must be _ increased _ by _ raising _ the _ lower _ integration bound , and so the range of integration extends to the _ upper _ stability bound . this gives the condition@xmath363 with @xmath354 the largest @xmath223 for stable solutions . we can then find the range of @xmath301 for which the locked state disappears by a hopf or by a stationary bifurcation , fig . [ hopf_stationary_1 ] . this is constructed from fig . [ hopf_stationary ] , giving the conditions for the instabilities at @xmath243 and @xmath354 to be hopf or stationary , and the result just determined for which of these instabilities limits the range of the fully locked solution . numerical results show that for the top - hat distribution the condition @xmath358 occurs only for a restricted range of parameters : large @xmath8 and @xmath240 near @xmath364 , the region to the right of the dashed line in fig . [ hopf_stationary_1 ] . combining these results yields a hopf bifurcation from the fully locked state in the shaded region of fig . [ hopf_stationary_1 ] . for other oscillator distribution shapes we do not have a criterion for the nature of the instability without a detailed solution of the self consistency condition for each width . so far the solution has been developed in terms of @xmath301 . we now determine the magnitude @xmath24 of the order parameter from the parallel self - consistency condition , the real part of eq . ( [ selfconsistency ] ) which can be written in the form@xmath365 where the range of integration is that determined from the transverse self - consistency condition , see fig . [ upper_lower ] . from @xmath366 we can calculate @xmath247 at the boundary of locked solutions . if the dependence @xmath367 is smooth and monotonic , we can map the results depending on @xmath368 onto functions of @xmath369 however , discontinuities in @xmath367 might well occur due to the jumps in the stability range , for example when the stationary bifurcation disappears as described before . this might lead to values of @xmath10 for which no prediction , e.g. for @xmath370 , has been yielded by the algorithm . we do not have results for such cases . [ tbh ] for @xmath371 , @xmath372 , @xmath373 ( top panel ) , and @xmath371 , @xmath374 , @xmath375 ( lower panel ) corresponding to the boundary of the fully locked solution : solid curve - stable solutions ; dashed curve - unstable solutions . the dash - dotted lines straddle the range of @xmath376 of locked oscillators , between @xmath355 and @xmath377 . the values of the order parameter are 0.92 ( upper panel ) and 0.93 ( lower panel).,width=268 ] two examples of the construction of locked solutions for the triangular distribution , defined below in eqs . ( [ triangular ] ) , are shown in fig . [ twobeta ] : the first for small @xmath248 , where a solution @xmath216 exists for all @xmath160 ; and the second for larger @xmath240 where there are ranges of @xmath160 for which no physical solution for the magnitude ( i.e. real and positive ) exists . the value of @xmath240 for which eq . ( [ transverse ] ) is satisfied in each case was found by simple bisection applied to the numerical results of the integration . in both the cases shown the boundary of the fully locked region occurs when @xmath378 . for the smaller value of @xmath240 this condition corresponds to the limit of the existence range of solutions . for the larger value of @xmath240 the limit occurs at the instability of the locked oscillator solution , which develops via a hopf bifurcation ( @xmath379 ) . results for the boundary of the fully locked state for a top - hat distribution with @xmath99 are shown in fig . [ th_pd ] , and for the triangular distribution with @xmath99 in fig . [ tri_pd ] . in this section we present detailed results for the case of a lorentzian distribution of frequencies . we concentrate on a distribution with @xmath99 , but also present some results for a wider distribution @xmath380 , which shows some novel features . we begin this section by calculating the linear stability of the unsynchronized state . since the lorentzian distribution is unbounded , there is no fully locked solution . for the purposes of analysis we choose , without loss of generality , a lorentzian distribution @xmath6 centered about zero frequency@xmath381 the half width at half height is @xmath7 . in terms of the shifted frequency @xmath149 the distribution is@xmath382 with@xmath383 the integrals eqs . ( [ integrals ] ) are @xmath384 the imaginary part of the self consistency condition eq . ( [ perp_integrals ] ) reduces to @xmath385 this serves to fix the frequency of the order parameter @xmath72 at the onset of synchronization via eq . ( [ define_delta ] ) in terms of the parameters of the system @xmath386 . there are _ two _ solutions for @xmath387@xmath388 for large @xmath8 or small @xmath7 the approximate solutions are @xmath389 giving a locking frequency near the center of the band of the free running oscillators , and @xmath390 giving a locking frequency far in the tails . note that the requirement that @xmath387 is real means that @xmath8 must be sufficiently large @xmath391 with@xmath392 for @xmath393 the unsynchronized state is linearly stable for all values of @xmath10 . the critical value of @xmath10 is determined from eq . ( [ par_integrals ] ) and evaluates to@xmath394 where the expression eq . ( [ delta ] ) for @xmath387 is to be substituted . given a width @xmath7 of the oscillator distribution , for each @xmath395 there are two critical values of @xmath10 : @xmath396 and @xmath397 [ corresponding to the minus and plus signs in the expression eq . ( [ delta ] ) for @xmath387 ] , such that the unsynchronized state is _ unstable _ for @xmath399 . it is remarkable that for very strong coupling , @xmath10 large , the unsynchronized state remains stable . however , as we have already seen and will discuss in more detail , a large amplitude synchronized state is also stable in this regime . for @xmath400 the results for @xmath208 reduces to @xmath401 and @xmath402 . in the limit @xmath403 the former result reproduces the result @xmath404 expected from the reduction to the phase equation valid in this limit . we have used simulations to confirm the boundaries for instability of the unsynchronized state as well as to study the behavior subsequent to the instability . the numerical results use the cutoff lorentzian form introduced in eq . ( [ eq_para_lort ] ) . two widths are considered : a narrow one with a peak height of @xmath99 and one approximately twice as wide with @xmath405 . in all cases the distribution tails are removed above some large frequency . qualitative differences in the phase diagrams are observed for the two distribution widths . [ tbh ] . solid and dashed lines show analytical results of the linear stability of the unsynchronized state . numerics show the bifurcations are supercritical along the solid portions and subcritical along the dashed portion . dash - dotted lines are saddle - node bifurcations observed in numerical simulations . states are : @xmath406 - unsynchronized ; @xmath407 synchronized with small and large amplitude respectively.,width=316 ] the phase diagram for the ( narrow ) lorentzian distribution with @xmath99 is shown in fig . [ lorentzianpd ] . the solid and dashed lines are the analytically obtained stability boundaries of the unsynchronized solution . the numerical simulations show that over the dashed portion of the linear instability curve the bifurcation is subcritical , giving a jump in the order parameter magnitude @xmath24 at onset . in addition the sweeps yield a number of saddle - node bifurcations identified as discontinuous jumps in @xmath24 ; these are denoted by dash - dotted lines in fig . [ lorentzianpd ] ; we do not have closed form relations for these boundaries . thus along the dashed or dash - dotted portions of the boundaries , discontinuous jumps in @xmath24 occur , either between the unsynchronized state and a synchronized state , or between two synchronized states with different values of @xmath24 . the two synchronized states are labelled @xmath408 and @xmath409 in fig . [ lorentzianpd ] . when they coexist at the same @xmath8 and @xmath10 the state @xmath409 has the larger value of @xmath24 , but for both states @xmath24 may go to zero , connecting continuously with the unsynchronized state @xmath406 , for some values of @xmath410 . [ tbh ] showing the time averaged order parameter magnitude @xmath411 as a function of @xmath10 from numerical simulations of the cutoff lorentzian distribution with @xmath99 . in panel ( a ) the only solutions are the unsynchronized ( @xmath406 ) and large @xmath24 synchronized ( @xmath409 ) states , while in panels ( b ) and ( c ) the small @xmath24 synchronized solution is also stable . arrows denote discontinuous jumps that were observed as simulations followed the various solution branches.,width=316 ] representative phase diagram slices for the lorentzian case with @xmath99 , showing the time - averaged order parameter magnitudes @xmath115 as a function of @xmath10 at fixed @xmath8 , are presented in fig . [ fig_cuts_ahalf_d8 ] . in agreement with eq . ( [ eq_amin ] ) the unsynchronized state is stable for all @xmath10 provided @xmath412 . simulations find the @xmath409 solution is bistable with @xmath406 over this @xmath8 range at larger @xmath10 values , as shown in fig . [ fig_cuts_ahalf_d8](a ) . for @xmath413 the unsynchronized state is unstable over a range in @xmath10 . as shown in fig . [ fig_cuts_ahalf_d8](b ) for a range of @xmath8 near @xmath414 a subcritical bifurcation occurs at @xmath396 as the @xmath406 solution becomes unstable and @xmath409 forms . with increasing @xmath10 @xmath409 is the only stable solution until a region of bistability where both synchronized solutions coexist . as shown in the phase diagram there is a small region ( labelled @xmath415 ) over which all three solutions are simultaneously stable . this region of tristability can be observed in fig . [ fig_cuts_ahalf_d8](b ) , over the @xmath10 range of hysteric transition between @xmath408 and @xmath406 near @xmath416 . with increasing @xmath8 the system passes through a tricritical point where the subcritical bifurcation at @xmath396 becomes supercritical . the tricritical point has a codimension of 2 . an example slice at @xmath8 sufficiently large that the bifurcation is supercritical is shown in fig . [ fig_cuts_ahalf_d8](c ) , where with increasing @xmath10 regions of bistability between the synchronized solutions and then @xmath406 and @xmath409 are observed . [ tbh ] slices from numerical simulations of the cutoff wide lorentzian with @xmath405 . shown is the time averaged order parameter magnitude @xmath115 for a range of @xmath10 at three @xmath8 values : ( a ) @xmath417 , ( b ) @xmath418 , and ( c ) @xmath419 . these solutions are observed in an array of @xmath67 oscillators with a lorentzian frequency distribution with @xmath405 and a cutoff frequency of @xmath420 . the dashed box region in panel ( c ) is examined more closely in fig . [ fig_zig_zag].,width=316 ] we now consider a lorentzian frequency distribution of approximately twice the width . specifically , we take @xmath405 in eq . ( [ eq_para_lort ] ) or eq . ( [ lorentzian_analytic ] ) . there are a number of changes to the details of the phase diagram that we do not discuss in detail . several of these are straightforward consequences of the wider frequency spread : for example the instability of the unsynchronized solution moves to larger @xmath8 and @xmath10 . instead we focus on two particular features . the first new feature that is evident from the fixed @xmath8 cuts shown in fig . [ fig_soln_cuts_a1_d16 ] is the reconnection of the branches of synchronized solutions that occurs as @xmath8 is decreased . in fig . [ fig_soln_cuts_a1_d16](c ) , typical of larger values of @xmath8 , the synchronized state growing from the @xmath397 instability ends at a saddle - node bifurcation , with the order parameter jumping to larger values as @xmath10 is decreased . this is the same as the behavior for the narrower distribution , fig . [ fig_cuts_ahalf_d8 ] . on the other hand for smaller values of @xmath8 , as in fig.[fig_soln_cuts_a1_d16](a ) , this state merges continuously with the larger magnitude state . this change in the topology of the solution branches as @xmath8 increases occurs through the development of two additional saddle - nodes , as shown in fig.[fig_soln_cuts_a1_d16](b ) near @xmath421 , and the collision of one of these with the saddle - node terminating the large-@xmath10 upper branch . [ tbh ] synchronized solution for a lorentzian frequency distribution ( color online ) . shown are the results for @xmath67 , @xmath419 , and a cutoff lorentzian distribution with @xmath405 and @xmath420 . the arrows denote the direction of the @xmath10 sweep , and the states are colored red for an upward sweep , green for a downward sweep , and black if both directions lead to the same state . the + symbols denote the starting point for particular sweeps to traverse a state interior to the band ( green for the start of a downward sweep , red for an upward sweep ) . panel ( a ) is a magnification of the dashed box in fig . [ fig_soln_cuts_a1_d16](c ) showing the @xmath115 over a range of @xmath10 . in ( b ) the corresponding size of the synchronized cluster measured by the synchronization index is shown.,width=316 ] another apparent difference with the wider frequency distribution is the steps displayed by the large amplitude synchronized state at larger @xmath10 in fig . [ fig_soln_cuts_a1_d16 ] . while these steps also occur with the narrow frequency distribution , their increased amplitude with the wider frequency distribution facilitates the presentation of the behavior . as an example , the dashed region in fig . [ fig_soln_cuts_a1_d16](c ) is shown enlarged in fig . [ fig_zig_zag ] . panel ( a ) in this figure plots the time averaged order parameter magnitude @xmath411 , while ( b ) shows that the solution changes correlate exactly with changes in the number of oscillators locked to the synchronized cluster . interestingly , the solution follows different trajectories depending on the history of changing @xmath10 . for example , if @xmath10 is monotonically increased the system settles onto the upper trajectory of solution steps ( colored red ) while the lower trajectory of solution steps ( colored green ) is observed when @xmath10 is monotonically decreased . these solutions overlap at smaller @xmath10 ( colored black ) . the two solution trajectories are in fact the boundaries of a band of multistable solution branches . this structure is demonstrated in fig . [ fig_zig_zag](a ) by beginning the system on each of these two boundaries and reversing the direction of @xmath10 variation . now as @xmath10 changes the solution moves off its boundary along a continuous line to the other boundary where it follows the expected path for that sense of @xmath10 changes . starting solutions are represented by the + symbols colored for the direction that @xmath10 is to be varied . beginning on the lower boundary , that is observed when decreasing @xmath10 , and increasing @xmath10 from the lower ( red ) + the oscillators follow the solid ( black ) line in the direction of the upward arrow to the upper boundary . likewise , beginning on the upper boundary , that is observed when increasing @xmath10 , and decreasing @xmath10 the oscillators display a series of states beginning with the upper ( green ) + and follow the solid ( black ) line to the lower boundary in the direction of the downward arrow . in this way , the synchronized solution with large time - averaged order parameter magnitude turns out to be a band of solutions , with striations that can be accessed by appropriate changes in @xmath10 . the same is true for the narrow lorentzian , however the band is smaller . the steps in the magnitude of the order parameter are found to be associated with jumps in the number of oscillators locked to the phase of the order parameter . defining a synchronized cluster as the oscillators locked to @xmath70 we find the discrete steps in @xmath24 are due to a small group of oscillators leaving the cluster , and then recombining with the cluster one at a time as @xmath10 increases . in fig . [ fig_zig_zag](b ) the cluster size is measured by the phase synchronization index , defined by eq . [ eq_sync_ind ] , as the number of oscillators with @xmath422 , where @xmath56 is some small number , @xmath423 , to allow for some phase variation over the finite time of measurement . with increasing system size any small group of oscillators leaving the cluster can be expected to have decreasing influence on the order parameter . as the number of oscillators tends to infinity the solution band becomes a region densely populated by these striations as the discrete steps become closer and decrease in length . for the simulations in fig . [ fig_zig_zag ] @xmath67 oscillators were used . to investigate finite size effects we also studied this solution using @xmath424 and @xmath425 . with increasing system size the individual steps on each solution boundary occur over a more narrow range in @xmath10 , becoming shorter in length , and correspondingly more densely packed . thereby , while the discrete steps will disappear in the infinite size limit the width of the band of synchronized solutions remains nearly the same . in this limit synchronized states will move along one of the striations to the band boundary appropriate for the direction of the @xmath10 sweep . for bounded distributions we can calculate the linear stability boundaries of both the unsynchronized and the fully synchronized states . we first do this for a uniform bounded distribution , which we call a top - hat distribution . a top hat distribution centered on @xmath343 is given by@xmath102{cc}$w^{-1}$ & for $ \left\vert \omega\right\vert \leq w/2$\\ $ 0 $ & for $ \left\vert \omega\right\vert > w/2$\end{tabular } \ \ \ \ \right.\ ] ] the integrals eqs . ( [ integrals ] ) for this distribution are @xmath426 , \\ i_{2 } & = \frac{1}{w}\ln\left\vert \frac{w-2\delta}{w+2\delta}\right\vert , \\ i_{3 } & = \frac{1}{2w}\ln\left [ \frac{1+(w-2\delta)^{2}/16}{1+(w+2\delta ) ^{2}/16}\right ] , \\ i_{4 } & = \left\ { \begin{tabular } [ c]{cc}$\frac{\pi}{w}$ & for$~\delta < w/2$\\ $ 0 $ & for $ \delta > w/2$\end{tabular } \ \ \ \right . .\end{aligned}\ ] ] the self consistency equations ( [ perp_integrals ] ) and ( [ par_integrals ] ) can be solved numerically . the equations simplify in the limit of small @xmath7 , and the results here can be displayed in closed form . in small @xmath7 limit there is one solution of eq . ( [ perp_integrals ] ) giving a locked frequency within the band of ( shifted ) oscillator frequencies , @xmath427 , for which @xmath428 and @xmath429 may be neglected in the small @xmath7 limit . equation ( [ perp_integrals ] ) then gives the explicit expression for the frequency offset from mid band @xmath430 and eq . ( [ par_integrals ] ) gives the condition of the parameters at the onset of synchronization@xmath431 notice that for large @xmath8 the locking frequency is close to the center of the band , and the critical condition reduces to@xmath432 the value expected from the mapping onto the kuramoto model for large @xmath433 . for small @xmath8 on the other hand @xmath387 approaches @xmath434 giving us the result that the locking frequency approaches the upper edge of the band . in this case the onset occurs at @xmath435 . even for moderate values of @xmath8 such as @xmath436 , the locking frequency is far off the band center ( a fraction @xmath437 of the half band width for @xmath436 ) . the second solution to eq . ( [ perp_integrals ] ) in the small @xmath7 limit gives an order parameter frequency _ outside _ of the band , @xmath438 . for small @xmath7 the frequency is given by@xmath439 and the onset of synchronization occurs at@xmath440 the solution that grows from this instability is a remarkable state in that it is _ synchronized _ in the sense that there is a nonzero value of the order parameter @xmath441 , but there are no oscillators _ frequency locked _ to one another or to the frequency of the order parameter : a plot of the actual frequency distribution as in figs . [ fig_smr_plateau](a ) and [ fig_lgr_plateau](a ) shows a smooth curve with no plateau . numerical investigation of this state shows that instead , the distribution of oscillators @xmath442 is enhanced for phases near the phase of the order parameter : oscillators slow down in this vicinity ( i.e.@xmath443 becomes smaller ) , but do not come to rest relative to the order parameter . now the ordered oscillator frequency plot is continuous , with no plateau of locked oscillators . [ tbh ] such that @xmath99 . the solid and dashed lines are from the linear stability analysis of the unsynchronized state : the unsynchronized state is unstable for @xmath10 values between the two portions of these lines . the numerical results show that the solid portion is supercritical , whereas the dashed portion is subcritical . dash - dotted lines are saddle - node bifurcations from numerical simulations . the dotted line is the stability boundary of the fully locked solution calculated using the methods of sec . [ sec_fulllocking ] . states are as in fig . [ lorentzianpd ] with , in addition , @xmath444 denoting the fully locked state.,width=316 ] the linear instability boundaries for a top - hat distribution of full width @xmath445 ( giving @xmath99 ) are shown as the solid and dashed lines in fig . [ th_pd ] . the approximate solutions eqs . ( [ u - approx1]-[u - approx4 ] ) turn out to be quite accurate even for @xmath445 : the approximate curves are indistinguishable from the numerical curves in fig . [ th_pd ] . note that unlike the case of the lorentzian distribution , the linear instability persists for @xmath446 . in this limit the frequency of the order parameter is right at the edge of the distribution of oscillator frequencies . the discontinuity of the distribution of oscillator frequencies at the band edge seems to be responsible for the persistence of the instability to small values of the coupling constants . [ tbh ] over a range of @xmath10 is shown for constant @xmath8 : ( a ) @xmath447 , ( b ) @xmath448 , ( c ) @xmath449 , and ( d ) @xmath450 . in these simulations @xmath67 and the width @xmath445.,width=316 ] the boundary of the fully - locked solution calculated using the analysis of sec . [ sec_fulllocking ] is also shown on fig.[th_pd ] as the dotted line . in addition we have performed a careful numerical scan of the @xmath451 plane for @xmath67 or @xmath452 oscillators with a uniform distribution of full width unity . these results confirm the analytic predictions , and again show additional transitions that are inaccessible to our analytic calculations . the numerics shows that the linear instability from the unsynchronized state is subcritical over the dashed portion of the linear instability curves as shown in the figure . other saddle - node bifurcations are shown as dash - dotted lines . the complete phase diagram is quite complicated . some representative numerical sweeps are shown in fig . [ fig_cuts_tophat ] . [ tbh ] oscillators having a triangular frequency distribution with @xmath99 . the time - averaged order parameter magnitude @xmath115 is plotted over a range of @xmath10 at fixed @xmath8 : ( a ) @xmath453 , ( b ) @xmath454 , and ( c ) @xmath455.,width=316 ] we finally present results for when the oscillators have a triangular frequency distribution@xmath102{cc}$(4/w^{2})(w/2-\left\vert \omega\right\vert ) $ & for $ \left\vert \omega\right\vert \leq w/2$\\ $ 0 $ & for $ \left\vert \omega\right\vert > w/2$\end{tabular } \ \ \ \ \ \ .\ \right . \label{triangular}\ ] ] this is the case studied in reference @xcite . the triangular distribution is bounded , and so can have a fully locked state as for the top - hat distribution , but does not have the discontinuity in @xmath6 at the edge of the distribution leading to singular behavior as @xmath456 for that distribution . the results are compiled from the linear stability analysis of the unsynchronized and fully locked state , as well as numerical investigation , usually of @xmath457 oscillators , as shown in fig . [ fig_tri_cuts ] . the integrals eqs . ( [ integrals ] ) needed for the stability analysis of the unsynchronized state can again be done in closed form , but the results are too cumbersome to list here . [ tbh ] such that @xmath99 . symbols and lines are as in fig . [ th_pd].,width=316 ] the results for @xmath458 for @xmath99 ( a width @xmath375 ) given by numerically solving eqs . ( [ perp_integrals ] ) and ( [ par_integrals ] ) are shown as the full and dashed lines in fig . [ tri_pd ] . simulations show that this transition is subcritical over the dashed portions . in addition , as in fig . [ th_pd ] , saddle - node bifurcations identified by jumps in the order parameter magnitude @xmath24 in the simulations are shown by dash - dotted curves . as for the top - hat distribution , the large @xmath10 limit of the instability region corresponds to an order parameter frequency _ outside _ of the band of shifted oscillator frequencies , so that there is no oscillator with a frequency equal to that of the order parameter . this seems to be the typical result for a bounded distribution . for the unbounded lorentzian distribution on the other hand , there are some oscillators that lock to the order parameter frequency at the large @xmath10 instability . however even for this distribution , for most values of @xmath8 the order parameter frequency is far in the tails of the distribution , and there are few oscillators at this frequency . presumably the effect of these locked oscillators on the critical @xmath10 is small , so that the dominant physics at this transition for bounded and unbounded distributions is not very different . near the smallest @xmath8 where the upper and lower stability boundaries meet , the order parameter frequencies of the two solutions approach the edge of the band ( from either side ) . in summary , we have analyzed in detail a model for the synchronization of nonlinear oscillators where the ordering arises from the reactive coupling between the oscillators , combined with the nonlinear frequency pulling of the individual oscillators . such a model may be a more realistic description than previous models for a variety of physical systems where dissipation plays a relatively minor role , for example high - q mechanical oscillators and some optical systems . more generally , the model may give a more complete description of synchronization where the individual frequencies are internally tuned in response to a mismatch with other frequencies in the ensemble , such as might occur in biological systems . we have presented detailed analytic calculations for the onset of partial synchronization from the unsynchronized state , as well as the existence and bounds of the fully locked synchronized state at large coupling and nonlinearity for the cases of bounded frequency distributions . the analytical calculations together with numerical simulations have been used to construct detailed phase diagrams of the different synchronized states as a function of the two parameters of the model , the coefficient of the nonlinear frequency pulling @xmath8 and the coupling constant @xmath10 , for various frequency distributions . the intersections of the various synchronized states lead to rich phase diagrams . there are a number of interesting features of these phase diagrams . the instability of the unsynchronized state occurs over a limited range of @xmath10 at fixed @xmath8 , so that the unsynchronized state regains stability at very large values of the coupling strength , although a large amplitude synchronized state also occurs here , often terminated by a saddle - node bifurcation as @xmath10 is decreased . this large amplitude synchronized state may also survive down to @xmath447 , i.e. even in the absence of the nonlinear frequency pulling . for bounded distributions the state that develops from the linear instability of the unsynchronized state as @xmath10 is decreased is synchronized ( the order parameter is nonzero ) but there is no frequency locking . the phase diagrams also show a wide variety of multistability , with one or more synchronized states and the unsynchronized state coexisting over various parameter ranges , leading to hysteresis as the parameters are varied . the multistability may be particularly dramatic for the large amplitude synchronized states such as displayed in fig . [ fig_zig_zag ] where many synchronized states coexist , leading to a band of solutions in the large @xmath25 limit . this paper is based upon work supported by the national science foundation under grant no . dmr-0314069 , hrl laboratories , llc under contract no . sr04209 , the u.s .- israel binational science foundation grant no . 2004339 , and the physbio program with funds from the european union and nato .

the statistical mechanics of two polymers with constraints on their winding angle has been extensively studied in order to understand the behavior of physical polymer systems , like for instance biological macromolecules of dna @xcite or liquid crystals composed of stacks of disk - shaped molecules @xcite , see refs . a detailed review on the subject , together with interesting proposals of how to include in the treatment of topologically entangled polymer link invariants which are more sophisticated than the winding number , can be found in @xcite . up to now , however , despite many efforts , mainly ideal polymer chains or loops winding around each others have been considered , while the repulsive interactions between the monomers have been treated approximatively or exploiting in a clever way scaling arguments integrated by numerical simulations , as for instance in @xcite . here we concentrate ourselves on the case of two directed polymers interacting via a repulsive dirac @xmath0function potential @xcite . we are particularly interested in the average degree of entanglement of the system , which we wish to estimate by computing the square average winding angle of the two polymers . this quantity is also called second moment of the winding angle or simply second moment and is a special example of the topological moments first introduced in ref . @xcite . to achieve our goals , we develop an approach , which combines quantum mechanical and field theoretical techniques . with respect to previous works , we are able to obtain exact results even if the repulsive interactions are not switched off . in principle , the average of any observable like the squared winding angle can be derived once the partition function of the system is known , but in our case it turns out that the partition function is simply too complicated to obtain any analytical result . this happens essentially because the full @xmath0function potential is not a central potential , since it mixes both radial and angular variables . for this reason , the usual procedure of going to polar coordinates and then solving the differential equation satisfied by the partition function of the entangled polymers with the method of separation of variables @xcite , does no longer produce simple formulas as in the situations in which only central forces are present . to avoid these difficulties , one possibility is to approximate the @xmath0function potential with some radial potential , like for instance the hard core potential of ref . however , here we shall adopt a different strategy , based on field theories , which does not require any approximation . this strategy has been developed in @xcite ( see also ref . @xcite for more details ) to cope with ideal closed polymers whose trajectories are concatenated . also such systems are characterized by a non - central potential , which comes out as a consequence of the topological constraints imposed on the trajectories . in the field theoretical formulation of refs . @xcite the computation of the second moment is reduced to the problem of computing some correlation functions of a field theory . a bonus is provided by the fact that this computation requires just a finite number of feynman diagrams to be evaluated . in the present case , due to the presence of the @xmath0function potential , the field theory which we obtain is no longer free as that of refs . nevertheless , we will see that the theory is still linear and thus it can be exactly solved once its propagator is known . luckily , this propagator may be computed exactly using powerful non - perturbative techniques developed in the context of quantum mechanics to deal with hamiltonian containing @xmath0function potentials , see refs . basically , starting from the green function of a particle whose dynamics is governed by a given hamiltonian @xmath1 , these techniques provide an algorithm to construct the green function of a particle corresponding to a _ perturbed _ hamiltonian @xmath2 , where @xmath3 is the @xmath0function potential . one advantage of these methods is that there is a long list of potentials for which the green functions of the unperturbed hamiltonians @xmath4 are known . in this way , it is easy to generalize our treatment including new interactions , which could be relevant in polymer physics , like for instance the coulomb interaction . the price to be paid is that the quantum - mechanical algorithm works when the green functions are expressed as functions of the energy instead of the time . in the polymer analogy , assuming that the ends of the polymers are attached to two planar surfaces perpendicular to the @xmath5axis and located at the positions @xmath6 and @xmath7 , the role of time is played by the distance @xmath8 , while the energy corresponds to the chemical potential conjugated @xmath8 . to recover the original dependence on @xmath8 , one needs to calculate an inverse laplace transform of the field propagator with respect to the energy . in general , this is not a simple task . once the propagator of the linear field theory is known , the correlation functions which enter in the expression of the second moment may be calculated contracting the fields in all possible ways using the wick prescription . at the end , we get in this way an exact formula of the second moment as a function of the _ energy _ , which , we remember , has here the meaning of the chemical potential conjugated to the distance @xmath8 . in the @xmath8 space , due to the problems of computing the inverse laplace transform of the propagator mentioned above , only an approximated expression of the second moment will be given in the limit of large values of @xmath8 and assuming that the strength of the @xmath0function potential is weak enough to allow a perturbative approach . our results allow both a qualitative and quantitative understanding of the way in which the repulsive interactions affect the entanglement of two directed polymers . the corrections introduced by these interactions in the expression of the second moment of ideal polymers have been studied in some interesting limits . first of all , it has been examined the limit of long polymer trajectories , in which we show that repulsive interaction become particularly relevant . moreover , we have investigated also the perturbative regime and the strong coupling limit , which is important to recover the excluded volume interactions . while it is not a problem to take the strong coupling limit within our exact treatment of the repulsive interactions , it turns out that , in this case , the expression of the second moment is particularly complicated from the analytical point of view . for this reason , in the conclusions we will discuss the application of a powerful perturbative method to study field theories at strong coupling due to kleinert @xcite . finally , the consistency of our results with the previous ones has been checked by studying the limit of ideal polymers . the material presented in this paper is divided as follows . in the next section , the problem of computing the second moment of the winding angle of two directed polymers interacting via a @xmath0function potential is briefly illustrated using the path integral approach . a constraint on the winding angle is imposed by coupling the trajectories of the polymers with a suitable external magnetic field , following the strategy of previous works like for instance @xcite . in section [ sec : three ] the second moment is expressed in the form of a finite sum of amplitudes of a linear field theory . these amplitudes may be computed once the propagator of the theory is constructed . in our case , the propagator coincides with the green function of a particle diffusing in a @xmath0function potential . the derivation of this green function in the energy representation using non - perturbative techniques developed in the context of quantum mechanics @xcite is the subject of section [ sec : four ] . the @xmath0function potential is responsible of the appearance of singularities in the propagator at short distances , which have been regulated here with the introduction of a cut - off . this procedure is motivated by the fact that in polymer physics there is no point in considering distances which are smaller than the dimensions of a monomer . a comparison with the more rigorous method of renormalization is made , showing the consistency of the two procedures . the propagator derived in section ( [ sec : four ] ) has a particularly nice form , in which the contributions coming from the repulsive forces can be separated from the free part of the propagator , which is related to the random walk of ideal polymers . this splitting of the propagator is used in section [ sec : five ] to discuss qualitatively and qualitatively the effects of the @xmath0function interactions on the entanglement of the system . the results of sections [ sec : three ] and [ sec : four ] provide in principle all the ingredients of the second moment . however , the amplitudes of the linear field theory derived in section [ sec : three ] should still be evaluated . in this task one encounters the typical problems occurring in the evaluation of the analytical expressions of feynman diagrams . in the case of the second moment there are just tree diagrams , but still one has to perform complicated integrations over the spatial coordinates which are transverse to the @xmath5axis . even assuming that polymers are ideal , the analytical evaluation of these integrations requires drastic approximations , see for instance @xcite . to avoid these difficulties , we average the second moment with respect to the positions of the endpoints of the two polymers . this averaged version of the second moment can be computed without any approximation in the energy representation . this is done in section [ sec : six ] . the expression of the averaged second moment in the @xmath9space is provided instead only at the first perturbative order in the strength of the repulsive potential and assuming additionally that the value of @xmath8 is large . we give also an exact formula of the second moment without performing any averaging procedure as a function of @xmath8 . this formula is however explicit only up to the calculation of the inverse laplace transform of the propagator derived in section [ sec : four ] . in section [ sec : seven ] we consider the situation in which the polymers are not interacting in order to allow the comparison with previous results . finally , the discussion of the obtained results and ideas for further developments are presented in the conclusions . our starting point is the action of two directed polymers @xmath10 and @xmath11 : @xmath12 \label{actnorm}\ ] ] where @xmath13 is the potential : @xmath14 the sign of @xmath15 has been chosen in such a way that the interaction associated to the potential @xmath16 is repulsive . the parameters @xmath17 and @xmath8 determine the average length of the trajectories of the polymers . the ends of the polymers are supposed to be fixed on two surfaces perpendicular to the @xmath5axis and located at the heights @xmath6 and @xmath7 . both polymers have a preferred direction along the @xmath18 direction . the vectors @xmath19 and @xmath20 , @xmath21 , measure the polymer displacement along the remaining two directions of the space . the action of eq . ( [ actnorm ] ) resembles that of two quantum particles in the case of imaginary time @xmath18 . to stress these analogies with quantum mechanics , the @xmath5variable will be treated as a pseudo - time and renamed using from now on the letter @xmath22 instead of @xmath18 . in the system of the center of mass : @xmath23 the action ( [ actnorm ] ) becomes : @xmath24 \label{cmact}\ ] ] the motion of the center of mass , which is a free motion described by the coordinate @xmath25 , will be ignored . we consider the partition function of the above two - polymer system with the addition of a constraint on the entanglement of the trajectories : @xmath26 } \delta(m-\chi ) \label{totparfun}\ ] ] @xmath27 is the so - called winding angle . its expression is given by : @xmath28 where @xmath29 is a vector potential with components : @xmath30 in the above equation we have represented the vector @xmath31 using cartesian coordinates @xmath32 . i. e. @xmath33 . moreover , from now on , middle latin indices @xmath34 will label the directions which are perpendicular to the @xmath35axis . the definition of the partition function @xmath36 is completed by the boundary conditions at @xmath37 and @xmath38 : @xmath39 the quantity in eq . ( [ wnum ] ) becomes a topological invariant if the polymer trajectories are closed . in the present case , in which the trajectories are open , @xmath27 just counts the angle with which one polymer winds up around the other . thus , the partition function @xmath36 gives the formation probability of polymer paths winding up of an angle @xmath40 exploiting the fourier representation of dirac @xmath0functions @xmath41 eq . ( [ totparfun ] ) can be rewritten as follows : @xmath42 where @xmath43 the lagrangian @xmath44 is that of a particle immersed in the magnetic potential associated to the vector field ( [ vecpot ] ) : @xmath45 the fourier transformed partition function @xmath46 is the grand canonical version of the original partition function @xmath36 , in which the number @xmath47 is allowed to take any possible value . @xmath46 coincides with the propagator @xmath48 , which satisfies the following pseudo - schrdinger equation : @xmath49{\cal g}_\lambda(l;{{\mbox{\bf r}}}_1,{{\mbox{\bf r}}}_0 ) = 0 \label{pseudseq}\ ] ] @xmath50 is the hamiltonian of the system , computed starting from the lagrangian ( [ newaction ] ) : @xmath51 eq . ( [ pseudseq ] ) is completed by the boundary condition at @xmath52 : @xmath53 the average degree of entanglement of the two polymers can be estimated computing the topological moments of the winding angle @xmath54 , @xmath55 @xcite . once the partition function is known , the @xmath54 may be expressed as follows : @xmath56 the quantities @xmath54 depend on the boundary conditions @xmath57 and , of course , on the parameters @xmath17 and @xmath8 . for practical reasons , we will also consider the following _ averaged _ topological moments : @xmath58 as eq . ( [ aderr ] ) shows , the average is performed with respect to the relative positions @xmath57 of the endpoints . this is equivalent to an average over the positions of the endpoints @xmath59 at the instants @xmath37 and @xmath38 , because the coordinates of the center of mass have been factored out from the partition function and thus they do not play any role . the advantage of the averaged topological moments is that , a posteriori , it will be seen that their computation is easier than that of the topological moments given in eq . ( [ sectopmomdef ] ) . from the physical point of view , the averaged topological moments measure the entanglement of two polymers , whose ends at the instants @xmath37 and @xmath38 are free to move . here we will be interested only in the second moment @xmath60 and in the averaged second moment @xmath61 , i. e. in the case @xmath62 of eqs . ( [ sectopmomdef ] ) and ( [ aderr ] ) . the second moment is in fact enough in order to estimate the formation probability of entanglement with a given winding angle and to determine how the winding angle grows with increasing polymer lengths . in the following it will be useful to work in the so - called energy representation , i. .e considering the laplace transformed of the partition function @xmath48 with respect to @xmath8 : @xmath63 the new partition function @xmath64 describes the probability of two entangled polymers of any length subjected to the condition that the relative positions of the polymer end at the initial and final instants @xmath65 and @xmath66 are given by the vectors @xmath67 and @xmath68 . with respect to the formulation in the @xmath9 space , however , the distance @xmath69 is no longer exactly equal to @xmath8 , but is allowed to vary according to a distribution which is governed by the boltzmann - like factor @xmath70 . thus , @xmath71 plays the role of the chemical potential conjugated to the end - to - end distance of the polymer trajectories in the @xmath35direction . it is worth to remember that , roughly speaking , small values of @xmath71 correspond to large values of @xmath8 , while large values of @xmath71 correspond to small values of @xmath8 . starting from eq . ( [ pseudseq ] ) and recalling the boundary conditions ( [ bbdc ] ) , it is easy to check that @xmath64 satisfies the stationary pseudo - schrdinger equation : @xmath72{\cal g}_\lambda(e;{{\mbox{\bf r}}}_1,{{\mbox{\bf r}}}_0)=\delta({{\mbox{\bf r}}}_1-{{\mbox{\bf r}}}_0 ) \label{pseenerep}\ ] ] where @xmath50 is always the hamiltonian of eq . ( [ ham ] ) . in this section we wish to evaluate the expression of the second moment as a function of the energy @xmath71 using a field theoretical formulation of the polymer partition function . the starting point is provided by the formula of the second moment in the @xmath9space suitably rewritten in the following way : @xmath73 for consistency with eq . ( [ sectopmomdef ] ) , the numerator @xmath74 and the denominator @xmath75 appearing in eq . ( [ secmomnumden ] ) must be of the form : @xmath76 and @xmath77 using eq . ( [ laplrep ] ) , it is now straightforward to compute the laplace transform of @xmath74 and @xmath75 : @xmath78 @xmath79 once the functions @xmath80 and @xmath81 are known , one can construct the ratio : @xmath82 which is nothing but the second moment of the winding angle expressed as a function of the chemical potential @xmath71 . we remark that the green function @xmath64 is related to the feynman propagator of the spin@xmath83 aharonov - bohm problem in quantum mechanics . in principle , this green function can be computed exactly starting from eq . ( [ pseenerep ] ) @xcite , but its final expression is too complicated for our purposes . moreover , the method used in @xcite to renormalize the singularities coming from the presence of the @xmath0function potential is valid only in a restricted region of the domain of @xmath84 . this is incompatible with our requirements , because , to derive the second moment , one has to integrate @xmath64 with respect to @xmath84 over the whole real line . for this reason , we prefer here to use a field theoretical representation of this green function . this is achieved by noting that @xmath85 coincides with the inverse matrix element of the operator @xmath86 : @xmath87 and may be expressed in a functional integral form in terms of replica fields : @xmath88 in the above equation @xmath89 are multiplets of replica fields : @xmath90 with action @xmath91\psi \label{erepact}\ ] ] the symbol @xmath92 in eq . ( [ ftrep ] ) denotes summation over the replica index . for example @xmath93 . below it will be used also the convention @xmath94 . the details of the derivation of eq . ( [ ftrep ] ) can be found in previous publications on the subject @xcite and will not be provided here . in order to proceed , it will be convenient to expand the action ( [ erepact ] ) in powers of @xmath84 : @xmath95 where we have put : @xmath96 \label{szero}\ ] ] @xmath97 \label{sone}\ ] ] @xmath98 at this point we come back to the computation of the quantities @xmath80 and @xmath81 appearing in the expression of the second moment . exploiting the new form of the partition function given by eqs . ( [ ftrep][stwo ] ) , together with the relation @xmath99 and the fact that @xmath100 , it is possible to rewrite eqs . ( [ momno ] ) and ( [ momnt ] ) as follows : @xmath101e^{-s_0(\psi^*,\psi ) } \label{nsdo}\ ] ] @xmath102 the right hand sides of eqs . ( [ nsdo ] ) and ( [ nsdt ] ) represent vacuum expectation values of a field theory governed by the action @xmath103 of eq . ( [ szero ] ) . in the formulation in terms of quantum operators we have : @xmath104 @xmath105 the correlation functions have a subscript @xmath106 to remember that , according to the replica method , they should be computed first assuming that the number of replicas @xmath106 is an arbitrary positive integer and then taking the limit for @xmath106 going to zero . the above correlators may be evaluated using standard field theoretical methods . one could be tempted to use a perturbative approach assuming that the value of @xmath15 appearing in the action @xmath103 of eqs . ( [ nsdo ] ) and ( [ nsdt ] ) is small , but this is not necessary . as a matter of fact , if it is true that @xmath103 does not describe free fields because of the presence of the @xmath0function potential , it is also true that it is just quadratic in the fields . as a consequence , one is allowed to define a propagator @xmath107 associated with this action . it is easy to check that @xmath107 satisfies the equation : @xmath108g(e;{{\mbox{\bf x}}},{{\mbox{\bf y}}})=\delta({{\mbox{\bf x}}},{{\mbox{\bf y } } } ) \label{fdfjh}\ ] ] using the above propagator , one can evaluate the amplitudes in eqs . ( [ nsdoqop ] ) and ( [ nsdtqop ] ) exactly by contracting the fields in all possible ways according to the wick theorem . after straightforward calculations , one finds : @xmath109 @xmath110 @xmath111 where @xmath112\label{amptwoone}\\ i_2({{\mbox{\bf r}}}_1,{{\mbox{\bf r}}}_0 ) & = & + \frac1{2c^2}\int d^2xd^2y\left [ a_i({{\mbox{\bf x}}})(\nabla_{\mathbf x}^ig(e;{{\mbox{\bf x}}},{{\mbox{\bf r}}}_1))g(e;{{\mbox{\bf y}}},{{\mbox{\bf x } } } ) a_j({{\mbox{\bf y}}})(\nabla_{\mathbf y}^jg(e;{{\mbox{\bf r}}}_0,{{\mbox{\bf y}}}))\right]\label{amptwotwo}\\ i_3({{\mbox{\bf r}}}_1,{{\mbox{\bf r}}}_0)&= & + \frac1{2c^2}\int d^2xd^2y\left [ a_i({{\mbox{\bf x}}})g(e;{{\mbox{\bf r}}}_0,{{\mbox{\bf x}}})(\nabla_{\mathbf x}^i \nabla_{\mathbf y}^jg(e;{{\mbox{\bf x}}},{{\mbox{\bf y}}}))a_j({{\mbox{\bf y}}})g(e;{{\mbox{\bf y}}},{{\mbox{\bf r}}}_1)\right]\label{amptwothree}\\ i_4({{\mbox{\bf r}}}_1,{{\mbox{\bf r}}}_0 ) & = & -\frac1{2c^2}\int d^2xd^2y \left [ a_i({{\mbox{\bf x}}})g(e;{{\mbox{\bf r}}}_0,{{\mbox{\bf x}}})(\nabla_{\mathbf x}^ig(e;{{\mbox{\bf x}}},{{\mbox{\bf y } } } ) ) a_j({{\mbox{\bf y}}})(\nabla_{\mathbf y}^jg(e;{{\mbox{\bf y}}},{{\mbox{\bf r}}}_1))\right]\label{amptwofour}\end{aligned}\ ] ] from the physical point of view , the above equations may be interpreted in the following way . the fields @xmath113 and @xmath114 contain operators which , inside each replica sector , create and annihilate segments of the two polymers , whose relative positions are given by the vector @xmath115 . the two polymer system has been projected in the two - dimensional plane perpendicular to the @xmath35axis . for this reason , there appear only the transverse coordinates @xmath115 . the only remnant of the third dimension is the dependence on the energy @xmath71 . the correlation functions ( [ ampone ] ) and ( [ amptwo ] ) describe the fluctuations of the two polymers immersed in the @xmath0function potential and subjected to the interactions represented by the vector potential ( [ vecpot ] ) . we recall that the origin of the latter interactions is the presence of the constraint on the winding angle in the partition function ( [ totparfun ] ) . to evaluate the correlation functions ( [ ampone ] ) and ( [ amptwo ] ) , one needs to consider only a finite number of feynman diagrams , corresponding to the relevant processes with which the two polymers interact together . the result , after the analytical evaluation of these diagrams , is provided by eqs . ( [ amptwoone][amptwofour ] ) . let us note that in these equations the repulsive interactions due to the @xmath0function potential are hidden in the propagators @xmath107 . the feynman diagrams related to the amplitudes of eqs . ( [ amptwoone])([amptwofour ] ) are all topologically equivalent to the diagram of fig . [ processone ] . ) ([amptwofour ] ) . the two polymers @xmath10 and @xmath11 start at a distance @xmath116 from each other and interact twice with the the external field @xmath117 . at the end , the relative position of the end points at the instant @xmath38 is given by @xmath68 . the three - vertices appearing in the figure are related to the interaction described by eq . ( [ sone ] ) . , title="fig : " ] + the amplitude of eq . ( [ amponek ] ) is related instead to the feynman diagram of fig . [ processtwo ] . ) . the two polymers @xmath10 and @xmath11 start at a distance @xmath116 from each other and interact with the the external field @xmath117 . at the final instant @xmath38 , the relative position of the end points is given by @xmath68 . the four - vertex appearing in this figure is related to the interaction described by eq . ( [ stwo ] ) . , title="fig : " ] + the vectors @xmath118 and @xmath119 denote the relative positions of the end points of the two polymers at the initial and final instants , as already mentioned . the integration variables @xmath120 and @xmath121 appearing in eqs . ( [ amponek][amptwofour ] ) may be regarded as the vectors which give the relative positions of the trajectories of the two polymers at the instants in which they interact together via the external vector potential @xmath122 of eq . ( [ vecpot ] ) . there is no restriction on the domain of integration of @xmath120 and @xmath121 , so that the components of these relative vectors are allowed to take any value . this implies that the distance between the polymer segments when the interaction with @xmath122 occurs can be arbitrarily large . now that the correlation functions which are present in the expressions of @xmath123 and @xmath124 given in eqs . ( [ nsdoqop ] ) and ( [ nsdtqop ] ) have been evaluated , see eqs . ( [ amponek][amptwofour ] ) , we may put everything together and give to the second moment of eq . ( [ secmomerep ] ) a more explicit form : @xmath125 in conclusion , the initial problem of computing the second moment of the winding angle @xmath60 has been reduced to the evaluation of a finite number of integrals , which are given in eqs . ( [ amponek][amptwofour ] ) . of course , to make these integrals really explicit , we still need to derive the propagator @xmath107 , which is so far the only missing ingredient . this will be done in the next section . let @xmath126 be the solution of the differential equation : @xmath127 for a given hamiltonian @xmath4 . when @xmath52 , @xmath126 satisfies the boundary condition : @xmath128 in the case of a hamiltonian @xmath50 , obtained by adding to @xmath1 a @xmath0function potential as a perturbation : @xmath129 we consider the analogous differential problem : @xmath130 @xmath131 we wish to compute @xmath132 starting from the green function @xmath126 , which is supposed to be known . it is possible to show that @xmath132 and @xmath126 are related by the integral equation @xcite : @xmath133 we see that in the right hand side of the above equation the presence of the @xmath0function forces us to consider the functions @xmath126 and @xmath132 evaluated at the points @xmath134 and/or @xmath135 . usually , at these points green functions may be not well defined due to the presence of singularities . a concrete procedure to remove these singularities will be indicated later . for the moment , we go further with formal manipulations , assuming that some kind of consistent regularization of the possible divergences has been introduced . first of all , we perform the integration over @xmath136 in eq . ( [ gkk ] ) : @xmath137 the integral in @xmath138 appearing in the right hand side of eq . ( [ beflap ] ) is a convolution which can be better treated after a laplace transform . thus , we transform both sides of this equation with respect to @xmath8 : @xmath139 where @xmath140 and @xmath141 at this point , it is easy to extract from eq . ( [ lapltragf ] ) the expression of @xmath107 : @xmath142 the above formula may be used in order to solve eq . ( [ fdfjh ] ) . in this case , @xmath4 coincides with the free action : @xmath143 and the function @xmath144 is given by : @xmath145 here @xmath146 denotes the modified bessel function of the second kind of order zero . clearly , we can not apply directly eq . ( [ finald ] ) without introducing a regularization . as a matter of fact , if not treated , the naive denominator in the second term of the right hand side is equal to infinity , i. e. @xmath147 . this is due to the fact that @xmath146 diverges logarithmically in the limit @xmath148 : @xmath149 a natural regularization is suggested by the fact that , in polymer physics , it has no sense to consider lengths which are smaller than the size of the molecules which compose the polymers . thus , it seems reasonable to regulate ultraviolet divergences by introducing a cut - off @xmath150 at short distances . the length @xmath150 is comparable with the molecular size . according to this prescription , by inserting the green function of eq . ( [ gzspec ] ) in eq . ( [ finald ] ) , we obtain : @xmath151 the symbol @xmath152 means that the quantity in the left hand side of an equation has been replaced in the right hand side with its regulated version . the above green function is what we need in order to evaluate explicitly the amplitudes of eqs . ( [ ampzero][amptwo ] ) . the infinities coming from the @xmath0function potential should be treated with some care in order to avoid ambiguities . for this reason , we would like to compare the naive prescription used here to derive eq . ( [ gffin ] ) with the more rigorous procedure of renormalization . it is known in fact that the renormalization of the infinities coming from @xmath0function interactions produces physically sensible results @xcite . the divergences will be regulated introducing a cut - off @xmath153 in the momentum space . as a consequence , it will be convenient to express the free green function of eq . ( [ gzspec ] ) in momentum space . to this purpose , we use the following formula : @xmath154 to evaluate the green function at the singular point @xmath155 we need to compute the following divergent integral : @xmath156 using the above cut - off prescription to eliminate the ultraviolet singularities we get , in the assumption @xmath157 : @xmath158 now , according to the spirit of renormalization , we subtract the infinities from the bare parameters of the theory . in our case , after choosing an arbitrary mass scale @xmath159 , which gives the renormalization point , we renormalize the bare coupling constant @xmath15 . actually , it will be better to call it @xmath160 instead of @xmath15 in order to distinguish it from the effective coupling constant @xmath15 appearing in eq . ( [ gffin ] ) . the subtraction of infinities is performed in such a way that the quantity : @xmath161 becomes finite . we choose a sort of minimal subtraction scheme , in which the renormalized coupling constant @xmath162 is related to the bare coupling constant @xmath160 as follows : @xmath163 applying the last two above equations back to eq . ( [ finald ] ) , we get as a result : @xmath164 eqs . ( [ gffin ] ) and ( [ gffinren ] ) are reciprocally compatible . in fact , since @xmath150 is very small , because it is the smallest possible length scale in our polymer problem , one can use the following approximation ( see eq . ( [ fdsf ] ) ) in the denominator of the second term of eq . ( [ gffin ] ) : @xmath165 comparing with the analogous denominator in eq . ( [ gffinren ] ) , it is possible to relate @xmath150 with the inverse of the mass @xmath159 : @xmath166 moreover , the effective coupling constant @xmath15 of eq . ( [ gffin ] ) may be identified with the renormalized coupling constant @xmath162 , which gives the strength of the repulsive interaction ( [ dirpot ] ) at distance scales of order @xmath150 . before concluding this section , we make a small digression about the translational invariance of the free hamiltonian ( [ freeham ] ) and consequently of the free green function ( [ gzspec ] ) . clearly , this is not the same translational invariance that was already present in the original action ( [ actnorm ] ) due to the translational invariance of the potential ( [ dirpot ] ) . this new invariance is rather related to the fact that the physics of the two polymer system in the absence of any interaction does not change if we modify the relative positions of the polymer ends at @xmath37 and @xmath38 in a symmetric way . an example of such transformations is the translation of both ends of polymer @xmath10 at the initial and final points by a constant vector @xmath167 : @xmath168 as a result of the translations ( [ strazerexone][straelexone ] ) , the relative vector @xmath169 of eq . ( [ cmcoor ] ) at the instants @xmath37 and @xmath38 changes as follows : @xmath170 clearly the propagator ( [ gzspec ] ) is invariant under the above transformations . this kind of invariance can be explained as follows . as far as the two polymers @xmath10 and @xmath11 do not interact , each of them may be treated as an independent system . if we translate for instance both ends of polymer @xmath10 at @xmath37 and @xmath38 in the symmetrical way shown by eqs . ( [ strazerexone ] ) and ( [ straelexone ] ) , the number of configurations of polymer @xmath10 and consequently the configurational entropy of the whole system do not change , because the transformation is equivalent to a translation of polymer @xmath10 in the space . of course , this invariance disappears as soon as the two polymers start to interact or if they are entangled together . indeed , if one adds to the free hamiltonian ( [ freeham ] ) a @xmath0function potential , the propagator stops to be translational invariant as shown by the green function of eq . ( [ finald ] ) , which does not depend on the difference @xmath171 . in principle we have at this point all the ingredients which are necessary to compute the second moment of eq . ( [ secmomerep ] ) . in eqs . ( [ nsdoqop ] ) and ( [ nsdtqop ] ) , in fact , the quantities @xmath80 and @xmath81 are written as linear combinations of the amplitudes of eqs . ( [ ampzero][amptwo ] ) , which can be explicitly evaluated using the propagator @xmath172 given in eq . ( [ gffin ] ) appearing in the definition of the propagator of eq . ( [ gffin ] ) are replaced by the vectors @xmath173 . this is to avoid confusions with the notation of eqs . ( [ amponek][amptwofour ] ) , where the same pair of vectors @xmath174 has been exploited to denote dummy integration variables . ] and the formulas of eqs . ( [ amponek][amptwofour ] ) . the remaining task is to perform the integrations over the coordinates @xmath115 and @xmath175 in eqs . ( [ amponek][amptwofour ] ) . from the analytical point of view , the evaluation of these integrals poses severe technical problems , which can be solved only with the help of drastic approximations . however , the difficulties become milder if we average the second moment over the endpoints of the polymers as shown in eq . ( [ aderr ] ) . in the energy representation , which we are using , this means that we have to consider the following averaged version of the second moment in eq . ( [ secmomerep ] ) : @xmath176 where @xmath177 accordingly , we need to integrate the quantities @xmath178 and @xmath179 , @xmath180 of eqs . ( [ amponek][amptwofour ] ) with respect to @xmath68 and @xmath67 . putting : @xmath181 we obtain from eqs . ( [ ampone ] ) and ( [ amptwo ] ) the following expressions of @xmath182 and @xmath183 : @xmath184 it will also be convenient to split the propagator @xmath185 of eq . ( [ gffin ] ) into two contributions : @xmath186 where @xmath187 is the free propagator of eq . ( [ gzspec ] ) , which is invariant with respect to the transformations ( [ strainvzer ] ) and ( [ strainvone ] ) , while @xmath188 in the above equation we have isolated in the expression of @xmath189 the factor : @xmath190 it is clear that the origin of the term @xmath189 in the propagator is due to presence of the @xmath0function interaction ( [ dirpot ] ) in the polymer action ( [ actnorm ] ) . in fact , if @xmath191 , this term vanishes identically . thus , using the splitting of the propagator of eq . ( [ propsplit ] ) , it is now possible to separate in eqs . ( [ amponek][amptwofour ] ) the contributions to entanglement given by repulsive forces . it seems natural to expand the quantities @xmath183 , @xmath192 and @xmath193 defined in eqs . ( [ deee ] ) , ( [ koneee ] ) and ( [ ionee ] ) with respect to @xmath194 as follows : @xmath195 where the superscript @xmath196 , with @xmath197 , denotes the order in @xmath194 . there are no higher order terms with @xmath198 , so the above expansions are exact . it is easy to show how @xmath192 and the @xmath193 s depend on the pseudo - energy @xmath71 . after a rescaling of the integration variables @xmath199 and @xmath175 in eqs . ( [ ennee ] ) and ( [ deee ] ) , one finds in fact that : @xmath200 @xmath201 where the factors @xmath202 s and the @xmath203 s are functions of the parameters @xmath150 and @xmath17 , but not of @xmath71 or @xmath15 . in fact , the coupling constant @xmath15 appears only inside the powers of @xmath194 . let us note in eqs . ( [ behaone ] ) and ( [ behatwo ] ) the presence of the overall factor @xmath204 in eqs . ( [ behaone ] ) and ( [ behatwo ] ) . looking at eq . ( [ enneeexp ] ) , it is clear that the whole function @xmath182 is characterized by the leading scaling behavior @xmath205 . in the @xmath9space , after an inverse laplace transform , this behavior corresponds to the following scaling law , which is typical of ideal polymers : @xmath206 . the powers of @xmath194 , appearing in the expressions of @xmath207 and @xmath208 , introduce corrections to this leading behavior that are at most logarithmic in @xmath71 . as a matter of fact , if the condition @xmath209 is satisfied , we have that : @xmath210 naively , the above seems the only logarithmic correction which is possible in the expressions of @xmath182 and @xmath183 when @xmath71 is small . however , that this is not true . in fact , in deriving eqs . ( [ behaone ] ) and ( [ behatwo ] ) , we have not considered the divergences which arise in some of the integrations over the variables @xmath211 and @xmath68 . after regulating these divergences with some prescription , as for instance the ultraviolet cut - off @xmath150 used in eq . ( [ gffin ] ) , we will see in section [ sec : six ] that the naive rescaling of variables exploited in order to obtain eqs . ( [ behaone ] ) and ( [ behatwo ] ) does no longer work and one should add extra logarithmic factors to these equations . ( [ behaone ] ) and ( [ behatwo ] ) may be also useful to study the case of polymers in confined geometries . as a matter of fact , for large values of @xmath71 , one recovers the limit of small values of @xmath8 , in which the region between the initial and final height is very narrow . looking at eqs . ( [ behaone ] ) and ( [ behatwo ] ) , it is clear that the only interesting corrections when @xmath71 is large come from the powers of @xmath194 . to evaluate these corrections , one should note that the modified bessel function of the second kind @xmath146 goes very fast to zero for large values of @xmath18 . as a consequence , already in the domain of parameters in which @xmath212 , it is possible to make the very interesting approximation @xmath213 unfortunately , it turns out that the values of the energy for which the above equation is satisfied are not physical , as it will be shown below . other useful information on the influence of repulsive forces on the winding angle can be obtained studying the form of the function @xmath189 of eq . ( [ nontrainvpro ] ) . we remember in fact that all the effects of the repulsive forces are concentrated in this component of the propagator . supposing for example that the value of @xmath214 is very large , i. e. : @xmath215 we have the following approximate expression of @xmath216 : @xmath217 a relation analogous to eq . ( [ appexplx ] ) may be written also for the variable @xmath218 . in practice , ( [ appexplx ] ) means that the repulsive interactions do not play any particularly relevant role in polymer configurations in which the ends of the trajectories at some point are very distant . this is not a surprise . if the trajectories at some height @xmath22 are very far from each other , they will have little or no chance to interact together via the repulsive interactions of eq . ( [ dirpot ] ) , which are of short range . ( [ appexplx ] ) gives the concrete law with which the contributions of the repulsive forces are suppressed in configurations of this kind . in particular , if the distance between the trajectories is much greater than the _ characteristic length scale _ @xmath219 the influence of the repulsive forces ceases to be relevant . of course , even if at some points the trajectories are very distant , polymers will always have a chance to get near enough to be able to interact if they are sufficiently long . as a consequence , we expect that the characteristic length @xmath220 increases with the increasing of the lengths of the trajectories . it is easy to check that this is exactly the case . to show that , let us consider the dependence of @xmath220 on the polymer length . one parameter which determines this length is the distance @xmath8 between the ends of the polymers along the @xmath35axis . indeed , a trajectory connecting the two ends of a polymer must be very long if these ends are located at very distant heights . in the energy representation , large values of @xmath8 correspond to small values of @xmath71 . for example , in the limit @xmath221 , which corresponds to infinite polymer lengths , we have that @xmath222 , confirming our intuitive expectations . another confirmation comes from eq . ( [ randisell ] ) below , where a rough estimation of the behavior of @xmath220 with respect to the distance @xmath8 is given . the dependence on @xmath8 is not the whole story . as a matter of fact , during their random walk in the @xmath35direction from @xmath37 to @xmath38 , polymers are also allowed to wander in the remaining two directions . loosely speaking , the variations in the length of the trajectories associated to the fluctuations in these transverse directions are taken into account by the parameter @xmath17 . smaller values of @xmath17 correspond to longer trajectories and vice - versa , see @xcite . it is now easy to realize from eq . ( [ chalensca ] ) that the characteristic length @xmath220 increases when @xmath17 decreases as expected . taking into account all the above considerations , it is possible to conclude that repulsive forces give relevant contributions to the second moment only in the case of configurations of the system in which the trajectories of the two polymers are not too far from each other . as a matter of fact , in the propagators appearing in the amplitudes of eqs . ( [ amponek][amptwofour ] ) all configurations in which the distance between the trajectories at some height in the @xmath35axis is bigger than a few characteristic lengths @xmath220 are exponentially suppressed according to eq . ( [ appexplx ] ) . one may also add that this suppression becomes milder in the case of long polymers , because we have seen that the characteristic lengths @xmath220 grows with the length of the polymers with a law which has been given in eq . ( [ randisell ] ) . in the rest of this section we will analyze some interesting limiting cases , in which repulsive interactions become particularly weak or strong . to this purpose , it would be appealing to consider the quantity @xmath223 , where @xmath194 has been given in eq . ( [ lambe ] ) , as an energy dependent effective or running coupling constant of the repulsive interactions . this could be suggested by the expansions of eqs . ( [ dexp][expioe ] ) and by the fact that the quantity @xmath223 has the right dimension to be a coupling constant . indeed , we will see that there are cases in which the strength of @xmath194 really determines the strength of the repulsive forces . however , this is not true in general , as it should be because @xmath194 is just a parameter which has been factored out from the expression of @xmath189 and thus its meaning does not coincide with that of a running coupling constant . keeping that in mind , we start to study the perturbative regime , in which @xmath15 is very small . in the part of the propagator in which there are the contributions of the repulsive forces , i. e. the function @xmath189 , @xmath15 is present only inside @xmath194 . expanding this quantity in powers of @xmath15 , we obtain : @xmath224 we see that , at the leading order in @xmath15 , @xmath194 is proportional to @xmath15 and thus , as it could have been expected , @xmath189 may be treated as a small perturbation with respect to the free propagator @xmath225 . let s now go back to eq . ( [ fapp ] ) . in that equation it turns out that @xmath194 has the same behavior as in the perturbative regime , even if eq ( [ fapp ] ) has been derived in the hypothesis that @xmath212 , but without supposing that @xmath15 is small . before dwelling upon the physical meaning of this coincidence , let s see what is the significance of the condition @xmath212 . to this purpose , we make the following approximations : @xmath226 as mentioned before , it is quite reasonable to assume that the length @xmath8 is proportional to the inverse of the energy @xmath71 , while the second approximation implies that polymers are very flexible . for example , in polyethylene the kuhn length@xmath227 is of the order of molecular sizes . exploiting eq . ( [ draapp ] ) , it turns out that the condition @xmath212 is equivalent to the condition @xmath228 . this would mean that our system is squeezed in a volume whose height @xmath8 is smaller than the size of a monomer . clearly , this situation is not very physical . since we have been able to compute the exact form of the propagator @xmath229 , it is not difficult to study also the strong coupling limit @xmath230 . as in the perturbative case , the only affected part of the propagator ( [ gffin ] ) is the factor @xmath194 appearing in @xmath189 . after a trivial calculation , one finds that , in the strong coupling limit , the form of @xmath189 is given by : @xmath231 assuming that polymers are very long , let us study the left hand side of the above equation . this is a ratio of modified bessel functions of the second kind . since @xmath150 is a very small quantity and these functions have a logarithmic singularity if their argument is small , see eq . ( [ fdsf ] ) , it is licit to suppose that @xmath232 unless @xmath233 and/or @xmath234 . on the other side , we know from eq . ( [ cslm ] ) that , if @xmath235 decays exponentially . in other words , in the left hand side of eq . ( [ scnontrainvpro ] ) the denominator will dominate over the numerator whenever the distance between the polymer trajectories is not of the order of a few molecular sizes . thus , if @xmath15 is large , the major contributions to winding angle coming from the repulsive interactions occur when the trajectories are very near to each other . this could be expected from the fact that , in the strong coupling limit , one recovers the excluded volume interactions . finally , let us study the domain of the parameters @xmath71 and @xmath17 in which the condition @xmath236 is verified . we will see that this domain is particularly interesting , because if condition ( [ condd ] ) is verified , the corrections of the repulsive interactions to the entropy dominated behavior of ideal polymers become relevant . it has been already shown that under the assumption made in eq . ( [ condd ] ) , the parameter @xmath194 is approximated as in eq . ( [ logcorr ] ) . even if it is not strictly necessary , we suppose here that @xmath15 has some finite value , while polymers are so long that the following inequality is satisfied : @xmath237 this further assumption is to eliminate the dependence on @xmath15 , which could introduce confusion in the following discussion due to possible interferences of condition ( [ condd ] ) with those of the perturbative and strong coupling regimes . in the @xmath9space , ( [ fjfsfd ] ) corresponds to the inequality @xmath238 . now @xmath189 may be approximated as follows : @xmath239 as promised , the above equation does not contain the parameter @xmath15 . we see from the left hand side of eq . ( [ domhigene ] ) that the function @xmath189 is logarithmically suppressed , due to the presence of @xmath240 in the denominator . this suppression effect is counterbalanced only at short distances by the two modified bessel functions of the second kind appearing in the numerator , which diverge logarithmically whenever @xmath241 and/or @xmath242 . the total result of these opposite effects in the expression of the averaged second moment will be presented in section [ sec : six ] after performing the explicit computation of the amplitudes of eqs . ( [ amponek][amptwofour ] ) . to conclude this section , let us give some concrete values of the involved parameters . first of all , let us estimate the values of @xmath8 , for which the two polymer system is in the regime ( [ condd ] ) . using the approximations made in eq . ( [ draapp ] ) , we may conclude that , if the relation ( [ condd ] ) is satisfied , the length @xmath8 needs such that @xmath243 , i. e. @xmath8 is at least of the order of hundred molecular lengths or more : @xmath244 . moreover , it is possible to give a rough estimation of the maximum distance of the end points , after which the two polymers are too far from each other to allow a relevant contribution to the winding angle due to repulsive interactions . using eq . ( [ cslm ] ) , in fact , it turns out that the repulsive interactions are relevant only in the range of distances : @xmath245 finally , the situation opposite to condition ( [ condd ] ) is not realistic , because it leads to the constraint @xmath246 . this would corresponds to the case of a polymer which is shorter than the size of the molecules composing it . at this point we are ready to compute the quantities @xmath182 and @xmath183 of eqs . ( [ ennee ] ) and ( [ deee ] ) . we start with @xmath183 . using eqs . ( [ deeeexp ] ) , ( [ dexp ] ) and the splitting ( [ propsplit ] ) of the propagator , one has at the zeroth order in @xmath194 : @xmath247 after a shift of variables , the above equation gives : @xmath248 where @xmath249 is the total surface of the system in the two dimensional space , which is transverse to the @xmath35axis . using the identity @xmath250 one finds : @xmath251 this expression of @xmath252 has the following interpretation : we are performing here an average of the second moment with respect to all possible initial and final positions of the endpoints of the polymers and @xmath183 _ counts _ the number of these configurations . the component @xmath252 of @xmath183 depends only on the free propagator @xmath144 , which is translational invariant in the sense discussed after eq . ( [ gzspec ] ) . this invariance explains why the number of configurations grows proportionally to the surface @xmath253 . the reason is that , for each configuration of the polymers , one can obtain other equivalently probable configurations by the symmetric translation of their ends on the surface @xmath253 at the initial and final instants . let us now apply to @xmath252 an inverse laplace transform , in order to go back to the @xmath9space . after a simple calculation we obtain : @xmath254 i. e. @xmath255 does not depend on @xmath8 . the next and last contribution to @xmath183 is given by : @xmath256 exploiting eq . ( [ intbas ] ) to integrate out the variables @xmath119 and @xmath257 , we get : @xmath258 we remark that the above contribution to @xmath183 vanishes in the limit @xmath191 . this could be expected due to the fact that @xmath259 collects all contributions coming from the repulsive interactions . these interactions break explicitly the translational invariance of the free part of the action and , as a consequence , @xmath259 is no longer proportional to the surface @xmath253 as @xmath252 . unfortunately , it is not easy to compute the inverse laplace transform of @xmath259 without making some approximation . to this purpose , we assume that the repulsive interactions are weak , i. e. @xmath260 , and that the value of @xmath8 is large . in this case , since we are in the domain of small @xmath71 s , it is possible to expand @xmath259 up to the second order in @xmath15 as follows : @xmath261 in order to obtain the above equation we have used both eqs . ( [ fdsf ] ) and ( [ pertapp ] ) . the inverse laplace transform of eq . ( [ deeoneexp ] ) gives : @xmath262l+\frac c{2\pi } v_0 ^ 2l\log l \label{deeonel}\ ] ] where @xmath263 is the euler constant . putting eqs . ( [ deezerointe ] ) and ( [ contdeeone ] ) together , we obtain : @xmath264 this is an exact result . an approximated expression of @xmath265 can be derived instead from eqs . ( [ dzerol ] ) and ( [ deeonel ] ) . now we turn to the derivation of @xmath182 . we start by computing order by order in @xmath194 the contributions to the quantities @xmath192 and @xmath193 of eqs . ( [ expke ] ) and ( [ expioe ] ) respectively . at the zeroth order we have for @xmath192 : @xmath266 after performing an easy integrations over the coordinates @xmath57 , one obtains : @xmath267 the remaining integral with respect to the @xmath115 coordinate is both ultraviolet and infrared divergent and needs to be regulated . we have already seen that the singularities in the ultraviolet domain may consistently be eliminated with the introduction of the small distance cut - off @xmath150 . a large distance cut - off is instead motivated by the fact that the size of a real system is necessarily finite . implicitly , we have already used this kind of infrared regularization in eq . ( [ deezero ] ) , where we have assumed that the total surface @xmath253 of the system in the directions which are transverse to the @xmath35axis is finite . supposing that the shape of @xmath253 is approximately a disk of radius @xmath268 , so that @xmath269 , we may write : @xmath270 substituting eq . ( [ polinfultint ] ) in eq . ( [ kzerot ] ) , one obtains the following expression of @xmath271 : @xmath272 the inverse laplace transform of @xmath271 gives : @xmath273 we have now to compute the quantities @xmath274 , with @xmath180 . the expressions of the @xmath275 s may be obtained from eqs . ( [ ionee ] ) and ( [ amptwoone][amptwofour ] ) , by substituting everywhere the propagator @xmath107 with its free version @xmath144 . it is easy to show that : @xmath276 this vanishing , which is actually a double vanishing , is due to the fact that each of the @xmath275 s contains an integral of a total divergence together with an integral which is zero for symmetry reasons . for some values of @xmath277 , like for instance when @xmath278 , to isolate such integrals it is necessary to perform some integrations by parts . this is allowed because the @xmath274 s are not affected by divergences , contrarily to @xmath192 . as an example , we work out explicitly the case of @xmath279 . the first vanishing integral is the following : @xmath280 this is of course zero due to symmetry reasons . the second vanishing integral in @xmath279 is of the form : @xmath281 after performing the integration over @xmath68 with the help of a shift of variables and of eq . ( [ intbas ] ) , we have , apart from a proportionality factor : @xmath282 since @xmath283 is a divergenceless vector potential , i. e. @xmath284 , @xmath285 can be rewritten as the integral of a total divergence : @xmath286 clearly , the left hand side of the above equation is zero . this fact can be also checked passing to the fourier representation . exploiting eq . ( [ ftsmbf ] ) and the formula @xmath287 in eq . ( [ deeint ] ) , one obtains : @xmath288 thus @xmath289 because @xmath290 . in an analogous way one shows that also @xmath291 and @xmath292 are identically equal to zero . we are now ready to compute the contributions to @xmath182 , which are linear in @xmath194 . first of all , we treat the term @xmath293 , which is given by : @xmath294 \label{konee}\ ] ] the integrations over @xmath67 and @xmath68 may be easily performed using eq . ( [ intbas ] ) and give as a result a factor which is proportional to @xmath204 . after that , only the following integral in @xmath115 remains to be done : @xmath295 here the ultraviolet divergence , which is present in the left hand side , has been regulated in the usual way with the introduction of the short distances cut - off @xmath150 . infrared divergences are absent . going to polar coordinates , the right hand side of the above equation becomes : @xmath296 putting everything together , one arrives at the final result : @xmath297 if the quantity @xmath298 is small , it is possible to derive the following asymptotic expression of @xmath293 : @xmath299 to go from eq . ( [ konefin ] ) to eq . ( [ koneeer ] ) , we have used the asymptotic formula : @xmath300 which is valid for small values of @xmath298 . we see from eqs . ( [ konefin ] ) and ( [ leaordaasyexp ] ) that the presence of ultraviolet divergences , together with the needed regularization , has modified the naive form of @xmath293 as a function of the pseudo - energy @xmath71 given in eq . ( [ behaone ] ) . the modification consists in the appearance of the factor @xmath301 , which exhibits a square logarithmic singularity in the limit @xmath302 . the inverse laplace transformed of @xmath293 can be derived only making some approximation . as in the case of @xmath259 , we will work in the double limit , in which @xmath15 is very small and @xmath8 is very large . after a few calculations we obtain : @xmath303\left ( \log s+c \right)+\right.\nonumber\\&&\left . \frac14l\log^2(2ca^2)+\frac 12 \log(2ca^2 ) \left [ ( c-1)l - l\log l \right ] \right\ } \label{ddd}\end{aligned}\ ] ] at this point we have to compute the expressions of the @xmath304 s , @xmath305 . it is possible to show that these contributions vanish identically , i. e. : @xmath306 the motivations of this vanishing are similar to the motivations for which there are no contributions at the zeroth order : all terms which appear in the quantities @xmath304 contain at least one integral of a total divergence or one integral , which is zero for dimensional reasons . as in the case of the @xmath274 s , there are some values of @xmath277 for which it is necessary to perform an integration by parts in order to isolate these vanishing integrals . once again , this is allowed because the @xmath304 s do not contain divergences . at the next order in @xmath194 , we have the last contribution to @xmath192 : @xmath307 after performing the integrations in @xmath68 and @xmath67 with the help of eq . ( [ intbas ] ) , eq . ( [ lcke ] ) becomes : @xmath308 the integral in @xmath115 is divergent and needs a regularization . going to polar coordinates , we obtain the result : @xmath309 also in this case , we note that the presence of the regularization modifies the dependence of @xmath310 on the pseudo - energy @xmath71 with respect to the naive formula of eq . ( [ behaone ] ) . the correction consists in the factor @xmath311 . in the limit @xmath302 , this factor diverges as powers of @xmath240 . to conclude the analysis of the contribution to @xmath182 at the second order in @xmath194 , we show that the @xmath312 s are identically equal to zero . as a matter of fact , it is easy to verify that for @xmath313 each @xmath312 contains terms of the kind : @xmath314 these terms vanish identically because of the following identity : @xmath315 substituting eq . ( [ auxfskj ] ) in eq . ( [ blmn ] ) and using the explicit expression of the vector potential @xmath283 of eq . ( [ vecpot ] ) , we get : @xmath316 clearly , the right hand side of the above equation is zero because @xmath317 . if @xmath278 , instead , the vanishing function @xmath318 of eq . ( [ blmn ] ) may be isolated in the expression of @xmath319 only after an integration by parts . finally , at the third order in @xmath194 we have only the quantities @xmath320 s , since @xmath192 has at most quadratic powers of @xmath194 . it is easy to realize that : @xmath321 because the following relations hold : s is separately equal to zero , because these quantities contain terms of the kind given in eq . ( [ blmn ] ) . ] @xmath322 as a consequence of eq . ( [ threezero ] ) , it is clear that there are no contributions to @xmath182 at this order . using eqs . ( [ kzerfin ] ) , ( [ konefin ] ) and ( [ ktwofin ] ) , we arrive at the final result for @xmath182 : @xmath323 we can now insert in the formula of the second moment of eq . ( [ secmomaveerep ] ) the functions @xmath183 and @xmath182 given in eqs . ( [ deeefin ] ) and ( [ nefin ] ) respectively . the outcome is : @xmath324 } { s+\frac \pi c\lambda(e)e^{-1 } } \label{mefin}\ ] ] in the @xmath9space , the already mentioned difficulties with the computation of the inverse laplace transform of @xmath183 and @xmath182 allow an analytical result only in the double limit of weak coupling constant @xmath15 and of large values of @xmath8 . at the first order in @xmath15 , the expression of @xmath61 reads as follows : @xmath325 where @xmath326 has been given in eq . ( [ ddd ] ) . so far , we have considered the averaged second moment of eq . ( [ secmomerep ] ) , corresponding to the case in which the polymer ends are not fixed . in the energy representation , we have seen that this version of the second moment can be exactly computed . to conclude this section , we would like to show that it is possible to provide also an exact expression of the second moment @xmath327 in the @xmath9 space and with fixed polymer ends up to an inverse laplace transform of the propagator given in eq . ( [ gffin ] ) . the starting point is the exact formula of the second moment @xmath328 of eq . ( [ feesm ] ) . all the ingredients of this formula are defined in eqs . ( [ secmomerep ] ) , ( [ nsdoqop])([nsdtqop ] ) and ( [ ampzero])([amptwofour ] ) . looking at eq . ( [ feesm ] ) , it is clear that : @xmath329 and @xmath330 let us note that the functions @xmath179 are all equal up to integrations by parts , which can shift the differential operators @xmath331 and @xmath332 in eqs . ( [ amponek])([amptwofour ] ) . this fact will be used in order to simplify the expression of the inverse laplace transformed of @xmath80 in the @xmath9space . to compute the inverse laplace transforms of both @xmath80 and @xmath81 , we use the following property of the inverse laplace transform of the product of two functions @xmath333 and @xmath334 : @xmath335 applying eq . ( [ convilp ] ) to evaluate the inverse laplace transforms of @xmath192 and of the @xmath179 in eqs . ( [ nefull ] ) and ( [ defull ] ) , we obtain after some calculations : @xmath336 the second term in the right hand side of eq . ( [ nlfull ] ) is the contribution given by the functions @xmath179 , @xmath180 , while the first term comes from @xmath178 . remembering the definition ( [ secmomnumden ] ) of the second moment in terms of @xmath74 and @xmath75 , we get : @xmath337^{-1}\left[\frac 2c \int d^2x { { \mbox{\bf a}}}^2({{\mbox{\bf x}}})\int_0^l ds g(l - s;{{\mbox{\bf r}}}_1,{{\mbox{\bf x}}})g(s;{{\mbox{\bf x}}},{{\mbox{\bf r}}}_0)\right.\nonumber\\ & & \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\ ! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\ ! \left.-\frac{2}{c^2}\int d^2x\int d^2 y\int_0^lds g(l - s;{{\mbox{\bf x}}},{{\mbox{\bf r}}}_1)\int_0^s ds ' \partial_x^i\partial_y^jg(s - s';{{\mbox{\bf y}}},{{\mbox{\bf x } } } ) g(s';{{\mbox{\bf r}}}_0,{{\mbox{\bf y } } } ) a_i({{\mbox{\bf x}}})a_j({{\mbox{\bf y}}})\right]\label{formtl}\end{aligned}\ ] ] if we knew how to compute the propagator @xmath132 starting from its laplace transformed ( [ gffin ] ) , we could evaluate explicitly the expression of the second moment in the @xmath9space . unfortunately , it is too complicated to perform the inverse laplace transform of the propagator @xmath107 . due to this technical difficulty , eq . ( [ formtl ] ) is only formal . progress can be made however in the limit @xmath191 , in which the propagator is given by the green function @xmath144 of eq . ( [ gzspec ] ) . this will be done in the next section . in order to allow the comparison with previous results , this section is dedicated to the case of ideal chains in which @xmath191 . first of all , we discuss the formula of the averaged second moment derived in the previous section , eq . ( [ mlfin ] ) . in the limit @xmath191 , ( [ mlfin ] ) becomes : @xmath338 the presence of a geometrical factor like the surface @xmath253 of the system in the expression of @xmath339 has been already related to the translational symmetry of eqs . ( [ strainvzer ] ) and ( [ strainvone ] ) . assuming that this surface has approximately the shape of a disc of radius @xmath268 , we can put @xmath340 as in eq . ( [ polinfultint ] ) . ( [ mlzerofin ] ) predicts that the average degree of entanglement scales as follows with respect to the distance @xmath268 : @xmath341 the meaning of eq . ( [ scalerisr ] ) is the following . we remember that the averaged second moment @xmath339 describes the entanglement of two closed polymers whose ends on the surfaces at @xmath37 and @xmath38 are not fixed . in this way , the polymers are allowed to move freely and it is natural to suppose that , the bigger will be the volume @xmath342 in which the polymers fluctuate , the bigger will be also the average distance between them . thus , if the surface @xmath253 increases its area , the probability of entanglement must decrease . the exact law of this decreasing is given by eq . ( [ scalerisr ] ) . on the other side , one would expect that the probability of getting entangled is higher for long polymers than for short polymers . ( [ mlzerofin ] ) gives a result which is in agreement with the above expectation , because the second moment @xmath343 scales as follows with respect to the parameters @xmath8 and @xmath17 , which determine the polymer length : @xmath344 in particular , one can show that the total length of a polymer increases proportionally to @xmath8 and it is inversely proportional to the square root of @xmath17 @xcite . accordingly , we see from eq . ( [ scalereslc ] ) that @xmath343 increases proportionally to @xmath8 and inversely proportional to @xmath17 . at this point we wish to study the second moment @xmath345 of polymers with fixed endpoints . the subscript @xmath346 has been added to the symbol of the second moment to remember that we are working in the limit @xmath191 . since we are dealing with ideal polymers , we have to substitute everywhere in eq . ( [ formtl ] ) the full propagator @xmath132 with the free one . the result of this operation is : @xmath347^{-1}\left[\frac 2c \int d^2x { { \mbox{\bf a}}}^2({{\mbox{\bf x}}})\int_0^l ds g_0(l - s;{{\mbox{\bf r}}}_1,{{\mbox{\bf x}}})g_0(s;{{\mbox{\bf x}}},{{\mbox{\bf r}}}_0)-\right.\nonumber\\ & & \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\ ! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\left.\frac{2}{c^2}\int d^2x\int d^2 y\int_0^lds g_0(l - s;{{\mbox{\bf x}}},{{\mbox{\bf r}}}_1)\int_0^s ds ' \partial_x^i\partial_y^jg_0(s - s';{{\mbox{\bf y}}},{{\mbox{\bf x } } } ) g_0(s';{{\mbox{\bf r}}}_0,{{\mbox{\bf y } } } ) a_i({{\mbox{\bf x}}})a_j({{\mbox{\bf y}}})\right]\label{mlfullvez}\end{aligned}\ ] ] we notice that , as it could be expected , eq . ( [ mlfullvez ] ) coincides with the expression obtained in @xcite for the second moment of one polymer winding up around an infinitely long straight wire lying along the @xmath5axis . luckily , the propagator @xmath348 can be explicitly constructed upon computing the inverse laplace transform of the propagator @xmath349 of eq . ( [ gzspec ] ) : @xmath350 it is easy to check that the second term in the right hand side of eq . ( [ mlfullvez ] ) , which is associated with the contributions coming from the @xmath179 s , does not grow with increasing values of @xmath8 . as a matter of fact , after a rescaling of variables , the numerator of this term gives : @xmath351\nonumber\\ & & \frac 1{t ' } e^{-\frac c{2(t')}\left|{\mbox{\mbld{y}}}'-\frac{\mathbf r_0}l\right|^2 } \label{mlsmfullvez}\end{aligned}\ ] ] in the limit @xmath352 , the quantity in the right hand side of the above equation scales as @xmath353 , where @xmath10 is a constant . moreover , the propagator ( [ freelprop ] ) , which is in the denominator , scales as @xmath354 . thus , the ratio between the right hand side of eq . ( [ mlsmfullvez ] ) and the propagator ( [ freelprop ] ) does not depend on @xmath8 . this completes the proof of our statement . as a consequence of this statement , as far as the scaling of @xmath345 for large values of @xmath8 is concerned , it is possible to make the following approximation : @xmath355^{-1 } \int d^2x { { \mbox{\bf a}}}^2({{\mbox{\bf x}}})\int_0^l ds g_0(l - s;{{\mbox{\bf r}}}_1,{{\mbox{\bf x}}})g_0(s;{{\mbox{\bf x}}},{{\mbox{\bf r}}}_0 ) \label{star}\ ] ] unfortunately , despite of the fact that we are treating ideal polymers , the integral in @xmath356 appearing in the above equation is still complicated and requires some approximation to be evaluated analytically . we will apply to this purpose the strategy used in ref . @xcite to compute the second moment of three dimensional polymers , adapting it to our two - dimensional case . first of all , let us note that the integral in ( [ star ] ) is ultraviolet divergent . however , the infrared divergences which appeared in the energy representation are absent . this is due to the behavior of the propagator @xmath126 , which is much milder at infinity than the behavior of the green function @xmath144 . to regulate the singularities at small distances , we proceed as usual by introducing the cut - off @xmath150 . after a rescaling of all variables similar to that of eq . ( [ mlsmfullvez ] ) , we get : @xmath357^{-1}\times \nonumber\\ & & \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\ ! \int_{|x'|\ge\frac{a\sqrt c}{\sqrt l } } \frac{d^2x'}l \frac 1{\mathbf x'^2 } \int_0 ^ 1 \frac{ds'}{s'(1-s ' ) } \left ( \frac c{2\pi } \right)^2 e^{-\frac 1{2(1-s')}\left|\mathbf x'-\mathbf r_1\sqrt{\frac cl}\right|^2 } e^{-\frac 1{2s'}\left|\mathbf x'-\mathbf r_0\sqrt{\frac cl}\right|^2 } \label{starst}\end{aligned}\ ] ] to go further , following @xcite , we assume that the relevant contribution to the integral in @xmath358 comes from a narrow region around the singularity in @xmath359 . thus , we may put @xmath360 after making the above crude approximation , we obtain : @xmath361^{-1}\log\left(\sqrt{\frac lc}a\right ) \int_0 ^ 1ds'\left [ \frac 1{1-s ' } e^{-\frac 1{2(1-s')}\mathbf r_1 ^ 2\frac cl } e^{-\frac 1{2s'}\mathbf r_0 ^ 2\frac cl}+ \right.\nonumber\\ & & \left . \frac 1{s ' } e^{-\frac 1{2(1-s')}\mathbf r_1 ^ 2\frac cl } e^{-\frac 1{2s'}\mathbf r_0 ^ 2\frac cl}\right ] \label{starstar}\end{aligned}\ ] ] in deriving the above equation we have used the simple relation @xmath362 . let us now study the integral @xmath363 the other integral in @xmath364 appearing in ( [ starstar ] ) can be treated in the same way after the change of variables @xmath365 . it is not to allowed to take in the right hand side of eq . ( [ inttrian ] ) the limit @xmath352 because in this way the integral will not be convergent due to the singularity in @xmath366 . for this reason , we split the domain of integration as follows : @xmath367 where @xmath368 . clearly , the second integral converges after performing the limit @xmath352 in the integrand and gives : @xmath369 the first integral instead diverges logarithmically with growing values of @xmath8 . however , now it is possible to expand the exponential @xmath370 in powers of its argument , because the singularity in @xmath371 lies outside the interval @xmath372 $ ] . keeping only the leading order term with respect to @xmath8 , we get : @xmath373 where ei@xmath374 is the exponential - integral function . when @xmath8 is large , this function may be approximated as follows : ei@xmath375 and , as a consequence : @xmath376 the second integral which we have left in eq . ( [ starstar ] ) gives the same result . putting everything together in the expression of the second moment of eq . ( [ starstar ] ) , we obtain the final result : @xmath377 this is exactly the behavior of the second moment derived in ref . @xcite . in this article we have studied the entanglement of two directed polymers from a non - perturbative point of view . our formulas of the second moment , a quantity which measures the degree of entanglement of the two polymers , take into account the repulsive forces acting on the segments of the polymers and are exact . the averaged second moment defined in eq . ( [ aderr ] ) , a version of the second moment corresponding to the situation in which the end points of the polymers are free to move , has been computed in eq . ( [ mefin ] ) as a function of the chemical potential @xmath71 conjugated to the distance @xmath8 between the end points in the @xmath35direction . the case of free ends is relevant in the treatment of nematic polymers and polymers in a nematic solvent @xcite . let us note that also the expression of the second moment without any averaging and in the @xmath8 space can be computed . this has been done in eq . ( [ formtl ] ) . however , this equation is explicit only up to the inverse laplace transform of the propagator ( [ gffin ] ) , which is too hard to be obtained in closed form . eq . ( [ mefin ] ) shows that the averaged second moment is of the form @xmath378 . the overall factor @xmath379 coincides with the scaling power law of two ideal polymers . the correction @xmath333 to this fundamental behavior due to the repulsive interactions is a complicated function of @xmath71 , whose analysis would require numerical methods . nevertheless , it is possible to identify a dominance of the repulsive interactions in the domain of parameters in which the condition @xmath380 is satisfied . this corresponds roughly speaking to the situation in which polymers are very long . in this region , the scaling laws with respect to the energy @xmath71 of the numerator and denominator appearing in the right hand side of eq . ( [ mefin ] ) are corrected by factors which are logarithmic powers of @xmath240 , see for instance eq . ( [ leaordaasyexp ] ) . one advantage of our approach is that it is easy to separate within the expression of the second moment the contributions of purely entropic origin which are typical of free polymers from the contributions coming from the presence of the @xmath0function potential in the polymer action . this is essentially due to the splitting ( [ appexplx ] ) of the propagator @xmath229 appearing in the amplitudes ( [ amponek][amptwofour ] ) . the component @xmath187 of the propagator coincides with the propagator of ideal polymers , while the component @xmath189 takes into account the effects of the interactions . thanks to the splitting ( [ appexplx ] ) , it has been possible to study the way in which the repulsive forces affect the average degree of entanglement of the two polymers . this has been done in section [ sec : five ] . our results are in agreement with the intuition . the precise law with which the effects of the repulsive forces on the entanglement decrease when the distance between the trajectories increases is given by eq . ( [ appexplx ] ) . in section [ sec : five ] it has been discussed also the strong coupling limit , which should be taken to recover the limit of excluded volume interactions . in our exact approach , it is not difficult to consider the case in which the coupling constant @xmath15 is large . for instance , the component @xmath189 of the propagator , which is responsible of the effects due to the repulsive interactions , has been given in the strong coupling limit in eq . ( [ scnontrainvpro ] ) . studying the form of this component assuming that polymers are very long , it has been argued that , at strong coupling , the major contributions to the winding angle coming from the repulsive interactions occur when the trajectories are very near to each other . many other qualitative and quantitative characteristics of the behavior of the two polymer system under consideration have been presented in section [ sec : five ] . the case of ideal polymers , in which @xmath191 , has been discussed at the end of section [ sec : seven ] in order to make comparison with previous works . the scaling of the averaged second moment for large values of @xmath8 obtained in eq . ( [ scalereslc ] ) is in agreement with the results of @xcite , if one takes into account the fact that , after the averaging procedure of eq . ( [ aderr ] ) and the infrared regularization of eqs . ( [ deezero ] ) and ( [ polinfultint ] ) . one is effectively treating a system of polymers confined in a cylinder of finite volume @xmath342 . in section [ sec : seven ] we have evaluated the second moment , always of two ideal polymers , using the approach of ref . the outcome of this calculation , namely the scaling behavior of @xmath381 at the leading order in @xmath8 , is reported in eq . ( [ lllsecmomfreefin ] ) . this result is in agreement with the square logarithmic behavior obtained in @xcite , but not with the logarithmic behavior predicted in @xcite . however , this discrepancy can be expected due to the fact that , in section [ sec : seven ] , we have assumed , following ref . @xcite , that the most relevant contribution to the second moment coming from the integral in eq . ( [ starst ] ) is concentrated in a narrow region near the singularity in @xmath359 . this clashes with the assumptions of ref . @xcite , in which instead it is argued that the main increase in the winding angle does not occur when the polymer trajectories are near , but rather when they are far one from the other . finally , there is also an apparent discrepancy between the linear scaling with respect to @xmath8 of the averaged second moment @xmath339 and the square logarithmic scaling of the second moment @xmath345 . this disagreement is explained by the fact that , in the first case , the ends of the polymers are free to fluctuate , while in the second case they are fixed . it is therefore licit to expect that two polymers with free ends are more likely to entangle than two polymers whose ends are constrained . concluding , we would like to discuss possible further developments of this work , together with some problems which are still left open . first of all , the number of entangling polymers has been limited to two . to go beyond this restriction , one should explore the possibility of replacing the external vector potential @xmath283 of eq . ( [ vecpot ] ) with chern simons fields . abelian chern - simons field theories have been already successfully applied in order to impose topological constraints to the trajectories of an arbitrary number of closed polymer rings in @xcite . we hope to extend those results also to the case of directed polymers in a forthcoming publication . of course , if the polymer trajectories are open , the constraints among them are no longer of topological nature as in @xcite , so that the application of chern - simons field theory to directed polymers should be considered with some care . we have also not made any attempt to introduce in the treatment of polymer entanglement more sophisticated constraints than those which can be imposed with the help of the winding angle . this is in effect still an unsolved problem , despite the fact that two powerful and strategies have been proposed for its solution @xcite . in the first approach , pioneered independently by kleinert , kholodenko and one of the authors @xcite . the constraints are expressed via the wilson loop amplitudes of non - abelian chern - simons field theories . some progresses toward a concrete realization of this program in polymer physics have been made in refs . @xcite . in the second approach , developed by nechaev and coworkers , see @xcite and references therein . polymer trajectories are mapped on a complex plane with punctures . the link invariants necessary to impose the constraints are then constructed using the properties of conformal maps . another possible development is the treatment of attractive interactions , in which the strength @xmath15 in eq . ( [ dirpot ] ) takes negative values . in this case , the @xmath0function potential admits a bound state @xcite and the propagator of eq . ( [ gffin ] ) develops a singularity , in which @xmath382 , at the energy corresponding to this bound state . it would be extremely interesting to investigate how these facts affect the polymers entanglement . another issue which deserves attention is that of hairpin turns . hairpins are important in nematic solvents @xcite and can be included with the help of field theories @xcite . we note also that in our formalism it is also possible to study the entanglement of polymers in confined geometries . for example , values of @xmath71 which are near to @xmath383 ( @xmath384 ) correspond roughly speaking to the situation in which polymers fluctuate in a quasi two - dimensional environment , in which the height in the @xmath35direction is of the order of a few molecular sizes . finally , an open problem , which has not be discussed here because we were mainly interested in the second moment . is the derivation of the full partition function @xmath64 of eq . ( [ laplrep ] ) . as anticipated in the introduction , it is not an easy task to compute @xmath64 because the repulsive potential of eq . ( [ dirpot ] ) is not central . we note however that the expression of @xmath64 coincides with the green function of a spin @xmath385 aharonov - bohm problem in the imaginary time formulation of quantum mechanics . this green function has been already derived in @xcite using sophisticated techniques developed in refs . @xcite , which bypass all the difficulties of dealing with a non - central potential . thus , in principle , the expression of the partition function @xmath64 is known . unfortunately , some of the consistency conditions imposed on the parameters in ref . @xcite seem to be incompatible with the requirements of our physical problem , as noted in section [ sec : three ] . for these reasons , the computation of the full partition function @xmath64 is still a problem which needs further investigations . luckily , the knowledge of the partition function is not necessary if one is interested to study the excluded volume interactions , which arise in the strong coupling limit . in fact , in this case it is possible to apply a powerful method due to kleinert @xcite . this method turns the weak coupling expansion into a strong coupling expansion which is convergent for large values of @xmath15 and is able to accommodate also the anomalous dimensions of quantum field theories . the convergence of this strong coupling expansion is mostly very fast , so that only a few coefficients of the weak coupling expansion must be known , see refs . @xcite for more details . these coefficients can be easily computed starting from the well known partition function of the aharonov - bohm problem without the insertion of the @xmath0function potential @xcite and treating this potential as a small perturbation assuming that the value of @xmath15 is small . the application of kleinert s method in order to complete the brief analysis of the strong coupling limit made in this paper is work in progress . 99 s. a. wasserman and n. r. cozzarelli , _ science _ * 232 * ( 1986 ) , 951 ; s. d. levene , c. donahue , t. c. boles and n. r. cozzarelli , _ biophys . * 69 * ( 1995 ) , 1036 ; f. livolant and y. bouligand , * j. phys . * ( paris ) * 47 * ( 1986 ) , 1813 ; r. podgornik , d. c. rau and v. a. parsegian , _ macromolecules _ * 22 * ( 1989 ) , 1780 . s. chandrasekhar , b. k. sadashiva and k. a. suresh , _ pramana _ * 9 * ( 1977 ) , 471 ; nguyen huu tinh , h. gasparoux and c. destrade , _ mol . _ * 68 * ( 1981 ) , 101 ; t. k. attwood , j. e. lyndon and f. jones , _ liq . crys . _ * 1 * ( 1986 ) , 499 . c. blisle , _ ann . * 17 * ( 1989 ) , 1377 . j. w. pitman and m. yor , _ ann . * 14 * ( 1986 ) , 733 . a. comtet , j. desbois and c. monthus , _ jour . * 73 * ( 1993 ) , 433 . a. grosberg and s. nechaev , _ polymer topology _ , advances in polymer science * 106 * , ( springer verlag , 1993 ) , 1 . b. drossel and m. kardar , in _ theoretical and mathematical models in polymer research _ , a. grosberg ( editor ) , ( academic press , 1998 ) , 187 ; cond - mat/9610019 ; _ phys . rev . _ * e53 * ( 1996 ) , 5861 . f. tanaka , _ prog . * 68 * ( 1982 ) , 148 . d. r. nelson , in _ observation , prediction and simulation of phase transitions in complex fluids _ , m. baus , l. f. rull and j. p. ryckaert ( eds . ) , ( kluwer , the netherlands , 1995 ) . j. d. moroz and r. d. kamien , _ nucl . _ * b506 * [ fs ] ( 1997 ) , 695 . a. l. kholodenko and t. a. vilgis , _ phys . _ * 298 * ( 1998 ) , 251 . f. ferrari and i. lazzizzera , phys . b * 444 * ( 1998 ) 167 ; f. ferrari , h. kleinert and i. lazzizzera , _ eur . phys . j. _ b * 18 * ( 2000 ) , 645 . m. otto , _ j. phys . a : math . * 34 * ( 2001 ) , 2539 ; _ topological interactions in systems of mutually interlinked polymer rings _ , cond - mat/0402104 . m. g. brereton and s. shah , _ a _ : math . * 15 * ( 1982 ) , 989 . s. edwards , _ proc . soc . _ * 91 * ( 1967 ) , 513 ; _ a_1 ( 1968 ) , 15 . f. spitzer , _ trans . _ * 87 * ( 1958 ) , 187 . s. prager and h. l. frisch , _ j. chem . * 46 * ( 1967 ) , 1475 . n. saito and y. chen , _ * 59 * ( 1973 ) , 3701 . h. kleinert , _ path integrals in quantum mechanics , statistics , polymer physics , and financial markets _ , ( world scientific publishing , 3nd ed . , singapore , 2003 ) . m. kardar and y .- c . zhang , _ phys . * 58 * ( 1987 ) . 2087 ; m. kardar , g. parisi and y .- c . zhang , _ phys . * 56 * ( 1986 ) , 889 . r. d. kamien , p. le doussal and d. r. nelson , _ phys . rev . _ * a 45 * ( 12 ) ( 1992 ) , 8727 . k. iwata and t. kimura , _ phys . _ * 74 * ( 1981 ) , 2039 . h. bethe and r. peierls , _ proc . ( london ) * 148a * ( 1935 ) , 146 ; e. fermi , _ ricerca scientifica _ * 7 * ( 1936 ) , 13 , [ english translation : e. fermi , _ collected papers _ vol italy 19211938 ( university of chicago press , chicago , 1962 ) ] . f. berezin and l. faddeev , _ sov . * 2 * ( 1961 ) , 372 . s. albeverio , f. geztesy , r. hoeg - krohn and h. holden , _ solvable models in quantum mechanics _ ( springer verlag , berlin , 1988 ) ; s. albeverio , z. brzeniak andl . dabrowski , _ jour . * 27 * ( 1994 ) , 4933 . p. gerbert , _ phys . * 40 * ( 1989 ) , 1346 . c. r. hagen , _ phys . * 64 * ( 1990 ) , 503 ; _ int . _ a*4 * ( 1991 ) , 3119 . c. grosche , _ phys . * 71 * ( 1993 ) , 1 ; _ ann . phys . _ * 3 * ( 1994 ) , 283 . d. k. park , _ jour . * 36 * ( 10 ) ( 1995 ) , 5453 . r. p. feynman and a. r. hibbs , _ quantum mechanics and path integrals _ , ( mcgraw - hill , new york , 1965 ) . l. s. schulman , _ techniques and applications of path integrals _ , ( wiley , new york , 1981 ) . r. jackiw , _ delta function potential in two- and three - dimensional quantum mechanics _ , in m. a. bg memorial volume , a. ali and p. hoodbhoy ( eds . ) , ( worlds scientific , singapore , 1991 ) . h. kleinert , _ phys . rev . _ * d57 * ( 1998 ) , 2264 . h. kleinert , _ phys . lett . _ * b434 * ( 1998 ) , 74 , [ arxiv : cond - mat/9801167 ] . h. kleinert and v. schulte - frohlinde , _ critical properties of @xmath386theories _ , world scientific , singapore 2001 . f. ferrari , _ phys . lett . _ * a323 * ( 2004 ) , 351 , cond - mat/0401104 . f. ferrari , h. kleinert and i. lazzizzera , int . j. mod . phys . b*14 * ( 2000 ) 3881 [ arxiv : cond - mat/0005300 ] . h. kleinert , _ pfadintegrale in quantenmechanik , statistik un polymerphysik _ , spektrum akademischer verlag , heidelberg , 1st edition 1993 , 850p . , isbn 3 - 86025 - 613 - 0 . a. kholodenko , _ trends in chem . 3 , research trends inc . trivandrum , india ( 1994 ) , pp . s. nechaev and r. voituriez , _ conformal geometry and invariants of 3-strand brownian braids _ , [ arxiv : hep - th/0407008 ] . f. ferrari , _ chern - simons field theories with non - semisimple gauge group of symmetry _ _ jour . math . * 44 * ( 1 ) ( 2003 ) , 138 , hep - th/0210100 . f. ferrari , _ topological field theories with non - semisimple gauge group of symmetry and engineering of topological invariants _ , published in _ trends in field theory research _ , o. kovras ( ed . ) , nova science publishers ( 2004 ) . j. cardy , _ jour . phys . a _ * 16 * ( 1983 ) , l355 .

recently , x - shaped radio galaxies ( xrgs)a peculiar class of radio galaxies with two pairs of misaligned lobes @xcite have received significant attention as new observations have explored and critiqued the competing models for the origin of their odd morphology . the distinctive `` x '' shape occurs due to the intersection of two centro - symmetric synchrotron - emitting structures at a common nucleus ( e.g. figure [ 3c403 ] ) . one of these structures is associated with an active relativistic jet ( the `` primary '' lobes ) whereas the other is fainter , more ragged , and does not appear to harbor a jet ( the `` secondary '' lobes or `` wings '' ) . the wings can be long , collimated , and symmetric about the nucleus , and may even exhibit z - shaped morphology of their own @xcite . the origin of the wings is not presently understood , but xrgs share other characteristics which indicate a common origin . x - shaped sources make up about 510% of fanaroff - riley type ii ( fr ii ; * ? ? ? * ) radio galaxies ( e.g. * ? ? ? the consensus sample is fairly small , but with the identification of @xmath0 `` candidate '' xrgs @xcite , a search for trends is now possible . based on the combined consensus and candidate sample , xrgs tend to have radio powers and host optical luminosities near the fr i / ii `` break '' in a ledlow - owen radio power visual magnitude @xcite type plot @xcite , and reside in elliptical galaxies with larger - than - average black hole masses ( inferred from the @xmath1 relation and broad lines from the active galactic nucleus ; * ? ? ? xrgs do not seem to be in galaxies currently undergoing a merger based on inferred starburst history @xcite and a spectroscopic search for broad emission lines and dusty nuclei @xcite , but the dynamic ages of their active lobes are younger than the age of the starburst @xcite . the wings are preferentially co - aligned with the _ minor _ axis of the host , whereas the primary lobes tend to be co - aligned with the major axis @xcite ; this alignment has been confirmed in the x - rays which trace the distribution of hot gas in and around the host galaxies that provides the working surface which shapes the radio galaxy @xcite . the radiative age of the wings does not seem to follow a clear trend , as some wings have steeper spectral indices than the primary lobes ( implying ageing ) whereas others do not @xcite . notably , some wings are actually longer than the primary lobes , and wings are only found in radio galaxies with strong bridges . a critical review of the observational data and existing models is found in @xcite ; here we briefly summarize the main threads . the origin of the secondary lobes has been attributed to ( i ) a rapid reorientation of the spin axis of the supermassive black hole ( smbh ) powering the jet ( i.e. the wings are fossils ) , either due to a smbh merger ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) or rapid precession ( e.g. * ? ? ? * ) ; ( ii ) redirection and collimation of `` backflow '' ( spent jet plasma flowing back from the terminal shocks ) @xcite ; ( iii ) a binary agn @xcite ; and ( iv ) interaction of the jet with disturbed morphology ( e.g. stellar shells , phase - wrapped remnants of a merged disk galaxy ; * ? ? ? since xrgs are usually strongly bridged sources and are apparently aware of their environments , it is worth examining closely the hypothesis that the x - shaped morphology originates from an interaction between the radio galaxy and its environment . in this paper , we seek to test the viability of the _ backflow model _ with three - dimensional hydrodynamic simulations in elliptical atmospheres . [ 3c403 ] in the backflow model , wings are produced by a single agn outburst with powerful jets as the backflow is diverted along the steepest pressure gradient of the surrounding atmosphere ( i.e. the minor axis ) . the wings either rise buoyantly or are driven in this direction @xcite or form as supersonic outflows along the direction of least resistance from a ruptured overpressured cocoon @xcite . these scenarios naturally explain the observed correlation between the wings and the minor axis as well as the dearth of strictly fr i xrgs ( the weaker fr i sources do not produce powerful back - flows ) . however , the backflow model faces significant challenges , most notably the very long wings in many xrgs ( often longer than the primary lobes ) . whereas powerful jets are expected to drive through the surrounding medium supersonically , wings which expand buoyantly would do so at most transonically for most of their lives @xcite . because , in this model , fluid effects are primarily responsible for the `` x '' shape , hydrodynamic simulations are an ideal proving ground . although recent hydrodynamic simulations of radio galaxy morphology exist in the literature , ( e.g. * ? ? * ; * ? ? ? * ; * ? ? ? * ) , most are not concerned with the formation of lateral asymmetries such as wings . in view of advances in simulating radio galaxies in the past decade , the mounting evidence that xrgs constitute a population demands a critical look at the backflow model with new simulations . in this paper , we discuss 3d hydrodynamic models of radio galaxies ignited in initially relaxed , elliptical atmospheres and examine how wings form . we begin by discussing our simulation setup and strategy for exploring wing production in section 2 , then present our model runs and describe the evolution of a generic winged source in section 3 . in particular , we find that favorable pressure gradients are necessary but not sufficient to produce an x - shaped source ; the character and time evolution of the jet is equally important . in section 4 , we briefly discuss the missing and artificial physics in our simulations , and in section 5 we assess the backflow model in view of our results as well as discuss the implications for the broader picture of radio lobe morphology . finally , in section 6 we summarize our main results and conclusions . throughout this paper we use the terms `` secondary lobes '' and `` wings '' interchangeably . we also use the term `` winged '' source to refer to any radio source with substantial symmetric off - axis distortions whereas `` x - shaped '' sources are a subset of winged sources with an axial wing - to - lobe length ratio of more than @xmath2 . this nomenclature reflects the view of the hydrodynamic model in which xrgs are indeed a subset of a broader category of distortions . finally , we use the term `` back - flows '' to refer specifically to actual fluid flows heading back to the nucleus from the jet heads , whereas we use `` backflow '' to refer to the material in these flows . we use a parallelized version of the zeus code @xcite for our hydrodynamic simulations . zeus is a second - order ( spatial accuracy ) eulerian finite - differencing magnetohydrodynamic ( mhd ) code which solves the standard equations of hydrodynamics , @xmath3 for an ideal compressible fluid and introduces artificial viscosity for shocks . the version we use , zeus - mpv2 , is based on the national center for supercomputing applications ( ncsa ) code described in @xcite . we use spherical polar coordinates ( @xmath4,@xmath5,@xmath6 ) in the purely hydrodynamical mode for all our runs ; the jets are therefore injected from an inner boundary sphere with some small , but finite , radius @xmath7 . we outline our basic simulation setup below , followed by our strategy for exploring winged sources and a description of the evolution of a standard double - lobed source for comparison . hydrodynamic and mhd simulations of radio galaxies interacting with their surroundings are common due to the possibility that energy deposited by the lobes is a mode of heating in agn feedback scenarios . we do not incorporate relativistic jet physics since we are primarily concerned with lobe mixing and evolution , but the robustness of this assumption is explored in section 4.1 . we note here that non - relativistic , light , hypersonic hydrodynamic jets reproduce several essential features of jets . these include recollimation shocks , terminal shocks at the jet head ( associated with radio `` hot spots '' ) and back - flows of spent material from the terminal shocks which sheath the jet and produce lobes . since we are interested in the evolution of back - flowing fluid in anisotropic environments , we adopt a non - relativistic purely hydrodynamic scheme . we inject the jets as bi - directional flows into an initially hydrostatic , ellipsoidal atmosphere . the atmospheres we set up are ellipsoids with a polytropic equation of state ( @xmath8 ) and a 3d @xmath9-model density profile @xcite given by @xmath10 where @xmath11 , @xmath12 is the core radius , and @xmath13 is the core density . to adapt this model for elliptical atmospheres , we make the radial density profile dependent on angle by adding a major - axis position angle pa and ellipticity @xmath14 to the atmosphere , which we denote as @xmath15 for an elliptical atmosphere . to easily allow for triaxial atmospheres , we define the ellipsoid in cartesian space ( @xmath16 , @xmath17 , @xmath18 ) and break up @xmath14 along each axis , then transform back to spherical coordinates to obtain the radius @xmath15 using the grid coordinates ( @xmath4 , @xmath5 , @xmath6 ) : @xmath19\\ y_e & = & r \sin(\theta)\biggl[\sin(\phi)cos(\text{pa})-\cos(\phi)\sin(\text{pa})\biggr]\\ z_e & = & r \cos(\theta)\biggl[\cos(\text{pa } ) + \sin(\text{pa})\biggr],\end{aligned}\ ] ] where pa is nominally measured clockwise from the @xmath20-axis ( in practice , it is only the difference between the jet axis and the major axis of the ellipsoid that matters ) . to normalize the size of the atmosphere , @xmath21 in equation 5 is defined as the largest value of @xmath14 along each axis . for example , an atmosphere with large @xmath22 is elongated along the @xmath20-axis . equations 4 and 5 may be understood as saying that introducing @xmath14 changes the effective core radius along a given axis : @xmath23 . for example , if @xmath24 , @xmath25 , and @xmath26 , then along the @xmath20-axis @xmath27 and along the @xmath28-axis @xmath29 . this phrasing is convenient when comparing the major and minor axes rather than the quantities along the full range of angles @xmath5 and @xmath6 and we use it hereafter . the atmospheres are initially set up in hydrostatic equilibrium , @xmath30 assuming that the background dark matter potential is static and dominant so that the gas self - gravity is not important . we define the adiabatic sound speed @xmath31 in code units and set up an isothermal atmosphere so the potential @xmath32 is @xmath33 since xrgs are powerful radio galaxies and exist at low redshift , we assume a smooth intergalactic / intracluster medium ( igm / icm ) and gas - poor systems , i.e. no disk of colder ( i.e. atomic or molecular ) material in the host ( c.f . * ? ? ? we note that in several exploratory runs , additional atmospheric complexity is overlaid onto a smooth atmosphere with no change in the gravitational potential @xmath32 , i.e. not initially in hydrostatic equilibrium . the jets in our simulations are hypersonic ( @xmath34100@xmath35 ) light ( @xmath36 ) flows injected in pressure equilibrium with the ambient material from back - to - back circular footpoints on the small inner boundary sphere at the origin . the kinetic luminosity of the jets is given by @xmath37 where @xmath38 is the area of the footpoint at @xmath7 , and ignoring higher order contributions from the gravitational energy or thermal flux . to tune @xmath39 we primarily vary @xmath40 because ( i ) @xmath39 is most sensitive to changes in @xmath40 , ( ii ) the jets must be `` light '' to ensure that the kelvin - helmholtz instability growth rate at the boundary of the cocoon is approximately the same as in the relativistic case @xcite , and ( iii ) jets are highly collimated . the maximum jet width is constrained by high - resolution x - ray observations of jet knots ( e.g. * ? ? ? * ) and very long baseline interferometry observations of transverse structures ( e.g. * ? ? ? . however , we note that in section 3.2 we vary @xmath41 and @xmath38 in a limited range . we insist that the jet cover enough grid zones ( @xmath42 ) to resolve transverse structures such as the oblique shocks which collimate the jet . the jet is rapidly precessed around a small angle ( @xmath43 ) at @xmath44 rad s@xmath45 ( code time ) to break the axisymmetry of the simulation setup and simulate helical instabilities , thus spreading the thrust out over a larger working surface ( c.f . ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? the location of the jet on the inflow sphere is thus given by @xmath46 . although the angular quantities are free parameters , they are constrained by the observed collimation of jets . typical values are @xmath47 and @xmath48 for @xmath49 . in agreement with other work , these jets develop a cylindrical core of fast - moving material sheathed in a slower concentric shell continuous at the boundaries with the surrounding material and core . we use a grid with zones spaced according to a geometric series in @xmath4 and @xmath5 ( 256 bins each ) and uniformly in @xmath6 ( 64 bins ) . this grid resolves the internal jet structure and important processes near the injection footpoints . we choose @xmath50 $ ] and @xmath51 $ ] to avoid the polar singularity and begin with reasonably sized grid zones . the @xmath5 grid is broken into two symmetric 128-bin components with the smallest zones near the poles where the jets are injected . the adequacy of our grid is demonstrated by a run with double the resolution in each direction which produces similar internal jet structure . mixing is not substantially different at the lobe boundaries . we use periodic boundary conditions in the @xmath6 direction and reflecting boundary conditions in @xmath5 . the outer @xmath4 boundary at @xmath52 has outflow conditions ( material leaving the grid ) ; likewise , outflow conditions exist for all @xmath7 zones except where the jets are injected . the outer boundary is far from the jet activity so negligible material leaves the grid there . we discuss the importance of the inner boundary sphere in section 4 . code units are transformed to physical units by choosing appropriate values for @xmath12 , @xmath35 , and @xmath13 . for instance , @xcite defined @xmath53 kpc , @xmath54 km s@xmath45 and @xmath55 g @xmath56 for a rich cluster and @xmath57 kpc , @xmath58 km s@xmath45 and @xmath59 g @xmath56 for a group or poor cluster . in the former scheme a code unit of time ( derived from the crossing time ) corresponds to 50 myr and in the latter 10 myr . in our runs , we vary @xmath12 and @xmath13 but fix @xmath35 ; a value of @xmath60 km s@xmath45 appears appropriate for xrgs based on temperatures derived from spectral fitting @xcite . a jet injected at @xmath61 would then have a physical speed between @xmath62 ( a lorentz factor of @xmath63 ) . however , as noted by @xcite , comparing classical and relativistic jet simulations requires careful matching of parameters , in particular the mass - energy density content of the jet . hence , the jet velocities chosen should not be taken directly as assumptions of true jet velocity . our runs were parallelized and used variety of processors . many of the runs were conducted on quad - core intel@xmath64 core@xmath65 @xmath66 and @xmath67 ghz workstations . we now outline our guiding strategy to determine whether x - shaped sources can result from the interaction of radio galaxy lobes with their environment . first , we only model powerful fr ii sources which produce strong back - flows . neither the physical origin of the fr i / ii dichotomy nor the differences between jets in radio loud and radio quiet sources are understood , but the higher luminosity fr iis exhibit the hot spots , well defined lobes , and bridges exploited by the backflow model . in accordance with the observations @xcite , we inject jets along the major axis of the surrounding atmosphere . we first use unrealistic atmospheres with very favorable pressure gradients for wing production ( as in * ? ? ? * ) , then explore jet and atmosphere parameters to study the production and characteristics of xrgs . in particular , we explore the dependence of wing formation on jet width , density , and kinetic luminosity as a function of time , and initial atmosphere parameters ( core radius , density , ellipticity , and position angle ) . although some parameter combinations are degenerate , this is a very large parameter space because the @xmath68 may include dead time and intermittency . compared to 2d modeling , our 3d simulations eliminate the requirement of axisymmetry which enhances the jet head advance speed @xcite and the coherence of the back - flows . the effect of turbulence is also more realistically explored in three dimensions . motivated by the expectation that pressure gradients affect the wings fundamentally the same way in different systems , we start from the ansatz that jet and atmospheric parameters are orthogonal . in other words , the jet parameters may be tuned in some fiducial atmosphere and the atmosphere parameters may be tuned with some fiducial jet such that the behavior of an arbitrary jet in an arbitrary atmosphere can be inferred . on the basis of this method we will evaluate the factors key to wing prominence and the viability of the backflow models . [ coevol ] if wings are distortions to a generic double - lobed source , we expect our simulated radio galaxies to strongly resemble the double - lobed sources produced by earlier models . hydrodynamic and mhd models of jets are commonly employed to study either the phenomenology of the jets and lobes ( with relativistic and nonrelativistic fluids ) or the energy deposited by the lobes into their environment . hence , a large body of work concerned with phenomenological models of these sources exists and has produced key insights into the life of a powerful radio source . these studies include work on early jet / lobe evolution in a relaxed atmosphere ( e.g. * ? ? ? * ; * ? ? ? * ) , the influence of the @xmath69 density contrast and internal mach number ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , jet stability in nonrelativistic and relativistic conditions ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , the importance of intermittency @xcite or intrinsically spreading out the jet thrust to slow the advance of simulated jets ( e.g. * ? ? ? * ) , the importance of the jet environment to morphology or energy deposition ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , and physics beyond mhd ( e.g. * ? ? ? naturally , there is overlap between the phenomenological studies and those motivated by the challenge to produce jets which prevent a cooling catastrophe in the icm of a host galaxy cluster by isotropizing energy distribution . the viability of radio galaxies as agn feedback mechanisms is beyond the scope of this work , but jet lifetime and the passive evolution of `` dead '' radio galaxies @xcite place important constraints on wing prominence . based on this work , we understand a powerful double - lobed source to have three distinct phases of evolution : ( i ) _ ignition _ , in which a spheroidal cocoon of relativistic plasma is formed around the nascent jets , ( ii ) the _ active phase _ , wherein the jet produces the cigar - shaped lobes associated with fr ii radio galaxies , and ( iii ) the _ passive phase _ where lobes evolve after the jet is turned off . below we outline these phases for a preliminary hydrodynamical simulation of a fast ( @xmath70 ) , light jet with large @xmath39 ( a `` fr ii '' source ) to provide a framework for discussing the evolution of winged sources ( left - hand panels of figure [ coevol ] ) . we note that @xcite have recently found good agreement between the structures observed in hydrodynamic simulations of hypersonic jets and laboratory experiment . [ [ ignition ] ] _ ignition _ + + + + + + + + + + the jet is injected in pressure equilibrium with the initially relaxed surrounding medium . the flow quickly forms and drives a spheroidal bow shock into the surrounding medium ( figure [ coevol ] , top left panel ) . at the location where the jet impacts the shock , back - flows develop which fill the space evacuated by the bow shock with light , hot spent jet plasma . this plasma forms a cocoon which sheaths the jet . at very early times the jet expands laterally , since it is unconfined by the cocoon plasma ( in part due to the initially conical shape of the jet ) and the expansion of the nascent radio galaxy is nearly self - similar @xcite . however , as a result , the early back - flows acquire a circulatory motion and flow along the inner edge of the bow shock ( figure [ backflow ] , top panel ) . by the time these flows reach the midplane between the two jets , they have velocity vectors pointing radially inwards and do not collide with back - flows from the counter - jet @xcite . in our models , the back - flows are also prevented from interacting with their counterparts from the counter - jet because our jets are injected into pristine atmospheres such that at very early times , each jet inflates its own cocoon . the two cocoons are both bounded by a bow shock , and the two shocks meet at the midplane between the jets and form an interstice of dense material that prevents early mixing ( figure [ backflow ] , top panel ) . whether such an interstice is present in real sources is not clear ( we do not resolve the very earliest jet stages such as the flood - and - channel phase seen in a clumpy , warm disk in * ? ? ? * ) , but the interstice is ablated and the cocoon is unified by the active phase . as the cocoon pressure builds due to backflow confined by the bow shocks , the jet becomes azimuthally confined and a recollimation shock appears near the injection point ; by the end of the ignition phase , the cocoon is strongly overpressured relative to its environment . at the same time , the terminal shock becomes increasingly distinct from the bow shock , and the jet head takes on the characteristic double - pronged appearance of a 3d hydrodynamical jet . while the back - flows still follow the bow shocks during the late ignition phase , they are increasingly straight . we note that this early evolution is seen in all our runs and does not depend much on atmosphere or jet parameters , nor on the boundary conditions ( as long as the core radius of the atmosphere @xmath12 is much larger than @xmath7 ) . [ [ active - phase ] ] _ active phase _ + + + + + + + + + + + + when the jet head overtakes and pierces the initial bow shocks ( driving a bow shock contiguous with the initial burst ; * ? ? ? * ) , the active phase begins and a classical radio galaxy develops , with a cigar - shaped cocoon , hot spots , bow shocks ( middle - left panels of figure [ coevol ] ) , and straight back - flows ( figure [ backflow ] , bottom panel ) . at this point , the cocoon ceases to be substantially overpressured and the cocoon s lateral expansion falls close to the sound speed of the ambient medium , lagging behind the bow shock ( the cocoon expands due to buoyancy , kelvin - helmholtz , and rayleigh - taylor instabilities ) . this weak shock ( @xmath71 ) sweeps up a large amount of material as it expands and becomes better described as a strong sound wave ( figure [ coevol ] ) . meanwhile , the jet develops internal oblique collimating shocks along its length , retaining a high velocity out to the terminal shock at the jet head ( i.e. the angular size of the jet decreases with increasing radius ) where the cocoon becomes momentum - driven . less powerful jets do not produce cocoons @xcite , but we focus on the cocoon - bounded case ( the cocoon pressure also depends on the internal mach number ; * ? ? ? [ [ passive - phase ] ] _ passive phase _ + + + + + + + + + + + + + + + once the jets are turned off , the back - flows cease and the `` dead '' radio galaxy rises buoyantly in the atmosphere , lifting large amounts of material to great heights and mixing with the surrounding medium @xcite . although the radio galaxy is relatively efficient at depositing its energy irreversibly in the surrounding medium , it is difficult to isotropize this energy deposition on short enough timescales to avoid a cooling catastrophes ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? the cocoon separates into bubbles which pinch off along the direction of the jets and rise in opposite directions ; the evolution of these bubbles may be quite complex ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? this very late stage evolution is unlikely to be important for winged galaxy evolution , as winged galaxies appear to be related to sources with strong bridges @xcite , and is not shown in figure [ coevol ] . with the scheme described in section 2 , we produce winged radio galaxies by the deflection of backflow into channels perpendicular to the jets . in our models , wings are produced in strongly asymmetric atmospheres when the jet is driven near the major axis ; the wings are produced along the minor axis as cocoon material escapes the central regions of the atmosphere and evolve buoyantly , collimated by the surrounding stratified atmosphere which promotes expansion along the steepest pressure gradient . hence , these wings are similar to those produced by @xcite and @xcite . without introducing additional complexity to the atmosphere , hydrodynamic wings are formed in two stages . first , during the ignition phase ( section 2.3 ) , the overpressured cocoon expands faster along the minor axis due to the steeper pressure gradient and forms small channels ( proto - wings ) perpendicular to the jet into which back - flowing plasma flows . the proto - wings produced in this way can account for 2040% of wing length at the end of the active phase depending on the gradient . second , during the active phase , the wings rise buoyantly , fed by turbulent flows near the midplane . although the back - flows near the jet heads are initially laminar and supersonic ( relative to the lobe material ) , they quickly decelerate and do not enter the wings as coherent flows ; the wings expand subsonically . during the active phase the wings evolve almost independently of the cocoon . hence , once wings have developed , their length depends only on the properties of the native atmosphere whereas the length of the jet - fed primary lobes is dominated by the properties of the jet ( in particular the kinetic luminosity as a function of time ) for powerful jets . therefore , _ decaying _ jets produce prominent wings . in this section , we explore these ideas in detail , first comparing the life of a winged source to a canonical double - lobed one ( figure [ coevol ] ) and then looking at the dependence of wing prominence and longevity on various tunable parameters in our models . winged galaxies experience the same life stages outlined in section 2.3 . we describe how wings fit into this process in detail below , referencing our standard ( extremal ) atmosphere ( run standard , tables [ atmostable ] and [ jettable ] ) . while unrealistic , this atmosphere provides important insights into the backflow model and can be directly compared to prior work @xcite . [ [ ignition-1 ] ] _ ignition _ + + + + + + + + + + during the ignition phase of a winged source , the anisotropic expansion of the overpressured cocoon in an aspherical atmosphere produces channels which will later become the wings ( the proto - wings ) . this pressure - driven expansion is supersonic but brief , since the cocoon quickly reaches pressure equilibrium . although these channels need not be produced by an overpressured cocoon , some channels must exist for wings to form . the degree of cocoon expansion in a given direction depends on the pressure gradient experienced in that direction , and hence on the atmospheric parameters . in particular , the eccentricity of the atmosphere effectively changes the core radius @xmath12 seen by the cocoon in different directions . for example , along the @xmath28-axis ( @xmath72 ; the jet is driven along @xmath73 ) equation 4 gives : @xmath74 where @xmath75 as above . since the pressure gradient is steeper for smaller @xmath76 , higher ellipticity along other axes promotes wing expansion : for @xmath25 and @xmath77 , @xmath78 along the @xmath28-axis . conversely , if @xmath79 and @xmath80 ( the jet is pointed along the minor axis ) , @xmath81 and wings are suppressed . of course , wing expansion depends on the actual pressure gradients rather than just the ratio along different axes ; the base values @xmath12 and @xmath13 determine whether wings can form ( i.e. a highly elliptical atmosphere can have shallow pressure gradients along the minor axis if it is very large ) . note that at very small radii @xmath82 ( i.e. during the ignition phase ) , the pressure gradient steepens linearly with increasing @xmath4 and at large radii ( during the active phase , see below ) the pressure gradient becomes shallower as @xmath83 . hence , the ignition phase is the time at which atmospheric asymmetry has the strongest effect on the ultimate morphology of the source . and @xmath84 velocity vectors overlaid on density slices ( taken at @xmath85 ) at different times for the simulation tx_e75_ve3_b5.0 ( table 3 ) . we have chosen ( arbitrarily ) @xmath58 km s@xmath45 and @xmath86 kpc ( corresponding to a code @xmath87 ) to represent a large elliptical galaxy . white vectors point radially outwards and cyan vectors point radially inwards ( i.e. the back - flows ) . vectors are only shown when the magnitude of @xmath88 exceeds the sound speed of the lobe material ( @xmath89 ) . note that early on , the back - flows follow the contact discontinuity and are directed towards the inner boundary , but later become straight as in @xcite . ] -axes ; at @xmath90 in the left panels , a new jet is reinjected . _ negative _ growth for the lobes indicates collapse during the passive phase . ] [ [ active - phase-1 ] ] _ active phase _ + + + + + + + + + + + + + + once the cocoon has reached pressure equilibrium , the proto - wings are no longer pressure - driven . as the back - flows become straight , they merge and fill the channels ( figure [ backflow ] , bottom panel ) to form structures more closely resembling observed wings . however , the flows into the wings are turbulent and transonic or subsonic ( relative to the internal lobe sound speed @xmath89 ) . in other words , wings are not _ driven _ during the active phase . rather , because they are structures filled with light fluid , they rise buoyantly in the atmosphere , collimated by the stratified elliptical atmosphere . since the wings rise buoyantly , their growth rate is subsonic . to see this , it is instructive to look at the simplified case of a spherical bubble ( of fixed radius ) rising buoyantly in a dense fluid . the ( terminal ) buoyant velocity at a given height for such a bubble is @xmath91 where @xmath92 , @xmath93 , and @xmath94 are the volume , cross - sectional area , and radius of the bubble respectively , @xmath95 is the `` drag '' coefficient , and we have used the background dark matter potential @xmath32 to compute @xmath96 in terms of @xmath4 and effective core radius @xmath23 . naturally , this follows the pressure gradient ( equation 12 ) , so when the bubble is at very small radii its velocity _ increases _ as @xmath97 , whereas outside @xmath76 it _ decreases _ as @xmath83 . for typical values in our simulations , the peak value of @xmath98 is approximately @xmath35 . hence , in a highly elliptical atmosphere , a bubble rising along the minor axis will rise subsonically with decreasing velocity for most of its lifetime . our wings do not fit this simplified case because they are large relative to @xmath76 , variable in size , aspherical , and connected to the cocoon . nonetheless , their expansion is subsonic for the same reason : the wings exceed @xmath76 early on when their growth is pressure - driven rather than buoyant , and @xmath98 monotonically declines thereafter . this transition from supersonic to subsonic wing expansion is plainly seen in figure [ wingvel ] . therefore , wings in our models may not intrinsically exceed the length of the primary lobes unless the primary lobes advance at some average speed @xmath99 . since powerful jets are required to produce the proto - wings in an overpressured cocoon phase , prominent wings require decaying jets , intermittent jets , or jets which deposit their thrust over an increasingly large area with time . the profile of the fiducial run is shown along with its wing prominence in figure [ standardgrowth ] and compared to a triaxial atmosphere in figure [ wingvel ] . note that in the fiducial run ( left panel of figure [ wingvel ] ) the wings eventually overtake the lobes ( in part due to the collapse of the lobes when the jet is very weak ) , but clearly move into a subsonic regime early and remain there . the bow shock remains mildly supersonic through most of both simulations . in figure [ diffwings ] we show wings produced by decaying jets during the active phase in several environments ( the fiducial run is shown in the top row ) . [ [ passive - phase-1 ] ] _ passive phase _ + + + + + + + + + + + + + + + once the jets are turned off , the cocoon disintegrates as the lobes either rise buoyantly as bubbles or collapse under the relaxing atmosphere . if the radio galaxy is within a few @xmath12 in a dense atmosphere , the fall - back of displaced material shreds the lobes into small bubbles . in either case , the wings pinch off and rise . these bubbles do not survive long , but their behavior might be substantially altered in the presence of magnetic fields ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? because observed winged sources have strong bridges , we expect them to be in the active phase . [ [ reinjection ] ] _ reinjection _ + + + + + + + + + + + + + a powerful jet reignited during the passive phase before the cocoon has disintegrated may significantly enhance the wings . if the old jet channel is somewhat broken up , the reinjected jet forms a new terminal shock and bow shocks _ inside _ the old cocoon . these new shocks do not form spheroidal structures ( as in a relaxed atmosphere ) but instead produce strong , straight back - flows near the midplane . hence , the wings receive a large influx of fresh supersonic plasma directly after reinjection ( figure [ coevol ] , bottom panels ) . this brightens the wings substantially and reinforces their structures . there is a relatively narrow window of time where this process is effective . if the reinjection occurs while the jet channels are largely intact ( i.e. during the active phase ) , the jet simply follows these channels . on the other hand , if the reinjection occurs when the wings have already separated from the lobes as individual bubbles , the new jet can not feed them . even if the reinjection occurs at the `` right '' time , the efficacy of the restarted jet at promoting wings is short - lived . however , this mechanism can produce wings that are intrinsically _ longer _ than the jet - driven lobes ( figure 3 , bottom panels ) . if the reinjected jet decays , this extreme axial ratio can be maintained for most of the lifetime of the restarted radio galaxy and a bona - fide x - shaped source results , although this source would only have the `` fr ii '' primary lobes for a short period of time ( @xmath100 myr for a source 40 kpc across ) . ( crosses ) in timestep intervals of @xmath101 for the standard simulation with a reinjection at @xmath90 . we have overplotted the velocity of the jet as the black line ( right @xmath102-axis ) . the roman numerals indicate ( i ) the overpressured cocoon phase , ( ii ) the active phase ( powerful jet ) , ( iii ) the active phase ( weak jet ) , ( iv ) the passive phase ( with cocoon collapse ) , ( v ) a second active phase due to the reinjected jet , and ( vi ) the second passive phase with further cocoon collapse . during cocoon collapse , the source does not resemble a winged source ( instead it is a `` dead '' radio galaxy ) . a source appears winged during the active phase when the jet is most like a weak fr ii or strong fr i. ] assuming equipartition ) of these sources . ] in section 3.1 , we describe the life stages of a winged radio source . we now describe the dependence of wing prominence on our tunable parameters . following the strategy laid out in section 2.2 , we begin with the fiducial simulation standard which is characterized by an atmosphere with a large ellipticity @xmath14 and a jet with a velocity profile @xmath103 $ ] ( figures [ coevol ] and [ wingvel ] ) . taking the jet from the fiducial run , we individually vary parameters in the atmosphere in order to see their influence on wings ( figure [ atmosplots ] , table [ atmostable ] ) , and then vary jet parameters in the fiducial atmosphere ( figure [ jetplots ] , table [ jettable ] ) . we then synthesize the information gleaned from these single - parameter curves to attempt to find atmospheres and jets which can explain real winged sources , including more complex behavior as well ( table [ synthesistable ] ) . all of the runs proceed to a code time of 1.0 or more , where most have entered the passive phase . reinjections , where used , occur near the beginning of the passive phase . to quantify wing prominence we use the wing - to - lobe axial ratio ( @xmath104 ) . this quantity is an imperfect measure because @xmath104 intrinsically varies with azimuthal angle @xmath6 and time , and observable sources are seen in projection so the observed @xmath104 will differ from the intrinsic value . @xmath104 is also essentially meaningless during ignition or the passive phase , but a source in one of these stages would not be classified as `` winged . '' for the parameter exploration in figures [ atmosplots ] and [ jetplots ] ( values in tables [ atmostable ] and [ jettable ] ) , we adopt as a fiducial value @xmath105 : the maximum intrinsic @xmath104 for any pair of azimuthal angles [ @xmath6,@xmath106 at a code time @xmath107 . for the standard jet ( @xmath103 $ ] ) , @xmath107 represents the transition from a powerful to weak jet in the active phase , and by this time the wings have begun to grow buoyantly . hence , a code time of @xmath107 is a reasonable place to measure the influence of the atmosphere on wings . we also use @xmath105 in varying the jet parameters , noting that while @xmath107 is no longer special , all of the runs in table [ jettable ] are in the active phase at this time . @xmath105 is not predictive of wing length later in the same simulation , but is a good measure of relative wing prominence between simulations due to the subsonic growth of wings . from figure [ atmosplots ] one can get a broad sense of the dependence of wing prominence on the size and shape of the atmosphere . it is immediately obvious that there is a strong dependence on the ellipticity @xmath14 of the atmosphere , and that smaller , denser atmospheres are the most conducive to wing formation ( although `` small '' could be physically quite large depending on the jet ) . it is also notable that relatively high values of @xmath104 can be achieved by @xmath107 ; because we use a decaying jet , the wings will only become more prominent during the later active phase . while these results are not surprising ( equation 12 ) , the particular form of the curves depends on both the jet and the atmosphere . as expected from @xcite and @xcite , long wings require jets co - aligned close to the major axis . the trends are less clear when varying jet parameters ( figure [ jetplots ] ) . in the top panels , the jet velocity is allowed to vary freely without conserving integrated kinetic luminosity between runs . if the jet is injected at constant velocity throughout the simulation ( figure [ jetplots]a ) , a faster jet is slightly better than a slower one at making wings . however , a decaying jet ( figure [ jetplots]b ) is better still , with pure exponential decay more effective at increasing @xmath105 than `` gaussian '' jets of the form @xmath108 $ ] . in the bottom panels of figure [ jetplots ] , we hold the kinetic luminosity constant between runs by modifying @xmath109 . from figure [ jetplots]c , it is evident that increasing the width of the jet ( at the cost of a slower jet ) is important . changing the density of the jet material has an extreme effect on @xmath105 ( figure [ jetplots]d ) ; above @xmath110 , the jet ceases to be `` light '' relative to the background and no longer forms lobes resembling radio galaxies @xcite . at very low densities , mixing becomes very efficient at disrupting the radio galaxy . we now discuss these results in detail , beginning with the properties of the fiducial run . lccccccc + + standard & y & 1.0 & 3.0 & 0.0 & 0.75 & 0.0 & 0.62 + + + + sj_e20 & n & 1.0 & 3.0 & 0.0 & 0.20 & 0.0 & 0.42 + sj_e30 & n & 1.0 & 3.0 & 0.0 & 0.30 & 0.0 & 0.43 + sj_e50 & n & 1.0 & 3.0 & 0.0 & 0.50 & 0.0 & 0.49 + sj_e60 & y & 1.0 & 3.0 & 0.0 & 0.60 & 0.0 & 0.54 + + + + sj_r0.5 & y & 0.5 & 3.0 & 0.0 & 0.75 & 0.0 & 0.73 + sj_r0.75 & y & 0.75 & 3.0 & 0.0 & 0.75 & 0.0 & 0.61 + sj_r2.0 & n & 2.0 & 3.0 & 0.0 & 0.75 & 0.0 & 0.45 + + + + sj_d1.5 & y & 1.0 & 1.5 & 0.0 & 0.75 & 0.0 & 0.52 + sj_d2.0 & y & 1.0 & 2.0 & 0.0 & 0.75 & 0.0 & 0.57 + sj_d4.0 & y & 1.0 & 4.0 & 0.0 & 0.75 & 0.0 & 0.65 + sj_d5.0 & y & 1.0 & 5.0 & 0.0 & 0.75 & 0.0 & 0.68 + + + + sj_pa5 & y & 1.0 & 3.0 & 0.0 & 0.75 & 5.0 & 0.62 + sj_pa10 & y & 1.0 & 3.0 & 0.0 & 0.75 & 10.0 & 0.60 + sj_pa15 & y & 1.0 & 3.0 & 0.0 & 0.75 & 15.0 & 0.58 + sj_pa20 & y & 1.0 & 3.0 & 0.0 & 0.75 & 20.0 & 0.51 + + + + tx_e75_1 & y & 1.0 & 3.0 & 0.375 & 0.75 & 0.0 & 0.65 + tx_e75_2 & y & 1.0 & 3.0 & 0.50 & 0.75 & 0.0 & 0.72 [ [ fiducial - run ] ] _ fiducial run _ + + + + + + + + + + + + + + our fiducial run ( standard , tables [ atmostable ] and [ jettable ] ) is the combination of our fiducial atmosphere ( @xmath26 , @xmath111 , @xmath112 , @xmath113 ) with our fiducial jet ( @xmath114 $ ] , @xmath115 , @xmath116 , aligned along the @xmath20-axis ) . like most of our runs , this simulation proceeds to a code time of @xmath117 , at which point the radio galaxy is in the early passive phase . unsurprisingly , this run produces some of the most prominent wings of our suite ( figure [ diffwings ] ) ; the shallow pressure gradient along the @xmath20-axis and declining velocity profile combine to produce stalled lobes which begin to collapse at the end of the active phase . at the same time , wings quickly escape the core region , and by @xmath118 are easily recognizable ( right - hand panels of figure [ coevol ] ) . after reaching @xmath119 at @xmath120 , the cocoon collapse leaves the wings as the most notable features ( figure [ standardgrowth ] ) . at @xmath90 , we re - inject a jet with @xmath121 $ ] ( figure [ standardgrowth ] and bottom panel of figure [ coevol ] ) . since the jet is expanding into tenuous material from the old cocoon , the expansion quickly re - establishes a cocoon . however , the reinjection also drives a bow shock inside the collapsing cocoon , allowing initially strong back - flows to feed the wings directly . hence , between @xmath122 and @xmath123 , an x - shaped radio galaxy is apparent ( @xmath124 ) . after @xmath123 , the jet again weakens to the point where collapse begins . if the jet is instead reinjected at the inception of the passive phase at @xmath125 , the end result is very similar ( @xmath124 at @xmath126 ) . on ( a ) the ellipticity of the atmosphere @xmath14 , ( b ) the core density @xmath13 , ( c ) the core radius @xmath12 , and ( d ) the angle between the jet and the major axis @xmath127 . each plot represents varying one parameter while holding the other standard atmosphere parameters ( @xmath128 , @xmath111 , @xmath26 , and @xmath129 ) steady and using the standard jet . in panel ( a ) the square and triangle represent runs with different @xmath130 for @xmath113 , demonstrating the effect of triaxiality on @xmath104 ( table [ atmostable ] ) . @xmath104 is measured at @xmath107 for all cases ; for the standard jet , this represents the transition during the active phase from a powerful to a weak jet ( this time and the `` wing threshold '' vary with choice of jet ) . values taken from table [ atmostable ] . ] the atmosphere in this run is similar to the 2d simulations presented in @xcite and 3d simulations in @xcite . striking asymmetries develop despite the axisymmetric atmosphere because of turbulent mixing between different slices in @xmath6 and because the back - flows form three - dimensional structures within the lobes . if these flows are slightly misaligned on opposite sides of the midplane , asymmetries develop . on ( a ) jet velocity for a constant - velocity jet , ( b ) e - folding time for a decaying jet ( diamonds represent jets of the form @xmath131 $ ] and triangles @xmath132 $ ] ) , ( c ) area of the jet nozzle , and ( d ) density of the jet material . all runs are conducted in the standard atmosphere ( table [ atmostable ] ) . in ( a ) and ( b ) , the jet kinetic luminosity is _ not _ conserved between runs , whereas in ( c ) and ( d ) we adjust the initial velocity @xmath109 to conserve @xmath39 . as in figure [ atmosplots ] , we show @xmath104 as measured at a code time @xmath107 ; all sources are in the active phase but not at the same place . ] [ [ atmosphere ] ] _ atmosphere _ + + + + + + + + + + + + in our exploration of different atmosphere properties ( figure [ atmosplots ] ; table [ atmostable ] ) , we vary @xmath14 , @xmath13 , @xmath12 , and @xmath127 individually while keeping the other atmosphere parameters steady and using the standard jet ( @xmath103 $ ] , @xmath133 , @xmath116 , and @xmath134 ) . except when we vary @xmath127 , the jet is coaligned with the major axis . it is clear from figure [ atmosplots ] that wing prominence depends strongly on the core radius @xmath12 , core density @xmath13 , and ellipticity @xmath14 . because the lobes and wings evolve almost independently , these dependencies can be understood by breaking the axial ratio @xmath104 into its constituent parts : the wing length @xmath135 and lobe length @xmath136 . for the purposes of this discussion , @xmath135 and @xmath136 are both taken to be measured at @xmath107 ( table [ atmostable ] ) . we find that @xmath136 varies only slightly as a function of @xmath12 ( less than 10% ) but does peak between @xmath137 and @xmath87 . at very low @xmath12 , the lobes are not confined by the atmosphere and spread out laterally , causing more turbulent mixing and dissipating the thrust of the jet . at very high @xmath12 , the lobes are completely confined by the atmosphere and the density gradient is so shallow that as the jet weakens it is increasingly resisted by the atmosphere . the importance of this effect on the lobes is more pronounced in a comparison between @xmath136 and @xmath13 where the lobe length declines by 50% between @xmath138 and @xmath139 . on the other hand , the wing length @xmath135 depends much more strongly on the core radius @xmath12 ( declining precipitously with increasing @xmath12 ) and only weakly on the core density . this behavior occurs because the wings respond primarily to the pressure gradient along the minor axis ( equation 12 ) which depends on @xmath12 and @xmath13 as @xmath140 . thus , a large @xmath104 requires small , dense atmospheres . in this context , the importance of ellipticity is clear : @xmath14 determines the ratio of core radii along the major and minor axes . for high @xmath14 , the lobes expand into material which does not vary much in density during the active phase whereas the wings quickly escape the core region . this behavior is unsurprising ; @xcite and @xcite make the same basic argument for wings produced when the jet is coaligned with the major axis of the atmosphere . however , the interplay between the parameters rules out predicting the wing length from first principles . for instance , one might imagine that we could produce wings in an atmosphere with @xmath141 ( figure [ atmosplots ] ) by making @xmath12 tiny and @xmath13 large . although this _ does _ increase the wing length , the radio galaxy expands beyond the dense part of the atmosphere during the ignition phase so the jet does not face as much resistance as might be expected . it is possible to produce wings in atmospheres with smaller @xmath14 by making the atmospheres triaxial ( using prolate ellipsoids ) instead of axisymmetric . @xcite expect triaxial atmospheres ( as opposed to the axisymmetric simulations of * ? ? ? * ) to produce longer wings because the backflow is collimated along a single minor axis as opposed to forming a torus at the midplane . we find that while a triaxial atmosphere does collimate the wings , it does not by itself make them substantially longer relative to an axisymmetric atmosphere with the same maximum @xmath14 because the backflow in our simulations is a compressible fluid and wing growth is driven by buoyancy ( the pressure gradient is the same along the minor axis ) . rather , triaxial atmospheres produce better - defined proto - wing channels during the ignition phase . these channels are then reinforced during the active phase . in other words , triaxiality increases @xmath105 at smaller values of @xmath14 because the ignition cocoon is less axisymmetric . finally , we test the sensitivity of wing formation to the degree of alignment between the jet and major axis @xmath127 ( figure [ atmosplots]d ) . long wings are difficult to produce when the jet is misaligned with the major axis , and the wings produced differ in character . this is because the ignition cocoon produces proto - wings perpendicular to the jet instead of along the atmosphere s minor axis . therefore , the longer wings which develop during the active phase do not benefit as much from the initial supersonic expansion . wings also become increasingly associated with the lobe on their side of the major axis as @xmath127 increases because the lobes are bent by the atmosphere and the back - flows no longer make it all the way to the midplane before flowing into the wings . hence , the lobe - wing pairs are mirror - symmetric about the midplane . the angle between the lobe and the wing in each pair is also largely determined by @xmath127 . lcclccc + + standard & y & 0.01 & @xmath142 $ ] & @xmath143 & @xmath144 & 0.62 + + + + sa_v30 & n & 0.01 & 30 & @xmath143 & @xmath144 & 0.39 + sa_v50 & y & 0.01 & 50 & @xmath143 & @xmath144 & 0.43 + sa_v80 & y & 0.01 & 80 & @xmath143 & @xmath144 & 0.44 + sa_v100 & y & 0.01 & 100 & @xmath143 & @xmath144 & 0.47 + sa_ve1 & y & 0.01 & @xmath145 $ ] & @xmath143 & @xmath144 & 0.48 + sa_ve6 & y & 0.01 & @xmath146 $ ] & @xmath143 & @xmath144 & 0.67 + sa_vg3 & y & 0.01 & @xmath147 $ ] & @xmath143 & @xmath144 & 0.47 + sa_vg6 & y & 0.01 & @xmath148 $ ] & @xmath143 & @xmath144 & 0.51 + + + + sa_d0.005 & y & 0.005 & @xmath149 $ ] & @xmath143 & @xmath144 & 0.73 + sa_d0.05 & n & 0.05 & @xmath149 $ ] & @xmath143 & @xmath144 & 0.19 + sa_d0.10 & n & 0.10 & @xmath149 $ ] & @xmath143 & @xmath144 & 0.09 + + + + sa_b30 & y & 0.01 & @xmath149 $ ] & @xmath143 & @xmath150 & 0.51 + sa_b7.5 & y & 0.01 & @xmath149 $ ] & @xmath143 & @xmath151 & 0.76 + sa_b5.0 & y & 0.01 & @xmath149 $ ] & @xmath143 & @xmath152 & 0.87 [ [ jet ] ] _ jet _ + + + + + in contrast to the atmosphere parameter exploration , determining the dependence of @xmath105 on the character of the jet is difficult because of the potential for time dependence in the jet power . this time dependence , along with the small width of the jets , also makes it more difficult to test our findings observationally . we insist that the jets be light , hypersonic flows ( section 2 ) which produce sources which resemble double - lobed radio galaxies , and within these constraints test the dependence of @xmath105 on jet power as a function of time , jet width , and density of jet material ( figure [ jetplots ] ) . for each of the simulations in figure [ jetplots ] we use the standard atmosphere with @xmath26 , @xmath111 , @xmath128 , and @xmath153 . note that figures [ jetplots]a and figures [ jetplots]b do _ not _ conserve kinetic luminosity between runs , whereas figures [ jetplots]c and figures [ jetplots]d do . the exploration of jet velocity is conducted with jets with the standard width of @xmath134 ( @xmath154 code units wide ) and density of @xmath133 . figures [ jetplots]a and [ jetplots]b demonstrate the importance of decaying jets to long wings . in figure [ jetplots]a , we use jets with differing velocities but no time dependence , finding that none produce a large @xmath105 . @xmath105 actually _ increases _ with increasing @xmath40 even though the faster jets also punch through the atmosphere more quickly . this is because the overpressured cocoon produced by a weaker jet during the ignition phase is less overpressured and thus produces smaller proto - wings . below @xmath155 , the sources are not cocoon - bounded at all . since the jets do not decay , @xmath104 obviously decreases with time . powerful jets which decay ( figure [ jetplots]b ) are effective at producing wings because the proto - wings form during the ignition phase and the jet generates strong back - flows early on when the jet head is close to the midplane . since the lobes grow increasingly slowly , the subsonically expanding wings keep pace with the active lobes more easily . as is clear in figure [ jetplots]b , the slower the decay , the smaller @xmath105 ( diamonds represent exponentially decaying jets and triangles jets with @xmath108 $ ] ) . jets with _ increasing _ velocity do not produce long wings because a powerful jet is necessary at the inception of activity to produce proto - wings . in figures [ jetplots]c and figures [ jetplots]d we investigate the dependence of @xmath105 on jet width and density . we vary @xmath9 and @xmath41 while conserving @xmath39 by varying the initial velocity of the jet @xmath109 , using the standard form of @xmath156 $ ] . we limit these runs to those with @xmath109 capable of producing an overpressured cocoon that seeds proto - wings . the nozzle width of the jet clearly has a strong effect on @xmath105 ( figure [ jetplots]c ) . increasing the jet width makes the ignition cocoon larger and more overpressured during the ignition phase , promoting expansion along the minor axis . the lobe width during the active phase increases with increasing @xmath9 , dissipating the jet thrust over a larger solid angle . short , fat lobes attached to fat wings result . on the other hand , if @xmath9 is tiny , the jet drills through the surrounding atmosphere quickly , and thin , long lobes result . it is not clear whether jet widths actually vary substantially between sources and what determines the width of the jet ; all jets are very narrow . thus , our results are more generally a statement that jets which dissipate their thrust over a wider area produce longer wings . figure [ jetplots]d shows @xmath105 as a function of @xmath41 . the sharp decline with increasing @xmath41 is due to the thrust carried by the jet . even at moderate velocities , denser material drives the lobes forward much faster than light material while at the same time driving weaker back - flows . hence , wing formation is not favored . we note that when @xmath157 , the sound speed of the lobe material is not much greater than that of the ambient medium and the jet does not develop the usual kh instabilities . these jets are therefore not `` light '' as required to reproduce realistic radio sources with non - relativistic hydrodynamics ( section 2 ) . however , these runs are shown along with the light ones to illustrate the importance of dissipating jet thrust to long wings . lccccclccc[t ] + + wide_ve1 & 1.5 & 4.0 & 0.0 & 0.75 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath151 & 0.58 + wide_ve3 & 1.5 & 4.0 & 0.0 & 0.75 & 0.01 & @xmath142 $ ] & @xmath158 & @xmath151 & 0.73 + tx_e75_ve1 & 1.0 & 3.0 & 0.375 & 0.75 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath151 & 0.63 + tx_e75_ve1_big & 1.5 & 4.0 & 0.375 & 0.75 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath151 & 0.68 + ( restart at @xmath159 ) & & & & & & & & & 0.78 + tx_e75_ve1_b7.5 & 0.75 & 3.0 & 0.375 & 0.75 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath151 & 1.02 + tx_e75_ve3_b5.0 & 1.5 & 4.0 & 0.375 & 0.75 & 0.01 & @xmath142 $ ] & @xmath158 & @xmath160 & 0.82 + tx_e60 & 0.75 & 3.0 & 0.30 & 0.60 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath151 & 0.44 + tx_e60_wide & 0.75 & 3.0 & 0.30 & 0.60 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath160 & 0.67 + ( restart at @xmath159 ) & & & & & & & & & 0.97 + tx_e60_d5.0 & 0.75 & 5.0 & 0.30 & 0.60 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath151 & 0.71 + tx_e45_d5.0_wide & 0.75 & 5.0 & 0.225 & 0.45 & 0.01 & @xmath142 $ ] & @xmath158 & @xmath160 & 0.58 + + + + disk_ve1 & 2.0 & 3.0 & 0.25 & 0.50 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath151 & 0.60 + & 0.25 & 8.0 & 0.0 & 0.90 & & & & & + disk_ve3 & 2.0 & 3.0 & 0.25 & 0.50 & 0.01 & @xmath142 $ ] & @xmath158 & @xmath151 & 0.83 + & 0.25 & 8.0 & 0.0 & 0.90 & & & & & + turbulent & 1.0 & 3.0 & 0.0 & 0.75 & 0.01 & @xmath142 $ ] & @xmath158 & @xmath151 & 0.64 + shell_ve1 & 1.0 & 3.0 & 0.0 & 0.75 & 0.01 & @xmath145 $ ] & @xmath158 & @xmath151 & 0.53 + shell_ve3 & 1.0 & 3.0 & 0.0 & 0.75 & 0.01 & @xmath142 $ ] & @xmath158 & @xmath151 & 0.71 + intermittent & 2.0 & 2.0 & 0.0 & 0.75 & 0.01 & @xmath161 $ ] & @xmath158 & @xmath151 & 0.44 + intermittent2 & 1.0 & 3.0 & 0.0 & 0.75 & 0.01 & @xmath162 $ ] & @xmath158 & @xmath151 & 0.47 + slow_start_1 & 1.0 & 1.5 & 0.0 & 0.75 & 0.01 & @xmath163 _ to _ & @xmath158 & @xmath151 & 0.42 + & & & & & & @xmath145 $ ] & & & + slow_start_2 & 1.0 & 1.5 & 0.0 & 0.75 & 0.01 & @xmath163 _ to _ & @xmath158 & @xmath160 & 0.45 + & & & & & & @xmath145 $ ] & & & + slow_start_3 & 1.0 & 5.0 & 0.0 & 0.75 & 0.01 & @xmath163 _ to _ & @xmath158 & @xmath160 & 0.55 + & & & & & & @xmath145 $ ] & & & [ [ synthesis - runs ] ] _ synthesis runs _ + + + + + + + + + + + + + + + + we have used the insights gained from examining the dependence of wing prominence on various atmosphere and jet parameters to new runs with complex atmospheres and jets in an attempt to produce long , realistic wings . these simulations ( listed in table [ synthesistable ] ) are not a systematic exploration of any phenomenon . we offer a few brief observations here . triaxial atmospheres offer the best hope of making long wings at lower @xmath14 ( e.g. figure [ diffwings ] ) , but still can not produce substantial wings in our simulations for @xmath164 . several runs from a large suite of such simulations are listed in table [ synthesistable ] with the prefix tx . generally , they follow the same trends as described in above ; the importance of both favorable atmosphere and jet parameters to wings is obvious in the simulation with @xmath165 ( tx_e45_d5.0_wide ) , where a wide , decaying jet in a small , dense atmosphere is required to produce wings comparable to that in the standard atmosphere at lower ellipticity . embedded disks of hot material ( i.e. a thick disk ; runs disk_ve1 and disk_ve3 ) , however , alleviate the problem , allowing prominent wings to grow in atmospheres with globally small @xmath14 . these disks are somewhat denser than the larger ism , so the ignition cocoon effectively encounters a small , dense , highly elliptical atmosphere during the ignition stage . the disks are blown apart by the blast wave from the ignition stage and might not be observable once the radio galaxy has turned on . in most of our models , the ignition stage occurs in a smooth , relaxed medium . in real galaxies , the ignition stage would occur in the galactic center where mergers , dynamical effects , and other phenomena associated with agn can significantly disturb the ism . we do not attempt to model real galaxies , but find that adding kolmogorov spectrum turbulence to our atmospheres or small bar - like perturbations to the underlying gravitational potential near the nucleus does not have a large effect on the ignition stage . as long as there is a sufficient amount of ism distributed around the nucleus , an overpressured cocoon can form . on the other hand , in runs where we begin with a weak jet of @xmath166 ( twice the sound speed of the lobe material ) , and no cocoon is formed , wings do not form ( the slow_start runs in table [ synthesistable ] ) . the backflow model has difficulty producing wings when a weak agn suddenly becomes more powerful because an overpressured cocoon only forms when the jet is completely confined by the atmosphere . these runs can , however , produce something akin to z - shaped morphology by virtue of the lobes escaping the densest regions of the atmosphere before the powerful jet turns on . we have briefly investigated the potential for gas - rich `` stellar shells '' from minor mergers located periodically along the major axis to produce wings ( motivated by * ? ? ? we use a simple model in which only the positive peaks of two - dimensional sine / sinc waves are added to the density of an underlying elliptical atmosphere ; the center is cut out . the amplitudes are set at a maximum of @xmath13 . the shells do not significantly impact the formation of wings emanating from the center of the galaxy , although they do resist the jet , allowing for a higher @xmath104 ( the shell runs in table [ synthesistable ] ) . in the model of @xcite , wings are instead produced near the site of the shell , and we can not reproduce this . obviously , these runs are not a thorough exploration of the effect of stellar shells , especially if they actually bend jets ( in which case they are beyond the scope of our models ) . motivated by the importance of a reinjected jet ( section 3.1 ) , we examine the role of rapid intermittency ( runs intermittent in table [ synthesistable ] ) . by this we mean jets which experience multiple outbursts during the run and which are essentially in an `` on '' state or an `` off '' state . generally , we find that rapid intermittency has the effect of suppressing wing formation and results in much more regular cocoons ( in terms of their projected morphology ) than single outbursts . on the other hand , ragged cocoons are produced by long periods of dead time . in neither case are wings promoted , as each successive brief outburst deposits most of its thrust at the ends of the lobes , far from the midplane . the chief reason why such intermittency does not produce prominent wings is that wings expand subsonically . since the duty cycles of the intermittent jets are much shorter than the crossing time , the cocoon expands as if a moderately powerful jet of constant velocity were powering it . unlike in single decaying outbursts , the intermittent jets deposit most of their thrust far from the midplane . on the other hand , intermittency on the timescale of an @xmath167-folding time of a decaying jet can be effective . because we have only solved the equations of hydrodynamics in evolving our simple models , it is worth considering the impact of additional complexity . our models do not include magnetic fields , special relativity , or radiative losses , and our simple setup does not take into account complex jet or atmospheric structure , feedback , and other processes that may be important to radio galaxy morphology . in practice , we do not believe these omissions invalidate our results . for example , special relativity and magnetic fields must play crucial roles in determining the character and transverse structure of the jet , but we only insist that our simulated jets reproduce the collimation , hot spots , and back - flows of fr ii sources . in other respects ( e.g. radiative efficiency ) , the jet and agn are inside a `` black box . '' because we are concerned with the behavior of lobe material , we believe these omissions are justified . likewise , radiative losses ( depleting of lobe energy via synchrotron emission and inverse compton scattering ) only become important relative to adiabatic losses at late times and are not in a position to influence the formation of wings . radiative losses from the icm are also irrelevant on the timescales of agn outbursts , so feedback ( the connection of jet power to the amount of material crossing the inner boundary ) resulting from cluster cooling is unimportant on these timescales . as to feedback from _ backflow _ @xcite , there is no obvious recipe to describe what happens to the mass flowing across the inner boundary , since the inner boundary radius is much larger than the nuclear engine , but as the back - flows straighten , less material will have velocity vectors pointing towards the origin and eventually the agn would cease to be `` fed '' by backflow . this is consistent with our entirely artificial recipe for decaying jets . on the other hand , magnetic fields ( that must be injected with the jet ) and relativistic jets ( which have a higher thrust for a given mass - energy density ) may strongly influence the behavior of lobe material . toroidal magnetic fields may help collimate the backflow and retard its mixing ( e.g. * ? ? ? the apparent continuity of fields in some wings ( e.g. ngc 326 in * ? ? ? * ) is suggestive , especially considering that in our models the lack of collimation leads to wide wings . however , @xcite find in simulations of fr ii sources that turbulence sets in within the cocoon even when ordered fields existed earlier , so it is unclear that the magnetic fields in the wings are actually toroidal . as to the realism of the _ jet _ , it is not our goal to understand the jet physics in detail , but we note the persistence of experimental hypersonic fluid jets in recent work by @xcite , which the authors use to argue that magnetic collimation is only required near the base of the jet . the importance of relativistic jets to the cocooon morphology is unclear . @xcite find that , for jets matched by velocity , pressure , radius and power , relativistic jets produce preferentially wider cocoons than classical ones . the authors suggest that this effect may be accounted for by the higher thrust in the relativistic case . however , @xcite find the opposite when using the same matching conditions : relativistic jets tend to produce narrower cocoons compared to classical jets ( more specifically , jets with a higher lorentz factor have smaller cocoons ) . without matched relativistic simulations it is impossible to directly test the importance of special relativity on our cocoon sizes , but it is remarkable that @xcite and @xcite agree that the qualitative nature of the cocoon is unchanged . including relativity may modify @xmath104 but would not fundamentally change the appearance of our winged sources because relativistic physics in the lobes would not change the processes which produce wings . in other words , we believe that special relativity is a second - order effect akin to magnetic fields . it is worth asking whether wings form only because our atmospheres are `` smooth '' , i.e. because we have not included turbulence in the icm or structure near the center ( where the environment is presumably complex ) . this question is particularly important given the observed interactions between jets and molecular clouds ( e.g. * ? ? ? * ) , stellar shells ( e.g. * ? ? ? * ) , gas in companion galaxies @xcite , and more generally the complex nuclear environments of active galaxies ( e.g. * ? ? ? to this end , we have introduced density perturbations mimicking stellar shells along the major axis at various intervals , tested the impact of a single ring near the nucleus , and introduced kolmogorov - spectrum turbulence to the icm . none of these structures alter the same basic evolution of winged sources . additional jet physics would need to be in place in order to determine whether the jet itself can be bent by interaction with high density pockets @xcite . all of these tests use density perturbations with a maximum amplitude of the core density @xmath13 . finally , bulk flows in clusters are clearly important to radio galaxy structure @xcite , although most xrgs are large , strongly bridged sources in which this may not be a defining effect . it is encouraging that our simulated radio galaxies reproduce the basic features of other simulations in the literature ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , but it is worth considering the impact of boundary conditions and the jet structure on the morphology of the lobes our setup is designed to produce lobes like those of radio galaxies with physics and conditions inherently different from those encountered in nature . the mixing experienced by the lobes is also artificial ( naturally set by the size of the grid zones , which vary along @xmath4 and @xmath5 ) , but mixing is too slow a process to suppress the formation of wings . the inner boundary is the most significant artifact of our simulation because it represents no physical analog but rather allows us to hide the agn and the jet collimation mechanism . in addition , backflow crossing the boundary slows down and disappears . this has two consequences : not all the backflow can be harnessed , and flows may be effectively directed around the boundary by eddies as slowing backflow crosses the boundary . in the first case , the amount of material is to small to influence the wings . in the second , we find that the inner boundary must be small compared to @xmath12 ( keeping the physical size of the jets fixed ) in order to prevent the initial cocoon from being unduly influenced by the sphere . inner boundary spheres which are too large tend to promote material flowing around them , hence promoting early wings , but suppress midplane mergers and hence the prominence of later wings . the initial interstices also survive for longer for oversized @xmath7 . for typical atmospheres ( @xmath168 ) , @xmath7 must be somewhat smaller than 0.1 code units ; @xmath49 is a compromise between a small impact on the initial cocoon and reasonably sized grid zones . runs with @xmath169 do not appear substantially different from those with @xmath49 , and the timestep is unreasonably small . we precess the jet injection footpoints rapidly ( @xmath44 hz ) in a very tight circle around the poles in order to break up the jet symmetry and spread its thrust out over a larger working surface . we only require that our jet reproduce internal features seen in other hydrodynamic simulations along with the terminal shocks that give rise to the back - flows and lobes . without breaking up the jets , the jet head travels very quickly and produces unrealistic lobes ( e.g. * ? ? ? * ) ; in mhd simulations we would expect helical instabilities to break up the jet . our precession scheme is thus an artifice that is motivated by observables but whose recipe is not an attempt to model a physical process ( c.f . however , one might worry about its influence on backflow and wings because the precession angle is a free parameter ( subject to the condition that @xmath43 ) . in other words , we can tune the width of the lobes and the rate of growth of the radio galaxy within some narrow range of parameters . since stalling the jet head contributes to wing prominence via shortening the active lobes , is the axial ratio @xmath104 artificially high ? we do not believe so . even with tiny precession angles , we form long wings in highly eccentric atmospheres ( the only atmospheres where long wings form ) . moreover , the jet head advance speed also depends on the kinetic luminosity as a function of time . we have successfully produced winged and x - shaped sources solely by redirecting back - flow with static environmental pressure gradients , and are now in a position to compare our simulated sources to observed xrgs ( figure [ gallery ] ) in order to critique proposed wing formation mechanisms involving back - flows . in these models , wings are produced as plasma flowing back from the jet heads is deflected into a direction misaligned with the jets . how this deflection occurs is uncertain @xcite , and until now the backflow model has not been rigorously investigated with three - dimensional simulations . here we outline a series of predictions ( expectations ) for properties which are , at least in principle , observable if our simulations are an adequate representation of wing formation . these predictions are also summarized in table [ predictions ] . we then assess the backflow model in light of these predictions and briefly discuss the implications of our work for the wider sample of distortions to the canonical double - lobed morphology of bridged radio galaxies . [ [ x - shaped - and - winged - sources - are - one - family ] ] _ x - shaped and winged sources are one family _ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + the most natural consequence of the backflow model is that short and long wings are produced in the same way . short wings should then be more common than long ones since the long wings are harder to produce . although this is an obvious point , we note it as a potential observational test because sources with shorter wings and the candidate x - shaped sources @xcite have not been tested for several of the trends seen in xrgs ( most importantly , the jet major axis correlation * ? ? ? * ; * ? ? ? our simulated sources predict that winged sources will fall in line , albeit with a greater spread in parameter space . llll 1 . & short wings common ; ( intrinsically ) long & no in ss09 sample & ss09 + & wings rare & ( yes including c07 ? ) & + 2 . & projection tends to enhance @xmath104 & undetermined & + 3 . & @xmath104 correlated with @xmath170 & sample size too small & hk10 + & ( weaker correlation with host @xmath14 ) & no ? & c02,ss09 + 4 . & wings require a jet aligned near & yes & c02,ss09,hk10 + & the major axis & & + 5 . & @xmath104 enhanced by higher pressure , & yes ? & l10 + & small atmospheres & yes ? & ss09 + 6 . & most xrgs should be weak fr iis & yes & c09,l10 + & ( and hence strongly bridged sources ) & yes & lw84 + 7 . & intermittency on scales of 310 myr & undetermined & + & produces longer wings & & + 8 . & wings are fainter than lobes & yes & lw84 + 9 . & wings have flat spectral indices & mixed results & lr07 + 10 . & backflow follows existing channels & yes ? & ss09 + & ( collimation requires existing channels ) & & + 11 . & wings grow subsonically & no ? ( large wings ) & + & ( hence wing agn outbursts are old ) & no ? & dt02 + 12 . & flow speeds in wings should be & undetermined & + & transonic for lobe material & & + & backflow model piggybacks on other hydrodynamic & yes ? & ss09 + & models & & [ [ projection - almost - always - enhances - wl ] ] _ projection almost always enhances @xmath104 _ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + wings in our models ( even those produced by triaxial atmospheres ) tend to be wider than the lobes and hold their shape better in rotation and projection . hence , we predict that projection almost always enhances @xmath104 . in other words , some observed sources with high @xmath171 probably have an intrinsically smaller aspect ratio ( e.g. figure [ gallery ] ) . if the projection angle can be worked out for a sufficient number of xrgs and winged sources , we expect the lobes in a number of xrgs to be shortened via projection , i.e. the intrinsic distribution of @xmath104 is shifted from the observed distribution ( an idea of the observed distribution may be found in * ? ? ? if this does not turn out to be the case , the backflow model would need a collimation mechanism ( see subsonic expansion below ) to be consistent with observations . unfortunately , this prediction makes it difficult to quantitatively compare the simulated population to the observed on because in most xrgs the projection angle is unknown . [ [ long - wings - require - high - ellipticity ] ] _ long wings require high ellipticity _ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + our simulations predict that intrinsically long wings can only be produced in very elliptical atmospheres ( figure [ atmosplots ] ) . the lowest @xmath14 for which we can make convincing wings is @xmath172 , and very long wings require @xmath173 . these values are much higher than the average powerful radio galaxy host ( @xmath174 ; * ? ? ? * ) and even most xrgs . a high initial atmospheric ellipticity ( or higher order asymmetry ) is a clear prediction of our models , but confirming this behavior in observed sources is difficult . for instance , a few sources with very small host @xmath14 and large @xmath104 ( the best example is ngc 326 ) appear to have wings originating _ outside _ the ism , so it is not evident that the ellipticity of the ism is always the relevant value for comparison . further , the ellipticity of the ism may differ substantially from that of the host galaxy on small scales @xcite even when in broad agreement on galactic scales @xcite . in this case , the best example in the literature is 3c 403 ( figure [ 3c403 ] , where @xmath175 and @xmath176 @xcite . indeed , we find that elliptical atmospheres of moderate @xmath14 with embedded disks are as effective at producing wings as single - component atmospheres with extreme @xmath14 , but the disks are largely destroyed by the radio galaxy . high resolution imaging may be required to see any disk remnants . the triaxiality of elliptical galaxies also plays a role , since we may not measure a true ellipticity . finally , because projection tends to enhance @xmath104 , extreme values of @xmath177 ( as we use in the standard atmosphere ) are not required to produce observed long wings . all else being equal , we still expect the intrinsic @xmath104 to be correlated with @xmath14 , but the caveats outlined above make it a difficult proposal to test presently . [ [ wings - require - a - jet - pointed - nearly - along - the - major - axis - of - an - anisotropic - environment - delta - textpa - sim-0 ] ] _ wings require a jet pointed nearly along the major axis of an anisotropic environment ( @xmath178 ) _ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + as expected ( based on the studies in * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , wings require a strongly asymmetric environment in which the jet is stalled by progression along the long axis ( figure [ atmosplots ] ) . the wings grow in the favorable pressure gradients along the minor axis or axes . our models expect a fairly strict requirement for the jet to be within @xmath179 of the major axis for substantial wings to be produced via the overpressured cocoon channel ( the usual method for seeding wings in our models ) . the requirement is even more stringent in the event that a disk is present , indicating that disks are likely to be important only in a minority of cases . as @xmath127 increases , the wings become shorter ( figure [ atmosplots ] ) as well as increasingly associated with the lobe on the same side of the major axis . these lobe - wing pairs thus have mirror symmetry about the midplane . this is in agreement with radio maps of such sources . since wings long enough to qualify the source as `` winged '' can be produced for a @xmath179 range of major axis jet separation and truly x - shaped sources are only ( intrinsically ) produced for @xmath180 , we would expect @xmath181 of winged sources to have lobe - wing pairs , with acute angles between a lobe and its wing produced for smaller angles and obtuse angles for larger ones . thus , we suggest that the existence of a large number of sources with lobe - wing pairs rather than true x - shaped morphology is consistent with the backflow model . however , we would also expect that wing length would decrease with increasing @xmath127 ; this has not yet been measured . [ [ longer - wings - are - produced - in - higher - pressure - smaller - atmospheres ] ] _ longer wings are produced in higher pressure , smaller atmospheres _ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + steeper pressure gradients occur for smaller core radius and higher core density / pressure ( equation 12 ) . since the steepness of the pressure gradient influences the rate at which wings grow , these small , dense atmospheres ( e.g. the ism ) are better at producing wings . moreover , wings can only form via the overpressured cocoon channel if the cocoon can escape the central core region and its high density before it comes into pressure equilibrium . the top panels of figure [ coevol ] demonstrate this point ( see also figure [ atmosplots ] ) . finally , a high core density also resists the jet advance along the major axis , allowing the wings to grow longer . thus , our models expect that the hosts of xrgs ( where the jet is presumably pointed along the major axis ) have higher pressure , on average , than those normal radio galaxies with jet geometry favorable to wings . @xcite find that the nuclear regions of xrgs have high temperatures ( @xmath182 k ) indicating that these regions may be overpressured . although this does not directly correspond to our models of relaxed , isothermal atmospheres , our requirement for a small , high pressure environment is in agreement with their work . [ [ most - xrgs - should - be - weak - friis ] ] _ most xrgs should be weak fr iis _ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + another key ingredient in our models is a decaying jet that begins as a powerful fr ii source and decays to luminosities more typical of fr i sources . a powerful jet is required to produce the overpressured cocoon and drive wings early on , whereas once wings begin expanding subsonically , a weakening jet ( which advances increasingly slowly along the major axis of the atmosphere ) allows the wings to become quite prominent . if the atmosphere is much smaller than the radio galaxy ( as in many cases ) , a decaying jet will still grow more and more slowly , allowing subsonically growing wings to keep up . this finding naturally explains the observations that xrg radio powers tend to lie near the fr i / ii `` break '' @xcite while also possessing strong bridges associated with powerful fr ii sources @xcite . @xcite argue that xrgs are the archetypal transition population between fr i and fr ii sources , with about half the xrgs in their sample having weak emission lines from the agn ( weak - lined fr iis are otherwise uncommon ) . this suggests that there is indeed an evolutionary progression in the agn and jet luminosity of xrgs . our models predict a fast - rise exponential - decay profile in which the xrg spends most of its lifetime as a weak fr ii source ( figure [ jetplots ] ) . however , we note that @xcite find in an sdss study that fr i and ii sources are not as obviously separated as in @xcite and that fr ii morphology occurs for a variety of radio powers and host galaxy masses . intermittency is another important prediction of our models for observed xrgs , but the action of intermittency in our simulations may represent a more complex underlying process . intermittency on timescales similar to the @xmath167-folding time of a decaying jet ( near the early passive phase ) is effective in our simulations because it allows the jet to bypass the overpressured cocoon stage and form a bow shock within the radio lobe itself , thereby depositing a large amount of backflow into the wings from a very powerful jet close to the wing bases . reinjection is effective in a relatively narrow window of time ( 310 myr depending on the size of the radio galaxy , e.g. figure [ standardgrowth ] ) between the active and passive phases , but this may be plausible : the ripples in the perseus cluster @xcite and abell 2052 @xcite do imply an agn duty cycle of about 10 myr . this is consistent with our simulated sources between 50100 kpc across . thus , it is conceivable that regular intermittency is a viable mechanism for enhancing wings , although we note that if the intermittency is very rapid or very slow , it is ineffective ( section 3.2 ) . on the other hand , intermittency could be replaced in our models by any mechanism which drives back - flows nearer the base of the wings such as the motion of denser ism into the path of the jet . we note that intermittency may also be effective for wings attached to the primary lobes farther from the nucleus ; a reinjected jet quickly re - establishes the cocoon and drives strong back - flows along its length as it rapidly reaches the prior hot spot . if the jet reinjected is similar to the outburst which initially formed the wings , we would still expect xrgs mostly to be weak fr ii sources . [ [ wings - should - be - fainter - than - the - primary - lobes ] ] _ wings should be fainter than the primary lobes _ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + the active lobes of xrgs are typically brighter than the wings . this behavior persists at lower ( mhz ) frequencies and is therefore not obviously attributable to spectral ageing , although spectral ageing may be in play for some xrgs @xcite . our models expect the wings to be fainter because they are substantially wider than the collimated primary lobes yet contain ( in most cases ) less lobe material . we therefore suppose that the wing material spreads out and decreases the magnetic field energy density @xmath183 . assuming that @xmath183 decreases with increasing volume ( i.e. assuming something like equipartition conditions , although precise equipartition is not required ) , the synchrotron emissivity in the wings will be substantially smaller than that in the lobes . in other words , `` spherical '' wings a factor of @xmath184 wider than the active lobes will have a factor of @xmath185 smaller @xmath183 than in the lobes . in optically thin conditions ( a good assumption ) , the surface brightness obtained from integrating through the wings only recovers a factor of @xmath184 in the wings , so we expect them to be a factor of @xmath186 dimmer than the active lobes ( more generally , the square of the ratio between the wing width and the lobe width ) for electrons of the same @xmath187 . of course , in projection the wings may appear even dimmer since the lobes will tend to be shortened and therefore increase in surface brightness ( ignoring relativistic dimming of the counterjet ) whereas the wings will not . clearly , this prediction rests on a number of assumptions not included in our models and is therefore somewhat weak . since the wings are filled with turbulent plasma , it is also possible that the magnetic field strength is increased by winding up of fields as the material mixes . moreover , the jet power ( and therefore the back - flow speed ) is time dependent in our simulations , so the ratio between the surface brightness of the lobes and the wings would be time dependent in our model as well . nonetheless , the diffusion of material in the wings and the aforementioned projection effects seem likely to dim the wings in the backflow model . [ [ backflow - follows - existing - channels ] ] _ backflow follows existing channels _ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + most of the wings in our simulated sources are seeded by the pressure - driven expansion of an overpressured cocoon early in the source s life . this cocoon expands asymmetrically because of the asymmetric pressure gradients and forms proto - wings which are later bolstered by the merger of laminar back - flows near the midplane sending material into the wings during the active phase . the laminar back - flows themselves are ineffective at drilling new channels . during the overpressured phase they acquire a vorticity near the terminal shocks that leads them to follow the contact discontinuity of the cocoon and ultimately flow back towards the agn ( figure [ backflow ] ) . during the active phase , they are mostly straight and flow towards the midplane . lobe material is also ineffective at producing channels because it follows the path of least resistance : in a confined ellipsoidal cocoon , it will simply spread out to increase the pressure throughout the cocoon rather than break out in a particular direction . this can be seen in the false synchrotron maps in figure [ coevol ] . these maps assume equipartition and are not indicative of what the sources would look like in the ghz bands . rather , they effectively trace pressure in the cocoon , and it is easy to see that the pressure in the confined cocoon on the left side of the figure remains high relative to the x - shaped source ( the false radio maps use the same scale ) . apart from proto - wings produced by the early pressure - driven expansions , kelvin - helmholtz and rayleigh - taylor instabilities produce channels which backflow reinforces . given sufficient time and a powerful jet , these whorls and fingers can become wings in their own right , but likewise this is due to the growth of the instabilities and buoyancy rather than redirected laminar back - flows . notably , some sources exhibit wings which do _ not _ emanate from the center or are unlikely to have been produced via an overpressured cocoon ( e.g. ngc 326 ) . because it is difficult to channel the back - flows , we hypothesize that if the backflow model is correct , some seed proto - wing was necessary to produce such sources . the origin of these proto - wings is unclear . [ [ wings - expand - subsonically ] ] _ wings expand subsonically _ + + + + + + + + + + + + + + + + + + + + + + + + + + + related to the expectation that back - flows are not directly responsible for drilling channels is the prediction that the wings of radio galaxies expand subsonically for most of their lifetimes ( figures [ backflow ] and [ wingvel ] ) . this expectation is contrary to the @xcite `` overpressured cocoon '' and @xcite or @xcite `` buoyant backflow '' proposals in the literature ( section 1 ) and poses a serious problem for the backflow model in terms of explaining long wings . we therefore discuss this point in some detail below . the problem of subsonic expansion can be broken into two distinct objections : ( 1 ) if the wings grow subsonically while the primary lobes grow supersonically , how can wings be longer than the lobes ? and ( 2 ) subsonic expansion implies wing lifetimes of more than 100 myr for sources hundreds of kpc across . the first objection is easier to reconcile with our models because our jets provide a natural mechanism for the subsonic expansion of the lobes , and it is not clear whether most radio galaxies ( even powerful ones ) really do expand very supersonically . further , as we have noted , projection tends to enhance wings . however , the lifetimes of large sources are more difficult to explain . to see this , consider that for a rich cluster atmosphere with @xmath188 kev ( this corresponds to a @xmath189 km s@xmath45 or , conveniently , @xmath117 kpc myr@xmath45 ) , subsonic growth implies that wings 100 kpc from base - to - tip would be at least 100 myr old . in group environments , where @xmath190 km s@xmath45 and monotonically declines at large radii @xcite , the longest observed wings ( e.g. 3c 315 whose wings span @xmath191 kpc ) could be almost 500 myr old ! even if agn activity persisted this long , synchrotron cooling might set in : if the bulk flows of replenishing backflow travel at a lobe sound speed @xmath192 , they would take between 20100 myr to replenish the leading edges of the wings . depending on the break frequency and assuming magnetic fields in equipartition ( @xmath193 of a few @xmath194 g ) and a lorentz factor @xmath187 of several@xmath195 , this may easily exceed the radiative cooling time . to solve this problem , variations of the backflow model require that the wings actually do expand supersonically . in the `` overpressured cocoon '' model @xcite , supersonic wing growth is achieved through sustained pressure - driven expansion , whereas in the `` buoyant backflow '' model @xcite , wings are driven by the nearly free expansion of diverted hypersonic back - flows . our models reproduce overpressured cocoons and hypersonic back - flows , but the overall wing advance speed is subsonic . for the overpressured cocoon model to produce long wings , a sustained overpressured state must be maintained to drive outflows . however , in our models the highly overpressured cocoon that forms soon after ignition quickly expands to reach pressure equilibrium : the supersonic expansion phase lasts at most around 15% of the lifetime of the radio galaxy regardless of the jet we inject ( e.g. figure [ wingvel ] ) . as the radius of the ( isobaric ) cocoon grows , its pressure falls much more rapidly than does that of the surrounding atmosphere at a similar radius ( see also * ? ? ? * figure 2 in their paper ) , so the advance speed of the lateral cocoon expansion falls precipitously from a peak of about mach 2 . thus , overpressured expansion appears to be relevant only very early in the life of the source and by itself can produce only short wings . indeed , this conclusion is supported by earlier simulations @xcite where even in advantageous triaxial atmospheres , the overpressured cocoon produced at most an intrinsic @xmath104 of @xmath196 ( the jets in their work were held at constant velocity and @xmath104 would thus decrease over time ) . @xcite suggest that the overpressured state may instead result from backflow piling up upon reaching the central ( dense ) region of the galaxy . our models do not support this scenario , since the cocoon is contiguous and nearly isobaric throughout the active phase . the buoyant backflow model argues that a combination of buoyancy forces and wings driven by hypersonic flows produce large x - shaped sources . for instance , @xcite estimate that the current outburst in 3c 403 ( figure [ 3c403 ] ) started 16 myr ago . if this outburst was solely responsible for generating the @xmath0 kpc wings hydrodynamically ( as preferred by * ? ? ? * ) , the average expansion speed of the wings would have to be @xmath197 km s@xmath45 ( @xmath198 ) . in contrast , a typical sound speed for galaxy groups is @xmath190 km s@xmath45 , a factor of @xmath199 smaller . likewise , @xcite prefer hypersonic expansion . what could drive such wings ? the answer invoked by these authors is redirected back - flows , which can be accelerated at the terminal hot spots up to a few percent of @xmath200 ( based on the spectral ageing distance method of * ? ? ? * ) . the observational inference of high backflow speeds is consistent with our models and others ( e.g. * ? ? ? * ) , which find back - flow speeds of about twice the lobe sound speed ( @xmath89 ) . further , @xcite argue that the collimated morphologies of some wings requires some driving force . however , it is not evident that these laminar back - flows can be harnessed to drive wing expansion ( see the prior prediction ) . subsonic wing growth in our simulations is a natural consequence of the tendency of backflow ( lobe material ) to mix and expand . the fast back - flows from the terminal shocks merge near the midplane , dissipating their speed and driving flows into the wings which are subsonic relative to the low - density material in the lobes ( figure [ backflow ] ) , although they are still supersonic relative to the background . these flows then expand and decelerate further in the wings , dissipating their thrust over a very large area . hence , detection of @xmath201 km s@xmath45 flows in the wings would be insufficient evidence for hypersonic wing expansion . supersonic wing expansion would require collimation to prevent thrust dissipation . can a collimation mechanism be found ? an obvious possibility we have not included is ordered magnetic fields , since particles can move along them much more easily than across , but @xcite find that in fr ii sources , lobe magnetic fields become turbulent on timescales of @xmath202 myr . more generally , we note that ( weak ) shocks found around double - lobed radio sources often imply cocoon expansion near mach 1 . if a powerful jet can not drive highly supersonic expansion , it seems unlikely that less powerful , more disorganized backflow could do so . lastly , it is important to emphasize that subsonic expansion is only a serious problem for very large wings or very young agn with a large @xmath104 ratio . our simulations can reproduce most winged and x - shaped sources with subsonic expansion if the agn outburst is between 10100 myr old instead of 110 myr . these simulations are also not the final word , as they only solve the equations of hydrodynamics in a relaxed , non - dynamic atmosphere . our results , taken in context , strongly implicate a hydrodynamic origin for x - shaped sources and a common origin for winged and x - shaped sources . we can produce bona fide xrgs ( i.e. with intrinsic @xmath203 ) with a single outburst from a jet with a plausible time dependent kinetic luminosity on timescales broadly consistent with observed sources . the radio galaxies we produce are consistent with prior simulations , and faithful reproductions of observed radio galaxies can be generated by tuning the viewing angle ( figure [ gallery ] ) . these sources are also consistent with most observed properties of the xrg population outlined above , and our simulations reproduce elements of the overpressured cocoon and buoyant backflow models while relying solely on the interaction of the radio lobe material with an anisotropic environment . hence , the backflow model remains a strong contender for the origin of xrgs it would be surprising if some xrgs were _ not _ produced in this manner . however , our simulations also present new challenges to the proposals in the literature : we find that long wings require unusually elliptical atmospheres and expand subsonically , making very long wings difficult to explain . having examined the deficiencies of our simulations , we can not identify an obvious internal remedy ; we note that pre - existing channels are necessary ( in the backflow model ) for some sources in the literature and may be generally important . the final prediction of our simulations is therefore that the xrg population is heterogenous and _ the backflow model works in tandem with other hydrodynamic models rather than solely on its own . _ in short , this is because our models make it too difficult to produce xrgs given their observed frequency . consider that winged and x - shaped sources make up 510% of double - lobed radio sources . if , as in our models , x - shaped sources are produced by a fortunate coincidence of jet geometry , ism ellipticity , and jet power , we might imagine that the fraction of x - shaped sources is some function of each of these variables . in the simplest case , where each of these factors contribute independently to promoting wings , the fraction of xrgs might look like @xmath204 where @xmath205 is the fraction of double - lobed sources which are x - shaped and the other @xmath206 values represent the fraction of sources for each variable which meet the threshold criterion for wings . the exponents are unknown measures of the relative importance of each variable . obviously , the fraction of xrgs must be smaller than the fraction of sources which meet any one criterion . now , consider that virtually all double - lobed radio galaxies emanate from elliptical hosts , and have apparently random jet major axis orientations ( as opposed to the very weakest radio - emitting agn whose jets seem to be pointed along the minor axis of their hosts ; * ? ? ? our models suggest that a jet within @xmath179 of the major axis of its host is required to produce wings . in a uniform distribution of radio jet position angles , we would thus expect that @xmath207 . assuming all powerful jets have a kinetic luminosity function conducive to forming xrgs , we still require about half of all sources where the jet is co - aligned with the major axis to have high ellipticity ( assuming the exponents are all equal to unity ) . given that ellipticity is clearly a very sensitive parameter and that the peak @xmath14 of powerful radio galaxy hosts is far below the wing threshold @xcite , it is clear that our simulations ( in this oversimplified formulation ) underpredict the observed frequency of winged sources . in other words , wings are _ too difficult _ to make in our models . to reconcile this result with the observations , either our models must be fundamentally deficient with respect to the behavior of backflow or they require alternate mechanisms to form proto - wing channels such as the jet stellar shell interaction model of @xcite or the jet merging ism explanation for z - shaped sources in @xcite . as we have seen , backflowing material reinforces any proto - wings and can turn them into full - fledged wings if the original mechanism fails to do so ; making the channels in the first place is what the backflow model can not easily do . thus , we propose that the backflow model has a commensal relationship with other hydrodynamic mechanisms for forming wings by reinforcing and growing any pre - existing channels accessible to a jet pointed along the major axis of its host galaxy . these channels would naturally grow most easily along the steepest pressure gradient . in cases where the jet power and atmosphere size match appropriately , these channels would be produced by the expansion of an overpressured cocoon as in our simulations . in other cases , channels could be produced by the interaction of the jet with structure in the ism or by minor mergers . in this scenario , we would still expect to see most of the predictions of our simulations listed above for the case of pure backflow in a relaxed atmosphere . at this point , it is worth revisiting the buoyant backflow model as phrased in @xcite , where xrgs are unified with other distortions to the canonical fr ii model ( about two thirds of strongly bridged radio galaxies show central distortions ) by the deflection of backflow around a denser medium . these distortions may also be related to radio galaxies with interrupted bridges by the `` superdisk '' model @xcite , in which gas displaced by a galaxy merger is believed responsible for docking the lobes in an asymmetric manner . because the backflow model as expressed in our simulations depends only on a fortuitous combination of atmosphere morphology and jet behavior to produce winged sources , we can indeed reproduce the basic bridge distortions of @xcite . however , we can not reproduce long z - shaped sources and some asymmetric distortions which do not appear to be the result of bulk motion or cluster turbulence . these include sources with a single wing on one side , sources with wings which themselves bend dramatically , sources with one fr i lobe and one fr ii lobe ( hymors ; * ? ? ? again , the main hindrance seems to be the inability to channel backflow ; jets which bend may solve this problem . although there are elements of the behavior of backflow which we do not presently understand , we come to the same conclusion as with x - shaped sources : the backflow model can account for other bridged distortions only in tandem with another hydrodynamic mechanism . we have conducted a series of three - dimensional hydrodynamic simulations of light , hypersonic jets to study the viability of the backflow model for the formation of wings in x - shaped radio galaxies . the xrgs seem to be a population unto themselves , with characteristic environmental geometry , radio power , black hole mass , etc . , and any successful model must account for these peculiarities . our main results follow . 1 . the jets , back - flows , and lobes in our simulations are similar to those in the recent literature , giving us confidence in the usefulness of our simulations as probes of the backflow model . 2 . wings in our models form in two stages : the establishment of channels or proto - wings and then buoyant ( usually subsonic ) expansion . specifically , our models corroborate the overpressured cocoon model of @xcite early on , but as cocoons quickly come to pressure equilibrium with their surroundings , most of the wing length represents subsonic expansion ( figure [ backflow ] ) . we have produced prominent wings by geometry and radio power alone , proving that the backflow model can , in principle , make x - shaped sources ( figure [ coevol ] and [ gallery ] ) . both the atmosphere and jet kinetic luminosity as a function of time are crucial to forming x - shaped sources . long wings are produced in a relatively small portion of parameter space , requiring galaxies with high ellipticity , decaying jets , proper jet orientation , and appropriate atmosphere size ( figures [ atmosplots ] and [ jetplots ] ) . 5 . the main challenges to the backflow model are the requirement for high ellipticity and subsonic wing growth . adding additional physics is not obviously helpful . the backflow model seems to require an additional mechanism to make proto - wings which the backflow reinforces ; in our models , we make these channels by the initial expansion of an overpressured cocoon in an anisotropic environment , but this can not explain every xrg . we can not form new channels solely by deflecting back - flows . if the backflow model can overcome the issues noted above , it is a very strong candidate for explaining x - shaped and other disturbed radio galaxies . our models naturally reproduce many of the characteristics of the xrg population ( table [ predictions ] ) . there are several natural extensions of this work which promise to be fruitful . first , adding magnetic fields and investigating other potential collimating mechanisms may rule out or boost the backflow model depending on the wing expansion speeds attained . second , our models rely on the formation of channels misaligned with the jets ; these have been proposed in several other contexts as well to explain radio galaxy morphology . the origin of these channels is not known , and identifying and testing candidates would be important for wing formation models . third , more realistic jets could determine whether the backflow model can support both long wings and active fr ii lobes with hot spots ; the bending of the jets in particular is an important issue . finally , models which produce the wings hydrodynamically but not explicitly by the deflection of backflow ( e.g. the recently proposed stellar shell model * ? ? ? * ) are worth exploring by making the atmospheres more complex in tandem with more realistic jets . , s. s. , & falle , s. a. e. g. 1996 , in astronomical society of the pacific conference series , vol . 100 , energy transport in radio galaxies and quasars , ed . e. hardee , a. h. bridle , & j. a. zensus , 173+

dileptons produced in heavy ion collisions interact only electromagnetically with their hadronic surroundings . especially the early phases of such collisions , when dilepton production rates are largest , can therefore be explored with lepton pairs . some of the initial ideas on photon @xcite and dilepton @xcite @xcite emission have met with considerable interest , and this area has now expanded into a major component of research in ultrarelativistic heavy ion collisions @xcite . on the experimental side , dilepton experiments for hadronic collisions @xcite@xcite have been complemented only recently by nucleus - nucleus experiments at very high energies @xcite@xcite . since the spectrum of lepton pairs and of real photons has been proposed as a signature of the quark - gluon plasma @xcite , it is imperative quantitatively to understand the emission mechanisms from the confined sector of qcd , which in this sense constitutes the background . guided by theoretical estimates , we can roughly divide the lepton pair production cross section , expressed in terms of its invariant mass @xmath1 , into a number of distinct regions . for @xmath2 , the drell - yan contribution dominates , while pairs arising from the semileptonic decay of @xmath3 are also important @xcite . for lepton pair invariant masses around the light vector meson masses ( @xmath4 , @xmath5 , and @xmath6 ) , two - body @xcite and three - body @xcite reactions dominate . the latter processes also extend their influence to the region @xmath7 , where the quark - gluon plasma may become visible for sufficiently high initial temperatures @xcite . in this paper , we concentrate on the regions of low invariant masses , the so - called `` soft '' limit @xmath8 mev . there , production via bremsstrahlung is expected to play an important role and many authors have recently concerned themselves with such reactions @xcite@xcite . since particle production plays an important role in ultrarelativistic heavy ion collisions , we consider only events that arise from microscopic meson - meson dynamics . in high energy heavy ion experiments , the meson to baryon ratio is such that the absence of baryons in our treatment is not expected to constitute a major hurdle to phenomenology . more specifically , we restrict our studies to those of pion - pion bremsstrahlung . the production of photons or dileptons necessarily involves 3- or 4-particle final states . except for special cases such as cross sections at fixed @xmath9 @xcite , such final states have universally been handled under some approximation . it is our goal in this paper to go beyond these by presenting an essentially exact formulation for the @xmath10 cross section ( as a function of dilepton invariant mass ) and , in the process , to test how far these approximations may be trusted . in order to achieve these goals , we formulate the pion - pion interaction in terms of a relativistic one boson exchange model ( obe ) . note that we do not attempt to compare directly to experimental data because the main point at issue is a comparison of theoretical approaches to bremsstrahlung . to do justice to small-@xmath1 data would require spacetime integration over the fireball region as well as inclusion of dalitz decays and the landau - pomeranchuk effect , a substantial task which will be postponed to future work . our paper is organized as follows : in the next section , we discuss the set of approximations known under the collective banner of the soft photon aproximation ( spa ) . next , we derive exact formulae for the bremsstrahlung generation of lepton pairs . section [ sec : threesp ] is devoted to technical issues having to do with phase space ; sections [ sec : curint ] and [ sec : xsexact ] then derive cross sections based respectively on various spa variants and the exact formulation . our obe model is presented in section [ sec : fullobe ] ; finally , section [ sec : results ] contains our numerical results in the form of cross sections @xmath11 and rates @xmath12 and a discussion on the issues raised . the soft photon approximation ( spa ) has been widely used in calculating bremsstrahlung dilepton spectra . it is based on the early realization @xcite that the cross section for production of low - energy real photons is dominated by the corresponding hadronic amplitude . in practice , this means neglecting the photon momentum in the hadronic matrix element as well as all photon emission from vertices and internal lines ( see section [ sec : spa ] ) . in order for the spa to be a valid approximation , two conditions @xcite must be met : 1 . the photon energy must be much smaller than the energy of any one of the hadrons participating in the scattering , @xmath13 2 . the process of radiating the photon must be separable from the hadronic interaction process , i.e. the hadronic and electromagnetic time and distance scales must be sufficiently different to permit separate treatment . the range of the hadron - hadron interaction is given roughly by @xmath14 , the inverse mass of the exchange boson , while the distance the hadron may propagate with off - shell energy @xmath15 before emitting the photon is about @xmath16 . the spa is valid only when @xmath17 , or @xmath18 implicit in these equations is , of course , a specific lorentz frame with respect to which the energies are measured . for the special case where two hadrons of equal mass @xmath19 collide , these conditions can be re - written in their cms frame as @xmath20 in a simple bremsstrahlung experiment , these limits are easily satisfied by selecting only photons or dileptons of low energy in the laboratory frame . an early dilepton paper by rckl , for example @xcite , calculated spectra for @xmath21 and @xmath22 gev while restricting transverse momentum to below 500 mev . in the complex multiparticle systems formed in the course of nucleus - nucleus collisions , however , the situation is more complicated : there are many binary collisions , and their respective cms frames do not generally coincide either with one another or with the overall nucleus - nucleus cms frame . a photon that is soft in a particular hadron - hadron cms therefore does not have to be soft in the laboratory and vice versa . ( this fact has to be taken into account when addressing issues related to the spectrum of real photons @xcite . ) in a situation where there are many cms frames , it therefore becomes necessary to look at relativistically invariant quantities . for lepton pairs , the extra degree of freedom provided by a non - vanishing invariant mass is a suitable variable of choice . looking at invariant masses means , however , that @xmath23 is no longer fixed but must vary over its full kinematic range , which for our example of colliding equal - mass hadrons is given by ( see section [ sec : curint ] ) @xmath24 once the upper kinematic limit for @xmath23 approaches or exceeds the bounds set by either eq . ( [ iad ] ) or ( [ iae ] ) , the soft photon approximation clearly fails . in figure 1 , we show the three functions ( [ iad ] ) and ( [ iae ] ) and ( [ iah ] ) for the case @xmath25 mev , @xmath26 mev and dilepton invariant masses @xmath27 and @xmath28 mev . it is immediately clear that the assumptions underlying the spa are not fulfilled even for small @xmath1 : the kinematic range accessible to @xmath23 is never much smaller than the limits set by the spa . the situation becomes even worse for larger @xmath1 . closely related to the above is a common misconception that the spa must necessarily be valid for small invariant masses . the confusion stems from the fact that , because @xmath29 a small value for @xmath23 necessarily implies small @xmath1 . however , the reverse is not true : both @xmath23 and @xmath9 can be large even while adding up to small @xmath1 . an extreme counterexample is of course a hard real photon with zero invariant mass but large energy . one can not , therefore , a priori expect spa calculations to coincide with full hadronic calculations even when @xmath1 is small . section [ sec : results ] provides examples of such deviations ; see especially the ratios between approximations and the exact calculation in figures 12 and 13 which do not approach unity for small @xmath1 . this point has been emphasized before @xcite . we are therefore led to conclude that the spa does not generally apply to rates or cross sections measured in terms of the dilepton invariant mass . correspondingly , doubt must be cast on the validity of the spa in calculations pertaining to such measurements . that said , one may still ask _ by how much _ spa - based calculations differ from the more complicated true situation : is it a matter of a few percent , or orders of magnitude ? full calculations such as presented in section [ sec : fullobe ] below are cumbersome and time - consuming , and it is therefore of value to understand quantitatively to what percentage the spa approximation can be trusted . it is this _ quantitative _ question that we attempt to cast some light on by making a model calculation capable both of `` true '' answers and an spa approximation . clearly , a single model calculation such as presented below can not answer this quantitative question exhaustively . nevertheless , its results provide at least an indicator of the reliability of the spa : if , within our simple obe , dileptons can or can not reasonably be described within the spa , it is possible or even likely that the same or similar conclusions would emerge from other models . rather than being just a single step , the soft photon approximation as implemented in a practical context can consist of a number of stages . starting with the most fundamental , these are : 1 . the virtual photon producing the lepton pair is emitted only by external legs of the hadronic reaction ; emission from internal parts of the hadronic reaction is neglected because the radiation from internal lines forms a sub - leading contribution @xcite . 2 . the dependence of the hadronic matrix element @xmath30 on the photon momentum @xmath31 is neglected . 3 . the photon momentum @xmath31 is neglected in the phase space delta functions . in addition , we shall be considering in this paper the effect of 1 . approximating the virtual photon current by the real photon current , ( @xmath32 ; see eqs . ( [ brk ] ) and ( [ brl ] ) ) . \(a ) and ( b ) have formed the basis for most bremsstrahlung dilepton calculations as they are hard to avoid . the restriction of phase space ( c ) , on the other hand , can be circumvented fairly easily by inserting later a phase space correction factor ( section [ sec : ruckl ] ) or by avoiding ( c ) altogether using eqs . ( [ brv ] ) , ( [ fcm ] ) and sections [ sec : threesp ] and [ sec : xsexact ] . the change from virtual to real photon currents in step ( d ) is not directly dependent on ( c ) and can be implemented either jointly or separately . we shall be investigating both alternatives and determining their impact on the overall dilepton rate . in section [ sec : fullcross ] and section [ sec : fullobe ] and beyond , we shall refrain from using even the basic assumptions ( a ) and ( b ) and calculate exact cross sections and rates within the same lagrangian . we wish to find the cross section for a dilepton pair @xmath33 emitted in the semi - elastic scattering of two pions @xmath34 and @xmath35 into @xmath36 and @xmath37 , @xmath38 the lepton pair forms via a virtual photon @xmath39 with four - momentum @xmath40 and mass @xmath41 emitted from either the central blob or external legs of the hadronic reaction @xmath42 the leading contribution will come from photon emission from external pion legs . first , we follow lichard @xcite in a leading - terms derivation but place more emphasis on the role of phase space ; the full formalism supplanting the leading - terms derivation is presented in section [ sec : fullcross ] . four diagrams contribute to the leading - term cross section , one of which is shown in figure 2 . in the notation of figure 2 , the total matrix element can be written as products of purely hadronic reactions @xmath30 with currents @xmath43 and a leptonic part @xmath44 @xcite , @xmath45\ ; l_\mu \;,\end{aligned}\ ] ] where @xmath46 and , in slightly condensed notation , the terms of the pionic current are @xmath47 with @xmath48 the charge of pion @xmath49 in units of the proton charge . note that @xmath43 contains the propagator for the off - shell pion . making approximation ( b ) , one now neglects the dependence on @xmath31 of the four hadronic matrix elements ; for example @xmath50 permitting the usual factorization of @xmath51 into a purely hadronic matrix element and a photon - plus - lepton part : @xmath52 where now the virtual - photon current @xmath53 contains the sum over the four diagrams . since approximations ( a ) and ( b ) are valid only for photons of small @xmath31 and by implication for small @xmath1 , the extra terms in numerator and denominator are usually neglected ( in what we called approximation ( d ) ) to form the real - photon current @xmath54 squaring the total matrix element and summing over lepton polarizations @xmath55 yields , using either current @xcite , @xmath56 \,,\end{aligned}\ ] ] where @xmath57 and the leptonic tensor is @xmath58 at this point , @xmath31 is still defined only as the sum of lepton momenta @xmath59 . with @xmath60^{1/2}$ ] the incoming flux , the unpolarized cross section is an integral over phase space of _ four _ outgoing particles , @xmath61 where we write @xmath62 for short . this can be reduced to a _ three_-particle phase space integral by changing variables and integrating over redundant degrees of freedom in the dilepton cms , @xmath63 where the positron solid angle @xmath64 must be kept due to the @xmath65 term , and @xmath66 is the lepton mass . the dependence of @xmath65 on @xmath64 is integrated out , @xmath67 d\omega_+ = { 8\pi\over 3}\left(1 + { 2\mu^2\over m^2}\right ) ( -j^2 ) \,,\ ] ] so that , using eqs . [ brm][brs ] , we can express the total cross section in terms of a leptonic factor @xmath68 and a `` hadronic plus virtual photon '' cross section @xmath69 , @xmath70 with @xmath71 and where @xmath72 is a function of _ three_-particle phase space @xmath73 with the understanding that its derivation is valid under approximations ( a ) and ( b ) only , eq . ( [ brt ] ) is exact and makes no assumptions with respect to phase space . in particular , we note that @xmath31 is still contained in the delta function constraining the phase space in eq . ( [ brw ] ) . next , we shall show that eq . ( [ brt ] ) is valid even when approximations ( a ) and ( b ) are not made but that @xmath69 takes on a form different from eq . ( [ brv ] ) . when calculating dilepton cross sections for all possible diagrams , including internal emission of photons , it is no longer possible to factorize the matrix element into a hadronic part and a current as in eq . ( [ bre ] ) . instead , we consider , for a given pion - pion reaction , the full matrix element @xmath74 where @xmath19 runs over all contributing diagrams . the sum of diagrams is gauge invariant , @xmath75 but not necessarily the individual terms . unlike in the previous sections @xmath76 contains , besides the hadronic interaction , the photon vertex and where applicable the external pion propagator , both of which were previously part of the current . starting again with the cross section in four - phase space ( [ bro ] ) and transforming from @xmath77 to @xmath78 using eq . ( [ brqq ] ) , the unpolarized dilepton cross section is @xmath79 squaring and summing over lepton polarizations , we obtain in analogy to eq . ( [ brm ] ) @xmath80 where as before @xmath81 . in the dilepton cms , @xmath82 and the second term in @xmath83 integrates to @xmath84 since @xmath85 in this frame , @xmath86 and hence in the dilepton cms @xmath87 so , using gauge invariance ( [ fcc ] ) , @xmath88 hence @xmath89 and the full cross section becomes @xmath90 which is identical to eq . ( [ brt ] ) , but with the virtual photon cross section reading @xmath91 comparing to eq . ( [ brv ] ) , we have the correspondence @xmath92 @xmath93 . equations ( [ fcl ] ) and ( [ fcm ] ) , together with use of the 3-phase space variables of section [ sec : threesp ] provide a fully covariant formalism capable of exact treatment of both hadronic and electromagnetic sectors of a given model . within the physical limitations implicit in a given model , it makes no further assumptions or approximations . in particular , none of the approximations listed in section [ sec : spa ] are needed . in the context of nuclear collisions , one is interested more in overall dilepton rates than in fixed-@xmath94 cross sections . not attempting to account fully for either quantum statistical or flow effects here , we shall make the simplest assumption of a locally thermal boltzmann gas of pions at a ( local ) temperature @xmath95 . the rate of pair production of invariant mass @xmath1 per unit of four - volume then is @xcite @xmath96 where @xmath97 is the basic three - particle kinematic function @xcite , which for equal masses @xmath98 has the more familiar form @xmath99 , and @xmath100 is the spin degeneracy factor which for pions is unity . @xmath101 is the modified bessel function . to complete our preparatory work , we summarize the treatment of three - particle phase space in terms of relativistic invariants . the relations of this section will be useful first in 3-phase space integration of the currents ( [ brk ] ) and ( [ brl ] ) and later in the exact treatment of bremsstrahlung dilepton production . three - particle phase space has been studied extensively and we here merely outline the procedure ; for details see ref . @xcite . as shown in figure 3 , the kinematics for a reaction @xmath102 can be described in terms of the five invariants @xmath103 , @xmath104 , @xmath105 , @xmath106 , and @xmath107 . for our purposes , we shall be using @xmath108 and @xmath31 interchangeably as referring to the virtual photon , i.e. @xmath109 and @xmath110 . the other momenta refer , as usual , to the incoming and outgoing hadrons . inverting , one gets @xmath111 the corresponding phase space integral is @xmath112 where the weighting is given in terms of the cayley determinant @xmath113 the cayley determinant is quadratic in any of its arguments ; specifically , we use the form @xmath114 where the kinematic limits on @xmath115 are given by @xmath116^{1/2 } , \end{aligned}\ ] ] written in terms of yet another determinant and the basic four - particle kinematic function @xmath117 the kinematic limits on @xmath118 are @xmath119 the limits on @xmath120 and @xmath121 are given by the chew - low plot : for values of @xmath121 bounded by @xmath122[s + m_1 ^ 2 - ( m_2 + m)^2 ] \nonumber\\ & \pm & { 1\over 2s } \lambda^{1/2}(s , m_a^2,m_b^2)\ ; \lambda^{1/2}(s,(m_2{+}m)^2,m_1 ^ 2 ) \,,\end{aligned}\ ] ] the limits on @xmath120 are @xmath123[m_a^2 + m_1 ^ 2 - t_1 ] \nonumber\\ & & \ \ \ \ + { 1\over 2m_a^2 } \lambda^{1/2}(s , m_a^2,m_b^2)\ ; \lambda^{1/2}(t_1,m_a^2,m_1 ^ 2 ) \ , , \\ \label{fvkm } s_2 ^ - & = & ( m_2 + m)^2 \,.\end{aligned}\ ] ] although not directly relevant for pions , we include for completeness the case where the hadronic masses @xmath124 and @xmath125 are unequal . in this case , @xmath121 may become larger than @xmath126 given above , and the corresponding additional phase space is delimited by @xmath127 and for this additional domain , @xmath120 is limited by @xmath128 of eq . ( [ fvk ] ) and the corresponding lower branch , given by a minus sign before the @xmath129-factors . clearly , @xmath121 resembles the 2-phase space invariant @xmath130 , but they are distinct quantities due to the presence of the additional momentum @xmath31 . integration over pure phase space yields @xmath131 where @xmath132 and we have written @xmath133 for short . this yields identical results to the pure phase space formula ( [ icg ] ) below , as it should . we note in passing that the ratio of the kinematic domains of ( 3-space ) @xmath121 to ( 2-space ) @xmath130 is exactly the phase space reduction factor found in ( [ icj ] ) below . all the elements for our calculation are now in place . in this section , we look at three approximations of increasing sophistication . we proceed first to find explicit expressions for the pion - pion - photon cross section @xmath69 of eq . ( [ brv ] ) based on the factorization into a current and hadronic matrix element of section [ sec : external ] . these approximations are based on an angular average of @xmath134 in the @xmath135 center - of - momentum frame . using three - space invariants , we integrate the current covariantly in section [ sec : xs3sp ] . in 1976 , rckl proposed @xcite that the dilepton cross section be written in terms of the _ real photon _ cross section : @xmath136 this ansatz , now in common use , was utilized by haglin et al . @xcite and others previously @xcite to derive dilepton production rates . in our notation , the rckl formula reads @xmath137 which becomes , using the formulas of the previous section , and approximating @xmath138 and @xmath139 , @xmath140 the real - photon cross section @xmath141 can be written down immediately from the virtual - photon cross section @xmath69 of ( [ brv ] ) , @xmath142 neglecting @xmath31 in the delta function of @xmath143 , the 3-phase space integral of eq . ( [ brw ] ) factorizes into the usual 2-phase space and the @xmath9-integration @xmath144 with @xmath145 leading all in all to @xmath146 where @xmath147 is the elastic hadronic cross section . integrating @xmath134 first with respect to the photon phase space with @xmath148 leads to @xmath149 where @xmath150/2\sqrt{s}$ ] is the upper kinematic boundary . the angular average over @xmath151 was found by haglin et al . for the real - photon current ( [ brl ] ) in terms of velocities @xmath152 in the @xmath153 ) cms as @xcite @xmath154 \,,\end{aligned}\ ] ] where @xmath155 \ , , \\ d_{ij } & \equiv & \left [ ( \bbox{\beta}_i - \bbox{\beta}_j)^2 - ( \bbox{\beta}_i { \times } \bbox{\beta}_j)^2 \right]^{1/2 } \,.\end{aligned}\ ] ] to convert these angular averages into functions of 2-phase space invariants @xmath156 , it is necessary to invoke approximation ( c ) restricting phase space : while @xmath157 in the chosen frame , the presence of @xmath9 means that @xmath158 , i.e. @xmath159 and @xmath160 are not true cms velocities . ignoring therefore @xmath9 at this point , the velocities and hence the angular averages can be written @xcite in terms of @xmath94 , @xmath130 and @xmath23 , @xmath161 \,,\end{aligned}\ ] ] with @xmath162 as usual . having expediently eliminated @xmath31 at two points , it is very important in this formulation to correct for phase space , i.e. to alleviate the effects of approximation ( c ) . pure phase space integration of @xmath143 ( _ including _ @xmath31 in the delta function ) yields exactly @xmath163 where @xmath164 integration over pure two - particle phase space , on the other hand , yields @xmath165 so @xmath166 , unconstrained by @xmath31 , grossly overestimates @xmath167 , making it necessary to insert _ a posteriori _ the factor @xcite @xmath168 into the @xmath23-integration ( [ rrh ] ) . the `` rckl approximation '' for dilepton production therefore reads , in its final form , @xmath169 this is the form used in ref . @xcite . starting not from the rckl formula but rather from lichard s improved formalism @xcite of section [ sec : external ] , the derivation proceeds along the same lines as the above . one finds correspondingly for the real - photon current @xmath170 differing from the above by a constant factor @xmath171 and the extra factors entering the exact version ( [ bru ] ) of @xmath68 . the corresponding virtual - photon current yields the third approximation within the angle - averaged current family , @xmath172 where the angular average of ( [ brk ] ) is @xmath173 where @xmath174 ) reduces to eq . ( [ hgk ] ) for @xmath175 . we shall be testing the effect of using the virtual over real - photon currents within the present approximation in section [ sec : results ] below . we note in passing that eq . ( [ hgh ] ) would be preferable to eq . ( [ hgk ] ) for very small hadron masses @xmath19 , since then the latter is logarithmically sensitive to @xmath176 while the former is regulated by the dilepton invariant mass @xmath1 . three approximations were needed for the above angle - averaged current cross sections : the rckl formula , converting cms velocities @xmath152 to invariants @xmath94 and @xmath130 leading to eqs . ( [ hge ] ) and ( [ hgf ] ) , and correcting for phase space using eq . ( [ icj ] ) . at the cost of increased computing time , all three can be avoided by utilizing the formalism of three - particle phase space invariants . within 3-phase space , and writing the flux as @xmath177 , the bremsstrahlung cross section of eq.([brv ] ) becomes ) be possible to restore @xmath31 to @xmath30 and then to include it with @xmath178 in the inner integral , but because of missing contact terms this procedure is not gauge invariant and shall not be pursued . ] @xmath179 \,.\ ] ] this differs from eq . ( [ icklo ] ) in the way the current is treated : by an angular average plus @xmath23-integration for the earlier , by full 3-phase space invariants in the present case . note that @xmath30 is here a function of @xmath121 rather than @xmath130 . in all the above approximations , neglecting the dependence on @xmath31 of the squared hadronic matrix element @xmath180 is unavoidable because this had been a precondition to factorizing into electromagnetic and hadronic parts in eq . ( [ bre ] ) . there have been attempts go beyond this on - shell matrix element approximation by expanding @xmath30 in @xmath31 ( see refs . @xcite ) , but this procedure involves successively more complicated derivatives without guaranteeing convergence . using 3-phase space invariants and the exact for the emission of real photons and to order @xmath181 for dilepton emission . ] expressions ( [ fcm ] ) for the pion - pion - photon cross section and ( [ fcl ] ) for the dilepton cross section , such expansions in @xmath31 become superfluous . to calculate explicit dilepton cross sections , we merely insert from our favorite model for pion - pion scattering the complete set of matrix elements , and write the elements of the squared sum of ( [ fcm ] ) in terms of these invariants : @xmath182 this casting in terms of invariants means that @xmath31 and all consequent off - shell effects in @xmath183 are fully taken care of . can be implemented within 2-phase space when only the _ differential _ cross section @xmath184 need be calculated ; see e.g. ref.@xcite ; for @xmath185 , this becomes impossible . ] putting it all together , we obtain @xmath186 \,.\ ] ] while the cross sections may be exact , microscopic models for the pion - pion interaction are far from perfect . in the next section , we make use of a simple model to illustrate the present cross section and the approximations enumerated before . the main point of this section , however , is to stress that , apart from model - induced ailments , the exact treatment of any model for the reaction @xmath187 is possible . to compare dilepton cross sections under the exact formula and various approximations introduced so far , we must make a model : little if any experimental data for the differential pion - pion - photon cross section is available . to this purpose , we return to a simple lagrangian with @xmath188 , @xmath4 and @xmath189 exchange used previously @xcite which , when @xmath130-integrated , successfully reproduced experimental data for elastic @xmath190 scattering @xcite : @xmath191 here , we include the full momentum dependence of the @xmath192 propagator which now reads @xcite @xmath193 the @xmath194 field also has a @xmath195-dependent term in the unitary gauge propagator @xcite , @xmath196 \,,\ ] ] but this term contributes to matrix elements only in the form @xmath197 etc . and is thus zero for equal pion masses . for @xmath130- and @xmath198-channel exchange of an @xmath199-meson ( @xmath200 ) with momentum @xmath195 by the pions , we implement monopole strong form factors @xcite @xmath201 in figure 4 , we show the results of fitting this model to the elastic scattering data as before @xcite but with the different propagator ( [ rtd ] ) . again , the fit is reasonably good , but the strong dependence on momentum - dependent terms in the @xmath192 propagator shows up in the region above the @xmath192 peak . best - fit values for the free parameters now are : @xmath202 mev , @xmath203 mev , @xmath204 mev ; @xmath205 , @xmath206 , @xmath207 . throughout , we take @xmath208 mev . for the current - based approximations of section [ sec : curint ] , the above lagrangian is all we need ; for the complete cross section of section [ sec : xsexact ] , it must be augmented by internal photon emission pieces . this can be achieved by minimal substitution , thereby guaranteeing gauge invariance ( for a similar approach , see the `` quantum hadrodynamics '' lagrangians of ref . @xcite ) . to the hadronic lagrangian ( [ rtb ] ) we therefore add @xmath209 details regarding these pieces and corresponding vertex factors can be found in the appendix . the @xmath210 interaction results in a modified strong form factor for @xmath130- or @xmath198-channel charged @xmath4 exchange . if @xmath211 and @xmath212 are the usual form factors at the @xmath213 and @xmath214 vertices respectively , with @xmath215 , electromagnetic gauge invariance requires that the form factor for the emission of a photon by the @xmath4 be @xmath216 where @xmath217 and @xmath218 are the respective @xmath4 propagators . for the monopole form ( [ rth ] ) of @xmath219 this becomes @xcite @xmath220 note that we do not include strict vector meson dominance coupling of the photons to vector mesons . this is equivalent to the introduction of electromagnetic form factors . for the specific application we are considering , small @xmath1 s , those would be @xmath221 1 . for previous obe models , see e.g. refs.@xcite and references therein . to summarize : the differential cross section for dilepton production as a function of invariant mass , @xmath222 , can be found in a number of approximations and an exact way ( within the limitations of the hadronic reaction model ) . the approximations are : the rckl approximation ( [ ick ] ) , two approximations , eqs . ( [ icklz ] ) and ( [ icklo ] ) , based on a more careful derivation by lichard , using the real and virtual photon currents respectively , and an approximation integrating the current covariantly in 3-phase space , eq . ( [ bvp ] ) . the `` exact '' formula is given by eq . ( [ xxc ] ) . of these , the rckl and lichard cross sections can be found easily and quickly within 2-phase space . the 3-phase space current approximation is somewhat more difficult but still uses 2-phase space hadronic matrix elements . implementing the exact formulation ( [ xxc ] ) involves much more effort : if there are @xmath223 diagrams contributing to a given reaction , the sum in ( [ xxc ] ) contains @xmath224 terms . since an individual term @xmath225 in ( [ xxb ] ) is itself a dirac sum , only partial factorization of the overall sum is possible . further complications arise from the presence of imaginary pieces in the matrix elements . to quantify the differences between the approximations and the exact formulation , we have studied all five distinct pion - pion reactions . writing @xmath226 as shorthand for the reaction @xmath227 and so on , we have calculated within our obe model cross sections for @xmath226 , @xmath228 , @xmath229 , @xmath230 and @xmath231 . a total of 36 diagrams contribute at tree level to the first two reactions ( 24 for @xmath39 emitted by external pion lines , 12 by vertices and the exchange meson ) ; the remaining reactions are made up of 16 diagrams each ( 8 external , 8 internal ) . numerical results were checked by performing the following consistency checks : setting @xmath232 , @xmath233 , @xmath234 , @xmath235 and @xmath236 in the _ hadronic _ part of the exact matrix elements ( while keeping these variables where they enter the `` current '' part ) and using only the external - emission diagrams reproduces exactly the results of the 3-phase space integrated current ( [ bvp ] ) . gauge invariance in the @xmath188 , @xmath4 and @xmath192 fields provided another sensitive test . figures 59 show the cross sections for these five reactions as a function of @xmath237 . in every case , the left and right panels show @xmath238 for @xmath27 mev and 300 mev respectively . final - state symmetrization factors were included where appropriate . initial - state symmetrization was also included for @xmath228 and @xmath230 in order to facilitate use within a thermal pion gas environment . when the initial - state pions are considered distinguishable , these two cross sections should be multiplied by a factor 2 for these reactions . because they are identical in structure to their charge - conjugate versions , cross sections for the reactions @xmath231 and @xmath228 were doubled this , too , should be corrected for when desired . to prevent overcrowding , only the rckl plus the virtual-@xmath39 current approximations ( [ icklo ] ) and ( [ bvp ] ) are shown , together with the exact calculations . all these were computed using the same obe model and parameter values specified in section [ sec : fullobe ] above . in the complicated structures and deviations in figs . 59 , the following points are of interest : 1 . the @xmath4 peak in the reaction @xmath226 is overestimated by a factor 2.5 and 1.5 by the rckl and lichard approximations respectively for @xmath27 mev ; for @xmath239 mev , the overestimation is much greater ( figure 5 ) . only the 3-space current approximation does an adequate job here . similarly , the @xmath192 peak of the exact result lies well below the corresponding approximations . overestimation ( 1.75.0 ) of the @xmath4 peak also occurs for approximations in @xmath231 . 2 . in the reaction @xmath231 , the 3-space current approximation , on the other hand , lies well below the exact result ( figure 6 ) . overestimation factors for @xmath238 are of order 0.54 for the reactions @xmath230 and @xmath229 , depending on approximation and cms energy . here , too , the 3-space current approximation generally does better than the others but again tends to underestimate in some parts ( figs . 7 , 8) . the largest discrepancy between approximations and the exact result occur for the reaction @xmath228 ( figure 9 ) : for the rckl approximation , factors 3 ( for 10 mev ) to 30 ( for 300 mev ) arise , while the lichard approximation yields corresponding overestimation factors of 1.9 and 1420 . since , however , this reaction contributes only little when contributions from all reactions are added up , this effect is not of much interest in a heavy - ion context . 5 . there is generally not much difference between using the real - photon ( not shown ) or virtual - photon currents in the lichard approximations ; this is in agreement with the conclusions of ref . @xcite . for @xmath239 mev , some differences can be discerned , but both deviate in general more from the exact result than from each other . figures 10 and 11 show the dilepton production rates per unit spacetime summed over all seven reactions , calculated using the boltzmann formula ( [ rtk ] ) , for temperatures @xmath240 and 200 mev respectively . these temperatures probably lie below and above an expected transition to a new quark - gluon phase and so represent extrema in their behavior . corresponding ratios of approximation over exact rates are shown in figures 12 and 13 . the rckl approximation is the worst , as expected ; overestimation is in the range 24 for @xmath240 mev , and 28 for @xmath241 mev . the lichard approximations overestimate by factors 1.44 , depending on temperature and @xmath94 . here , the _ real-@xmath242 _ 3-phase space current , the lower of the two dotted lines , does surprisingly well , deviating from the exact result by less than 20% throughout . comparing this fact , however , with the substantial deviations shown in the @xmath94-differential cross sections of figs . 59 , one must conclude that much of the `` agreement '' seen on the level of rates is due to averaging over differences seen in the @xmath94-dependent cross sections and that such `` agreement '' therefore does not validate the approximations . we also note the fact that none of the approximations approaches the exact result for small values of @xmath1 : even for the smallest value shown ( @xmath27 mev ) , the discrepancy is still above 40% for the lichard approximations and larger than a factor 2 for the rckl approximation . figure 14 shows the fractional contributions to the total @xmath243 mev rate for the rckl , virtual-@xmath242 current lichard , and exact cross sections . as discussed , the @xmath228 contribution virtually falls away in the exact calculation ; by contrast the reaction @xmath231 plus charge conjugate becomes more important , in second place behind the dominant reaction @xmath226 . finally , we ask where the discrepancies between approximations and exact cross section originate : is this due to the inclusion of diagrams with photon emission from vertices and propagators ? a partial answer is provided by figure 15 , where the cross sections for reactions @xmath244 and @xmath229 are plotted . the upper lines correspond , as before , to the rckl and two lichard approximations , while the solid line again represents the exact result . the lowest dash - dotted line , on the other hand , represents the exact result but excluding all internal diagrams and their cross terms with external ones . the difference between this lower line and the exact result ( solid line ) therefore represents the contribution of the internal diagrams ; while the difference between the lower dash - dotted line and the upper lines ( approximations ) represents the change in cross section due to inclusion / exclusion of @xmath31 in the _ external_-emission diagrams . we see that the contribution of internal emission is not all that large , albeit nonnegligible . by far the most important effect on @xmath238 is the inclusion of the full dependence of @xmath245 on the photon momentum @xmath31 . in fact , the effect is so large that for the reaction @xmath228 the cross section would become negative if internal diagrams were not included ! in other words , among the list of approximations listed in section [ sec : spa ] , approximation ( b ) is the most far - reaching . the importance of including @xmath31 fully in the matrix element can be understood qualitatively : as illustrated in figs . 59 for one reaction , the elastic pion - pion cross sections vary considerably within the region of interest . adding or subtracting a sizeable amount of energy - momentum in the form of @xmath31 therefore naturally leads to a substantial change in cross section also . this change is , however , neglected in all the approximations listed : there are no resonances in the cross section which could provide some amount of cancellation by pion - pion reactions being `` shifted into '' and `` shifted out of '' a resonance peak . ] . we believe that this ( unavoidable ) neglect is at the heart of the considerable differences between approximations and exact result seen in figs . 59 . in summary , the various approximations for bremsstrahlung from pion - pion collisions can not in general be believed beyond at best a factor 2 or more , depending on invariant mass and temperature . we have shown in sections [ sec : fullcross ] and [ sec : xsexact ] how 3-phase space can be fully taken into account . the most important effect , it turns out , is not the inclusion of photon emission from vertices or propagators but the strong dependence of the cross section on the inclusion of the photon momentum @xmath31 into the _ hadronic _ collision part . since this calculation neglected baryonic degrees of freedom and resonances , and because of the technical difficulties of a further integration over spacetime , our results are not applied directly to experimental data ; this is a matter for the future . all we can say at this stage is that it is likely that the approximations overestimate the dilepton yield throughout the mass region of interest , and that the degree of overestimation rises both with increasing invariant mass @xmath1 and temperature @xmath95 . since the bremsstrahlung channel competes with various dalitz decays , and since various experiments take much care in applying cuts to eliminate dalitz pairs as far as possible @xcite a quantitative estimate would necessarily have to take such experimental cuts into account also . in addition , attention should be paid to the issue of the landau - pomeranchuk effect , which is expected to suppress dilepton bremsstrahlung rates @xcite . our results , taken in conjunction with these calculations , therefore appear to indicate even larger suppression of bremsstrahlung dileptons than had been thought previously . we thank p. lichard for enlightening discussions and his careful derivation of dilepton formulas . thanks to a. drees , j. zhang , j. c. pan , p. lipa and b. buschbeck for useful discussions . this work was supported in part by the natural sciences and engineering research council of canada , by the qubec fcar fund , by a nato collaborative research grant , by the national science foundation under grant number 9403666 and by the austrian fonds zur frderung der wissenschaftlichen forschung ( fwf ) . besides the hadronic interaction lagrangian ( [ rtb ] ) with its vertex factors @xmath246 we need the electromagnetic interaction lagrangian linear in @xmath247 , @xmath248 with the pion current @xmath249 the @xmath194 field tensor , @xmath250 and @xmath251 the usual electromagnetic tensor , the individual contributions are @xmath252_3 + { 1\over 2 } e f_{\mu\nu } [ \bbox{\rho}^\mu \bbox{{\times } \rho}^\nu]_3 \\ & = & ie a_\mu \left [ \rho_{- \nu } ( \partial^\mu\rho_+^\nu - \partial^\nu\rho_+^\mu ) - \rho_{+ \nu } ( \partial^\mu\rho_-^\nu - \partial^\nu\rho_-^\mu ) \right ] + ie \partial_\mu a_\nu \left [ \rho_-^\mu \rho_+^\nu - \rho_-^\nu \rho_+^\mu \right]\ , , \nonumber\\ \label{ggi } { \cal l}_{\pi\pi\sigma\gamma } & = & - 2 e g_\sigma a_\mu j_\pi^\mu \ , , \\ \label{ggj } { \cal l}_{\pi\pi\rho\gamma } & = & - e g_\rho a_\mu \left [ \bbox{\pi { \times}}(\bbox{\pi { \times } \rho}^\mu ) \right]_3 \\ & = & e g_\rho a_\mu \left [ 2 \pi^+ \pi^- \rho_0^\mu - \pi^- \pi^0 \rho_+^\mu - \pi^+ \pi^0 \rho_-^\mu \right]\ , , \nonumber\\ \label{ggk } { \cal l}_{\pi\pi f\gamma } & = & - e g_f f_{\mu\nu } \left [ a^\mu j_\pi^\nu + a^\nu j_\pi^\mu \right]\,.\end{aligned}\ ] ] the corresponding vertex factors are @xmath253 & \mbox{\hspace*{5mm}for } \rho_\pm^\alpha(p ) \to \rho_\pm^\beta(p^\prime ) a^\mu(q ) \\ % \label{vtppsg } \gamma_{\pi\pi\sigma\gamma } & = & \pm 2ie g_\sigma ( p+p^\prime)^\mu & \mbox{\hspace*{5mm}for } \pi^{\pm}(p ) \to \pi^{\pm}(p^\prime ) \sigma a^\mu \\ % \label{vtpprg } \gamma_{\pi\pi\rho\gamma } & = & \left\ { \begin{array}{l } -2ie g_\rho g^{\mu\nu } \\ + ie g_\rho g^{\mu\nu } \\ + ie g_\rho g^{\mu\nu } \\ \end{array } \right . & \begin{array}{l } \mbox{\hspace*{4mm}for } \pi^{\pm}(p ) \to \pi^{\pm}(p^\prime ) \rho_0^\nu a^\mu \\ % \mbox{\hspace*{4mm}for } \pi^{\pm}(p ) \to \pi^0(p^\prime ) \rho_{\pm}^\nu a^\mu \\ % \mbox{\hspace*{4mm}for } \pi^{0}(p ) \rho_{\pm}^\nu \to \pi^{\pm}(p^\prime ) a^\mu \\ \end{array } \\ % \label{vtppfg } \gamma_{\pi\pi f\gamma } & = & \pm ie g_f \left [ g^{\alpha\mu}(p+p^\prime)^\beta + g^{\beta \mu}(p+p^\prime)^\alpha \right ] & \mbox{\hspace*{5mm}for } \pi^{\pm}(p ) \to \pi^{\pm}(p^\prime ) f^{\alpha\beta } a^\mu \ , . \end{array}\end{aligned}\ ] ] g. roche , talk presented at the nato advanced study institute , bodrum / turkey , september 1993 , and references therein ; in : proceedings of the nato asi on hot and dense nuclear matter , edited by w. greiner , h. stcker and a. gallmann , nato asi series b335 . serot and j.d . walecka , _ the relativistic nuclear many - body problem _ , advances in nuclear physics , vol . 16 , edited by j. w. negele and e. vogt , plenum , new york ( 1986 ) ; b.d . serot and j.d . walecka , rep . phys . * 55 * , 1855 ( 1992 ) . figure 1 : breakdown of the soft photon approximation : the spa is valid only when @xmath23 is much smaller than the two solid lines shown . the region between the dashed lines is the domain of integration for @xmath23 when calculating cross sections as a function of dilepton invariant mass . figure 2 : one of the four contributing diagrams following from spa factorization of the matrix element into leptonic tensor @xmath83 , photon emission current @xmath255 , and elastic hadronic matrix element @xmath30 . figure 5 : cross section for the reaction @xmath257 as a function of @xmath237 for fixed dielectron invariant masses @xmath27 and 300 mev . solid line : exact obe calculation eq . ( [ xxc ] ) . dash - dotted line : rckl approximation ( [ ick ] ) this is the commonly - used version of the spa . other approximations are : dashed line : angular - averaged virtual @xmath242 current ( [ icklo ] ) . dotted line : 3-phase space current ( [ bvp ] ) . figure 10 : total bremsstrahlung yield of @xmath266 for all seven pion - pion reactions , for a boltzmann gas with temperature @xmath240 mev . solid line : exact obe calculation , dash - dotted line : rckl approximation , dashed line : angular - averaged virtual-@xmath39 current , dotted line : 3-phase space current , virtual-@xmath39 current ( upper ) real-@xmath242 current ( lower ) . figure 12 : ratios of spa approximation calculations divided by exact obe rate for @xmath240 mev . dash - dotted line : rckl / exact , dashed line : ( angular - averaged virtual @xmath242 current)/exact . upper dotted line : ( 3-phase space virtual @xmath242 current)/exact . lower dotted line : ( 3-phase space real @xmath242 current)/exact . figure 14 : fractional contribution to the total dielectron rate at @xmath243 mev as a function of dilepton invariant mass @xmath1 for ( a ) the rckl , ( b ) virtual - photon lichard approximations and ( c ) the exact calculation . solid line ( upper ) : @xmath226 , dashed line : @xmath231 plus c.c . , solid line ( lower ) : @xmath230 , dash - dotted line : @xmath229 , dotted line : @xmath228 plus c.c . figure 15 : contribution of external - emission vs. external - plus - internal diagrams for the reactions @xmath229 ( left ) and @xmath228 ( right ) , both for @xmath239 mev . lines are as in figs . 7 and 9 . the new dash - dotted line below the ( solid line ) exact calculation represents contributions arising solely from emission of @xmath39 by an external pion line , but taking @xmath31 into account in @xmath245 , in contrast to the approximations ( upper lines ) which neglected @xmath31 .

manipulating quantum systems allows one to accomplish tasks far beyond the reach of classical devices . quantum information is paradigmatic in this sense : quantum computers will substantially outperform classical machines for several problems @xcite . though significant progress has been made recently , severe difficulties still remain , amongst which decoherence is certainly the most important . large systems consisting of many qubits must be prepared in fragile quantum states , which are rapidly destroyed by their unavoidable coupling to the environment . measurement - based feedback and coherent feedback are possible routes towards the preparation , protection and stabilization of such states . for coherent feedback strategy , the controller is also a quantum system coupled to the original one ( see @xcite and the references therein ) . this paper is devoted to measurement - based feedback where the controller and the control input are classical objects @xcite . the results presented here are directly inspired by a recent experiment @xcite demonstrating that such a quantum feedback scheme achieves the on - demand preparation and stabilization of non - classical states of a microwave field . following @xcite and relying on continuous - time lyapunov techniques exploited in @xcite , an initial measurement - based feedback was proposed in @xcite . this feedback scheme stabilizes photon - number states ( fock states ) of a microwave field ( see e.g. @xcite for a physical description of such cavity quantum electrodynamics ( cqed ) systems ) . the controller consists of a quantum filter that estimates the state of the field from discrete - time measurements performed by probe atoms , and secondly a stabilizing state - feedback that relies on lyapunov techniques . the discrete - time behavior is crucial for a possible real - time implementation of such controllers . closed - loop simulations reported in @xcite have been confirmed by the stability analysis performed in @xcite . in the experimental implementation @xcite , the state - feedback has been improved by considering a strict lyapunov function : this ensures better convergence properties by avoiding the passage by high photon numbers during the transient regime . the goal of this paper is to present , for a class of discrete - time quantum systems , the mathematical methods underlying such improved lyapunov design . our main result is given in theorem [ thm : obscontrollermeas ] where closed - loop convergence is proved in presence of delays and measurement imperfections . the state - feedback scheme may be applied to generic discrete - time finite - dimensional quantum systems using _ controlled _ measurements in order to deterministically prepare and stabilize the system at some pre - specified target state . the dynamics of these systems may be expressed in terms of a ( classical ) nonlinear controlled markov chain , whose state space consists of the set of density matrices on some hilbert space . the jumps in this markov chain are induced by quantum measurements and the associated jump probabilities are state - dependent . these systems are subject to a discrete - time sequence of positive operator valued measurements ( povms @xcite ) and we use these povms to stabilize the system at the target state . by controlled measurements , we mean that at each time - step the chosen povm is not fixed but is a function of some classical control signal @xmath0 , similar to @xcite . however , we assume that when the control @xmath0 is zero , the chosen povm performs a quantum non - demolition ( qnd ) measurement @xcite for some orthonormal basis that includes the target state . the feedback - law is based on a lyapunov function that is a linear combination of a set of martingales corresponding to the open - loop qnd measurements . this lyapunov function determines a `` distance '' between the target state and the current state . the parameters of this lyapunov function are given by inverting metzler matrices characterizing the impact of the control input on the kraus operators defining the markov processes and povms . the ( graph theoretic ) properties of the metzler matrices are used to construct families of open - loop supermartingales that become strict supermartingales in closed - loop . this fact provides directly the convergence to the target state without using the invariance principle . a common problem that occurs in quantum feedback control is that of delays between the measurement process and the control process @xcite . in this paper we demonstrate , using a predictive quantum filter , that the proposed scheme works even in the presence of delays . convergence analysis is done for perfect and imperfect measurements . for imperfect measurements , the dynamics of the system are governed by a nonlinear markov chain given in @xcite . in both the perfect and imperfect measurement situations we prove a robustness property of the feedback algorithm : the convergence of the closed - loop system is ensured even when the feedback law is based on the state of a quantum filter that is not initialized correctly . this robustness property , is similar in spirit to the separation principle proven in @xcite . we use the fact that the state space is a convex set and the target state , being a pure state , is an extreme point of this convex state space and therefore can not be expressed as a convex combination of any other states . one then uses the linearity of the conditional expectation to prove the robustness property . our result is only valid for target quantum states that are pure states . the paper is organized as follows . in section [ sec : openloop ] , we describe the finite dimensional markov model together with the main modeling assumptions in the case of perfect measurements and study the open - loop behavior ( theorem [ thm : first ] ) which can be seen as a non - deterministic protocol for preparing a finite number of isolated and orthogonal quantum states . in section [ sec : closedloop ] , we present the main ideas underlying the construction of these control - lyapunov functions @xmath1 : a metzler matrix attached to the second derivative of the measurement operators and a technical lemma assuming this metzler matrix is irreducible . finally , theorem [ thm : main ] describes the stabilizing state feedback derived from @xmath1 . the same analysis is done for the case of imperfect measurements in section [ sec : imperfect ] . for the state estimations used in the feedback scheme we propose a brief discussion on the quantum filters and prove a rather general robustness property for perfect measurements in section [ sec : closedloop ] with theorem [ thm : obscontroller ] and for imperfect ones in section [ sec : imperfect ] with theorem [ thm : obscontrollermeas ] . section [ sec : photonbox ] is devoted to the experimental implementation that has been done at laboratoire kastler - brossel of ecole normale suprieure de paris . closed - loop simulations and experimental data complementary to those reported in @xcite are presented . we consider a finite dimensional quantum system ( the underlying hilbert space @xmath2 is of dimension @xmath3 ) being measured through a generalized measurement procedure at discrete - time intervals . the dynamics are described by a nonlinear controlled markov chain . here , we suppose perfect measurements and no decoherence . the system state is described by a density operator @xmath4 belonging to @xmath5 , the set of non - negative hermitian matrices of trace one : @xmath6 to each measurement outcome @xmath7 , @xmath8 being the numbers of possible outcomes , is attached the kraus operator @xmath9 depending on @xmath10 and on a scalar control input @xmath11 . for each @xmath0 , @xmath12 satisfy the constraint @xmath13 , the identity matrix . the kraus map @xmath14 is defined by @xmath15 the random evolution of the state @xmath16 at time - step @xmath17 is modeled through the following dynamics : @xmath18 where @xmath19 is the control at step @xmath17 , subject to a delay of @xmath20 steps . this delay is usually due to delays in the measurement process that can also be seen as delays in the control process . @xmath21 is a random variable taking values @xmath10 in @xmath22 with probability @xmath23 . for each @xmath10 , @xmath24 is defined when @xmath25 . we now state some assumptions that we will be using in the remainder of this paper . [ assum : imp ] for @xmath26 , all @xmath27 are diagonal in the same orthonormal basis @xmath28 : @xmath29 with @xmath30 [ assum : impt ] for all @xmath31 in @xmath32 there exists @xmath33 such that @xmath34 [ assum : imo ] all @xmath35 are @xmath36 functions of @xmath0 . assumption [ assum : imp ] means that when @xmath26 the measurements are quantum non demolition ( qnd ) measurements over the states @xmath37 : when @xmath38 , any @xmath39 ( orthogonal projector on the basis vector @xmath40 ) is a fixed point of . since @xmath41 , we have @xmath42 for all @xmath43 according to assumption [ assum : imp ] . assumption [ assum : impt ] means that there exists a @xmath10 such that the statistics when @xmath44 for obtaining the measurement result @xmath10 are different for the fixed points @xmath45 and @xmath46 . this follows by noting that @xmath47 for @xmath43 . assumption [ assum : imo ] is a technical assumption we will use in our proofs . when the control input vanishes ( @xmath48 ) , the dynamics are simply given by @xmath49 where @xmath21 is a random variable with values in @xmath22 . the probability @xmath50 to have @xmath51 depends on @xmath52 : @xmath53 . [ thm : first ] consider a markov process @xmath52 obeying the dynamics of with an initial condition @xmath54 in @xmath5 . then with probability one , @xmath52 converges to one of the @xmath55 states @xmath56 with @xmath43 and the probability of convergence towards the state @xmath56 is given by @xmath57 this theorem is already proved in @xcite and also in @xcite in a slightly different formulation . a direct proof is based on the following lyapunov function : @xmath58 since @xmath59 is a martingale for any @xmath60 and since @xmath61 in subsection [ ssec : overview ] , we give an overview of the control method and then in subsections [ ssec : weights ] and [ ssec : stab ] , we prove the main results . finally in subsection [ ssec : qfilter ] , we prove a robustness principle that explains how we can ensure convergence even if the initial state @xmath54 is unknown . theorem [ thm : first ] shows that the open - loop dynamics are stable in the sense that in each realization @xmath52 converges ( non - deterministically ) to one of the pure states @xmath56 with probability @xmath62 . the control goal is to make this convergence deterministic toward a chosen @xmath63 playing the role of controller set point . we build on the ideas in @xcite to design a controller that is based on a strict lyapunov function for the target state . in this paper we assume arbitrary controlled kraus operators @xmath64 that can not be decomposed into qnd measurement operators @xmath65 followed by a unique controlled unitary operator @xmath66 with @xmath67 as assumed in @xcite where @xmath68 . it can be argued that the control we are proposing is non - hamiltonian control @xcite , as the control parameter @xmath0 is not necessarily a parameter in the interaction hamiltonian and could indeed be any parameter of an auxiliary system such as the measurement device . to convey the main ideas involved in the control design , we begin with the case where there are no delays ( @xmath69 ) and we assume the initial state is known . we wish to use the open - loop supermartingales to design a lyapunov function for the closed - loop system . by an open - loop supermartingale we mean any function @xmath70 such that @xmath71 for all @xmath72 . the lyapunov function underlying theorem [ thm : first ] demonstrates how we can construct such open - loop supermartingales . now , at each time - step @xmath17 , the feedback signal @xmath73 is chosen by minimizing this supermartingale @xmath74 knowing the state @xmath52 : @xmath75}{\text{argmin } } \bigg\ { { { \text{\normalsize$\mathbb e$}}\left(v(\rho_{k+1 } ) |\rho_k = \rho , u_k = u\right)}\bigg\}.\ ] ] here @xmath76 is some small positive number that needs to be determined . because @xmath77 $ ] and the control @xmath73 is chosen to minimize @xmath74 at each step , we directly have that @xmath74 is a _ closed - loop supermartingale _ , i.e. , @xmath78 for all @xmath72 . if this supermartingale @xmath74 is bounded from below then we can directly apply the convergence theorem [ thm : app ] in the appendix to prove that @xmath52 converges with probability one to the set @xmath79 what remains to be done is to choose an appropriate open - loop supermartingale @xmath74 so that the set @xmath80 is restricted to the target state @xmath81 . the lyapunov function @xmath82 defined in does not discriminate between the different basis vectors . we therefore use the lyapunov function @xmath83 where @xmath84 and @xmath85 is a small positive constant . the weights @xmath86 are strictly positive numbers except for @xmath87 . this function @xmath88 is clearly a concave function of the open - loop martingales @xmath59 and therefore is an open - loop supermartingale . moreover , the weights @xmath86 can be used to quantify the distance of the state @xmath4 from the target state @xmath89 ( c.f . figure [ fig : distance ] below which shows how @xmath86 are chosen for the experimental setting ) . a state @xmath4 is in the set @xmath80 if and only if for all @xmath90 $ ] , we have @xmath91 also from the fact that @xmath88 is an open - loop supermartingale , we have for all @xmath72 @xmath92 we prove in lemma [ lem : first ] below that given any @xmath93 , we can always choose the weights @xmath94 so that @xmath88 satisfies the following property : @xmath95 , @xmath96 has a strict local minimum at @xmath26 for @xmath97 and strict local maxima at @xmath98 for @xmath99 . this combined with equation then ensures that for any @xmath99 , there is some @xmath90 $ ] such that @xmath100 therefore using equation , we know that @xmath56 is in the limit set @xmath80 if and only if @xmath97 . this idea can easily be extended to the situations where the delay @xmath101 is non zero . take @xmath102 as state at step @xmath17 : denote by @xmath103 this state where @xmath104 stands for the control input @xmath0 delayed @xmath105 steps . then the state form of the delayed dynamics is governed by the following markov chain @xmath106 the goal is to design a feedback law @xmath107 that globally stabilizes this markov chain @xmath108 towards a chosen target state @xmath109 for some @xmath110 . in theorem [ thm : main ] , we show how to design a feedback relying on the control lyapunov function @xmath111 . the idea is to use a predictive filter to estimate the state of the system @xmath101 time - steps later . finally , we address the situation where the initial state of the system is not fully known but only estimated by @xmath112 . we show under some assumptions on the initial condition that , the feedback law based on the state @xmath113 of the miss - initialized filter still ensures the convergence of @xmath113 as well as the well - initialized conditional state @xmath52 towards @xmath114 . this demonstrates how the control algorithm is robust to uncertainties in the initialization of the estimated state of the quantum system . the construction of the control lyapunov function relies on two lemmas . [ lem : r ] consider the @xmath115 matrix @xmath116 defined by @xmath117 when @xmath118 , the non - negative @xmath119 is a right stochastic matrix . for @xmath31 , @xmath120 . thus @xmath116 is a metzler matrix . let us prove that the sum of each row vanishes . this results from identity @xmath121 . deriving twice versus @xmath0 the relation @xmath122 yields @xmath123 since for @xmath26 , @xmath124 the above sum corresponds to @xmath125 . therefore , the diagonal elements of @xmath116 are non - positive . if @xmath118 , then @xmath126 and the matrix @xmath119 is well defined with non - negative entries . since the sum of each row of @xmath116 vanished , the sum of each row of @xmath127 is equal to @xmath128 . thus @xmath127 is a right stochastic matrix . to the metzler matrix @xmath116 defined in lemma [ lem : r ] , we associate its directed graph denoted by @xmath129 this graph admits @xmath130 vertices labeled by @xmath131 . to each strictly positive off - diagonal element of the matrix @xmath116 , say , on the @xmath132th row and the @xmath133th column we associate an edge from vertex @xmath132 towards vertex @xmath134 [ lem : first ] assume the directed graph @xmath135 of the matrix @xmath116 defined in lemma [ lem : r ] is strongly connected , i.e. , for any @xmath136 , @xmath137 , there exists a chain of @xmath105 distinct elements @xmath138 of @xmath139 such that @xmath140 , @xmath141 and for any @xmath142 , @xmath143 . take @xmath144 . then , there exist @xmath145 strictly positive real numbers @xmath146 , @xmath147 , such that * for any reals @xmath148 , @xmath147 , there exists a unique vector @xmath149 of @xmath150 with @xmath151 such that @xmath152 where @xmath153 is the vector of @xmath150 of components @xmath148 for @xmath147 and @xmath154 ; if additionally @xmath155 for all @xmath147 , then @xmath156 for all @xmath147 . * for any vector @xmath157 , solution of @xmath158 , the function @xmath159 satisfies @xmath160 since the directed graph @xmath135 coincides with the directed graph of the right stochastic matrix @xmath127 defined in lemma [ lem : r ] , @xmath127 is irreducible . since it is a right stochastic matrix , its spectral radius is equal to @xmath128 . by perron - frobenius theorem for non - negative irreducible matrices , this spectral radius , i.e. , @xmath128 , is also an eigen - value of @xmath127 and of @xmath161 , with multiplicity one and associated to eigen - vectors having strictly positive entries : the right eigen - vector ( @xmath162 ) is obviously @xmath163 ; the left eigen - vector @xmath164 ( @xmath165 ) can be chosen such that @xmath166 . consequently , the rank of @xmath116 is @xmath145 with @xmath167 and @xmath168 where @xmath169 is the hyper - plane orthogonal to @xmath170 . since @xmath171 , @xmath172 , exists @xmath173 such that @xmath174 . since @xmath167 , there is a unique @xmath173 solution of @xmath175 such that @xmath87 . the fact that @xmath176 when @xmath177 for @xmath178 , comes from elementary manipulations of @xmath179 showing that @xmath180 . @xmath181 , set @xmath182 set @xmath183 then @xmath184 is given by @xmath185 and @xmath186 therefore @xmath187 and @xmath188 the main result of this section is expressed through the following theorem . [ thm : main ] consider the markov chain with assumptions [ assum : imp ] , [ assum : impt ] and [ assum : imo ] . take @xmath63 and assume that the directed graph @xmath135 associated to the metzler matrix @xmath116 of lemma [ lem : r ] is strongly connected . take @xmath189 , @xmath190 the solution of @xmath152 with @xmath87 , @xmath177 for @xmath191 , @xmath192 ( see lemma [ lem : first ] ) and @xmath193 take @xmath194 and consider the following feedback law @xmath195}{\text{argmin}}\big({{\text{\normalsize$\mathbb e$}}\left(w_\epsilon(\chi_{k+1})|\chi_k , u_k=\xi\right)}\big ) : = f(\chi_k)\ ] ] where @xmath196 then there exists @xmath197 such that , for all @xmath1980,u^*]$ ] and @xmath1990,\min_{n\neq \bar n } \left(\frac{\lambda_{n}}{r_{n , n}}\right)\right]$ ] , the closed - loop markov chain of state @xmath200 with the feedback law of theorem [ thm : main ] converges almost surely towards @xmath201 for any initial condition @xmath202^\tau$ ] . for the sake of simplicity , first we demonstrate this theorem for @xmath203 and thus for @xmath204 . we then explain how this proof may be extended to arbitrary @xmath205 . here , @xmath206 is given by @xmath207 that can also be presented as @xmath208 with @xmath209 : @xmath210}{\text{min}}\sum_{\mu } p_{\mu,\rho_k}^{\beta_{1,k}}v_\epsilon({{\mathbb k}}^{\xi}({{\mathbb m}}_\mu^{\beta_{1,k}}(\rho_k ) ) ) .\end{gathered}\ ] ] since @xmath211 for any @xmath212 , we have @xmath213 where @xmath214 is defined in . thus , @xmath215 is a supermartingale . according to theorem [ thm : app ] , the @xmath216-limit set @xmath80 of @xmath108 is included in @xmath217 ~|~ q_1(\rho,\beta_1)=0,~q_2(\rho,\beta_1)=0\right\}.\ ] ] we show in 3 steps that @xmath80 is restricted to the target state . + * step 1 . * for all @xmath218 , there exists @xmath219 such that @xmath220 ~|~ { \left < n|\rho|n\right > } \geq 1-\delta\bigg\}.$ ] + * proof of step 1 . * step 1 follows from @xmath221 for all @xmath222 and technical lemma [ lem : q1 ] of the appendix . + * there exist @xmath223 and @xmath224 such that @xmath225 ~|~ { \left<\bar{n}|\rho|\bar{n}\right > } \geq 1-\delta'\}.$ ] + * proof of step 2 . * if @xmath226 , the minimum of the function @xmath227\ni \xi \mapsto f_{\rho,\beta_1}(\xi)=\sum_{\mu } p_{\mu,\rho}^{\beta_1}v_\epsilon({{\mathbb k}}^{\xi}({{\mathbb m}}_\mu^{\beta_1}(\rho ) ) ) $ ] is attained at @xmath228 . by construction of @xmath88 , we know that , for @xmath229 , @xmath39 , @xmath230 if @xmath178 , and @xmath231 the dependence of @xmath232 versus @xmath4 and @xmath233 is continuous from assumption [ assum : imo ] . therefore , @xmath234 such that for all @xmath235 in @xmath236 ~|~ { \left < n|\rho|n\right > } \geq 1-\delta'\}$ ] we have @xmath237 therefore , such @xmath235 and @xmath228 , can not minimize @xmath238 and we have @xmath239 ~|~ { \left < n|\rho|n\right > } \geq 1-\delta'\ } = \emptyset . $ ] therefore we get , @xmath240 ~|~ { \left<\bar{n}|\rho|\bar{n}\right > } \geq 1-\delta'\},\ ] ] by setting @xmath241 in step 1 , for @xmath76 small enough . + * step 3 . * @xmath76 can be chosen small enough so that @xmath242 . + * proof of step 3 . * by construction of @xmath88 , we have for @xmath243 and @xmath244 , @xmath245 by continuity of @xmath232 with respect to @xmath4 and @xmath233 , we can choose @xmath246 and @xmath76 small enough such that for all @xmath247 $ ] and @xmath4 satisfying @xmath248 we have @xmath249 this implies that on the domain @xmath250 ^ 2 ~|~ { \left < n|\rho|n\right > } \geq 1-\bar{\delta}\},$ ] @xmath251 is a uniformly strongly convex function of @xmath252 $ ] . thus the argument of its minimum over @xmath252 $ ] is a continuous function of @xmath4 and @xmath233 . now we choose @xmath253 , where @xmath254 is as in step 2 . take a convergent subsequence of @xmath255 ( that we still denote by @xmath108 for simplicity sakes ) . its limit @xmath256 belongs to @xmath80 and also to @xmath257 ~|~ { \left < n|\rho|n\right > } \geq 1-\bar{\delta}\}.$ ] therefore , @xmath258}{\text{argmin } } f_{\rho_k,\beta_{1,k}}(\xi)$ ] converges to @xmath259}{\text{argmin } } f_{\rho_\infty,\beta_{1,\infty}}(\xi)=0 $ ] since @xmath260 . because @xmath261 , @xmath262 tends almost surely towards @xmath263 . we know that @xmath264 and @xmath265 tend also almost surely to @xmath263 ( theorem [ thm : app ] ) and by uniform ( in @xmath4 ) continuity with respect to @xmath262 , @xmath266 and @xmath267 tend almost surely to @xmath263 . since @xmath268 implies @xmath269 , @xmath52 converges almost surely towards @xmath89 . this completes the proof for @xmath270 . + * extension to @xmath205 . * for @xmath271 and @xmath103 , the proof is very similar . we still have with @xmath214 and @xmath272 given by formulae analogous to and : @xmath262 is replaced by @xmath273 ; @xmath274 is replaced by @xmath275 ; @xmath276 is replaced by @xmath277 since kraus maps are trace preserving . the analogue of technical lemma [ lem : q1 ] , whose proof relies on similar continuity and compactness arguments , has now the following statement : _ @xmath278 such that @xmath279^\tau $ ] satisfying @xmath280 , @xmath281 such that @xmath282 . _ + the last part of the proof showing that , for @xmath283 small enough , the control @xmath73 is a continuous function of @xmath108 when @xmath108 is in the neighborhood of the @xmath216-limit set @xmath284^\tau~|~q_1(\chi)=q_2(\chi)=0\}$ ] remains almost the same . when the measurement process is fully efficient ( i.e. , the detectors are ideal with detection efficiency equal to one and they observe all the quantum jumps ) and the jump model admits no error , the markov process represents a natural choice for estimating the hidden state @xmath285 indeed , the estimate @xmath286 of @xmath4 satisfies the following recursive dynamics @xmath287 where the measurement outcome @xmath21 is driven by . in practice , the control @xmath73 defined in theorem [ thm : main ] could only depend on this estimation @xmath113 replacing @xmath52 in @xmath108 . if @xmath288 , then @xmath289 and theorem [ thm : main ] ensures convergence towards the target state . otherwise , the following result guaranties the convergence of such observer / controller scheme when @xmath290 . [ thm : obscontroller ] consider the recursive equation and assume that the assumptions of theorem [ thm : main ] are satisfied . for each measurement outcome @xmath21 given by , consider the estimation @xmath291 given by with an initial condition @xmath112 . set @xmath292 where @xmath293 is given in theorem [ thm : main ] . then there exists @xmath197 such that , for all @xmath1980,u^*]$ ] and @xmath1990,\min_{n\neq \bar n } \left(\frac{\lambda_{n}}{r_{n , n}}\right)\right]$ ] , @xmath52 and @xmath113 converge almost surely towards the target state @xmath89 as soon as @xmath290 . set @xmath294 . the main idea is that @xmath295 ( where we take the expectation over all jump realizations ) depends linearly on @xmath54 even though we apply the feedback control . since @xmath19 is a function of @xmath296 , @xmath297 is given by @xmath298 , where @xmath299 . the linearity of @xmath295 with respect to @xmath54 is thus verified . now we apply the assumption @xmath290 and therefore one can find a constant @xmath3000,1[$ ] and a well - defined density matrix @xmath301 in @xmath302 such that @xmath303 applying the dominated convergence theorem : @xmath304 by the linearity of @xmath305 with respect to @xmath54 , @xmath306 is given by @xmath307 as @xmath308 and @xmath309 , we necessarily have that both of them converge to @xmath128 since @xmath310 : @xmath311 this implies the almost sure convergence of @xmath52 towards the pure state @xmath312 . we have also @xmath313 depending linearly on @xmath54 . thus @xmath314 is given by @xmath315 using the fact that when @xmath316 , @xmath317 for all @xmath17 , we conclude similarly that @xmath113 converges almost surely towards @xmath312 even if @xmath54 and @xmath112 do not coincide . we now consider the feedback control problem in the presence of classical measurement imperfections with the possibility of detection errors . this model is a direct generalization of the ones used in @xcite ( see also e.g. , @xcite for an introduction to quantum filtering ) . the imperfections in the measurement process are described by a classical probabilistic model relying on a left stochastic matrix @xmath318 , @xmath319 and @xmath33 : @xmath320 and for any @xmath10 , @xmath321 . the integer @xmath322 corresponds to the number of imperfect outcomes and @xmath323 is the probability of having the imperfect outcome @xmath324 knowing the perfect one @xmath10 . set @xmath325 since @xmath52 follows , @xmath326 is also governed by a recursive equation @xcite : @xmath327 where for each @xmath324 , @xmath328 is the super - operator defined by @xmath329 with @xmath330 , and where @xmath331 is a random variable taking values @xmath324 in @xmath332 with probability @xmath333 since @xmath323 is a left stochastic matrix , @xmath334 which is precisely the kraus map associated with the markov process @xmath52 . by assumption [ assum : imp ] , each pure state @xmath56 , @xmath131 remains a fixed point of the markov process when @xmath48 . we now consider the dynamics of the filter state @xmath335 in the presence of delays in the feedback control . similar to the case with perfect measurements , let @xmath336 be the filter state at step @xmath17 , where @xmath337 is the feedback control at time - step @xmath17 delayed @xmath338 steps . then the delay dynamics are determined by the following markov chain @xmath339 instead of assumption [ assum : impt ] we now assume [ assum : hatimpt ] for all @xmath31 in @xmath32 there exists @xmath319 such that @xmath340 assumption [ assum : hatimpt ] means that there exists a @xmath324 such that the statistics when @xmath48 for obtaining the measurement result @xmath324 are different for the fixed points @xmath45 and @xmath46 . this follows by noting that @xmath341 for @xmath43 . we now state the analogue of theorem [ thm : main ] in the case of imperfect measurements . [ thm : mainmeas ] consider the markov chain with assumptions [ assum : imp ] , [ assum : imo ] and [ assum : hatimpt ] . take @xmath63 . assume that the directed graph of metzler matrix @xmath116 of lemma [ lem : r ] is strongly connected . take @xmath189 , @xmath190 solution of @xmath152 with @xmath87 , @xmath177 for @xmath191 , @xmath192 ( see lemma [ lem : first ] ) and set @xmath342 take @xmath194 and consider the following feedback law @xmath343}{\text{argmin}}\big({{\text{\normalsize$\mathbb e$}}\left(w_\epsilon({{\widehat{\chi}}}_{k+1})|{{\widehat{\chi}}}_k , u_k=\xi\right)}\big ) = f({{\widehat{\chi}}}_k ) , $ ] where @xmath344 + then there exists @xmath197 such that , for all @xmath1980,u^*]$ ] and @xmath1990,\min_{n\neq \bar n } \left(\frac{\lambda_{n}}{r_{n , n}}\right)\right]$ ] , the closed - loop markov chain of state @xmath345 with the feedback law of theorem [ thm : mainmeas ] converges almost surely towards @xmath201 for any initial condition @xmath346^\tau$ ] . the proof of this theorem is almost identical to that of theorem [ thm : main ] with @xmath347 replaced by @xmath348 and @xmath349 replaced by @xmath350 and we do not give the details of this proof . + for estimating the hidden state @xmath335 needed for the feedback design of theorem [ thm : mainmeas ] , let us consider the estimate @xmath351 given by @xmath352 where @xmath331 corresponds to the imperfect outcome detected at step @xmath17 . such @xmath331 is correlated to the perfect and hidden outcome @xmath21 of through the classical stochastic process attached to @xmath318 : for each @xmath21 , @xmath331 is a random variable to be equal to @xmath319 with probability @xmath353 . in practice , the control @xmath73 defined in theorem [ thm : mainmeas ] could only depend on this estimation @xmath354 replacing @xmath326 in @xmath355 . the following result guaranties the convergence of the feedback scheme when @xmath356 . [ thm : obscontrollermeas ] consider the recursive equation and take assumptions of theorem [ thm : mainmeas ] . consider the estimation @xmath357 given by with an initial condition @xmath358 . set @xmath359 where @xmath293 is given by theorem [ thm : mainmeas ] . then there exists @xmath197 such that , for all @xmath1980,u^*]$ ] and @xmath1990,\min_{n\neq \bar n } \left(\frac{\lambda_{n}}{r_{n , n}}\right)\right]$ ] , @xmath52 and @xmath354 converge almost surely towards the target state @xmath89 as soon as @xmath356 . let us first prove that @xmath326 defined by converges almost surely towards @xmath89 . since @xmath360 , we have @xmath361 . thus , there exist @xmath362 and @xmath3000,1[$ ] , such that , @xmath363 . similarly to the proof of theorem [ thm : obscontroller ] , @xmath364 depends linearly on @xmath365 : @xmath366 is given by @xmath367 . moreover by theorem [ thm : mainmeas ] , @xmath366 converges almost surely towards @xmath128 , thus @xmath368 and @xmath369 converge also towards @xmath128 . consequently @xmath326 converges almost surely towards @xmath89 . since @xmath326 is the conditional expectation of @xmath52 knowing the past imperfect outcomes and control inputs and since its limit @xmath89 is a pure state , @xmath52 converges necessarily towards the same pure state almost surely . convergence of @xmath354 relies on similar arguments exploiting the linearity of @xmath370 versus @xmath365 . in this section , we give the explicit expression of the feedback controller which has been experimentally tested in laboratoire kastler - brossel ( lkb ) at ecole normal suprieure ( ens ) de paris . we briefly summarize how the control design elucidated in this paper is applied to the lkb experiment . we refer the interested reader to @xcite for more details . this feedback controller has been obtained by the lyapunov design discussed in previous sections . [ fig : setup ] figure [ fig : setup ] is a sketch of the experimental setup considered in @xcite . two - level atoms , with state space spanned by the vectors @xmath371 and @xmath372 ( excited and ground state respectively ) , act as qubits . the atoms , emitted from box b , interact individually with the electromagnetic microwave field stored in the cavity @xmath373 and get entangled with it . the cavity is placed in between two additional cavities , @xmath374 and @xmath375 , where the atomic state can be manipulated at will . together , they form an atomic interferometer similar to those used in atomic clocks . the atomic state , @xmath371 or @xmath372 , is eventually measured in the detection device @xmath376 . using the outcomes of those measurements , a quantum filter , run by a real - time computer , estimates the density matrix @xmath4 of the microwave field trapped in the cavity @xmath373 . it also uses all the available knowledge on the experimental setup , especially the limited detection efficiency @xmath377 ( percentage of atoms that are actually detected ) , the detection errors @xmath378 and @xmath379 ( @xmath380 is the probability to detect an atom in the wrong state ) . a controller eventually calculates ( state feedback ) the amplitude @xmath381 of the classical microwave field that is injected into @xmath373 by an external source @xmath382 so as to bring the cavity field closer to the target state . the succession of the atomic detection , the state estimation and the microwave injection define one iteration of our feedback loop . it corresponds to a sampling time of around @xmath383s . for a more detailed description see @xcite . the state to be stabilized is the cavity state . the underlying hilbert space is @xmath384 which is assumed to be finite - dimensional ( truncations to @xmath385 photons ) . it admits @xmath386 as an ortho - normal basis . each basis vector @xmath387 corresponds to a pure state , called photon - number state ( fock state ) , with precisely @xmath388 photons in the cavity , @xmath389 with notations of subsection [ ssec : model ] , @xmath390 and the index @xmath391 . the three main processes that drive the evolution of our system are decoherence , injection , and measurement . their action onto the system s state can be described by three super - operators @xmath392 , @xmath393 and @xmath394 , respectively . we refer to @xcite for more details . all operators are expressed in the fock - basis @xmath395 truncated to @xmath385 photons . after taking into account our full knowledge about the experiment , we finally get the following state estimate at step @xmath396 : @xmath397 with the following short descriptions for the super - operators @xmath392 , @xmath393 and @xmath398 * the decoherence manifests itself through spontaneous loss or capture of a photon to or from the environment which is described as follows : @xmath399 where @xmath400 , @xmath401 and @xmath402 with @xmath403 @xmath404 and @xmath405 photon annihilation and creation operators ( @xmath406 and @xmath407 ) and with @xmath408 the photon number operator ( @xmath409 ) . besides , @xmath410 is the mean number of photons in the cavity mode at thermal equilibrium with its environment . * the evolution of the state @xmath4 after the control injection is modeled through @xmath411 with @xmath412 in reality , the control at step @xmath17 , @xmath413 is subject to a delay of @xmath20 steps which corresponds to the number of flying atoms between the cavity @xmath373 and the detector @xmath376 . * in the real experiment , the atom source is probabilistic and is characterized by a truncated poisson probability distribution @xmath414 to have @xmath415 atom(s ) in a sample ( we neglect events with more than 2 atoms ) . this expands the set of the possible detection outcomes to @xmath416 values @xmath417 , related to the following measurement operators , @xmath418 @xmath419 , @xmath420 , + @xmath421 , @xmath422 and @xmath423 where @xmath424 @xmath425 and @xmath426 are physical parameters . + the real measurement process is not perfect : the detection efficiency is limited to @xmath427 and the state detection errors are non - zero ( @xmath428 . these imperfections are taken into account by considering the left stochastic matrix @xmath323 which is given in @xcite . consequently , the optimal state estimate after measurement outcome @xmath429 gets the following form : @xmath430 the measurement operators @xmath431 are diagonal in the fock basis @xmath432 , illustrating their quantum non - demolition nature with respect to the photon number operator and thus fulfilling assumption [ assum : imp ] . besides , assumption [ assum : hatimpt ] can also be fulfilled by a proper choice of the experimental parameters @xmath433 and @xmath426 . for @xmath434 ( no cavity decoherence ) the markov model of density matrix @xmath335 associated to the filter is exactly of the form with @xmath435 being fixed - points in open - loop . similarly , the underlying markov process of the true cavity state @xmath4 , which is unobservable in practice because of detection errors and delays , admits the same fixed points in open - loop . with parameters given in subsection [ ssec : exp ] ( except @xmath436 ) , these markov processes satisfy assumptions [ assum : imp]-[assum : impt]-[assum : imo ] for @xmath4 and assumptions [ assum : imp]-[assum : imo]-[assum : hatimpt ] for @xmath335 . moreover the metzler matrix @xmath116 of lemma [ lem : first ] is irreducible . consequently the assumptions of theorem [ thm : mainmeas ] are satisfied . the feedback law proposed in theorem [ thm : obscontrollermeas ] and relying on the filter state @xmath351 will stabilize globally the unobservable state @xmath4 towards the target photon - number state @xmath437 . numerous closed - loop simulations show that taking @xmath438 in the feedback law does not destroy stability and does not affect the convergence rates . this explains why in the simulations and experiments , we set @xmath438 despite the fact that theorem [ thm : obscontrollermeas ] guaranties convergence only for arbitrary small but strictly positive @xmath85 . for @xmath439 positive and small , the @xmath440s are no more fixed - points for @xmath4 and @xmath335 in open - loop . let us detail how to adapt the feedback scheme of theorem [ thm : obscontrollermeas ] . at each step of an ideal experiment the control @xmath73 minimizes the lyapunov function @xmath441 ( @xmath85 is set to zero ) calculated for state @xmath442 . in our real experiment however , we also take into account decoherence and @xmath101 flying not - yet - detected samples and therefore choose @xmath73 to minimize @xmath443 for @xmath444 with @xmath351 given by . here we have introduced the kraus map of the real experiment @xmath445 . the control @xmath0 minimizing @xmath446 is approximated by @xmath259}{\operatorname{argmax}}\left(a_1 \xi + a_2 \xi^2/2\right)$ ] with @xmath447\rho)}$ ] and @xmath448,{\text{\bf{a}}}^\dag\!-\!{\text{\bf{a}}}]\rho)}$ ] , where @xmath449 is the diagonal operator @xmath450 . the coefficients @xmath451 are computed using lemma [ lem : first ] where , for @xmath452 , @xmath453 are chosen negative and with a decreasing modulus versus @xmath388 in order to compensate cavity decay . for @xmath454 , we have compared in simulations different setting and selected the profile displayed in figure [ fig : distance ] . used in the feedback for the simulation of figure [ fig : simulation ] and for the experiment of figure [ fig : experimental].,height=226 ] closed - loop simulation of figure [ fig : simulation ] shows a typical monte - carlo trajectory of the feedback loop aiming to stabilize the 3-photon state @xmath455 . the experimental parameters used in the simulations are the following : @xmath456 , @xmath457 , @xmath458 , @xmath459 , @xmath460 , @xmath461 , @xmath462 , @xmath463 , @xmath464 , and @xmath465 . for the feedback , @xmath451 are given in figure [ fig : distance ] , @xmath438 and @xmath466 . the initial states @xmath54 and @xmath358 take the following form : @xmath467 . . ( a ) detection results . ( b ) control lyapunov function . ( c ) control input . ( d ) estimated photon number probabilities : @xmath468 in red , @xmath469 in thick green , and @xmath470 in blue . ( e ) cavity photon number probabilities : @xmath471 in red , @xmath472 in thick green , and @xmath473 in blue . , scaledwidth=50.0% ] the results of the experimental implementation of the feedback scheme are presented in figure [ fig : experimental ] . figure [ fig : experimental](e ) shows that the average fidelity of the target state is about @xmath474 . besides , the asymmetry between the distributions for @xmath475 and @xmath476 indicates the presence of quantum jumps occurring preferentially downwards ( @xmath477 . contrarily to the simulations of figure [ fig : simulation ] , the cavity photon number probabilities relying on @xmath4 are not accessible in the experimental data of figure [ fig : experimental ] since we do not have access to the detection errors and to the cavity decoherence jumps @xcite . nevertheless , green curves in simulations of figures [ fig : simulation](d ) and [ fig : simulation](e ) indicate that when @xmath469 exceeds @xmath478 , @xmath4 coincides , with high probability , with @xmath437 . . ( a ) detection results . ( b ) evolution of the control function . ( c ) control injection . ( d ) estimated photon number probabilities : @xmath468 in red , @xmath469 in thick green , and @xmath470 in blue . ( e ) estimated photon number probabilities averaged over 4000 experimental trajectories of the feedback loop.,scaledwidth=50.0% ] we have proposed a lyapunov design for state - feedback stabilization of a discrete - time finite - dimensional quantum system with qnd measurements . extensions of this design are possible in different directions such as * replacing the continuous and one - dimensional input @xmath0 by a multi - dimensional one @xmath479 ; * assuming that @xmath0 belongs to a finite set of discrete values ; * taking an infinite dimensional state space as in @xcite where the truncation to finite photon numbers is removed ; * considering continuous - time systems similar to the ones investigated in @xcite ; * ensuring convergence towards a sub - space @xcite instead of a pure - state and thus achieving a goal similar to error correction code as already proposed in @xcite . * acknowledgements : * the authors thank michel brune , serge haroche and jean - michel raimond for enlightening discussions and advices . 10 c. ahn , a. c. doherty , and a. j. landahl . continuous quantum error correction via quantum feedback control . , 65:042301 , mar 2002 . h. amini , m. mirrahimi , and p. rouchon . stabilization of a delayed quantum system : the photon box case - study . , 57(8):19181930 , 2012 . h. amini , p. rouchon , and m. mirrahimi . design of strict control - lyapunov functions for quantum systems with qnd measurements . in _ 50th ieee conference on decision and control and european control conference _ , pages 81938198 , 2011 . michel bauer and denis bernard . convergence of repeated quantum nondemolition measurements and wave - function collapse . , 84:044103 , oct 2011 . s. bolognani and f. ticozzi . engineering stable discrete - time quantum dynamics via a canonical qr decomposition . , 55(12):27212734 , 2010 . l. bouten , r. van handel , and m. r. james . an introduction to quantum filtering . 46(6):21992241 , 2007 . l. bouten and r. van handel . , chapter on the separation principle of quantum control . world scientific , 2008 . ( see also : arxiv : math - ph/0511021 ) l. bouten , r. van handel , and m. r. james . a discrete invitation to quantum filtering and feedback control . , 51:239316 , 2009 . m. brune , j. bernu , c. guerlin , s. deleglise , c. sayrin , s. gleyzes , s. kuhr , i. dotsenko , j - m . raimond , s. haroche . process tomography of field damping and measurement of fock state lifetimes by quantum nondemolition photon counting in a cavity . , 101 , 240402 , 2008 . i. dotsenko , m. mirrahimi , m. brune , s. haroche , j .- raimond , and p. rouchon . quantum feedback by discrete quantum non - demolition measurements : towards on - demand generation of photon - number states . , 80 : 013805 - 013813 , 2009 . c. w. gardiner and p. zoller . quantum noise . , 2000 . deterministic and nondestructively verifiable preparation of photon number states . , 97(073601 ) , 2006 . j. gough and m.r . james . the series product and its application to quantum feedforward and feedback networks . , 54(11):2530 2544 , 2009 . c. guerlin , j. bernu , s. delglise , c. sayrin , s. gleyzes , s. kuhr , m. brune , j - m . raimond and s. haroche . progressive field - state collapse and quantum non - demolition photon counting . , 448 : 889894 , 2007 . s. haroche and j.m . oxford university press , 2006 . james and j.e . gough . quantum dissipative systems and feedback control design by interconnection . , 55(8):1806 1821 , 2010 . k. kashima and n. yamamoto . control of quantum systems despite feedback delay . , 54(4):876 881 , 2009 . kushner . . holt , rinehart and wilson , inc . , 1971 . m. mirrahimi and r. van handel . stabilizing feedback controls for quantum systems . , 46(2):445467 , 2007 . nielsen and i.l . chuang . . , 2000 . k. nishio , k. kashima , and j. imura . effects of time delay in feedback control of linear quantum systems . , 79:062105 , jun 2009 . r. romano , d. dalessandro . incoherent control and entanglement for two - dimensional coupled systems . , 73:022323 , 2006 . r. romano , d. dalessandro . environment - mediated control of a quantum system . , 97:080402 , 2006 . c. sayrin . 2011 . c. sayrin , i. dotsenko , x. zhou , b. peaudecerf , th . rybarczyk , s. gleyzes , p. rouchon , m. mirrahimi , h. amini , m. brune , j.m . raimond , and s. haroche . real - time quantum feedback prepares and stabilizes photon number states . , 477:7377 , 2011 . doi:10.1038/nature10376 . a. somaraju , i. dotsenko , c. sayrin , and p. rouchon . design and stability of discrete - time quantum filters with measurement imperfections . , 50845089 , 2011 . a. somaraju , m. mirrahimi , and p. rouchon . approximate stabilization of an infinite dimensional quantum stochastic system . , 25:1350001,2013 . h. wang , m. hofheinz , m. ansmann , r. c. bialczak , e. lucero , m. neeley , a. d. oconnell , d. sank , j. wenner , a. n. cleland , and john m. martinis . measurement of the decay of fock states in a superconducting quantum circuit . , 101 : 240401 , 2008 . h.m . wiseman and a.c . optimal unravellings for feedback control in linear quantum systems . , 94(7):070405 , 2005 . h.m . wiseman and g.j . milburn . . , 2009 the following theorem is just an application of theorem @xmath128 in ( * ? ? ? [ thm : app ] let @xmath480 be a markov chain on the compact state space @xmath481 suppose , there exists a continuous function @xmath482 satisfying @xmath483 where @xmath484 is a non - negative continuous function of @xmath485 then the @xmath216-limit set @xmath486 ( in the sense of almost sure convergence ) of @xmath480 is contained by the following set @xmath487 the proof is the following . for all @xmath493 , condition @xmath494 implies that @xmath495 . taking the sum over all @xmath496 , we get @xmath497 for all @xmath60 and @xmath10 . since @xmath498 and @xmath499 are @xmath36-function of @xmath233 , these relations read @xmath500 where @xmath501 stands for @xmath502 and the scalar functions @xmath503 depend continuously on @xmath4 and @xmath504 . + let us finish the proof by contradiction . assume that for all @xmath489 , there exists @xmath505 $ ] satisfying @xmath506 , such that @xmath507 , @xmath508 . take @xmath373 tending towards @xmath509 . since @xmath510 and @xmath511 remain in a compact set , we can assume , up to some extraction process , that @xmath510 and @xmath511 converge towards @xmath512 and @xmath513 in @xmath5 and @xmath514 $ ] . since @xmath515 , we have @xmath516 . since latexmath:[\[\begin{gathered } \label{eq : rhocn } we have by continuity for @xmath373 tending to @xmath509 and for all @xmath60 and @xmath10 : @xmath500 . thus there exists @xmath518 such that @xmath519 ( see the proof of theorem [ thm : first ] ) . since @xmath520 , for @xmath373 large enough , @xmath521 and thus @xmath522 taking @xmath523 , by assumption [ assum : impt ] , there exists @xmath10 such that @xmath524 . replacing in yields : @xmath525 thus , there exists @xmath526 , such that for @xmath523 and @xmath373 large enough @xmath527 but @xmath528 . this is in contradiction with @xmath529 as soon as @xmath530 .

nonclassical light generation has attracted much attention over the last 40 years , partly because it not only provides a powerful probe for addressing fundamental issues of quantum theory such as complementarity , hidden variables , and other aspects central to the foundations of quantum mechanics @xcite ; but also holds promise for many potential applications to quantum information processing @xcite , quantum computation and communication @xcite , quantum cryptography @xcite , quantum imaging @xcite , quantum lithography @xcite , and quantum metrology @xcite . in particular , entangled photon pairs have already been established as a standard research tool in the field of quantum optics . traditionally , paired photons are produced from spontaneous parametric down conversion ( spdc ) @xcite , in which a strong pump laser drives the atomic oscillators in a noncentrosymmetric crystal into a nonlinear regime , and then two down - converted beams are radiated by these oscillators . the two down - converted photons from such a nonlinear process usually have very broad bandwidth ( typically in the thz range ) and very short coherence time ( typically around few picoseconds ) so that their waveforms are not resolvable by existing single - photon detectors ( which have resolution around the nanosecond range ) . in addition , the short coherence length of the spdc biphotons , roughly about @xmath0 @xmath1 m , also limits their application in long - distance quantum communication . the optical properties of these wide - band biphotons have been well studied both experimentally and theoretically for more than two decades @xcite . narrow - bandwidth biphotons are ideal for a number of recently proposed protocols for long - distance quantum communication based on coherent interaction between single photons and atomic ensembles @xcite , which require efficiently absorbing biphotons and storing the entanglement . other applications include the long - distance quantum state teleportation which requires long temporal coherence time of paired photons @xcite . using cavity - enhanced spdc in nonlinear crystals , paired photons with a bandwidth of about 10 mhz and coherence time of about @xmath2 ns have been reported @xcite . however , the spdc process is usually tuned to occur in the region of far - off - resonant atomic transitions , which limits the conversion efficiency and requires a large pump power . moreover , in the far - off - resonant pumping the nonlinear parametric coupling coefficient can be treated as a constant and the information about the material internal transition structure is unresolvable from the two - photon correlation measurement . the two most recently developed advanced technologies make the near- and on - resonance nonlinear processes possible . the first is laser cooling and trapping of neutral atoms @xcite . because cold atoms with a temperature of about 100 @xmath3 have negligible doppler broadening and very long lifetime , their atomic hyperfine structures can be resolved without need of a doppler - free setup . the second is the electromagnetically induced transparency ( eit ) @xcite which not only can eliminate absorption on resonance but also enhance nonlinear interactions at low light level @xcite . following the theoretical proposal in @xcite and early experiments by lukin and kimble groups @xcite , generation of nonclassical correlated photon pairs by exploring the writing - reading " procedure has attracted much research interest . however , the created paired photons are not maximally time - frequency entangled due to two separate operation processes . using cold atomic - gas media and eit - assisted four - wave mixing ( fwm ) , harris group at stanford university was the first to generate time - frequency entangled narrow - band biphotons with cw driving fields in a double-@xmath4 system @xcite . by controlling the eit at a sufficiently high optical depth ( od ) around 50 , biphotons with subnatural linewidth have been produced in a recent experiment , where the coherence time is up to about 1 @xmath1s in a two - dimensional magneto - optical trap ( mot ) @xcite . vuletic and his colleagues have also produced biphotons with a temporal coherence length of 100 ns by trapping cold atoms inside an optical cavity @xcite . , by driving a two - level system into the nonlinear region , have also created paired photons with correlation time of same order of magnitude @xcite . to describe the physics behind narrow - band biphoton generation from the eit - based multilevel atomic system , two different approaches have been presented in the literature . one approach is to use the heisenberg - langevin method by solving the coupled field operator equations in the heisenberg picture , and the results agree well with the experimental data in the two - photon correlation measurement @xcite . the advantage of this method is that it provides a sophisticated calculation of the effects from the langevin noises , the medium absorption and optical gain on the biphoton wavepacket . working in the interaction picture , the other approach developed by wen _ @xcite focuses on the description of the two - photon state vector after the nonlinear medium using perturbation theory . this state vector theory not only properly describes the experimental results but also offers a clear physical picture on biphoton generation mechanism . it is the intent of this paper to review and generalize this latter approach by taking into account both linear loss and gain , and provide more insights on the roles of nonlinear interaction and linear response to the biphoton amplitude . using this state vector picture , we show that the two - photon wave function is a convolution of the nonlinear and linear optical responses . according to this convolution , the two - photon temporal correlation is considered in two regimes , damped rabi oscillation and group delay . the paper is organized as the following . in sec . [ sec : biphotonstate ] , using perturbation theory we lay out a general formulism of the two - photon state and biphoton wavepacket ( or amplitude ) generated from the fwm parametric process . in sec . [ sec : opticalresponse ] , taking a four - level double-@xmath4 atomic system as an example , we show that the linear and nonlinear optical responses to the generated fields play an important role in determining the two - photon amplitude . by looking at the nonlinear susceptibility , we illustrate the mechanism of biphoton generation in such a system . in sec . [ sec : singleatomregime ] we describe the damped rabi oscillation regime where the biphoton amplitude is mainly characterized by the structure of the third - order nonlinear susceptibility . we show that the two - photon temporal correlation can exhibit damped and over - damped rabi oscillaitons . in sec . [ sec : groupdelayregime ] , we describe the group - delay regime where the biphoton wavepacket is mainly determined by the phase matching due to the linear optical response . in sec . [ sec : interference ] , we examine the two - photon interference using biphotons generated from such a cold atomic - gas medium . we show that by manipulating the linear optical response , it is possible to switch the two - photon anti - bunching - like effect to bunching - like effect . finally , we draw the conclusions and outlook in sec . [ sec : conclusion ] . ( color online ) biphoton generation via a four - level double-@xmath4 atomic system . ( a ) the level structure , where in the presence of a cw pump ( @xmath5 ) and coupling ( @xmath6 ) beams , paired stokes ( @xmath7 ) and anti - stokes ( @xmath8 ) photons are spontaneously created from the four - wave mixing processes in the low - gain regime . ( b ) the forward generation configuration . ( c ) the backward generation geometry . , width=302 ] a schematic of biphoton generation via a four - level double-@xmath4 atomic system is shown in fig . [ fig : fourlevelconfiguration ] , where in the presence of a cw pump ( @xmath5 ) and coupling ( @xmath6 ) lasers , phase - matched , paired stokes ( @xmath7 ) and anti - stokes ( @xmath8 ) photons are spontaneously produced from the fwm process in the low - gain limit . as shown in fig . [ fig : fourlevelconfiguration](a ) , the strong coupling laser forms a standard three - level @xmath4 eit scheme with the generated anti - stokes field . therefore , the role of the coupling laser here is that it not only assists the fwm nonlinear process but also creates a transparency window for the anti - stokes photons by the slow - light effect . with the use of eit , it is possible to generate the anti - stokes photons near atomic resonance @xmath9 , which greatly enhances the efficiency of fwm process . the phase - matching condition allows forward and backward two generation configurations , as shown in fig . [ fig : fourlevelconfiguration](b ) and ( c ) , respectively . in the backward generation configuration paired stokes and anti - stokes photons can propagate collinearly with the pump and coupling beams @xcite , and can also propagate in a right - angle geometry @xcite . these flexible generation setups depend on whether the relationship @xmath10 is well satisfied , where @xmath11 are wave vectors of the input pump and coupling fields . one advantage for the backward generation geometry is that it allows us to easily separate the generated weak fields from the strong inputs . in the following discussions , we assume that the cold atomic - gas medium consists of identical four - level atoms prepared in their ground state @xmath12 [ see fig . [ fig : fourlevelconfiguration](a ) ] . the idealized atoms or molecules are confined within a long , thin cylindrical volume with a length @xmath13 and atomic density is @xmath14 . for simplicity , in this paper we will not take into account the doppler broadening and polarization effects . in the two - photon limit , the quantum langevin noise introduces unpaired photons which are not of interest here and so is ignored . in addition , we concentrate on the two - photon temporal correlation . we start with the geometries shown in fig . [ fig : fourlevelconfiguration](b ) and ( c ) , where stokes - anti - stokes photon pairs propagate along the @xmath15 axis . we denote the electric field as @xmath16=\frac{1}{2}[e^{(+)}+c.c.]$ ] , where @xmath17 and @xmath18 are the positive- and negative - frequency parts . the input pump and coupling beams are assumed to be strong classical fields , so that @xmath19},\nonumber \\ e^{(+)}_c(z , t)&=&e_ce^{i[k_c z-\omega_c t]},\label{eq : pcfields}\end{aligned}\ ] ] where @xmath20 are pump and coupling field wave numbers . the @xmath21 sign in eq . ( [ eq : pcfields ] ) stands for the forward and backward propagation geometries for the pump laser . the single - transverse - mode stokes and anti - stokes fields are taken as quantized , @xmath22 } , \nonumber \\ \hat{e}^{(+)}_{as}(z , t)&=&\frac{1}{\sqrt{2\pi}}\int d\omega \sqrt{\frac{2 \hbar \omega}{c \varepsilon_0 a}}\hat{a}_{as}(\omega)e^{i[k_{as}(\omega)z-\omega t ] } , \label{eq : fieldoperators}\end{aligned}\ ] ] where @xmath23 are wave numbers of stokes and anti - stokes photons , and @xmath24 is the single - mode cross section area .. the annihilation operators @xmath25 and @xmath26 satisfy the commutation relation @xmath27=[\hat{a}_{as}(\omega),\hat{a}^\dagger_{as}(\omega')]= \delta(\omega-\omega ' ) . \label{eq : fcommutation}\end{aligned}\ ] ] in the interaction picture , the effective interaction hamiltonian for the four - wave mixing parametric process takes the form @xcite @xmath28 where @xmath29 means the hermitian conjugate , and @xmath30 is the third - order nonlinear susceptibility to the stokes ( or anti - stokes ) field defined by the nonlinear polarizability @xmath31 @xcite . with use of eqs . ( [ eq : pcfields ] ) and ( [ eq : fieldoperators ] ) , after the @xmath15 integration we rewrite the hamiltonian ( [ eq : hamiltonian0 ] ) as @xmath32 where @xmath33 is the phase mismatching for the forward ( @xmath34 ) and backward ( @xmath35 ) configurations . when the pump and coupling fields are incident with an angle respect to the @xmath15 axis , the phase mismatching can be generalized as @xmath36 , where @xmath37 is the unit vector along the @xmath38 axis . the integral over the length @xmath13 gives a sinc function @xmath39 , which determines the two - photon natural spectrum width . @xmath40 is the nonlinear parametric coupling coefficient , where @xmath41 and @xmath42 are central frequencies of the anti - stokes and stokes fields , respectively . perturbation theory gives the photon state at the output surface(s ) approximately as a linear superposition of @xmath43 , where @xmath44 is the vacuum state . the two - photon ( biphoton ) state @xmath45 can be expressed as @xmath46 ) in eq . ( [ eq : biphotonstate0 ] ) , the time integral yields a @xmath47 function , @xmath48 , which expresses the energy conservation for the process . if @xmath13 is infinite , then the sinc function @xmath49 in ( [ eq : hamiltonian ] ) becomes a @xmath47 function , @xmath50 . in this case , the conditions @xmath51 both hold and the phase matching is said to be perfect . the phase matching condition arises from the fact that the fwm process is a coherent process . now the two - photon state ( [ eq : biphotonstate0 ] ) becomes @xmath52 as seen from eq . ( [ eq : biphotonstate ] ) , the two - photon state is entangled in frequency and wave number , but is not entangled in polarization . in frequency space , the entanglement is the result of the frequency phase - matching condition , which implies that the detection of a photon at frequency @xmath8 requires the detection of the other photon at frequency @xmath53 . the frequency correlation has interesting consequences for the temporal behavior of the pair @xcite , as we shall see . the state is also entangled with respect to the wave number since the sinc function @xmath39 can not be factorized as a function of @xmath54 times a function of @xmath55 . in the general noncollinear case , the wave - number entanglement has implication for the spatial correlation of photon pairs @xcite . the polarization - entanglement generation has been discussed in @xcite . to study the optical properties of paired stokes and anti - stokes photons generated from such a four - level system , let us look at the simple experiments of two - photon joint - detection measurement as shown in fig . [ fig : fourlevelconfiguration](b ) and ( c ) , where anti - stokes photons go to detector d@xmath56 and stokes photons go to d@xmath57 . the two - photon temporal correlation is the quantity of primary interest in this paper . the field annihilation operators in time domain on the surfaces are expressed as @xmath58 } , \nonumber \\ \hat{a}_{as}(t)&=&\frac{1}{\sqrt{2\pi}}\int d\omega \hat{a}_{as}(\omega)e^{i[k_{as}(\omega)l/2-\omega t]}. \label{eq : operatorintime}\end{aligned}\ ] ] it can be shown that they obey the following commutation relation @xmath59=[\hat{a}_{as}(t),\hat{a}^\dagger_{as}(t')]=\delta(t - t').\label{eq : tcommutation}\end{aligned}\ ] ] the joint - detection probability is defined by glauber s theory , @xmath60 in general , @xmath61 is referred to as the _ two - photon amplitude _ , or _ biphoton wavepacket_. with the help of eqs . ( [ eq : biphotonstate ] ) and ( [ eq : operatorintime ] ) , the two - photon wave function ( or amplitude ) now is @xmath62 where @xmath63 with the relative time delay @xmath64 and @xmath65 . in eq . ( [ eq : biphotonwavefunction1 ] ) , @xmath66 is defined as the _ longitudinal detuning function _ , @xmath67 from eq . ( [ eq : biphotonwavefunction1 ] ) we see that the biphoton wave function is determined by both the nonlinear coupling coefficient ( @xmath68 ) and the longitudinal detuning function ( @xmath69 ) . in the time domain , @xmath70 is a convolution ( @xmath71 ) of @xmath72 and @xmath73 @xmath74=l\int d\tau'\tilde{\kappa}(\tau')\tilde{\phi}(\tau-\tau ' ) . \label{eq : convolution}\end{aligned}\ ] ] here , @xmath75 is fourier transform of the nonlinear coupling coefficient @xmath76 , @xmath77 and similarly , @xmath78 is fourier transform of the longitudinal detuning function @xmath66 , @xmath79 the photon pair emission rate @xmath80 can be evaluated , by using eqs . ( [ eq : biphotonwavefunction1 ] ) , as @xmath81 now let us consider a photon counting experiment . let the detectors be at equal distances from the source so that the relative time delay caused by the free - space propagation of electric fields are equal . suppose the stokes photon is detected by d@xmath57 at time @xmath82 , and after a controllable delay ( @xmath83 ) detector d@xmath56 is opened to detect the anti - stokes photon . assuming perfect detection efficiency , the averaged two - photon coincidence counting rate is determined by @xmath84 @xmath85 is the coincidence window function within the time bin width @xmath86 : @xmath87 for @xmath88 ; otherwise , @xmath89 . when the two - photon coherence time is much larger than the detector time bin width @xmath86 , eq . ( [ eq : rcc ] ) becomes @xmath90 , which is a measure of the two - photon correlation function . this is the essence of narrow - band paired photons measurement discussed in this paper . equation ( [ eq : biphotonwavefunction ] ) is the starting point for the discussions in the following sections . by extending wave numbers as complex variables , the effects of loss and gain inside the material can be properly included in the above theory , as we shall see . it is found that to be consistent with the heisenberg - langevin theory @xcite in the low - gain limit , the argument in @xmath69 should be replaced by @xmath91 , where @xmath92 is the conjugate of @xmath93 . before proceeding the discussion let us make a short summary . in this section , we have laid out the general formulism of the two - photon state and wave function generated in a fwm medium pumped by two cw classical fields using perturbation theory . the two - photon state ( [ eq : biphotonstate ] ) is a coherent superposition and it is entangled in frequency and wave numbers . the biphoton wave function , as shown in eq . ( [ eq : convolution ] ) , is a convolution of @xmath72 and @xmath73 . alternatively , eq . ( [ eq : biphotonwavefunction1 ] ) states that the two - photon spectrum is determined by both the nonlinear susceptibility and the phase matching . this implies two regimes for the temporal correlation measurement , as discussed in sec . [ sec : singleatomregime ] and sec . [ sec : groupdelayregime ] . the methodology explored here is equivalent to the heisenberg - langevin theory @xcite at the two - photon level . moreover , this method offers a clear physical picture of the paired photons , which makes it easier to understand the two - photon interference experiment @xcite . using the above theory new phenomena have been analyzed @xcite which are not easy to be interpreted from the heisenberg - langevin theory . for example , the two - photon correlation at a high optical depth @xcite , the simulation by numerically solving the heisenberg - langevin coupled - operator equations shows a good agreement in the leading edge spike with the experiment data , but the nature of optical precursors is much easier to be understood in the present biphoton state picture @xcite . as shown in eq . ( [ eq : biphotonwavefunction ] ) , the two - photon wavefunction is determined by both the nonlinear susceptibility and the longitudinal detuning function . therefore , in this section we look at the linear and nonlinear optical responses to the generated fields . on the one hand , the importance of the nonlinear susceptibility is that it not only characterizes the strength of the nonlinear parametric process but also is the mechanism of biphoton generation . as a consequence , it determines the feature of the two - photon temporal correlation in one regime . on the other hand , the longitudinal detuning function ( related to the linear susceptibilities ) sets the natural spectral width of biphotons , and governs the pattern of the temporal correlation in another regime . the optical response of a four - level double-@xmath4 eit system has been discussed in @xcite . et al_. have further systematically studied the optical responses in a three - level double-@xmath4 scheme @xcite and a two - level system @xcite . recently , the responses of a four - level inverted - y system have been analyzed in @xcite . for simplicity , here let us still choose paired stokes - anti - stokes generation shown in fig . [ fig : fourlevelconfiguration ] as an example to illustrate the physics behind fwm . assuming that the intensity of the off - resonant pump laser is chosen so that the atomic population remains primarily in the ground level @xmath12 , the third - order nonlinear susceptibility for the generated anti - stokes field is @xcite @xmath94}\nonumber \\ = \frac{-n\mu_{13}\mu_{32}\mu_{24}\mu_{41}/(\varepsilon_0\hbar^3)}{4(\delta_{p}+i\gamma_{14})(\omega-\omega_e/2+i \gamma_{e})(\omega+\omega_e/2+i\gamma_{e})},\label{eq : chi3}\end{aligned}\ ] ] in eq . ( [ eq : chi3 ] ) , @xmath95 are the electric dipole matrix elements , @xmath96 is the coupling rabi frequency , and @xmath97 are dephasing rates , respectively . @xmath98 is the pump detuning from the atomic transition @xmath99 . @xmath100 is the detuning of the anti - stokes photons from the transition @xmath9 , and we will take @xmath101 as the anti - stokes central frequency . @xmath102 is the effective coupling rabi frequency . @xmath103 is the effective dephasing rate . the third - order nonlinear susceptibility has two resonances separated by @xmath104 and each is associated with a linewidth of @xmath105 . the two - photon coherence time is inversely proportional to @xmath105 . more importantly , the two resonances here indicate two types of fwm processes . alternatively , two types of paired stokes and anti - stokes photons can be generated from these two fwm processes in the spontaneous emission region . one fwm process happens when the central frequency of the anti - stokes photons is @xmath106 and the central frequency of the correlated stokes photons is @xmath107 . the other fwm occurs as the anti - stokes photons have the central frequency at @xmath108 while the paired stokes photons have the frequency at @xmath109 . this is the generation mechanism of biphotons produced from the fwm in atomic - gas media . consequently , the interference between these two types of biphotons will appear in the two - photon temporal correlation , as shall be discussed in sec . [ sec : singleatomregime ] . the results obtained here agree well with the dressed - state picture @xcite . the linear susceptibilities at the stokes and anti - stokes frequencies are , respectively , @xmath110 where @xmath111 is the pump rabi frequency . the complex wave numbers of stokes and anti - stokes photons are obtained from the relations @xmath112 and @xmath113 , where the imaginary parts stand for the raman gain and eit loss , respectively . the linear susceptibility ( [ eq : chilinear ] ) to the anti - stokes field takes the form of standard eit . as expected , the coupling beam @xmath114 not only assists the propagation of the anti - stokes photons by creating an eit window , but also manipulates the group velocity using the slow - light effect . another role played by eit here is that it allows the nonlinear optics occurring near atomic resonance with small absorption and hence enhances the efficiency of the nonlinear interaction . it is instructive to examine the linear responses under some simplifying approximation . taking @xmath115 , eqs . ( [ eq : chis ] ) and ( [ eq : chilinear ] ) give @xmath116 and @xmath117 so that the wave - number mismatching is @xmath118 . here @xmath119 is the central wave number of the anti - stokes field , @xmath120 is its group velocity , and @xmath121 , the imaginary part of the anti - stokes wave number , characterizes the eit loss , respectively . now eq . ( [ eq : detuningfunction ] ) can be approximated as @xmath122 which sets a coherence time about @xmath123 . in eq . ( [ eq : phasematchingapp ] ) , @xmath124 where @xmath125 is the on - resonance absorption cross section in the transition @xmath9 . as discussed in this section , there are three important characteristic frequencies that determine the shape of the biphoton wave function . the first is the coupling effective rabi frequency @xmath104 , which determines the two - resonance spectrum of the nonlinear susceptibility . the second is the linewidth @xmath105 of the two resonances in the nonlinear susceptibility . the third is the full - width - at - half - maximum ( fwhm ) phase - matched bandwidth determined by the sinc function , @xmath126 , where @xmath127 is the anti - stokes group delay time . in time domain , they correspond to the rabi time @xmath128 , nonlinear coherence time @xmath129 and group delay time @xmath127 . the group delay time can be estimated from @xmath130 with the optical depth defined as @xmath131 . the competition between @xmath132 , @xmath133 and @xmath127 will determine which effect plays a dominant role in governing the feature of the two - photon correlation . therefore , in the following we will discuss the two - photon joint - detection measurement in two regimes , damped rabi oscillation and group delay . let us first look at the damped rabi oscillation regime , where optical properties of the two - photon amplitude ( [ eq : biphotonstate ] ) are mainly determined by the nonlinear coupling coefficient . this regime requires that the effective coupling rabi frequency @xmath104 and linewidths @xmath134 be smaller than the phase - matching bandwidth @xmath135 or @xmath136 and @xmath137 , so that we can treat the longitudinal detuning function as @xmath138 and @xmath139 . the biphoton spectral generation rate is proportional to @xmath140 . hence both biphoton spectrum intensity and emission rate are proportional to @xmath141 . if @xmath142 , the two - photon temporal correlation clearly exhibits a damped rabi oscillation with oscillation period @xmath143 ; if @xmath144 , the second and following oscillations are suppressed because of the short dephasing time , as experimentally demonstrated in @xcite . let us consider the case of @xmath145 first , which implies a real effective coupling rabi frequency @xmath104 . the two - photon wave function now is determined by fourier transform of the nonlinear coupling coefficient ( [ eq : nonlinearresponse ] ) , which is @xmath146\theta(\tau ) . \label{eq : f}\end{aligned}\ ] ] and @xmath147 . the heaviside step function here is defined as @xmath148 for @xmath149 , and @xmath150 for @xmath151 . combining with eq . ( [ eq : biphotonwavefunction ] ) , the biphoton wavepacket becomes @xmath152\theta(\tau),\label{eq : total}\end{aligned}\ ] ] where @xmath153 and @xmath101 have been used . the physics of eq . ( [ eq : total ] ) is understood as follows . because the two - photon state is entangled , it can not be factorized into a function of @xmath154 times a function of @xmath155 . @xmath156 , depends only on the relative time delay @xmath83 , which implies that the pair is randomly generated at any time . if the wave packets of the pump and the coupling beams are not plane waves , the term @xmath157 would become a wave packet with a coherence length determined by the pump and coupling . equation ( [ eq : total ] ) shows also destructive interference between the two biphoton channels . the first term in the bracket on the rhs of eq . ( [ eq : total ] ) represents the two - photon amplitude of paired anti - stokes at @xmath158 and stokes at @xmath159 ; while the second term is the two - photon amplitude of paired anti - stokes at @xmath160 and stokes at @xmath161 . to further see the interference , let us look at the two - photon glauber correlation function , @xmath162\theta(\tau),\label{eq : rabig2}\end{aligned}\ ] ] the physics of eq . ( [ eq : rabig2 ] ) shows the beating ( or interference ) between two types of paired photons generated from the two fwm processes . the two - photon correlation ( [ eq : rabig2 ] ) exhibits damped rabi oscillations of period @xmath143 . the damping rate is determined by the resonant linewidth @xmath105 in the doublet . at @xmath163 , @xmath164 vanishes and as @xmath165 , @xmath164 also approaches to zero . this indicates photon anti - bunching - like effect [ @xmath166 . in the counter - propagation geometry , if @xmath167 , @xmath168 becomes a symmetric distribution about @xmath163 , because spontaneously emitted paired photons can trigger both detectors d@xmath56 and d@xmath57 as experimentally verified in @xcite . figure [ fig : rabioscillation](a ) shows the first experimental demonstration of this type of biphoton source by bali _ _ in a cold @xmath169rb gas @xcite . the solid lines in fig . [ fig : rabioscillation](a1 ) and ( a2 ) are theoretical curves obtained from the heisenberg - langevin theory . in fig . [ fig : rabioscillation](b1 ) and ( b2 ) , the curves are theoretical simulations using eq . ( [ eq : rabig2 ] ) with the same experimental parameters in ( a1 ) and ( a2 ) . comparing with the heisenberg - langevin theory , the state vector picture also gives an agreement with the experimental data . from eq . ( [ eq : rabig2 ] ) , fig . [ fig : rabioscillation](a1 ) and ( b1 ) correspond to the case of @xmath170 where the oscillations can be clearly observed before damping out ; while fig . [ fig : rabioscillation](a2 ) and ( b2 ) correspond to the case of @xmath171 where only the first oscillation is observable and other disappear due to the fast dephasing rate @xmath172 . although in @xcite the lack of other oscillations in fig . [ fig : rabioscillation](a2 ) were interpreted due to the group delay , we believe that both cases shown in fig . [ fig : rabioscillation](a1)and ( a2 ) are in the damped rabi oscillation regime and the correlation time is well fitted with @xmath133 and @xmath132 . this can be further verified by reducing the optical depth by a factor of 1/2 , which is corresponding to a reduction of the group delay by a factor of 1/2 . however the correlation time widthes in fig . [ fig : rabioscillation](a2 ) hold the same . if @xmath173 , the effective coupling rabi frequency becomes purely imaginary , @xmath174 . the biphoton wave function becomes @xmath175 at the weak coupling limit , @xmath176 and eq . ( [ eq : overdamped1 ] ) takes the form @xmath177 this can be understood as follows . since the coupling beam is very weak , the atomic - level structure is not changed due to the input power . two types of biphotons are generated with the same central frequency but are associated with different linewidths , @xmath178 and @xmath179 , respectively . the interference between these two types of biphotons gives a manifested exponential decay . in such a limit , the eit disappears and the medium absorption decreases the biphoton generation efficiency . this case might be observable in the dilute atomic - gas medium with very low optical depth ( @xmath180 ) . suggested by balic _ _ et al.__@xcite and demonstrated by du _ _ et al.__@xcite , the group delay regime is defined as @xmath181 and the ( eit ) slow light effect can be used to dynamically control the biphoton temporal correlation time . the group delay condition is equivalent to @xmath182 , i.e. , the biphoton bandwidth is determined by the phase - matching spectrum . therefore , in this section we treat the third - order nonlinear susceptibility as a constant over the phase - matching spectrum . as a consequence , the mechanism of double resonance of biphoton generation is washed out by the longer group - delay time . the two - photon correlation in this regime looks more like that with conventional spdc photons . the coincidences tends to be a square - like pattern . the biphoton wave function can be approximated as @xmath183 where @xmath184 is the on - resonance nonlinear coupling constant . we find that in this regime , the eit profile modifies the biphoton wave function through transparency window and slow light . when the eit bandwidth ( @xmath185 ) is larger than the phase - matching bandwidth , the anti - stokes loss can be ignored and the two - photon wave packet approaches a rectangular shape . however , when the eit loss is significant , the biphoton wave packet follows an exponential decay . when @xmath186 , we can ignore the loss in eq . ( [ eq : phasematchingapp ] ) and rewrite it as @xmath187 its fourier transform gives the biphoton wave function @xmath188 the rectangular function @xmath189 , ranging from @xmath163 to @xmath190 , shows that the anti - stokes photon is always delayed with respect to its paired stokes photon because of the slow light effect . it is clear that the two - photon correlation time is determined by the group delay @xmath191 . the photon pair generation rate can be calculated from eq . ( [ eq : pairrate ] ) as @xmath192 thus even though the spectral generation rate , @xmath193 , scales as @xmath141 , the total rate of paired counts scales linearly as @xmath194 because the bandwidth reduces linearly with optical depth @xcite . as explained by rubin _ et al_. @xcite , the rectangular waveform may be understood as follows : the paired photons are always produced from the same , but unknown , space point in the nonlinear medium . if emitted from the front surface , the anti - stokes photon has no delay , and both photons arrive at detectors simultaneously . however , if emitted from the back surface , the anti - stokes photon is delayed relative to the stokes photon by @xmath127 . for conventional spdc photons , the rectangular - shaped biphoton wave packet has been well known and its sub - ps correlation time has been directly measured in @xcite . interestingly , we find that to observe the rectangular shape , the condition @xmath186 sets a lower bound for the optical depth . using @xmath185 and @xmath195 , we obtain @xmath196 . it can be shown that such an optical depth is necessary to have the eit delay - bandwidth product greater than two @xcite . ( color online ) two - photon correlation function in the group delay regime . ( a ) rectangular - shape correlation for an ideal spontaneous parametric down converter with a relative group delay @xmath190 and pair emission rate @xmath80 . ( b ) experimental data ( red ) and theoretical curve ( solid blue line ) obtained from the heisenberg - langevin theory in ref . ( c ) propagation delay of an anti - stokes pulse ( red ) with respect to a reference pulse ( black ) . data are taken from ref . ( d ) theoretical simulation using the perturbation theory fitted with same experimental parameters in ( b).,width=302 ] figure [ fig : groupdelay](a ) shows an ideal rectangular - shape correlation with a group delay of @xmath191 . in fig . [ fig : groupdelay](b ) the experimental demonstration of a near - rectangular - shape correlation is obtained at the optical depth around 53 @xcite , where experimental data are denoted by + " . the solid line is theoretical fitting using the heisenberg - langevin theory . as seen from fig . [ fig : groupdelay](b ) , the correlation time of about 300 ns is achieved and is consistent with the anti - stokes pulse delay measurement shown in the inset ( c ) , where the smaller curve represents the delayed signal and the higher one is the reference pulse . the experimental parameters used here are @xmath197 mhz , @xmath198 , @xmath199 , @xmath200 , @xmath201 , and @xmath202 , respectively . the eit transparency width is estimated around @xmath203 mhz and the phase - matching spectral width about @xmath204 mhz . the theoretical simulation using the perturbation theory [ eq . ( [ eq : biphotonwavefunction1 ] ) ] , plotted in fig . [ fig : groupdelay](c ) , also fits very well with the reported experimental data . it is found that the exponential - decay behavior at the tail is due to the finite eit loss , which alters the correlation function away from the ideal rectangular shape . in the experiment @xcite , _ have also experimentally verified that the total coincidence counts are linearly dependent on the optical depth . the sharp peak shown in the leading edge of the two - photon coincidence counts in fig . [ fig : groupdelay](b ) is the first observed sommerfeld - brillouin precursor @xcite at the biphoton level @xcite . mathematically , the integration in the two - photon amplitude ( [ eq : biphotonwavefunction1 ] ) is similar with the classical pulse propagation . in the stationary - phase approximation , starting from eq . ( [ eq : biphotonwavefunction1 ] ) one can show that the sharp peak is resulted from the interference of biphotons which are generated out of the eit opacity window . the detailed analysis of the precursor generation at the two - photon level has been presented in @xcite by using the theory laid out in sec . [ sec : biphotonstate ] . although by solving the coupled - operator equations the heisenberg - langevin theory predicts the correct pattern , the physics behind the precursor is not easy to be retrieved from this theory . suppose now the eit loss is not negligible , i.e , @xmath205 . in such a case , the longitudinal detuning function ( [ eq : phasematchingapp ] ) can be approximated as @xmath206 the two - photon wave function becomes @xmath207 the correlation function @xmath208 has an exponential decay time of @xmath209 . this means that the anti - stokes photon detected at @xmath83 is generated at the depth of @xmath210 such that the amplitude decreases by @xmath211 . in sec . [ sec : singleatomregime ] and sec . [ sec : groupdelayregime ] , we have discussed the two - photon temporal correlation in two regimes . in general , the two - photon wave function is a convolution of both the nonlinear and linear responses as shown in eq . ( [ eq : convolution ] ) . in both regimes , the correlation functions have non - zero values only at @xmath212 . the observed anti - bunching - like effect in the damped rabi oscillation regime is due to the destructive interference between two types of biphoton generation . in principle , if the detection system is frequency sensitive enough to resolve these two types of biphotons , the damped rabi oscillations will disappear and the square - wave - like pattern manifested by exponential decay will be observed . ( color online ) biphoton generation in a three - level system . ( a ) right - angle experimental configuration . ( b ) three - level double-@xmath4 energy level diagram.,width=302 ] in this section , we are going to briefly show that it is possible to switch the two - photon coincidence counting rate from photon anti - bunching - like effect to bunching - like effect , by manipulating the linear optical responses to both stokes and anti - stokes fields @xcite . we consider a case that stokes and anti - stokes photons are degenerate at both spectrum and polarization so that both stokes and anti - stokes experience slow light propagation . as shown in fig . [ fig : feyman](a ) and ( b ) , we consider paired stokes and anti - stokes photons produced from a three - level double-@xmath4 scheme in a right angle configuration . in the counter - propagation geometry , frequency - phase matching allows both stokes and anti - stokes fields to reach both detectors . the two generated fields both may experience slow - light effect inside the medium due to the linear optical responses . the biphoton wave function now is expressed as @xmath213\ast\pi(\tau ;- l / v_g , l / v_g).\label{eq : convolution2}\end{aligned}\ ] ] even though there is no temporal overlap between @xmath75 and @xmath214 , their convolution with @xmath215 mixes them together in the time domain . by manipulating the group delay , the two - photon interference in eq . ( [ eq : convolution2 ] ) can be switched from the destructive to the constructive . alternatively , the coincidence counting rate is changed from photon anti - bunching - like effect to bunching - like effect . this can be achieved by making the indistinguishability of their propagation pathways which gives rise to the biphoton interference . the interference comes from two indistinguishable paths for paired stokes and anti - stokes photons . in one possible propagation pathway , the pair is created at @xmath216 with stokes photon going to the left and anti - stokes photon to the right . in the other propagation pathway , the pair is created at @xmath217 but with stokes photon going to the right and anti - stokes photon to the left . the interference between the two paths then modifies the two - photon correlation by further altering the group delay . the details of this two - photon interference are presented in @xcite . recently , wen _ _ has proposed a scheme to switch the two - photon interference from photon anti - bunching - like effect to bunching - like effect by manipulating the linear optical responses to the paired photons in a strongly driven two - level system @xcite . in conclusion , using perturbation theory , we have discussed the nature of the two - photon state and its waveform for narrow - band biphoton generation near atomic resonance via fwm parametric process using eit and slow light . in general , the biphoton waveform is determined by a convolution of the nonlinear response ( determined by the third - order nonlinear susceptibility ) and the linear response ( determined by the longitudinal detuning function ) . therefore , the two - photon wave function can be analytically solved in two regimes , the damped rabi oscillation and the group delay . in the damped rabi oscillation regime , where the biphoton wave function is mainly determined by the third - order nonlinear susceptibility , the two - photon temporal correlation exhibits a damped rabi oscillation because of the interference between two types of fwms . in the group delay regime , where the two - photon amplitude is mainly determined by the phase - matching condition , the correlation approaches a rectangular shape manifested by an exponential decay . the theoretical simulations agree very well with recent experimental results @xcite and the heisenberg - langevin theory @xcite . although the discusses are based on a four - level double-@xmath4 eit system , the analysis can be easily extended to three - level and two - level systems . before concluding , we note that a number of other interesting avenues are presently being explored that are closely related with this new type of biphoton source . by manipulating the linear optical responses to the generated fields , bell inequality can be tested by two - photon beatings using biphotons generated from a two - level system @xcite . it is interesting to note that the correspondence principle fails in describing paired - photon generation in a two - level system . as mentioned in @xcite , it is not generally valid to simply replace the coupled - field maxwell s equations by coupled field - operator equations without changing the coefficients . but the underlying physics is not clear at this moment . it will be interesting to explore quantum networks with these narrow - band paired photons . for example , the feasibility of an eit - based quantum memory can be easily demonstrated using current biphoton source by looking at the two - photon interference , as suggested in @xcite . it may be interesting to look at the squeezing properties after an eit system if the input squeezed light is initially generated from the atomic - gas medium @xcite . as proposed in @xcite experimental demonstration of polarization - entangled narrow - band biphoton generation would be useful for quantum computation and communication . another problem is that it is not clear whether the maximally time - frequency entangled photon pairs can be generated in the room - temperature atomic - gas medium . since all the experimental and theoretical work now are focused on cold atomic ensembles , it would be interesting to know if the doppler broadening becomes dominant in the process , what kind of feature would be observed in the two - photon correlation measurement ? by optimizing the system , can we still observe the damped rabi oscillations ? apart from temporal correlation , a possible application of these narrow - band biphotons may involve the transverse correlation @xcite . it may be worth studying whether this new type of source is useful for quantum imaging and lithography . s. du acknowledges the start - up financial support from the department of physics at the hong kong university of science and technology . j .- wen and m. h. rubin were supported in part by the u.s . army research office under muri grant w911nf-05 - 1 - 0197 . a. n. boto , p. kok , d. s. abrams , s. l. braunstein , c. p. williams , and j. p. dowling , quantum interferometric optical lithography : exploiting entanglement to beat the diffraction limit , " phys . lett . * 85 * , 2733 - 2736 ( 2000 ) . a. migdall , r. datla , a. v. segienko , j. s. orszak , and y .- h . shih , measuring absolute infrared spectral radiance with correlated visible photons : technique verification and measurement uncertainty , " appl . opt . * 37 * , 3455 - 3463 ( 1998 ) . a. kuzmich , w. p. bowen , a. d. boozer , a. boca , c. w. chou , l .- m . duan , and h. j. kimble , generation of nonclassical photon pairs for scalable quantum communication with atomic ensembles , " nature ( london ) * 423 * , 731 - 734 ( 2003 ) . wen and m. h. rubin , transverse effects in paired - photon generation via an electromagnetically induced transparency medium . i. perturbation theory , " phys . a * 74 * , 023808 ( 2006 ) ; transverse effects in paired - photon generation via an electromagnetically induced transparency medium . ii . beyond perturbation theory , " phys . rev . * 74 * , 023809 ( 2006 ) . wen , m. h. rubin , and s. du , two - photon beatings using biphotons generated from a two - level system , " _ slow light and fast light conference _ ( utah , 2007 ) ; j .- wen , s. du , m. h. rubin , and e. oh , two - photon beatings using biphotons generated from a two - level system , " opt . express ( to be published ) .

observations of the cmb , supernovae , lensing and baryon acoustic oscillations have led to the cosmological standard model which requires an accelerated expansion of the late universe , driven by dark energy but despite many years of research its origin has not yet been identified . there are two major explanations for the origin and properties of dark energy . the first solution consists of introducing a cosmological constant @xmath0 with a constant energy density causing an effective repulsive force between cosmological objects at large distances . from the particle physics perspective the cosmological constant could correspond to the vacuum energy density . the theoretical expectations for the vacuum energy density caused by fluctuating quantum fields , however , exceeds the observational bounds on @xmath0 by up to 120 orders of magnitude . this discrepancy remains for more almost a century one of the most challenging puzzles in physics , @xcite . alternatively , the acceleration of the universe can be explained by introducing new dynamical degrees of freedom , either by invoking new fluids @xmath1 with negative pressure or by changing the geometrical part of einstein s equations . in particular , weakening gravity on cosmological scales could not only tackle the cosmological constant problem , but would also come hand in hand with new degrees of freedom which might be responsible for a late - time speed - up of the hubble expansion . such scenarios could arise in massive gravity or in higher - dimensional frameworks . in the higher dimensional picture , the dvali - gabadadze - porrati ( dgp ) model is one of the important large scale modified theories of gravity @xcite . in this braneworld model our universe is confined to a three - brane embedded in a five - dimensional bulk . on small scales , four - dimensional gravity is recovered due to an intrinsic einstein hilbert term sourced by the brane curvature , whereas on larger , cosmological scales gravity is systematically weaker as the graviton acquires a soft mass @xmath2 which limits its effective range . being a fundamentally higher dimensional theory , the effective four - dimensional graviton on the brane carries five degrees of freedom , namely the usual helicity-2 modes , two helicity-1 modes and one helicity-0 mode . whilst the helicity-1 mods typically decouple , the helicity-0 one can mediate an extra fifth force . in the limit @xmath3 , one recovers general relativity ( gr ) through the vainshtein mechanism : the basic idea is to decouple the additional modes from the gravitational dynamics via nonlinear interactions of the helicity-0 mode of the graviton , @xcite . as a result , at the vicinity of matter , the non - linear interaction for the helicity-0 mode become large and hence suppresses its coupling to matter . this decoupling of the nonlinear helicity-0 mode is manifest in the limit where @xmath4 and @xmath5 while the strong coupling scale @xmath6 is kept fixed . this limit enables a linear treatment of the usual helicity-2 mode of gravity while the helicity-0 mode @xmath7 is described non - linearly , which is the so - called decoupling limit . one of the successes of the dgp model is the existence of a self - accelerating solution , where the acceleration of the universe is sourced by the graviton own degrees of freedom ( more precisely its helicity-0 mode ) . unfortunately that branch of solution seems to be plagued by ghost - like instabilities @xcite , in the dgp model , but this issue could be avoided in more sophisticated setups , for instance including gauss - bonnet terms in the bulk @xcite . more recently , it has been shown that the decoupling limit of dgp could be extended to more general galilean invariant interactions @xcite . this galileon model relies strongly on the symmetry of the helicity-0 mode @xmath7 : invariance under internal galilean and shift transformations , which in induced gravity braneworld models can be regarded as residuals of the 5-dimensional poincar invariance . these symmetries and the postulate of ghost - absence restrict the construction of the effective @xmath7 lagrangian . there exist only five derivative interactions which fulfill these conditions . from the five dimensional point of view these galilean invariant interactions are consequences of lovelock invariants in the bulk of generalized braneworld models , @xcite . since their inception there has been a flurry of investigations related to self - accelerating de sitter solutions without ghosts @xcite , galileon cosmology and its observations @xcite , inflation @xcite , lensing @xcite , superluminalities arising in spherically symmetric solutions around compact sources @xcite , k - mouflage @xcite , kinetic gravity braiding @xcite , etc ... . furthermore , there has been some effort in generalizing the galileon to a non - flat background . the first attempt was then to covariantize directly the decoupling limit and to study its resulting cosmology @xcite . in particular it was shown in @xcite that the naive covariantization would yield ghost - like terms at the level of equation of motion but a given unique nonminimal coupling between @xmath7 and the curvature can remove these terms resulting in second order of equations of motion @xcite , which are also consistent with a higher - dimensional construction @xcite . in this paper , we will pursue the same strategy when constructing our proxy theory . the outcome of this kind of covariantization method was explored further in @xcite . while this covariantization is ghost - free , the galileon symmetry is broken explicitly in curved backgrounds . only recently there has been a successful generalization to the ( anti- ) de sitter background and ultimately to maximally symmetric backgrounds where it has been discovered that the de sitter galileon interactions acquire additional , potential - like terms being functions of @xmath7 and mixtures of @xmath7 and gradients of @xmath7 , all fulfilling a generalized galileon symmetry @xcite . there exists a parallel to theories centered on a massive graviton : galileon - type interaction terms naturally arise in gravitational theories using a massive spin-2 particle as an exchange particle , which has , in addition , been constructed to be ghost - free be it in three dimensions , @xcite or for an generalized fierz - pauli action in four dimensions @xcite . not only is the existence of a graviton mass a fundamental question from a theoretical perspective , it could also have important consequences both in cosmology and in solar system physics , @xcite . although solar system observations have confirmed gr to high accuracy and placed bounds on the graviton mass to be smaller than a few @xmath8ev , even such a small mass would become relevant at the hubble scale which corresponds to the graviton compton wavelength . in particular , it has been successfully shown that this massive gravity theory exhibits a stable self - accelerated solution in the decoupling limit since the scalar mode can generate a constant negative pressure density . in the decoupling limit , the expansion history of the universe in this self - accelerating branch was found to be indistinguishable from @xmath0cdm , @xcite . while the self - accelerating solutions in the above models yield viable expansion histories including late - time acceleration , they do not address the cosmological constant problem , the giant mismatch between the theoretically computed high energy density of the vacuum and the low observed value . a possible answer comes from the idea of degravitation , which asserts that the energy density could be as large as the theoretically expected value , but would not bear a large effect on the geometry . technically , gravity is less strong on large scales ( ir - limit ) and could act as a high - pass filter suppressing the gravitational effect of a potentially large vacuum energy . since such modifications of gravity in the ir naturally arise in models of massive gravity , they logically provide a possible mechanism to degravitate the vacuum energy density , @xcite , which was observed in bi - galileon models @xcite as well as in the decoupling limit of massive gravity , @xcite . analogously , the dgp braneworld model can be extended to higher dimensions to tackle the cosmological constant problem as well , @xcite . in this paper , we focus on the covariantization of a ghost - less extension of fierz - pauli massive gravity recently proposed in @xcite and show that this proxy model allows for a stable self - accelerating solution . hereby , we have performed the covariantization in the jordan as well as in the einstein frame . we discuss the differences between the two approaches and the consequences of choosing the wrong starting frame . in the well defined proxy theory we are able to tackle the puzzle about the self - acceleration of the universe but not the one about the cosmological constant problem . furthermore , we study the perturbations around the self - accelerating background and provide the expression for the modified evolution equation for these density perturbations . as expected , the @xmath7 field enhances the gravitational clustering resulting in a rapid growth of structures . this result is quite different from the one we had obtained in @xcite , where the self - accelerating solution was indistinguishable from a @xmath0cdm , but we emphasize that the theory used in this paper , is a proxy model distinct from massive gravity . at early times , the enhancement of clustering is restrained since the galileon self - interactions are the dominant ones suppressing their energy density relative to that of matter or radiation . once the matter density has dropped sufficiently , the galileon become an important contribution to the dark sector of the universe driving cosmic expansion . + the paper is organized as follows : we start in section [ sec : mg1 ] with a summary of the non - linear theory of massive gravity @xcite which we use for constructing our proxy theory , where we discuss the successful implementation of ghost - free massive gravity in the decoupling limit . we then move to our proxy theory in section [ sec : proxy ] where we first give some motivations on why we use a proxy theory rather than the exact non - linear theory . we then work out the consequences of this covariantization for cosmology and specifically for late - time acceleration in section [ sec : sa ] where we study also the stability conditions and the reason why a degravitating solution can not be constructed in our covariantized theory . we then move onto more general cosmology in section [ sec : cosmology ] and follow the helicity-0 mode contribution to the universe throughout its evolution , before quickly presenting the consequences for structure formation in section [ sec : structure ] . we use the same notation as @xcite , with @xmath9 , @xmath10 and @xmath11 where the covariant derivative is taken w.r.t . @xmath12 and square brackets @xmath13 $ ] represent the trace of a tensor @xmath14=\pi_{\mu\nu}\pi^{\mu\nu}$ ] and @xmath15 ^ 2=\pi^\mu_\mu\pi^\nu_\nu$ ] . the first theory of massive gravity was proposed by fierz and pauli in 1939 @xcite , but was shown to be unstable by boulware - deser ( bd ) @xcite , due to the non - propagation of the hamiltonian constraint at the non - linear level . this instability can also be seen in the language @xcite where in the decoupling limit , the helicity-0 mode typically has equations of motion with more than two derivatives , and hence does not possess a well defined cauchy surface . however it was shown in @xcite that the graviton potential could be built in such way as to remove any higher derivative term in the equations of motion , and obtain a galileon - type of action for the additional helicity-0 mode . to review this , let us start with a graviton of mass @xmath2 described by l=(r - u(g , h))[eq : full ] with the potential @xmath16 , where the tensor @xmath17 is constructed in terms of the metric @xmath18 and the four fields @xmath19 by @xmath20 . we can then split the fields into helicity-1 and -0 contributions , but the helicity-1 mode decouples in the decoupling limit and can be consistently set to zero . we therefore focus on the helicity-0 mode @xmath7 and write @xmath21 such that @xmath22 . defining the quantity @xmath23 the most generic potential that bears no ghosts in the decoupling limit is u(g , h)=-4@xmath24[eq : fullu ] where @xmath25 are two free parameters and _ 2&=&[]^2-[^2 ] + _ 3&=&[]^3 - 3 [ ] [ ^2]+2[^3 ] + _ 4&=&[]^4 - 6[^2][]^2 + 8[^3 ] [ ] + 3[^2]^2 - 6[^4 ] . notice that @xmath26 can be expressed in terms of @xmath27 and the tadpole @xmath28 $ ] , @xcite . it has then been shown that ghost - like pathologies in this theory of massive gravity theory disappear to all orders in the decoupling limit and at least up to quartic order beyond the decoupling limit , as well as completely non - linearly in some specific cases . while this theory is completely covariant , studying the cosmology as well as other non - trivial curved geometries can be extremely complicated . instead , we will here focus on its version in the decoupling limit and covariantize the theory directly from this limit . we emphasize that the resulting proxy theory will be distinct from the theory of massive gravity presented above , but presents nevertheless some interesting features for cosmology . in the decoupling limit , taking the scales @xmath29 and @xmath3 , while keeping the strong coupling scale @xmath30 fixed , the above lagrangian reduces to the more compact expression l=-12h^e^_h_+h^x^(1)_+h^x^(2)_+h^x^(3)_+h^t_[eq : massivegravity ] where @xmath31 stands for the helicity-2 mode canonically normalized , @xmath32 is the lichnerowicz operator , the coefficients @xmath33 are related to the free parameters @xmath25 and @xmath34 denote the interactions with the helicity-0 mode @xcite x^(1 ) _ & = & g_- _ [ eq : x1 ] + x^(2 ) _ & = & ^2_-_-12([^2]-[]^2)g _ + x^(3 ) _ & = & 6 ^ 3_-6[]^2_+3([]^2-[^2])_-g_([]^3 - 3[^2][]+2[^3]).[eq : x3 ] it is worth to mention that these interaction terms are all transverse and at most second order in time derivatives to ensure the absence of ghost . being ghost - free , these interactions are closely related to the galileon interactions and fulfill the same internal symmetry . in the next section we will covariantize these interaction terms @xmath35 and discuss their physical properties . + in the case of external sources there is a coupling between the metric @xmath31 and the stress energy tensor via @xmath36 but there is not such a direct coupling with the field @xmath7 . nevertheless , if one diagonalizes the first interaction term @xmath37 by a change of variables of the form @xmath38 , then the coupling between @xmath7 and external sources will become transparent , @xmath39 . instead of studying the cosmology in the exact non - linear covariantized theory which can be extremely hard , we use the alternative approach of covariantizing the lagrangian in the decoupling limit , and use the resulting theory as a proxy . we claim that the covariantized version of the above lagrangian is simply given by [ eq : totalaction ] s=(^2 r+^(,g)+l^matter(,g ) ) , where @xmath40 is the lagrangian for the matter fields @xmath41 living on the geometry , the lagrangian for @xmath7 is @xcite [ eq : covjor ] ^=@xmath42 . and the tensor @xmath43 stands for the dual riemann tensor l^=2r^+2(r^g^+r^g^-r^g^-r^g^ ) + r(g^g^-g^g^ ) . this form of tensor structure has been first discussed by horndeski @xcite in the context of the most general scalar - tensor theory , and more recently in @xcite . however we point out here that these interactions come as a direct outcome of massive gravity . we can show explicitly the following correspondences , h^x^(1 ) _ & & -r + h^x^(2 ) _ & & -__g^ + h^x^(3 ) _ & & - _ _ _ l^. [ eq : corres ] the einstein equation is given by [ eq.:einstein ] g_=t^_+t^matter _ with t^_=t^(1)_-t^(2)_-t^(3 ) _ and the structure of the einstein and riemann dual tensor ensure that @xmath7 enters at most with two derivatives in the stress - energy tensor , t^(1)_&=&x^(1)_+g _ [ eq : t1 ] + t^(2)_&=&x^(2)_+12l_^^+12g_()^2 [ eq : t2 ] + t^(3)_&=&x^(3)_+32l_^()^2 [ eq : t3 ] where @xmath44 are defined in ( [ eq : x1 ] - [ eq : x3 ] ) . furthermore , the fact that @xmath45 , @xmath46 , @xmath47 and @xmath48 have at most one time - derivative guarantees the propagation of constraints . since we are not in the einstein frame , these stress - energy tensor are only transverse on - shell , and satisfy the relation , @xmath49 where @xmath50 is the equation of motion with respect to @xmath7 . since both the einstein tensor and the riemann dual tensor are transverse , this equation of motion is also at most second order in derivative , _ & = & + & = & -r - g^-l^(__+r^___)=0,[eq : pij ] where we have used the fact that d_d_d_l^&=&-r^__l^=14 ^l_gb with @xmath51 . in the rest of this paper , we study the resulting cosmology in this proxy theory , starting with the existence of self - accelerating solutions . in what follows , we focus on the cosmology of the covariantized theory ( [ eq : totalaction ] , [ eq : covjor ] ) , and focus for that on a frw background with scale factor @xmath52 and hubble parameter @xmath53 . the resulting effective energy density and pressure for the field @xmath7 are then ^&=&(6h+6h^2-h^2 ^ 2-h^3 ^ 3)[eq : effdens ] + p^&=&3,[eq : effpress ] and the equation of motion for @xmath7 ( [ eq : pij ] ) in the frw space - time is equivalent to @xmath54+@xmath55=r . similarly as in @xcite , this expression can be rewritten more compactly + 3h - r=0 if we define @xmath56 as = h^2@xmath57 . now , we would like to study the self - acceleration solution with @xmath58const and @xmath59 . for the @xmath7 field we make the ansatz @xmath60 . furthermore , we assume that we are in a regime where @xmath61 so that we can neglect terms proportional to @xmath7 and consider only the terms including @xmath62 or @xmath63 . thus , the friedmann and field equations can be recast in h^2=(6q-9a_2 q^2 - 30a_3 q^3)[eq : modfrid ] + h^2(18a_2 q+54a_3 q^2 - 12)=0 . assuming @xmath64 , the field equation then imposes , q= while the friedmann equation ( [ eq : modfrid ] ) sets the hubble constant of the self - accelerated solution . similar to what has been found in @xcite our proxy theory admits a self - accelerated solution , with the hubble parameter set by the graviton mass . for the stability condition of this self - accelerating solution the first constraint we have is to demand @xmath65 . the other constrain comes from the stability condition for perturbations on the background which we discuss in the following subsection . in the last subsection we have shown explicitly that our proxy theory exhibits a self - accelerating solution with @xmath66 . now , we would like to study whether the perturbations on this background are stable and what the constraints are . for this purpose , consider perturbations on the background solution of the following form = _ 0(t)+(t , x , y , z)[eq : perturbations ] the second order action for the perturbations is l=-__g^-__^(0)_l^-__0_d_d_l^[eq : perlag ] , which can be written in the form l = k_tt(^2-()^2 ) where k_tt=-(a_2 + ) and c_s^2=13 ( 2 + ) . the condition for the stability is then given by @xmath67 , @xmath68 and @xmath69 , which are fulfilled if [ eq : stability ] a_2>0 0>a_3>-18 a_2 ^ 2 . to compare this result with the condition obtained in the decoupling limit , @xcite , we first mention that @xmath70 and @xmath71 . in terms of the parameters used there , we need to compare our conditions @xmath72 , @xmath73 to the conditions @xmath74 and @xmath75 . we see that our theory is less constraining but is still within the parameter space derived in @xcite . it is not surprising that the stability condition in the decoupling limit and in our covariantized theory do not coincide totally as we have explicitly broken the symmetry when getting the proxy and in particular , our solution does now spontaneously break lorentz invariance , which was not the case in the decoupling limit , @xcite . it is also worth pointing out that the self - accelerating solution by itself does not propagate any superluminal mode , since @xmath76 . we emphasize as well that the constant @xmath77 solution is a dynamical attractor . for this we just consider time dependent perturbations @xmath78 which is a special case of fulfilling the same stability conditions . the equation of motion for perturbations simplifies to _ t(a^3 ) = 0 the solution for @xmath79 is given by ( t)~a^-3 . thus , these perturbations @xmath80 redshift away exponentially compared to the @xmath81 self - accelerating solution . therefore , the self - accelerating solution is an attractor . more interestingly , one can wonder whether degravitation can be exhibited in these class of solutions . if one take @xmath82 and @xmath83 , it is straightforward to see that we obtain @xmath84 , so the field has absolutely no effect and can not help the background to degravitate . the situation is however different when setting @xmath85 , which was possible in the decoupling limit . in this proxy theory , such a behaviour will involve explicit space dependances in the equations of motion which should be dealt with specifically . interestingly , the interactions considered here are precisely of the same form as that studied recently in @xcite . there as well , in the absence of spatial curvature @xmath86 , the contribution from the scalar field vanishes if @xmath83 . comparing with @xcite , we can hence wonder whether the addition of spatial curvature @xmath87 in our proxy theory could help achieving degravitation , but relying strongly on spatial curvature brings concerns over instabilities which are beyond the scope of this study . in the following we would like to discuss in more detail the interplay of all the constituents of the universe . we assume that matter , radiation and the scalar field @xmath7 contribute to the total energy density of the universe . + ^mat ) consider the scalar field @xmath7 as a perfect fluid with the effective energy density and pressure given by ( [ eq : effdens ] , [ eq : effpress ] ) . thus , the equation of state parameter of this new field would be _ & = & + & & + . at this point one should mention that the energy density for the @xmath7-field is not conserved but rather given by @xmath49 ( where @xmath50 is the equation of motion for @xmath7 ) , which is not surprising since @xmath7 is non - minimally coupled to gravity in the jordan frame . therefore , we can have @xmath88 . + in the following we will first assume that at early times in the evolution history of the universe we can neglect the extra density coming from the helicity-0 @xmath89 . we will then check this assumption by plugging the solution for @xmath53 back in the equation of motion for @xmath7 . if we assume that at early times the radiation density dominates , we simply have h^2=_0^rada^-4 a~t^1/2=1/3[eq : raddom ] during the radiation era , the dominant terms in the equation of motion for @xmath7 are then @xmath90 which can be solved assuming the previous expression for @xmath53 ( [ eq : raddom ] ) _ rad~t^1.75 ^_rad~t^-1/4 , @xmath91 and @xmath89 during the epochs of radiation , matter and @xmath0-domination normalised to today @xmath89 . during the radiation domination the energy density for @xmath7 goes as @xmath92 and during matter dominations as @xmath93 and is constant for later times @xmath94.,scaledwidth=80.0% ] + at later times when the matter dominated epoch starts we have h^2=_0^mata^-3 a~t^2/3=0 now the dominant terms in the equation of motion for @xmath7 are @xmath95 . we get for @xmath7 this time ^mat~c_2t+ ^_mat = c_2 t^-1 + . summarizing , during radiation domination the effective energy density for the @xmath7-field goes like @xmath96 while during matter domination as @xmath97 and approaches a constant at late time . as shown in the figure @xmath98 can be neglected at early times where @xmath99 and @xmath100 . we end the cosmological analysis by looking at the evolution of matter density perturbations . the density perturbations follow the evolution _ m+2h_m= where @xmath41 is the newtonian potential . the effects of @xmath7 are all encoded in its contribution to the poisson equation . consider perturbations of @xmath7 around its cosmological background solution @xmath101 . @xmath102 gives a contribution to the newtonian potential of the form @xmath103 . in the newtonian approximation we have @xmath104 . the equation of motion for the scalar field in first order in @xmath102 is -g^ d_d_- l_(4_0^ d^d^+2r^__0^)=r which is equivalent to ( neglecting @xmath105 ) = r and last but not least we need the trace of einstein equation , ( [ eq.:einstein]-[eq : t3 ] ) . perturbing the trace to first order , we get -r&= & + & & + to reach that point , we have neglected the perturbations of the curvature of the form @xmath106 as they are negligible compared to @xmath107 since we work in the regime where @xmath108 . we have also ignored terms of the form @xmath109 , which could a priori be relevant , but their inclusion would require solving the full einstein equation , which is beyond the scope of this study . as a first approximation , such terms are hence ignored . the perturbations for the source is just given by @xmath110 for non - relativistic sources , thus we have = where @xmath111 stands for & & q1-(2 h _ 0+_0+(2 h + 3 h^2 ) ) + & & -@xmath112 . finally , the modified evolution equation for density perturbations is _ m+2h_m=(1 + 13q ) . knowing the background configuration it is then relatively straightforward to derive the effect on structure formation . we recover the usual result that when the field is screened , @xmath113 , the extra force coming from the helicity-0 is negligible and the formation of structure is similar as in @xmath0cdm . after having studied the cosmology and the structure formation of our proxy theory in the jordan frame , the natural question is whether we would expect similar results in a different frame . instead of covariantizing our lagrangian in the jordan frame , it is on an equal footing to go to the einstein frame where the ricci scalar is not multiplied by the scalar field @xmath7 and covariantize the theory at that stage . since it is unclear at first sight which frame is the physical frame we will consider both frames and discuss their differences . our starting lagrangian was l=-12h^e^_h_+h^x^(1)_+h^x^(2)_+h^x^(3)_+h^t _ now , when we do the following change of variables h_=|h_+_[eq : ef1 ] we can diagonalize the first mixed term @xmath37 such that the lagrangian takes the following form l=-12|h^e^_|h_+32+|h^x^(2)_-32()^2 + + |h^x^(3 ) _ -2()^2([^2]-^2 ) + ( |h_+_)t^ covariantizing this action is straightforward . we use again the correspondences in and it has been shown explicitly that the covariant equivalence to @xmath114 and @xmath115-\box\pi^2)$ ] are given by @xmath114 and @xmath116-\frac14(\partial\pi)^2r ) $ ] respectively , which do not yield any ghostlike instabilities ( @xcite , @xcite ) . thus , the covariantized action in the einstein frame is simply given by & = & ^2 r+32-__g^-32()^2 + & & - _ _ _ l^ + 2()^2(^2-[^2]-14()^2r ) [ eq : einstein ] + & & + l_m ( , ( 1+)g_)similarly as before , the properties of @xmath117 and @xmath43 ensure that their equations of motion lead at most to second order derivative terms . to find a self - accelerating solution we set again a pure de sitter metric , with @xmath118 . the friedmann and the field equations then take the form & & ^2h^2=3a_2 ^ 3q^2 + 10a_3 ^ 3q^3 - 12q^2 + 3a_2q^3 + 15a_3q^4 + & & ^3(-1 + 3q(a_2 + 4a_3q))+2(a_2 + 3a_3q)h^2=0[eq : frein ] when comparing the above friedmann and the field equations with the one we had in the jordan frame , we see significant differences coming from the extra terms which were not there in the jordan frame . these terms yield friedmann and field equations proportional to @xmath119 and @xmath120 which are more difficult to solve . + for fairness , we should compare both actions in the same frame . we do so by performing a conformal transformation on the action : g_=^2 g _ ^2=(1- ) . for simplicity we consider the case for which @xmath121 , so under this conformal transformation the covariantized action becomes _ j= ^2r-32 ^ -4()^2-(^^g_+32()^2 + 54()^4 ) in the limit where @xmath122 we have then finally the following expression _ j= ^2 r+32-__g^-32()^2 , which coincides with the theory obtained from the einstein frame , with @xmath121 . so within the regime of validity of our results , our conclusions are independent of the choice of frame . however beyond this regime the theory originally constructed from the jordan frame could violate the null energy condition from the term proportional to @xmath123 . common to generic massive gravity models is the presence of at least one additional helicity-0 degree of freedom which originates from a spin - two graviton field . in contrast to an arbitrary scalar field , its dynamics is not driven by a potential but rather by specific derivative terms fixed by symmetries . such a degree of freedom is experimentally testable . on solar system and galactic scales gravity is very well compatible with gr and correctional terms from a modified gravity model are excluded . on these scales , however , the effects of massive gravity might be cloaked by the vainshtein mechanism , where the helicity-0 mode interactions become appreciable to freeze out the field fluctuations . a number of screening mechanisms have been devised such as the chameleon and symmetron mechanisms . contrary to those the success of the vainshtein mechanism does not rely on some specific potential but instead on derivative interactions , which cause the helicity-0 mode to become decoupled from matter and light on short distances , which nevertheless could have observational signatures on larger scales in cosmic structure formation . a further benefit of the vainshtein mechanism would be a natural solution to the problem that the matter density and the cosmological constant are of similar magnitudes today . the cosmological application of the vainshtein mechanism works such that at early times , the galileon interactions are dominated by self - interaction , which suppresses their energy density relative to that of matter or radiation . if the matter density has decreased sufficiently by cosmic expansion , the galileon constitutes an important contribution to the energy density of the universe and drives cosmic expansion . although the precise cosmology of the theory of massive gravity proposed in @xcite has not been derived , most models of massive gravity give rise to alterations of the hubble function similar to dgp - gravity . experimental tests of the expansion dynamics of the universe include the distance - redshift relation of supernovae , and measurements of the angular diameter distance as a function of redshift , as in the case of the cosmic microwave background and the baryon acoustic oscillations . apart from these geometrical tests , the time - dependence of evolving cosmic structures can be investigated , and the influence of the gravitational theory on the geodesics of relativistic ( photons ) and nonrelativistic ( dark matter ) test particles . the first category includes gravitational lensing and the sachs - wolfe effects , which have been shown to differ from their gr - expectation in some modified gravity theories , and a similar result can be expected in massive gravity . the second category includes the homogeneous growth of the cosmic structure , and the formation of galaxies and clusters of galaxies by gravitational collapse . again , massive gravity would influence the time sequence of gravitational clustering and the evolution of peculiar velocities , as well as the number density of collapsed objects . in particular , massive gravity would enhance gravitational clustering since they tend to lower the collapse threshold for density fluctuations in the large - scale structure , leading to a higher comoving number density of galaxies and clusters of galaxies . naturally , these changes are degenerate with a different choice of cosmological parameters and with introducing non - gaussian initial conditions , which would be very interesting to quantify . recent discrepancies of @xmath0cdm with observational data on large scales include the number of very massive clusters , the strong lensing cross section , anomalous multipole moments of the cmb , the axis of evil , and large peculiar velocities . it is beyond the scope of this paper to address these issues using massive gravity , but we propose how to proceed from constructing a massive gravity theory to providing observationally testable quantities . from our point of view it is advisable to focus on probes of large scales , due to difficulties related to nonlinear structure formation and the influence of baryons on small scales . natural questions concern the homogeneous dynamics of the universe , the formation of structures and the shape of geodesics of relativistic and nonrelativistic particles . basically , using our proxy theory one should be able to make predictions concerning these four issues and the combination of the four should give insight into the nature of the gravitational sector . in our proxy theory , from the modified field equation the hubble - function can be derived easily , which allows the definition of distance measures , needed in the interpretation of supernova data . cosmic structure growth tests the newtonian limit for slowly - moving particles and describes the clustering of galaxies and the growth of structures which are investigated by e.g. gravitational light deflection . lensing , in turn , makes use of the geodesic equation for relativistic particles , and measures the correlation function of the matter density , weighted with the lensing efficiency function , which in turn is derived from distance measures . it would be quite interesting to study these observational consequences and constraints of our proxy theory in a future work . we would like to thank clare burrage , paul de fromont , gregory gabadadze , parvin moyassari and andrew tolley for useful discussion . furthermore , special thanks for robert caldwell ! this work is supported by the snf . g. r. dvali , g. gabadadze and m. porrati , phys . b * 485 * , 208 ( 2000 ) [ arxiv : hep - th/0005016 ] ; g. r. dvali and g. gabadadze , phys . d * 63 * , 065007 ( 2001 ) [ arxiv : hep - th/0008054 ] . a. i. vainshtein , phys . b * 39 * ( 1972 ) 393 . c. deffayet , g. r. dvali and g. gabadadze , phys . d * 65 * , 044023 ( 2002 ) [ arxiv : astro - ph/0105068 ] . k. koyama , phys . d * 72 * , 123511 ( 2005 ) [ arxiv : hep - th/0503191 ] . c. charmousis , r. gregory , n. kaloper , a. padilla , jhep * 0610 * , 066 ( 2006 ) . [ hep - th/0604086 ] . a. nicolis , r. rattazzi and e. trincherini , phys . d * 79 * , 064036 ( 2009 ) [ arxiv:0811.2197 [ hep - th ] ] . c. de rham and a. j. tolley , jcap * 1005 * , 015 ( 2010 ) [ arxiv:1003.5917 [ hep - th ] ] . f. p. silva and k. koyama , phys . d * 80 * , 121301 ( 2009 ) [ arxiv:0909.4538 [ astro-ph.co ] ] . n. chow and j. khoury , phys . d * 80 * , 024037 ( 2009 ) [ arxiv:0905.1325 [ hep - th ] ] . j. khoury , m. wyman , phys . rev . * d80 * , 064023 ( 2009 ) . [ arxiv:0903.1292 [ astro-ph.co ] ] ; t. kobayashi , h. tashiro and d. suzuki , phys . d * 81 * , 063513 ( 2010 ) [ arxiv:0912.4641 [ astro-ph.co ] ] ; t. kobayashi , arxiv:1003.3281 [ astro-ph.co ] ; m. wyman , j. khoury , phys . rev . * d82 * , 044032 ( 2010 ) . [ arxiv:1004.2046 [ astro-ph.co ] ] ; r. gannouji and m. sami , phys . d * 82 * , 024011 ( 2010 ) [ arxiv:1004.2808 [ gr - qc ] ] ; a. de felice , s. mukohyama and s. tsujikawa , phys . d * 82 * , 023524 ( 2010 ) [ arxiv:1006.0281 [ astro-ph.co ] ] . a. ali , r. gannouji and m. sami , phys . d * 82 * , 103015 ( 2010 ) [ arxiv:1008.1588 [ astro-ph.co ] ] ; a. de felice and s. tsujikawa , arxiv:1008.4236 [ hep - th ] ; a. de felice , r. kase and s. tsujikawa , phys . d * 83 * , 043515 ( 2011 ) [ arxiv:1011.6132 [ astro-ph.co ] ] ; k. hirano , z. komiya , h. shirai , [ arxiv:1103.6133 [ astro-ph.co ] ] ; k. hirano , [ arxiv:1012.5451 [ astro-ph.co ] ] ; p. brax , c. burrage , a. -c . davis , [ arxiv:1106.1573 [ hep - ph ] ] ; x. gao , [ arxiv:1106.0292 [ astro-ph.co ] ] . p. creminelli , a. nicolis and e. trincherini , jcap * 1011 * , 021 ( 2010 ) [ arxiv:1007.0027 [ hep - th ] ] ; l. p. levasseur , r. brandenberger , a. -c . davis , [ arxiv:1105.5649 [ astro-ph.co ] ] ; t. kobayashi , m. yamaguchi and j. yokoyama , phys . rev . lett . * 105 * , 231302 ( 2010 ) [ arxiv:1008.0603 [ hep - th ] ] ; t. kobayashi , m. yamaguchi and j. yokoyama , arxiv:1103.1740 [ hep - th ] . c. burrage , c. de rham , d. seery and a. j. tolley , jcap * 1101 * , 014 ( 2011 ) [ arxiv:1009.2497 [ hep - th ] ] ; p. creminelli , g. damico , m. musso , j. norena and e. trincherini , jcap * 1102 * , 006 ( 2011 ) [ arxiv:1011.3004 [ hep - th ] ] . s. mizuno , k. koyama , phys . rev . * d82 * , 103518 ( 2010 ) . [ arxiv:1009.0677 [ hep - th ] ] ; s. renaux - petel , [ arxiv:1105.6366 [ astro-ph.co ] ] . e. babichev , c. deffayet and r. ziour , int . j. mod . d * 18 * , 2147 ( 2009 ) [ arxiv:0905.2943 [ hep - th ] ] ; c. deffayet , o. pujolas , i. sawicki and a. vikman , jcap * 1010 * , 026 ( 2010 ) [ arxiv:1008.0048 [ hep - th ] ] ; r. kimura and k. yamamoto , arxiv:1011.2006 [ astro-ph.co ] . c. deffayet , g. esposito - farese and a. vikman , phys . d * 79 * , 084003 ( 2009 ) [ arxiv:0901.1314 [ hep - th ] ] ; c. deffayet , s. deser and g. esposito - farese , phys . d * 80 * , 064015 ( 2009 ) [ arxiv:0906.1967 [ gr - qc ] ] ; c. deffayet , x. gao , d. a. steer , g. zahariade , [ arxiv:1103.3260 [ hep - th ] ] . a. de felice and s. tsujikawa , phys . lett . * 105 * , 111301 ( 2010 ) [ arxiv:1007.2700 [ astro-ph.co ] ] ; s. nesseris , a. de felice and s. tsujikawa , phys . d * 82 * , 124054 ( 2010 ) [ arxiv:1010.0407 [ astro-ph.co ] ] . g. goon , k. hinterbichler , m. trodden , [ arxiv:1103.5745 [ hep - th ] ] ; g. goon , k. hinterbichler , m. trodden , [ arxiv:1103.6029 [ hep - th ] ] ; c. burrage , c. de rham , l. heisenberg , jcap * 1105 * , 025 ( 2011 ) . [ arxiv:1104.0155 [ hep - th ] ] . g. gabadadze , phys . b * 681 * , 89 ( 2009 ) [ arxiv:0908.1112 [ hep - th ] ] ; c. de rham , phys . lett . b * 688 * , 137 ( 2010 ) [ arxiv:0910.5474 [ hep - th ] ] . c. de rham and g. gabadadze , phys . b * 693 * ( 2010 ) 334 [ arxiv:1006.4367 [ hep - th ] ] . s. f. hassan , r. a. rosen , [ arxiv:1104.1373 [ hep - th ] ] . l. berezhiani , m. mirbabayi , [ arxiv:1104.5279 [ hep - th ] ] . g. dvali , g. gabadadze and m. shifman , _ published in * minneapolis 2002 , continuous advances in qcd * 566 - 581 _ ; arxiv : hep - th/0208096 ; g. dvali , g. gabadadze and m. shifman , phys . 044020 ( 2003 ) [ arxiv : hep - th/0202174 ] . n. arkani - hamed , s. dimopoulos , g. dvali and g. gabadadze , arxiv : hep - th/0209227 . g. dvali , s. hofmann and j. khoury , phys . d * 76 * , 084006 ( 2007 ) [ arxiv : hep - th/0703027 ] . g. gabadadze and m. shifman , phys . d * 69 * , 124032 ( 2004 ) [ arxiv : hep - th/0312289 ] . c. de rham , g. dvali , s. hofmann , j. khoury , o. pujolas , m. redi and a. j. tolley , phys . lett . * 100 * , 251603 ( 2008 ) [ arxiv:0711.2072 [ hep - th ] ] ; c. de rham , s. hofmann , j. khoury and a. j. tolley , jcap * 0802 * , 011 ( 2008 ) [ arxiv:0712.2821 [ hep - th ] ] ; c. de rham , j. khoury and a. j. tolley , phys . lett . * 103 * , 161601 ( 2009 ) [ arxiv:0907.0473 [ hep - th ] ] ; c. de rham , j. khoury and a. j. tolley , arxiv:1002.1075 [ hep - th ] .

the goal of cosmology is to describe the evolution history of the universe as a whole , tracing it as far back as possible from the present epoch . this seemingly daunting task is made tractable by the fact that the observed matter density in the universe appears to be approximately homogeneous when viewed on the largest scales . further , observations of the cosmic microwave background ( cmb ) radiation reveal it to be isotropic ( after subtracting a dipole contribution due to our local motion ) down to the level of one part in @xmath0 . the twin conditions of homogeneity and isotropy , if valid on all length scales , would immensely simplify the form of the metric and matter tensors used in the general relativistic description of cosmology . it can be shown @xcite that on purely geometric grounds , the most general metric describing a homogeneous and isotropic universe takes on the form of the friedmann - lema^ itre - robertson - walker ( flrw ) line element . the standard method of proceeding is to assume that the geometry of the universe can be described by a background metric with small perturbations , where the background is exactly homogeneous and isotropic and is hence of the flrw form . ( for a detailed discussion of the assumptions involved in the standard cosmology , see ref . @xcite . ) this approach is adopted not only for the early universe , for example in describing big bang nucleosynthesis and the dynamics of the cmb anisotropies , but is also assumed to be valid at recent epochs . in particular , the luminosity distances constructed to compare data from type ia supernovae ( sne ) with theory , assume a metric of the flrw form with its dynamics given by the einstein equations for a homogeneous and isotropic universe . we know , however , that the matter content of the universe today is _ not _ homogeneous on small scales - one finds inhomogeneity at the scale of individual galaxies for example ; and there is evidence @xcite that voids of order @xmath1 mpc is the conventional parameter which appears in the definition of the hubble constant @xmath2 kms@xmath3mpc@xmath3 . ] in diameter account for @xmath4-@xmath5 of the observed universe . there is also evidence for voids 3 - 5 times this size @xcite , as well as local voids on smaller scales @xcite . the evidence for homogeneity of the matter distribution in the universe only appears at scales larger than about @xmath6 mpc or so @xcite . it is then reasonable to assume that the metric describing our universe must also reflect the inhomogeneity of the matter on the corresponding scales . one may now ask whether it is at all justified to use the luminosity distances constructed using the flrw metric with its associated dynamics to compare theory with observational data from sne . in this paper we argue that the use of such luminosity distances is indeed justified , provided we are only interested in using such constructions in probing length scales larger than a certain minimum scale which is fixed by requiring that the assumption of homogeneity be valid ( a lower bound on such a scale could be @xmath6 mpc for example ) . more importantly , we propose that the _ dynamics _ of this ( effective or template ) macroscopic metric must be governed not by the standard einstein equations applied to the flrw metric , but by the effective equations obtained after spatial averaging on constant time slices , as developed by buchert @xcite . an outstanding open issue has been the following : how does one compare the predictions of the inhomogeneous , spatially averaged cosmology with observations , and should one be attempting an interpretation in terms of spatially averaged variables in the first place ? for instance , one could study light propagation and determine the luminosity distance versus redshift ( @xmath7 ) relation which best fits the observations , within an _ inhomogeneous _ model of the universe . following the important and interesting early work of dyer and roeder , where the angular - size and luminosity distances were obtained for beams propagating in `` lumpy '' flrw universes @xcite , detailed investigations have been carried out in @xcite . such a program has also received considerable attention in the context of the spherically symmetric inhomogeneous lematre - tolman - bondi ( ltb ) models @xcite . however , it is not clear to us that this is necessarily the best way to proceed , at least as far as distant sne are concerned . a realistic computation of light propagation in an inhomogeneous geometry seems difficult at best , and while the use of simplified toy models such as the ltb solutions in this endeavor may be justified on the grounds that we may inhabit a locally underdense region @xcite , it is not clear that the assumptions justifying the use of the toy model remain valid at large scales ( see also ref . @xcite for a recent review of the ltb approach ) . and after all , the concordance model infers the presence of a dark energy by comparing the predictions of the standard homogeneous and isotropic flrw cosmology against observations @xcite . if one decides to interpret the flrw cosmology as a template against which observations may be compared , then in order to decide whether cosmic inhomogeneities could be an alternative to the concordance model , one must construct a homogeneous and isotropic cosmology , after a suitable smoothing procedure , and compare its predictions with observations such as the @xmath7 relation for type ia sne . one should note that the effects of averaging within the standard perturbed flrw framework of cosmology , have been studied by several authors ( see , e.g. , refs . @xcite ) , and it has been argued that these effects must necessarily be small . the papers in refs . @xcite however , do not necessarily address the entire problem . for example , ref . @xcite deals with only super - horizon fluctuations , whereas work done by others ( see , e.g , ref . @xcite ) indicates that there may be nontrivial contributions due to averaging from _ _ sub__horizon fluctuations where the matter perturbations have gone nonlinear . and while the authors of ref . @xcite construct a cosmological model with negligible effects of averaging , there do exist toy models of inhomogeneous spacetimes in which these effects are large ( see , e.g. , ref . @xcite ) . furthermore , the possibility of _ nonperturbative _ effects of averaging has not , to our knowledge , been satisfactorily addressed , much less ruled out . it is important to realize that current @xmath8-body simulations for example , which are based on a newtonian approximation of gravity , consistently miss possible relativistic effects of averaging . this is because the modifications due to averaging in a newtonian setting can be shown to reduce to boundary terms in general @xcite , and to vanish in the case of spherical symmetry @xcite . these terms therefore do not contribute in simulations which either set periodic boundary conditions on the simulation box , or use a spherically symmetric simulation region . as we will see , the `` backreaction '' effects in buchert s framework ( i.e. , the modifications to the cosmological equations due to averaging ) are coupled to the average of the spatial curvature , this being a completely relativistic effect . nontrivial consequences of this coupling deserve more attention than they have received in the context of @xmath8-body simulations ( see ref . @xcite ) . we also note that an argument for describing the universe using flrw plus newtonian perturbations ( e.g. , as given in ref . @xcite ) , is based on the fact that it is newtonian perturbations _ and _ a negative pressure component , which describe the universe very well . in other words , one argues that the effect of averaging inhomogeneities must be small , but at the cost of adding an ill - understood dark energy component in the universe . on the other hand one should allow for the following possibility ( as of now neither confirmed nor ruled out ) : at an initial epoch , say at the end of inflation , the perturbations in the homogeneous metric would give a small `` backreaction '' as a result of averaging . as a result of its coupling with the curvature , this backreaction would grow and become significant around the time of structure formation , when matter inhomogeneities on small scales become nonlinear . the backreaction would also nontrivially affect the growth of the metric perturbations . such a possibility could only be tested within a fully nonperturbative treatment of the evolution of inhomogeneities . the goal of this paper is modest . we argue that the scale factor , defined in the spatially averaged cosmology via the volume of a domain , should correspond to an effective flrw metric , provided the size of the averaging domain coincides with the scale at which cosmological homogeneity arises . this allows us , in principle , to compare predictions of a spatially averaged cosmology with observations , in the standard manner , for instance by computing the @xmath7 relation . ( however , see ref . @xcite for an alternative approach to matching the spatially averaged dynamics with observations . ) the predictions of the spatially averaged cosmology would in general differ from standard flrw cosmology , because the scale - factor now obeys modified einstein equations . we will demonstrate that consistent models of spatially averaged flrw templates exist which can , at least in principle , explain observations such as those of the @xmath7 relation for supernovae . [ intro ] in a nutshell , our new results in this paper are that we propose a template metric which takes into account the kinematical effects of inhomogeneities in the universe . to implement this idea , we first suggest that in buchert s spatial averaging scheme , the size of the averaging domain should coincide with the cosmological scale at which homogeneity sets in . we then argue that , as a first step towards predicting the relation between inhomogeneities and observations , the expansion factor @xmath9 defined by buchert should be identified with the scale - factor @xmath10 in standard flrw cosmology . the template metric is then the flrw metric , with @xmath9 playing the role of @xmath10 . we then show , with the help of a few toy examples , that the luminosity distance - redshift relation in a model universe with an accelerating @xmath9 can be computed in the standard manner . to the best of our knowledge , such a template metric has not been proposed in earlier work by other researchers . at the length scales of cosmological interest , individual galaxies are small enough to be treated as point - like objects . given the large number ( @xmath11 ) of galaxies estimated to exist in the observed universe , it is also assumed that the matter content of the universe can be described to a good approximation as a continuum fluid . at the level of the matter distribution , this picture is comparable to that of a gas composed of a large number of molecules , wherein the internal structure of the molecules can be ignored when dealing with the macroscopic properties of the gas which is treated as a continuum fluid . further , the complex intermolecular interactions in the gas at the microscopic level can be dealt with in a mean field approximation by considering volume averages of the quantities of interest . this picture is made precise by introducing the concept of _ physically infinitesimal _ volumes @xcite or _ coarse - graining cells_. the fluid approximation is assumed to be valid provided the ( effectively ) infinitesimal volumes used in defining continuum quantities such as the matter density of the fluid , correspond to _ physical _ volumes which are large enough to contain a very large number of the particles constituting the fluid . at the same time , since these volumes are to be effectively infinitesimal , their physical size must also be much smaller than the scale of the fluid as a whole . we can then say that the length scales @xmath12 being probed by these physically infinitesimal volumes must lie in a certain range @xmath13 which is determined by the details of the fluid ( see also @xcite ) . clearly we must satisfy @xmath14 where @xmath15 is the size of the system . in the cosmological context , @xmath15 would be the hubble scale , and in order to deal with homogeneous and isotropic effective descriptions we would also have to satisfy @xmath16 where the homogeneity scale @xmath17 as mentioned earlier can be as large as @xmath6 mpc . the physical quantities we wish to deal with , such as the matter density , pressure , fluid velocity field , etc . are defined by suitably averaging or _ coarse graining _ over the physically infinitesimal volumes ( hence the terminology of coarse - graining cells ( cgcs ) ) . for example , the matter density of a given cgc is defined as the mass contained in it divided by its physical volume . the ( effectively ) continuum matter density is built up by repeating this for every cgc in the fluid . since we deal with the notion of a continuum fluid in cosmology , clearly we need a working description of coarse graining appropriate in this context . however , the underlying theory being general relativity , there are problems . the crucial issue is that along with the matter distribution ( which is now tensorial ) , the _ metric _ for the spacetime also needs to be coarse grained . as discussed by ellis @xcite , in defining a smoothing operation to go from the scale of stars ( scale 1 in ellis terminology , _ microscopic _ in ours ) to cosmological scales where homogeneity can be assumed to have set in ( scale 5 for ellis , _ macroscopic _ for us ) , one would employ two logically different smoothing operators , one each on the metric and the matter tensor . we denote the smoothing operator acting on the metric as @xmath18 and the one acting on the matter tensor as @xmath19 ( our notation differs from that used in ref . we assume that einstein s equations are valid on the microscopic scale of stars , which is reasonable since it is at these scales that solar system tests and binary pulsar studies probe general relativity . symbolically then ( see fig . [ fig1 ] ) , the metric @xmath20 at this scale is used to construct the einstein tensor @xmath21 which is equated to the matter tensor @xmath22 , giving einstein s equations @xmath23 . to obtain the corresponding description at the macroscopic scale , we use the smoothing operators to get a new metric tensor @xmath24 and a matter tensor @xmath25 . the macroscopic metric @xmath26 is used to construct the einstein tensor @xmath27 at the macroscopic scale . formally , one can think of the map @xmath18 as inducing a map @xmath28 which acts on the microscopic einstein tensor @xmath21 to yield @xmath29 . since the operators @xmath28 and @xmath19 are in general different , we have @xmath30 . hence , if the einstein equations are valid on a given length scale , then they are in general not valid on some other length scale connected to the first by appropriately defined smoothing operators . if we can re - express the new einstein tensor @xmath27 in terms of quantities obtained by acting the operator @xmath19 on _ both sides _ of the einstein equations at the microscopic scale , then it should be possible to write @xmath31 where @xmath32 represents corrections to the einstein equations arising from an explicit smoothing . in the standard approach , it is assumed that these corrections are small enough to be neglected . ( it should be emphasized that given the recent evidence for voids in the universe , one or more additional scales may eventually have to be included in the averaging scheme . ) added to all this is the complication that smoothing or averaging operations are notoriously difficult to define for tensor quantities in a gauge covariant manner . some examples of averaging techniques employed in general relativity ( and in cosmology in particular ) can be found in ref . @xcite . here we restrict ourselves to the formalism developed by buchert @xcite for volume averages of scalar quantities in a given choice of time - slicing . in the following sections we recall this formalism and argue that it is naturally adapted to the notion of the smoothing operator for the _ matter _ sector . we will reserve the term `` averaging '' for the matter smoothing defined in this way . as for the smoothing of the metric ( which we will refer to as `` coarse graining '' ) , we will take recourse to cosmological observations and argue that the result of such an operation on the inhomogeneous metric must yield the flrw metric on a suitable coarse graining scale . for simplicity we restrict our attention to a pressure - less fluid or `` dust '' , although a generalization to fluids with non - zero pressure is also possible ( see ref . @xcite ) . for a general spacetime containing irrotational dust , the metric can be written in synchronous and comoving gauge . ] , @xmath33 the expansion tensor @xmath34 is given by @xmath35 where the dot refers to a derivative with respect to ( proper ) time @xmath36 . the traceless symmetric shear tensor is defined as @xmath37 where @xmath38 is the expansion scalar . the einstein equations can be split @xcite into a set of scalar equations and a set of vector and traceless tensor equations . the scalar equations are the hamiltonian constraint and the evolution equation for @xmath39 , @xmath40 @xmath41 [ avg2 ] where @xmath42 is the ricci scalar of the 3-dimensional hypersurface of constant @xmath36 , @xmath43 is the rate of shear defined by @xmath44 and @xmath45 is the matter density of the dust . note that all quantities in eqns . generically depend on both position @xmath46 and time @xmath36 . eqns . and can be combined to give raychaudhuri s equation @xmath47 the continuity equation @xmath48 which gives the evolution of @xmath45 , is consistent with eqns . , . we only consider the scalar equations , since the spatial average of a scalar quantity can be defined in a gauge covariant manner within a given foliation of space - time . for the space - time described by , the spatial average of a scalar @xmath49 over a _ comoving _ domain @xmath50 at time @xmath36 is defined by @xmath51 where @xmath52 is the determinant of the 3-metric @xmath53 and @xmath54 is the volume of the comoving domain given by @xmath55 . spatial averaging is , by definition , not generally covariant . thus the choice of foliation is relevant , and should be motivated on physical grounds . in the context of cosmology , averaging over freely - falling observers is a natural choice , especially when one intends to compare the results with standard flrw cosmology . following the definition the following commutation relation then holds @xcite @xmath56 which yields for the expansion scalar @xmath39 @xmath57 introducing the dimensionless scale factor @xmath58 normalized by the volume of the domain @xmath50 at some initial time @xmath59 , we can average the scalar einstein equations , and the continuity equation to obtain @xcite @xmath60 @xmath61 @xmath62 [ avg7 ] here @xmath63 , the average of the spatial ricci scalar @xmath42 , is a domain dependent spatial constant . the ` kinematical backreaction ' @xmath64 is given by @xmath65 and is also a spatial constant over the domain @xmath50 . a necessary condition for to integrate to takes the form of the following differential equation involving @xmath64 and @xmath63 , @xmath66 [ avg ] the averager defined in eqn . is a natural candidate for the smoothing operator @xmath67 which must act on the matter tensor , although we have not yet specified the domain size . note that this averaging is defined on an inhomogeneous underlying geometry , i.e. it does not smoothen out the geometry itself . it remains to determine the operator @xmath18 that will coarse grain the metric . although we do not have an explicit definition for this operator , we can define a notional smoothing operator acting on the metric , by using physical inputs from the universe we observe . namely , we propose that since the universe has a matter content which is homogeneous on large scales , and since the cmb is highly isotropic around us , this notional operator must reflect these observations by acting on the general metric of eqn . to yield a flrw metric . crucially , the _ scale factor _ of this flrw metric must ( within our assumptions ) be given by buchert s functional @xmath68 , where the domain size to be used will be specified shortly . the argument leading to this proposal is as follows . we begin by making precise the notion of `` homogeneity on large scales '' , by requiring that in the range @xmath69 of cgc sizes allowed by the fluid approximation validity criterion , there exist a subset @xmath70 satisfying the following condition : when the microscopic matter density @xmath71 is averaged over a length scale @xmath72 , the resulting smoothed density _ must lose all information _ about the particular cgc over which the smoothing was performed . in other words , for @xmath72 , at any given time @xmath36 , the smoothed matter density must have the _ same _ value in all cgcs . it is in this sense that the smoothed matter density is homogeneous ( see however , the discussion at the end of this section concerning an explicit smoothing of the metric ) . this definition can be related to the concept of _ statistical _ homogeneity if we assume that the total volume of the observable universe is large enough that each cgc effectively samples one of an ensemble of configurations of matter distribution ( see ref . @xcite for a careful discussion of this point ) . it is possible that homogeneity sets in at a scale too large to be considered as the size of the cgcs , namely that the range @xmath73 is not contained in the range @xmath74 and that @xmath75 is not satisfied . we will assume that this does not happen ( for a more detailed discussion on this assumption , see sec . 2.1 of ref . we also make the crucial assumption that @xmath73 does not change with time . this will later allow us to consistently analyze light propagation using the effective flrw metric . this assumption needs some justification and we will return to this issue in the discussion . now define a coarse - graining cell to be an averaging domain @xmath50as used in buchert s formalism , with a characteristic comoving size @xmath72 . the averaged density is then @xmath76 , independent of the location of the particular cgc being considered . it also follows that the averaged expansion scalar @xmath77 is homogeneous in the above sense ( using eqn . ) and describes the evolution of the physical volume of any given cgc . thus , @xmath78 plays the role of the expansion scalar for a geometry whose infinitesimal spatial volume elements probe length scales of the size of our cgcs ( see however the discussion at the end of this section ) . it is therefore natural to demand that @xmath78 be the expansion scalar for our notionally smoothed out macroscopic metric . this is equivalent to demanding that the scale factor of this homogeneous and isotropic effective metric be identified with the functional @xmath68 ( up to a constant factor which we choose to be unity by an appropriate choice of units for the effective comoving coordinates ) . we are thus left with an effective metric whose line element can be written as @xmath79 which probes comoving length scales of size @xmath12 as chosen above . the 3-ricci scalar of this effective metric is given by @xmath80 and should be thought of as arising from the coarse graining of the geometry . following ellis ideas from sec . [ cgcpic ] then , @xmath81 is in general different from the averaged 3-ricci scalar @xmath63 which arises from applying the matter averager to the inhomogeneous scalar curvature ( @xmath82 ) . the constant @xmath83 has the usual interpretation as a parameter characterizing all possible homogeneous and isotropic 3-metrics . we can write @xmath84 in terms of @xmath85 by defining @xmath86 . so we are merely shifting focus from ( the _ a priori _ unknown ) @xmath84 to another unknown @xmath87 . as the notation suggests , we think of @xmath87 as being a correction ( a part of @xmath32 in the notation of sec . [ cgcpic ] , see eqns . below ) to the 3-ricci scalar of the flrw metric , although this correction need not be small compared to @xmath85 and need not evolve proportional to @xmath88 . we choose not to normalize the magnitude of the constant @xmath89 , retaining instead the normalization of the functional @xmath68 at the initial time @xmath59 . the evolution equations appropriate at this scale are buchert s equations and together with the integrability condition . for completeness , and in the spirit of ellis arguments we can rewrite equations and as follows @xmath90 [ smop2 ] where we have symbolically denoted the specific components of the general modified equations to which the various terms of buchert s equations belong . the integrability condition which supplements these equations is purely a consequence of having non - zero corrections @xmath32 to the standard flrw equations . as a quick check , we note that requiring the corrections @xmath32 to vanish implies both @xmath91 ( from eqn . ) and @xmath92 , i.e. @xmath93 ( from eqn . ) ; the integrability condition is identically satisfied and we recover the usual flrw solution . in the arguments leading to eqns . , we have ignored the effects of a virtual _ change _ in the averaging length scale @xmath12 . considerable effort has been spent by buchert and carfora @xcite in studying such effects . in particular , in the spirit of the real space renormalization group formalism , these authors derive a novel _ curvature backreaction _ effect analogous to the kinematical backreaction @xmath94 , together with a volume scaling effect . the inclusion of these effects would change the evolution equations satisfied by the scale factor @xmath95 . we also note that buchert and carfora s techniques ( involving ricci flow of the inhomogeneous geometry into a smooth geometry ) form one candidate for an explicit realization of the metric coarse graining operator @xmath18 . seen in this light , it is apparent that by accounting for the above effects , we would also have to appropriately modify our notion of large scale homogeneity ( to include the volume scaling effect and the residual information from the curvature fluctuations ) . in the arguments leading to the effective metric , the scale factor @xmath95 would be replaced by an appropriately `` dressed '' quantity characterizing the evolution of effectively infinitesimal volumes in the smoothed geometry . however , the main arguments leading to an effective metric which is flrw , will not be altered given the existence of the range @xmath96 . for simplicity , we ignore these additional effects in this paper , and hope to investigate them in future work . it should be noted that a notion of coarse - graining is always implicit in standard flrw cosmology . it is assumed that a coarse - graining ( while satisfying the conditions of the fluid approximation ) is possible on a length scale which coincides with the scale at which homogeneity is achieved ( in other words , the existence of @xmath96 is assumed ) . standard cosmology however ignores the possible modifications to flrw cosmology brought about by the process of explicitly smoothing over the inhomogeneities - an issue which buchert s approach addresses . our key proposal in this paper is that the size of the averaging domain should be taken to coincide with the scale of homogeneity , and that ( up to some corrections ) the quantity @xmath68 corresponds to the scale factor of a template flrw metric . we now illustrate how our proposal can facilitate the comparison of the predictions of the spatially averaged cosmology with observations , for example by computing the @xmath7 relation using the scale factor @xmath68 , which has a modified time evolution compared to the corresponding scale factor in standard cosmology . in our approach , we take the effective or template metric to be flrw , and since we are holding the size @xmath12 of the averaging domain fixed , there is nothing conceptually new to be done while studying its null geodesics . the important difference , of course , is that the evolution of the scale factor is now given by the modified friedmann equations ( subject to the caveats listed at the end of the previous section ) . we note that the use of the flrw metric is , strictly speaking `` unrealistic '' in the sense that in the real universe , light travels mostly in vacuum which has a non - zero weyl tensor whereas in the flrw cosmologies which are conformally flat , light travels through a geometry with vanishing weyl tensor ( see also refs . @xcite and @xcite ) . this issue is clearly relevant in the standard approach as well . while we have not addressed this issue in the context of averaging , we emphasize that our approach must be thought of as a first step in gauging the effects of inhomogeneities , by using an _ explicitly _ averaged construction of the template flrw model to be used . the arguments relating the redshift of a source to the scale factor @xmath95 and to its comoving distance @xmath97 from the observer , are exactly the same as in the standard approach . in particular , we have for the redshift @xmath98 of a source that emits light at time @xmath36 which is received by an observer at time will indicate the value of the quantity at the present epoch @xmath99 , and the subscript @xmath100 , the value at the initial epoch @xmath59 . ] @xmath99 , @xmath101 and the comoving distance @xmath97 to the source is determined by solving the equation @xmath102 a caveat to the above relations is that the `` step - size '' used in computing the integrals involved must be understood to be no smaller than the size @xmath12 of the cgcs . further , the entire construction must be interpreted as a `` fitting template '' that allows us to compare observations with theoretical predictions in a straight forward manner . with our choice of conventions , the magnitude of @xmath89 is a dynamically determined parameter of the model and determines the form of the luminosity distance @xmath7 as follows . solving eqn . for @xmath103 gives @xmath104 where @xmath105 is the hubble parameter and the function @xmath106is defined as @xmath107 in order to complete the picture we need the functional dependence of the hubble parameter @xmath108 on the scale factor @xmath95 . this requirement is complicated by the fact that the system of equations , , and is only consistent , it does not close . in order to obtain the required relation then , it is necessary to make certain assumptions about the evolution of the kinematical backreaction @xmath64 and the averaged 3-ricci scalar @xmath84 . _ @xcite have analyzed a class of scaling solutions of the form @xmath109 the case @xmath110 , @xmath111 is the only scaling solution with @xmath112 . for the cases @xmath113 , the backreaction is proportional to the averaged 3-curvature should not be confused with the comoving coordinate in the effective metric . ] , @xmath114 the forms of the integrability condition and the solutions indicate that the scaling solutions can be superposed to obtain new solutions . we will use two such superpositions to construct models which contain an accelerating scale factor @xmath95 and which also yield analytic expressions for the luminosity distance @xmath7 . the solutions we consider are of the form @xmath115 in keeping with our earlier discussion of the quantity @xmath87 . namely , we model @xmath87 by a single scaling solution . here @xmath97 and @xmath116 are related as in eqn . , and @xmath89 and @xmath117 are constants with dimensions of ( length)@xmath118 . we are interested in models which admit an accelerating scale factor @xmath95 , and will hence assume @xmath119 and @xmath120 ( the reason for which will be apparent shortly ) . inserting the relations in eqns . and we find @xmath121 @xmath122 although both models considered above must be regarded as unphysical ( since the luminosity distance of a source is a measurable quantity ) , consistent models also exist . for example , the case @xmath156 , @xmath157 corresponding to a constant backreaction , mimics the cosmological constant of the standard cosmology . further , it may be argued that current observations seem to indicate that we live in a universe with @xmath158 and hence the above examples are not observationally relevant in any case . to show that the formalism does admit consistent ( accelerating ) models with @xmath158 as well , we now give an example of such a model . + + _ case 3 _ : @xmath159 , @xmath158 . in this case , although the _ effective _ spatial curvature @xmath160 is zero , the average of the _ physical _ spatial curvature @xmath63 _ does not _ vanish , and is given by ( see eqn . ) @xmath161 . if we take @xmath119 , which guarantees that the scale factor will accelerate for @xmath162 ( see eqn . ) , then we have @xmath163 . the quantity @xmath136 is now given by @xmath164 since @xmath165 and @xmath166 are positive , this integral is well behaved for all @xmath98 . further , since @xmath158 , we have @xmath167 . defining @xmath168 and noting that @xmath169 and @xmath170 , we get @xmath171^{1/2 } } } \ , . \label{lumin16 - 2}\ ] ] in this notation , which is reminiscent of that in the standard lcdm model , if we use eqn . evaluated at present time we also see that @xmath172 , and that acceleration begins at a redshift given by @xmath173 as an example , in fig . [ fig2 ] we have plotted the behaviour of @xmath174 in units of @xmath175 in this model ( for redshifts between @xmath176 and @xmath177 ) , setting @xmath178 , @xmath179 . in this case the acceleration redshift is @xmath180 . for comparison we also show the @xmath7 curve in the standard lcdm model with @xmath181 and @xmath182 , for which the acceleration redshift is @xmath183 . [ interestingly , a model in which the backreaction @xmath64 behaves as @xmath184 at early times , has recently been studied by li and schwarz @xcite in the context of spatial averaging of a perturbed flrw cosmological model . ] + + we have therefore demonstrated the existence of models of the backreaction which can admit an accelerating scale factor . it is important to keep in mind though , that the models considered above were not physically well - motivated , they served only to demonstrate the construction of luminosity distances using the modified dynamics . a more realistic scenario would probably contain a backreaction which becomes significant only at late times , when structure formation has advanced sufficiently far . it will be an interesting exercise to numerically construct luminosity distances for such models and compare them with those of the concordance model . we have argued that for the purposes of cosmological study , an explicit coarse graining of the matter density in the universe must be performed using the formalism of spatial averaging developed by buchert . due to the observed large scale homogeneity of the matter density and the isotropy of the cmb radiation , the notionally smoothed out template metric obtained after this coarse graining must be described by a flrw line element , with the scale factor being given by buchert s functional @xmath68 . the procedure for constructing luminosity distances is therefore identical to that in the standard approach , except that the _ evolution _ of the effective scale factor must be given by the modified flrw equations supplemented by the integrability condition . ( the scale factor and its evolution will be further modified by including perfect fluid sources with non - zero pressure @xcite and the curvature backreaction and volume effects @xcite . ) such a construction of luminosity distances must be interpreted as a convenient way of comparing observations with theoretical models , while retaining the essence of the physical inhomogeneous geometry . it will be an interesting project to compare such luminosity distances in realistic backreaction scenarios with current supernovae observations in an attempt to constrain the magnitude and evolution of the backreaction . we end by discussing some additional issues concerning an explicit smoothing of the metric which we avoided in the present treatment , and also some potential problems with our construction . firstly , we return to the issue of holding the comoving range @xmath73 , and in particular the smoothing scale @xmath12 , constant . this is justified only so long as the condition @xmath185 is satisfied , where @xmath186 is the appropriate comoving scale associated with the observable universe at redshift @xmath98 . for example , if we choose a scale @xmath12 which satisfies this condition at the present epoch , the same scale will not in general be valid at the last scattering epoch of the cmb . however , it seems reasonable that for a small enough redshift range , the condition @xmath187 can be satisfied with a single constant @xmath12 . for epochs sufficiently far back in the past , the scale @xmath12 would have to be reduced , and the construction of the @xmath7 relation would no longer be valid . further , a change in the size of the averaging domain would in general affect both the scale factor @xmath95 and the `` constant '' @xmath89 in a non - trivial way . in particular , any change in @xmath89 would render the effective flrw metric meaningless . also , a change in @xmath89 need not necessarily be caused by a change in the averaging domain size . it seems likely @xcite that an explicit smoothing of the geometry using buchert and carfora s technique will yield a 3-space of different spatially constant curvatures at different times , this difference being independent of any change in the averaging domain size with time and being possibly incommensurate with our assumption of @xmath188 . this could be interpreted in terms of a change in @xmath89 with time , and in such a case , one would be dealing with a _ different _ flrw template at different times . the effects of such a construction are not clear at present . clearly this issue needs more careful study , and we hope to address this in the future . of course , if @xmath89 is made a function of time , care must be taken that the matter content is now so chosen that einstein s equations continue to be satisfied , as pointed out by rsnen @xcite in response to @xcite . also , as pointed out by rsnen @xcite , the use of a homogeneous and isotropic flrw template will necessarily ignore physical effects such as a non - trivial shear in the underlying geometry . a potential problem with our template construction is that effects of the non - trivial scalar curvature @xmath189 on light propagation are also simplified to a proportionality to @xmath88 . although the full scalar curvature will show up in the evolution of @xmath68 via its spatial average @xmath84 , it is possible that this construction may inaccurately model the real universe . a study of the evolution of the effective spatial curvature as mentioned in the preceding paragraph may help resolve this issue . there is also the related issue of degeneracy , namely that several underlying inhomogeneous models may reproduce the same backreaction effects when averaged . for example , it was shown @xcite in the context of the ltb solutions , that models with a vanishing spatial curvature ( the so called marginally bound models ) in the inhomogeneous geometry also have a vanishing backreaction @xmath94 after averaging . one is now faced with the situation wherein interpreting say luminosity distance observations within the inhomogeneous geometry may distinguish two inhomogeneous models , but interpreting the data _ after _ averaging would not make this distinction and would hence perhaps rule out _ both _ models . this issue has been recently highlighted by enqvist and mattsson ( the last paper in ref . @xcite ) who show that a marginally bound ltb model can fit the data from type ia sne without any spatial averaging . since the standard flrw matter dominated model is also ( a special , homogeneous case of ) a marginal ltb model which _ does nt _ fit the sne data @xcite , the formalism after spatial averaging would rule out both models by demanding a non - trivial backreaction instead . however , since the ltb ( and in fact any inhomogeneous ) models themselves face a similar degeneracy issue ( for example several different ltb models can currently claim to fit the sne data @xcite ) , this matter can not be considered as settled , and also deserves further attention . lastly , we note that buchert s averaging scheme itself has been criticized in the literature ( see , e.g. , ref . @xcite ) on the grounds that the effects of averaging in a noncovariant manner as done in this scheme , are likely to be gauge artifacts . it should be noted that within second - order perturbation theory , gauge invariance of the averaged equations has been demonstrated in @xcite and @xcite . it is also possible to approach the problem of determining modifications to the standard cosmological equations via a covariant method such as zalaletdinov s macroscopic gravity @xcite for example . such an attempt has in fact been made recently and has led to effective cosmological equations very similar to those of buchert ( see ref . the proposal in the present paper however , is attractive to us due to its simplicity in implementation , and may be treated as a phenomenological model of the more rigorous equations derived in ref . [ discuss ] we have greatly benefited from our correspondence with thomas buchert and syksy rsnen , and it is a pleasure to acknowledge their insightful comments and criticism of an earlier draft . ap also thanks rakesh tibrewala for useful and stimulating discussions . 1 s weinberg , _ gravitation and cosmology _ , john wiley and sons , inc . , 1972 ; + t padmanabhan , _ theoretical astrophysics , vol.iii_ , cambridge univ . press , 2002 ; + m trodden and s carroll , _ tasi lectures , introduction to cosmology _ , [ astro - ph/0401547 ] , ( 2004 ) . f hoyle and m s vogeley , _ astrophys . j. _ , * 566 * , 641 , ( 2002 ) [ astro - ph/0109357 ] ; _ astrophys . j. _ , * 607 * , 751 , ( 2004 ) [ astro - ph/0312533 ] ; + m s vogeley _ et al . _ , in _ proceedings iau colloquium no . 195 _ , ( 2004 ) , a. diaferio , ed . + s g patiri , _ et al . _ , _ mon . not _ , * 369 * , 335 , ( 2006 ) [ astro - ph/0506668 ] . _ , _ astrophys . * 642 * , 54 , 2005 [ astro - ph/0411197 ] ; + pietronero l and labini f s , _ proc . _ * 822 * , 294 , 2006 [ astro - ph/0406202 ] ; + joyce m , _ et al . _ , _ astron . & _ , * 443 * , 11 , 2005 [ astro - ph/0501583 ] ; + kerscher m , schmalzing j , buchert t and wagner h , _ astron . & _ , * 333 * , 1 , 1998 [ astro - ph/9704028 ] ; + kerscher m , _ et al . _ , _ astron . & . _ , * 373 * , 1 , 2001 [ astro - ph/0101238 ] ; + hikage c _ et al . _ ( sdss collaboration ) , _ publ . _ , * 55 * , 911 , 2003 [ astro - ph/0304455 ] . + see also ref . @xcite . c c dyer and r c roeder , _ astrophys . j. _ , * 189 * , 167 , ( 1974 ) ; + r c roeder , _ astrophys . j. _ , * 196 * , 671 , ( 1975 ) ; + m sasaki , _ mon . not . _ , * 228 * , 653 , ( 1987 ) . + an attempt to use this formalism in the spatial averaging context has been made by + d palle , _ nuovo cim . _ , * 117b * , 687 , ( 2002 ) [ astro - ph/0205462 ] . n mustapha , c hellaby , and g f r ellis , _ mon . not r. astron . _ , * 292 * , 817 , ( 1997 ) [ gr - qc/9808079 ] ; + m n celerir , _ astron . and astrophys . _ , * 353 * , 63 , ( 2000 ) [ astro - ph/9907206 ] ; + h iguchi , t nakamura and k nakao , _ prog . of theo _ , * 108 * , 809 , ( 2002 ) [ astro - ph/0112419 ] ; + j w moffat , _ jcap _ , * 10 * 012 , ( 2005 ) [ astro - ph/0502110 ] ; + h alnes , m amazguioui and o gron , _ phys . d _ * 73 * , 083519 , ( 2006 ) [ astro - ph/0512006 ] ; + r mansouri , [ astro - ph/0512605 ] , ( 2005 ) ; + r a vanderveld , e e flanagan and i wasserman , _ phys . d _ * 74 * , 023506 , ( 2006 ) [ astro - ph/0602476 ] ; + d garfinkle , _ class . * 23 * , 4811 , ( 2006 ) [ gr - qc/0605088 ] ; + k enqvist and t mattsson , [ astro - ph/0609120 ] , ( 2006 ) . t futamase , _ phys . _ , * 61 * , 2175 , ( 1988 ) ; + m kasai , _ phys . _ , * 69 * , 2330 , ( 1992 ) ; + t w noonan , _ gen . rel . and grav . _ , * 16 * , 1103 , ( 1984 ) ; + r m zalaletdinov , _ gen . rel . and grav . _ , * 24 * , 1015 , ( 1992 ) ; _ gen . rel . and grav . _ , * 25 * , 673 , ( 1993 ) ; + j p boersma _ phys . _ * d57 * , 798 , ( 1998 ) [ gr - qc/9711057 ] ; + a a coley , n pelavas and r m zalaletdinov , _ phys . rev _ , * 95 * , 151102 , ( 2005 ) [ gr - qc/0504115 ] ; a a coley and n pelavas , _ phys . rev . _ * d74 * , 087301 , ( 2006 ) [ astro - ph/0606535 ] . + several other references and a discussion on the issue of averaging in cosmology can be found in - a krasiski , _ inhomogeneous cosmological models _ , cambridge univ . press , 1997 . t buchert and m carfora , _ class . _ , * 19 * , 6109 , ( 2002 ) [ gr - qc/0210037 ] ; + t buchert and m carfora , _ phys . rev . _ , * 90 * , 031101 ( 2003 ) [ gr - qc/0210045 ] ; + t buchert and m carfora , _ proc . of 12th workshop on general relativity and gravitation ( jgrg 12 ) , tokyo , japan _ , ( 2002 ) , [ astro - ph/0312621 ] ; + m carfora and k piotrkowska , _ phys . d _ , * 52 * , 4393 , ( 1995 ) [ gr - qc/9502021 ] ;

the management of heat in dielectric beams and membranes is an important part of the design in micro - electro - mechanical systems ( mems ) @xcite . while the physics of thermal transport in dielectric materials is well understood , the effect of fabrication processes on the surface physics is less clear and widely treated as a hidden variable in the evaluation of the device performance . this uncertainty can prolong the design - test cycle , where a parameter such as beam geometry is iterated until the target thermal conductance @xmath3 is obtained @xcite . in our application , the mems is a cryogenic detector known as the transition - edge sensor ( tes ) @xcite . figure [ fig : q - band - detector ] shows a tes designed for measurements of the polarization state of the cosmic microwave background @xcite . the tes is comprised of a superconducting moau bilayer deposited on a single - crystal silicon substrate . the silicon is etched around the bilayer and other electrical components to form a membrane that is supported by dielectric beams . the total thermal conductance @xmath4 of the beams is a critical parameter that determines the sensitivity ( @xmath5 ) , response time ( @xmath6 ) , and saturation power of the tes detector ( @xmath7 ) ( saturation energy in the case of a tes micro - calorimeter ) @xcite . the response time is also a function of the total heat capacity @xmath8 of the detector , which is the sum of the heat capacity of the silicon membrane and normal metal films . the thermal conductance from the membrane to the frame is effectively set by the ballistic - dominated thermal conductance of the short beam ( magnified ) . the long silicon beams are @xmath10 @xmath1m@xmath9 , and merely support tes bias and rf leads . the thermal conductance of the long silicon beams is diffusive - ballistic and sub - dominant to the conductance of the ballistic beam . the pd layer deposited on the membrane sets the detector heat capacity . , width=321 ] at temperatures far below the debye temperature of the lattice ( @xmath11 k ) , anharmonic phonon collisions in the bulk of the dielectric are negligible , and the phonon - mean - free path @xmath0 is restricted by the cross sectional dimensions of the beam @xcite . in state - of - the - art far - infrared and millimeter tes detectors operating at sub - kelvin temperatures @xcite , the beam length @xmath12 is much larger than @xmath0 , and the thermal conductance is often described by the fourier law for phonon diffusion , @xmath13 , where @xmath14 and @xmath15 are the width and thickness of the beam , and @xmath16 is the bulk thermal conductivity of the dielectric . @xmath16 is determined empirically with the fabrication and testing of numerous beams of varying width and length , and the aspect ratio @xmath17 is tuned to achieve the desired conductance . however , this approach rarely provides a complete view of the thermal physics , and higher order effects such as surface roughness of the beams have been identified as the likely cause of the reported non - uniformity in @xmath3 across detector arrays @xcite . a simple analysis can readily show the sensitivity of phonon propagation in long beams ( @xmath18 ) to the details of the surface physics . for a random gaussian surface , the roughness can be characterized by the ratio of the standard deviation @xmath19 of the surface features to the peak thermal wavelength @xmath20 of the incident phonon field . when @xmath21 , phonons are reflected specularly from the surface @xcite . as the wavelength becomes much smaller than the scale size of the surface , @xmath22 , phonons are scattered over all angles , or diffused . assuming isotropic diffusion , it can be shown that for a rectangular beam , the phonon mean - free - path @xmath0 scales logarithmically with @xmath23 ( @xmath24 for @xmath25 @xcite ) . hence , in a beam where @xmath18 , phonons experience a large number of scattering events , @xmath26 , and the heat flux entering the beam at one end @xmath27 is reduced to @xmath28 at the exit port , where @xmath29 is the diffuse fraction . the derivative @xmath30 is a measure of the sensitivity of @xmath3 to changes in @xmath31 and proportional to the power - law function @xmath32 . in practice , the statistics of the surface roughness can be heavily biased by the fabrication and is likely non - gaussian . dielectric beams can be roughened in various forms depending on the compatibility and properties of the etchants with the dielectric material . preliminary tests can shed light on the roughening steps , however , fabrication processes can be difficult to control in practice . details such as base pressure and process chamber conditioning during dry - etch steps can vary between batches of devices , especially in shared facilities . as a result , the characteristics of the roughness of a single surface can change significantly over spatial scales shorter than the beam length , and between wafers . furthermore , not all surfaces of a beam are rough to the same degree . these effects add to the unpredictable variability in @xmath31 and therefore @xmath3 in beams dominated by diffusive thermal transport . the work in this paper is focused on understanding the effects of fabrication on the thermal conductance in single - crystal silicon beams . the beam geometries explored span several orders of magnitude in @xmath33 . the results demonstrate that @xmath3 in short ( @xmath34 @xmath1 m , @xmath35 1 ) uniform beams with ballistic - dominated phonon transport is insensitive to the fabrication conditions and can be realized with precision in devices on different wafers . this approach is suitable for tes detector arrays in ground - based and air - borne telescopes , where the detector noise equivalent power requirements is greater than 10@xmath36 w/@xmath37 . throughout this paper , the terms _ diffuse _ and _ ballistic _ are used to describe the scale size of the phonon transport in a silicon beam . in the presence of surface roughness , the thermal transport in a long beam with @xmath38 is dominated by diffuse reflections off the beam walls , and @xmath39 . for a short beam , ballistic transport dominates and @xmath40 . we emphasize that these limits bound the diffusive - ballistic conductance observed in practice . in sec . [ sec : method ] , we briefly describe the device fabrication and readout hardware employed for the low - temperature measurements of @xmath3 and @xmath41 . in sec . [ sec : model ] , we describe the theory of radiative phonon transport in diffusive - ballistic beams . in sec . [ sec : diff - ball ] , we present results that illustrate the effect of fabrication on the diffusive - ballistic nature of the thermal conductance . in sec . [ sec : controlg ] , we describe the superior performance of ballistic beams for the control of @xmath3 of tes detectors . finally , we demonstrate that because of the low specific heat capacity of single - crystal silicon , the heat capacity of the detectors fabricated on this substrate can be effectively controlled using a normal metal film with high specific heat . the fabrication of the tes detectors reported in this paper is described in detail in ref . @xcite , but for completeness , we provide a brief description here . fabrication starts with a float - zone - refined p - type ( 100 ) boron doped silicon - on - insulator wafer . heat flow is primarily in the [ 110 ] direction . the single - crystal silicon layer is 5 @xmath1 m thick , with a room temperature resistivity of 5 k@xmath42-cm and a carrier density of 5@xmath4310@xmath44 @xmath45 . a moau bilayer is deposited by a combination of dc plasma sputter deposition and electron beam evaporation . the mo and au are photolithographically patterned to define the superconducting sensor . the au layer is etched by ion milling and the mo layer is etched by reactive ion etching ( rie ) . further metallic components are deposited on the device layer to enable the required operation of the pixel at rf frequencies , and to enable on - chip diagnostic capabilities such johnson noise thermometry as needed . au resistors are deposited by electron beam evaporation and patterned by lift - off . following an in - situ rf ar plasma cleaning of the wafer , a nb microstrip layer is sputter deposited , photolithographically patterned , and etched in fluorine - based rie chemistry . the silicon membrane is defined and reactive ion etched down to the buried oxide layer . the silicon wafer is then bonded to a pyrex handle wafer using a sacrificial wax . the backside of the silicon wafer is patterned by photolithography and the bulk silicon is etched by inductively - coupled plasma deep reactive - ion etching ( drie ) . the wax is then dissolved in solvent to release the devices . measurements were performed in a mechanically - cooled system that reaches 3 k at the coldest stage . further cooling of the detector stage down to 100 mk was achieved using an adiabatic demagnetization refrigerator . the devices were mounted and thermally anchored to a gold - plated copper package , and wirebonded to a superconducting quantum interference device ( squid ) cold stage amplifier for johnson noise current measurements . further amplification was achieved with a squid series array mounted at the 3 k stage . the temperature of the 100 mk cold stage was read with a calibrated ruo@xmath46 thermometer @xcite with calibration traceable to the national institute of standards and technology . the precision of the temperature control was 32 @xmath1k rms at 150 mk . the room temperature electronics was comprised of a custom low - noise pre - amplifier , feedback electronics for flux - locked - loop operation , commercial low - noise high - gain amplifier and low - pass filter ( srs-650 ) , and commercial 16 bit digital - to - analog electronics ( ni pci-6120 ) for digitization of the amplified feedback signal . the method of johnson noise ( jn ) thermometry is frequently utilized for the characterization of the thermal properties of thin - films below 1 k @xcite . at audio frequencies ( 0.5 - 20 khz ) and sub - kelvin temperatures , the current noise spectral density of a resistor @xmath47 in series with the squid input coil takes the form @xmath48^{-1}/r$ ] , where @xmath49 is the electron temperature , @xmath50 is the characteristic frequency of the circuit , and @xmath12 is the total inductance of the input coil and stray inductance in the input circuit loop . this equation directly relates measurements of @xmath51 to @xmath49 . for the test devices that are described in ref . @xcite , the jn thermometer and heater are 240@xmath43230@xmath430.27 @xmath1m@xmath9 au resistors deposited on top of interdigitated nb leads . the kapitza conductance at the au - silicon interface and the electron - phonon conductance in the au are approximately 10 nw / k each at 100 mk @xcite . these conductances are a factor of 50 greater than the highest silicon beam conductance measured in our devices . we therefore assume the electron system in the thermometer is in thermal equilibrium with the phonons in the silicon membrane . the maximum expected temperature gradient across the membranes is less 0.5 mk across the 0.1 - 1 k temperature range . the random error in a temperature measurement from the jn current can be estimated from the radiometer equation @xmath52 @xcite , where @xmath53 khz is the bandwidth over which the current noise power spectral density was measured and @xmath54 s is the measurement time . hence , the fractional error throughout the temperature range of our measurements is less than 0.5% . measurements of the jn current as a function of bath temperature were used to fully calibrate the cryogenic and warm readout amplifiers . typical readout noise of 15 pa/@xmath37 referenced to the input circuit was achieved . the jn signal was 30 - 100 pa/@xmath37 in the range 0.1 - 1 k. to measure @xmath3 of the silicon beams , a small signal dc heater power was applied to the membrane and the increase in temperature was recorded . the relationship between heater power and membrane temperature at a given bath temperature @xmath55 can be parameterized by the power law function @xmath56 , where @xmath57 is the heater power , @xmath58 is the membrane temperature , @xmath59 is a constant that is proportional to the number of phonon modes , and @xmath60 is the temperature exponent that is indicative of the temperature dependence of the phonon mean - free - path . the first order derivative yields the thermal conductance @xmath61 . in our measurements , the change in temperature was small , @xmath6210% , and the approximation @xmath63 was valid . we have also utilized tes current - voltage curves to determine the thermal conductance of detectors in which normal metal films for jn thermometry were omitted . when a tes is voltage - biased on the normal - to - superconducting phase transition , the silicon membrane temperature is fixed at the critical temperature of the moau bilayer , @xmath64 mk . the thermal conductance is then determined from a fit to the joule power dissipated in the tes as a function of bath temperature . in this case , @xmath65 and @xmath66 , where the parameters @xmath59 and @xmath60 are equivalent to those derived from jn thermometry in the same temperature range . in building a model of heat flow in a silicon beam , it is important to first identify the thermal baths that define the boundary conditions in the model . a thermal bath is ideally a blackbody that provides a matched termination for phonons at all wavelengths . in a detector such as that shown in fig . [ fig : q - band - detector ] , each silicon beam is connected to two thermal baths , the frame and the membrane . the electron system in the pd film is strongly coupled to the phonon system in the membrane . at 150 mk , the electron - phonon conductance in the pd is 30 nw / k @xcite , and the kaptiza conductance between the pd and the silicon membrane is @xmath62 130 nw / k @xcite . for comparison , the highest beam conductance measured is 240 pw / k at 150 mk . the silicon frame is effectively a blackbody since its lateral dimensions are much larger than the peak thermal wavelength . at 100 mk , the peak thermal wavelength is @xmath67 @xmath1 m in silicon , where @xmath68 and @xmath69 have the usual definitions and @xmath70 m s@xmath71 is the speed of sound for transverse waves . a phonon scattered from a beam into the silicon frame has a low probability of coherently backscattering into the beam . given the 5 @xmath1 m thickness of the silicon layer , and the thermal wavelengths above 100 mk ( @xmath72 @xmath1 m ) , the membrane and the frame can be modeled as three - dimensional multi - moded thermal baths , where the number of modes is given by the planck distribution and is much larger than unity . the beams behave as three - dimensional multi - moded acoustic waveguides since the widths ( @xmath73 @xmath1 m ) and 5 @xmath1 m thickness are much larger than @xmath74 at 0.1 - 1 k. phonons emitted by the baths couple radiatively to the silicon beams , propagate unimpeded in the bulk of the beams , and scatter at the beam surfaces . when the surfaces of the beams are perfectly smooth , phonons coupled to the beams are scattered elastically without any loss of coherence in the beams . the conductance in this case is determined by the beam areas and is referred to as ballistic . in the case of rough surfaces , phonons scatter diffusively with a large probability of backscatter . the scattering is elastic but the phonon wavevectors are randomized at the surface . if the surface features form cavities much larger than @xmath74 , the surface can be thought of as a blackbody . the conductance is thus boundary - limited and was first described theoretically by casimir @xcite . an important assumption in casimir s formalism is the isotropic ( lambertian ) reemission of phonons by the beam surfaces . in practice , the scattered pattern of phonons from a rough surface is strongly dependent on the surface characteristics . when the roughness can be described statistically , the reflection pattern is averaged over the ensemble of the surface realizations . for a surface with a gaussian correlation function , the roughness can be characterized by the correlation length @xmath75 and standard deviation @xmath19 of the surface features above a mean height . figure [ fig : beampattern ] illustrates the scattered pattern as a function of @xmath76 @xcite . for a nearly smooth surface , @xmath77 , phonons are largely reflected in the specular direction . when @xmath78 , phonons incident on the surface are diffused in all directions , however , the flux is not lambertian . nonetheless , in numerous theoretical studies of boundary - limited scattering @xcite , the reflected pattern is decomposed into a diffuse lambertian fraction @xmath29 and a specular fraction @xmath79 , as illustrated in fig . [ fig : beampattern ] . the diffuse reflection is modeled as inelastic scattering that redistributes the phonon wavevector , although strictly speaking all reflections are elastic at the surface facets that define the roughness . the surface can then be thought of as a blackbody with emittance @xmath29 . hence , a completely lossless but rough surface can be modeled as a lossy surface @xcite . this approach is numerically less intensive than calculations of the reflection pattern based on the exact surface profile , or the statistics of the surface @xcite . ) and a very rough surface ( @xmath78 ) . the diffuse fraction of the realistic beam patterns are generally approximated as isotropic ( lambertian ) , which is relatively simple to integrate into models of thermal transport . , width=321 ] our model of heat flow in a silicon beam has the following assumptions : ( i ) the ends of the beam constitute thermal baths with constant temperature and unity diffuse fraction . the baths continue to distances much larger than the beam length as shown in fig . [ fig : modeldiagram](a ) . this construct is numerically convenient from the perspective of the heat transport model presented , and is similar to that adopted in ref . ( ii ) the surfaces of the beam reflect phonons with a diffuse fraction @xmath29 that is isotropic , and a specular fraction given by @xmath79 . ( iii ) the gradient of the temperature profile across a beam is linearly dependent on @xmath29 and constant as a function of length . this assumption is valid in multi - moded beams , where the number of modes is much larger than unity and the diffuse phonon reflection can be treated as diffuse emission @xcite . models that iteratively solve for the temperature profile in multi - moded beams show negligible deviations from a constant temperature gradient @xcite . when the phonon mean - free - path is much smaller than the beam length , or @xmath29 is close to unity , phonon diffusion is the dominant term of the total heat flux , and the constant temperature gradient is an excellent approximation . the approximation also holds when @xmath29 approaches zero , or the mean - free - path is comparable to the beam length , since the temperature gradient is linearly dependent on @xmath29 ( the mathematical description of this detail is in the appendix ) . ( iv ) the phonon modes that propagate are two orthogonal transverse modes with a sound speed of @xmath80 m / s , and one dilatational mode with a sound speed of @xmath81 m / s . each mode has a thermal spectrum . ( v ) phonons propagate unimpeded in the crystalline bulk of the silicon beams , and only scatter at the beam surfaces . ( vi ) the diffuse flux is reemitted according to lambert s law of thermal radiation . ( vii ) all modes scatter diffusively with the same probability @xmath29 , and there is no spectral dependence for @xmath29 . it is possible to map @xmath82 by fitting measurements of @xmath4 and calculating the peak thermal wavelength as a function of @xmath58 @xcite . for a gaussian surface @xcite , when @xmath77 , the specular power reflection coefficient is proportional to @xmath83 and the diffuse coefficient to @xmath84 ; when @xmath78 the reflection coefficient is independent of wavelength . hence , the thermal spectrum can be scattered very differently as temperature changes , and in recognizing the limitations of our model , we do not extrapolate the results to @xmath82 . ) that extend to distances much larger than @xmath12 . ( b ) each surface element of the beam radiates phonons into all solid angles . the radiated flux from the infinitesimal area @xmath85 that reaches the shaded beam port is reduced depending on the diffuse fraction of the beam surfaces . ( c ) using the method of images , the path of specularly scattered phonons can be re - imaged to one vector . the components of the vector map the number of scattering events at each surface of the beam . , width=321 ] figure [ fig : modeldiagram](b ) illustrates the radiative model . the beam is divided into infinitesimal areas @xmath86 that radiate heat into all solid angles . the phonons are scattered several times before reaching the beam exit port and thermalizing in a bath . during each scattering event , the phonon flux is reduced by a factor @xmath87 that is the diffuse fraction of the surface @xmath88 . the nature of the problem is then to simultaneously calculate the flux that is viewed by @xmath89 from @xmath85 and count the number of scattering events from each surface of the beam . for the latter , the complex path of the phonons can be unfolded , or re - imaged , to a single vector that reaches an equivalent beam exit port as shown in fig . [ fig : modeldiagram](c ) . the components of the vector then equate to the number of scattering events from each surface . the details of the radiative model are described in the appendix [ sec : app : a ] . we define an equivalent view factor for a rectangular waveguide @xmath90 , \label{eqn : fx - final}\end{aligned}\ ] ] where @xmath91 and @xmath92 are equivalent view factors between parallel and perpendicular surfaces of the beam respectively , and @xmath93 is the contribution of the direct ( unscattered ) flux from the surfaces of the beam to the cross sectional area at @xmath94 ( reference plane ) . @xmath95 is the diffuse fraction of the surface emitting the flux , and the index @xmath96 runs over all surfaces of the beam . the index @xmath97 runs over the surfaces perpendicular to @xmath96 . included in the terms @xmath91 and @xmath92 is a factor @xmath98 , where @xmath99 is the total number of scattering events from each surface as described in fig . [ fig : modeldiagram](c ) . for the baths , @xmath100 is obtained by setting @xmath101 in eq . [ eqn : fx - final ] . the total heat flux through the beam is @xmath102 , @xmath103dx_i , \label{eqn : px0}\end{aligned}\ ] ] where @xmath104 , and @xmath105 is an average group velocity . given the linear temperature profile @xmath106 , the approximation @xmath107 where @xmath58 is the average of the bath temperatures , and the relation @xmath108 , eq . [ eqn : px0 ] reduces to @xmath109 where @xmath110 is the net diffuse power from the beam surfaces , and @xmath111 is the net specular power that is exchanged between the baths . the conductance of the beam as a function of temperature when scaled by the maximum ( ballistic ) conductance @xmath112 is @xmath113 in the short beam limit , @xmath114 , and @xmath115 . for non - zero diffuse fractions and as the beam length becomes much larger than the cross - sectional dimensions , @xmath116 , and using @xmath117 where @xmath118 is the volume specific heat of the material , the mean free path takes the form @xmath119 . when the diffuse fractions of the surfaces of the rectangular beam are different , all terms in the summation over the index @xmath96 in eq . [ eqn : fx - final ] must be evaluated . the numerical evaluation of the problem is reduced by a factor of two if there is symmetry in the diffuse fraction of the surfaces , @xmath120 and @xmath121 . for the silicon beams discussed in this paper , the top surface and sidewalls are diffuse , and the bottom surface is specular . it is important to note that the diffuse fraction of the top and sidewall surfaces are uncorrelated . the terms @xmath91 and @xmath92 in eq . [ eqn : fx - final ] are evaluated using monte - carlo integration with 10@xmath122 uniformly distributed random points averaged 10@xmath123 times . @xmath93 is evaluated with an adaptive quadrature routine . figure [ fig : modelsimulation ] illustrates the accuracy of the model compared to an analytic expression for the mean - free - path @xmath0 in a rectangular beam @xcite . at the diffuse limit ( @xmath39 ) , the conductance scales as @xmath124 , and at the ballistic limit , the conductance is independent of length . the expression @xmath125 is remarkably accurate over the range of beam length explored in fig . [ fig : modelsimulation ] . this fact highlights the importance of the phonon mean - free - path , that a single parameter can capture a significant fraction of the heat transport in the ballistic - diffusive limit . figure [ fig : g - aol ] shows the thermal conductance of silicon beams as a function of beam area - to - length ratio . the cause of the variability in conductance ( a factor of @xmath625 for the same beam geometry ) is the roughening of the silicon beam surfaces during fabrication . the top surface of the beams are primarily roughened during a dry etch step that patterns a nb layer . the plasma chemistry , 75% cf4 , 15% o@xmath126 , etches silicon at a higher etch rate than nb . the exact processes that roughen the silicon are unclear , but the likely candidates are micro - masking due to non - volatile species in the plasma , and slower etch of native and in - situ nbo@xmath46 in the fluorine chemistry . the use of an etch stop such as 100 nm thick alo@xmath46 was successfully employed in early devices , however , these detectors exhibited very large uncontrolled heat capacity due to the amorphous nature of the etch stop @xcite . test devices were also fabricated with a lift - off process that left the beam surfaces smooth . the conductance of these beams were higher , suggesting boundary - limited phonon propagation as the cause of the variability in conductance of the beams with rough surfaces . ( excluding the lift - off points ) , and is the simplest interpretation of the data that is intended to merely guide the eye . details of this data are summarized in table [ tbl : data ] . , width=321 ] we have explored the relationship between thermal conductance and surface roughness with an extensive dataset of atomic force microscopy ( afm ) images . non - destructive afm of a silicon beam surface a millimeter in length is not practical . instead , the silicon device layer was imaged on the handle wafer , in close proximity to each device . for a silicon surface with a gaussian roughness , @xmath127 nm and @xmath128 nm depending on the processing conditions . however , we have found that a large fraction of the roughened surfaces did not satisfy the mathematical assumptions required to apply the theory of phonon scattering by nearly - smooth gaussian surfaces , and the measured thermal conductance differed from expectations of such a theory , even when @xmath19 was much less than the typical phonon wavelength at the temperatures explored ( 0.1 - 1 k ) . we observed power spectral densities of surface height that were approximately power laws over the range of observable wavevector values . in this case , we found an empirical relationship between the power - spectral - density @xmath129 of the surface profile and the measured magnitude and temperature dependence of the phonon mean - free - path in our silicon beams , @xmath130 , where @xmath15 is the silicon beam thickness and @xmath131 is the phonon wavevector . as expected from beckmann s theoretical treatment of scattering from non - gaussian surfaces @xcite , the observed relationship involves the correlation function of the local tilt angle of the surface facets , @xmath132 . it is challenging to precisely predict the phonon mean - free - path from information contained in the afm data due to practical sampling considerations . this occurs for several reasons : the statistics of the surface roughness are observed to vary significantly across the wafer and are sparsely sampled in @xmath133 @xmath1m@xmath123 patches set by the resolution and stability of the afm . in addition , for a multi - moded beam with thickness of order the width , details such as surface roughness of the beam sidewalls are significant , as is shown in sec . [ sec : controlg ] . non - destructive afm imaging of the sidewall geometry is extremely difficult to perform , and such a task becomes impractical when considering the number of silicon beams that constitutes a detector imaging array . however , we have found that afm images can provide qualitative insight to the thermal conductance in beams where the surface roughness relative to the thermal wavelength has a significant amplitude and is non - gaussian . suggests resonant phonon scattering from the rough surface . ( b ) afm images of 6.4@xmath436.4 @xmath1m@xmath123 patches representative of the top surface of the beams described in ( a ) . the scale size estimated from the peaks in @xmath134 is consistent with the roughness scales measured from these images . the rough surface is heavily pitted , with an average pit depth of 30 nm . the large distribution in pit size and shape is the likely cause of the observed broad resonance . since the scale size is much smaller for the smooth surfaces ( @xmath135 nm ) , the onset of the peaks occur at a much higher temperature ( shorter thermal wavelengths ) . the striping in the afm image of the smooth surface is due to residue and accumulation of particulates on the afm tip . beam dimension are @xmath136 mm , and @xmath137 @xmath1 m for devices labelled with squares , diamonds , and triangles respectively.,width=321 ] figure [ fig : cas - lim](a ) shows the conductance of smooth ( lift - off ) and roughened ( dry etch ) silicon beams when compared to the boundary - limited conductance in each case @xcite . the phonon propagation transitions from diffuse to ballistic at 300 mk , where the peak thermal wavelength is @xmath138 nm . the minimum in conductance suggests the transition is resonant for the rough beam @xcite . the inset in fig . [ fig : cas - lim](a ) shows the inverse phonon mean free path @xmath134 . the scale height of the scatterers @xmath139 can be estimated from the temperature at which @xmath134 peaks . using the rayleigh criterion , @xmath140 @xcite , where @xmath74 is the phonon wavelength at the peak temperature , and @xmath141 is the phonon angle of incidence to the surface . assuming an average over all angles , @xmath142 , the rayleigh criterion suggests @xmath143 nm for the rough beam and @xmath144 nm for the smooth beams . these estimates are consistent with the roughness measured from the afm images shown in figure [ fig : cas - lim](b ) . the rough ( dry etch ) surface is heavily pitted , with an average pit depth of 30 nm and average size of 250 nm that is comparable to @xmath74 at 300 mk . the scale size is much smaller for the smooth ( lift - off ) surfaces , @xmath135 nm , which leads to the onset of resonance at a much higher temperature . in addition to phonon scattering by rough surfaces of a beam , organic residue such as photoresist on the surfaces can enhance diffuse scattering . thin layers of metals and non - metals hundreds of angstroms thick have been shown to induce diffuse transport in silicon @xcite . in our devices , the reduction and variability in @xmath3 in a few cases were at least qualitatively attributed to residue accumulation on beams of order a millimeter in length . care must be taken when devices are cleaned since the cleaning process may introduce further residue , or roughness ( see example of silicon dipped in hf in ref . @xcite ) . table [ tbl : data ] summarizes the silicon beam geometries and conductances shown in fig . [ fig : g - aol ] . the exponent @xmath145 is inversely proportional to @xmath3 , indicating a transition from diffuse to ballistic transport . low @xmath3 devices have boundary - limited transport , and @xmath146 . high @xmath3 devices are approaching ballistic transport , and @xmath147 in the transition region . .conductance fit parameters @xmath148 and @xmath145 as a function of beam geometry , @xmath149 . the entries marked with an asterisk indicate beams with smooth top surfaces ( lift - off devices described in the text ) . [ cols="^,^,^,^,^ " , ] [ tbl : data ] in the design of prototype tes detectors , the choice of the beam length and width was guided by the data in fig . [ fig : g - aol ] . several strategies were adopted in order to control the conductance and produce uniform detectors for the focal plane array described in ref . @xcite . to minimize the top surface area of the silicon beams roughened by the dry etch , the width of the nb leads were extended to cover a larger fraction of this surface , and the beam widths were defined with a combination of drie using the standard bosch process @xcite followed by rie to remove the silicon down to a nb ground plane . the two processes are separated to reduce potential metallic contamination in the drie chamber while minimizing lateral silicon undercut from the isotropic rie . the bottom surface of the beams in contact with the nb ground plane are smooth , and scanning electron microscopy images of the sidewalls showed no significant features from the rie etching of the beams employed in test structures , fig . [ fig : sidewall - sem](a ) . the beam sidewalls were predicted to be nearly specular , with conductances comparable to the lift - off devices shown in fig . [ fig : g - aol ] . the top surface of the beams was expected to be slightly rough , @xmath150 , since the area exposed during the dry etch step was reduced by a factor of two . four beams 785 @xmath1 m in length and 13 @xmath1 m in width were integrated into the tes design . post fabrication , the beams had well - defined top surfaces , but the sidewalls were roughened by the drie steps , fig . [ fig : sidewall - sem](b ) . the total conductance was a factor of 5 below the predicted target range . in principle , the scale length and height of the sidewall features that include the drie scalloping as well as passivation residue present a very diffuse surface for phonons above 100 mk . figure [ fig : longbeam - model ] compares the total conductance of the four beams to simulations based on the diffusive - ballistic model described in sec . [ sec : model ] . the diffuse fractions of the sidewall and top surface of the beams vary across a wafer and between wafers . the inferred diffuse fraction of the top surface of the beams is @xmath151 . a possible source for the moderate roughness is the 200 nm thick nb lead sputter - deposited on each beam . the lead and beam are generally considered as decoupled systems , and the nb conductance is estimated to be far below that of the silicon beam @xcite . however , it is the diffusion process of the phonons in the silicon by the nb layer that is of importance and generally unknown . it has been shown that even in the presence of large acoustic mismatch , phonons in silicon can be diffused at surfaces with metallic layers tens to thousands of angstroms thick ( see fig . 8 in ref . @xcite for examples of phonon diffusion caused by al and ti films below 1 k ) . , devices marked with crosses , squares , diamonds , and circles have similar sidewall profiles . ( b ) to minimize the top surface area roughened by the dry etch of the nb lead layer , drie was used to define the beam width . the ridges at the top of the beam are scalloping from the drie cycles . possible sources for the bright features on the sidewall are leftover passivation layer , and charged silicon features . , width=321 ] it is evident from the results in fig . [ fig : longbeam - model ] that a solution to the problem of conductance control is a design that is least sensitive to the effect of fabrication conditions such as surface roughening and residue . our approach was to reduce the length @xmath12 of a beam , which increases the phonon mean - free - path @xmath0 by reducing the interaction probability of the phonons with the rough sidewalls of the beam . the conductance is less sensitive to @xmath0 in the limit @xmath152 . figure [ fig : shortbeam ] shows the simulated and measured total conductance when a ballistic beam 10 @xmath1 m long and 13 @xmath1 m wide is integrated into the detector as shown in fig . [ fig : q - band - detector ] . given our knowledge of the surface roughness after the dry etch step ( fig . [ fig : longbeam - model ] ) , a diffuse fraction @xmath153 was expected and subsequently obtained in practice . m long beams shown in fig . [ fig : q - band - detector ] . the measured conductance was a factor of @xmath625 below the design conductance ( hatched box ) . the cause of the variability is boundary - limited scattering . ( right ) simulations of the conductance as a function of beam wall diffuse fractions . the inferred diffuse fraction of the samples are moderately rough top surfaces , @xmath154 , and very rough sidewalls , @xmath155 . , width=321 ] . the shaded and hatched regions are defined by the measured range of conductance of the long diffuse beams shown in fig . [ fig : longbeam - model ] . ( right ) the measured conductances are within the predicted range of the simulations ( light grey ) . , width=321 ] the total conductance , which is largely set by the ballistic beam , was predicted prior to fabrication . the diffuse beams merely support microstrip and bias leads necessary for the operation of the detector . however , they contribute the largest fraction to the variance in the total conductance across a wafer and between wafers . the magnitude and variance in @xmath3 of these beams were estimated from the data in fig . [ fig : longbeam - model ] . the fractional standard deviation in total conductance is @xmath28% and within the requirement of the target instrument @xcite . the range of conductances that can be accessed with a ballistic beam is suitable for low - noise bolometric detectors operating at ground - based or air - borne telescopes where the signal power is of order 1 - 3 pw / k . in the ideal design , a single ballistic beam sets the conductance of the detector and carries all signal and bias leads . this approach is very well suited to the fabrication of kilo - pixel arrays , especially when the signal is coupled to the detector with free - space absorbers . for space applications where the signal power is less than 1 pw , the ballistic beam can be a phononic structure with phonon bandgaps that can be engineered to reduce the conductance @xcite . the heat capacity of the dielectric also plays an important role in the performance of thermal detectors . for a bolometer and calorimeter , @xmath41 affects the detector time constant , and in some cases anomalous heat capacity associated with the dielectric has been reported to increase thermodynamic noise of these detectors @xcite . the heat capacity of the silicon membranes were determined from @xmath156 , where the thermal time constant @xmath157 was measured from the decay in temperature of a silicon membrane to a pulse in heater power . compared to the debye specific heat @xcite , @xmath158 at 200 mk and 0.7 at 1 k. the systematic difference between the measurements and the debye approximation is due to the large contribution to the total heat capacity of the devices from the normal metal au heater / thermometer and pd - au termination on the silicon membranes . nonetheless , the measured specific heat are extremely small . for the detectors shown in fig . [ fig : q - band - detector ] , the estimated silicon heat capacity is @xmath622 fj / k at 150 mk , which is advantageous since the overall detector heat capacity can be controlled with a normal metal that can be reliably deposited on the membrane . we have successfully used pd to achieve a heat capacity of 2 pj / k @xcite . measurements indicate that the specific heat of pd below 1 k is 5.8 mj mol@xmath71 k@xmath159 , which is lower than the room temperature value of 9.4 mol@xmath71 k@xmath159 @xcite . we have successfully deposited 400 nm pd layers with a lift - off process . it is important to emphasize that there is no evidence for anomalous specific heat in the single - crystal silicon device layers . this is in contrast to the large excess specific heat of amorphous sin@xmath46 that is widely used in the fabrication of thermal detectors @xcite . the thermal conductance of single - crystal silicon devices can be precisely controlled using a short beam with ballistic - dominated phonon transport . the conductance of the short beam is largely set by its area and is insensitive to the details of the fabrication conditions that can vary over time . this approach is superior to the conventional use of long beams for the control of thermal conductance . phonon transport in long beams is diffusive - ballistic and very sensitive to the detailed surface physics of the beams . the conductance is thus boundary - limited , and has been observed in beams with modest surface roughness ( 5 - 10 nm rms in surface height ) . resonant phonon scattering has also been observed at 300 mk in beams with highly pitted surfaces with pit depths of order 30 nm . boundary - limited scattering has contributed to a large variance , up to a factor of 5 , in the conductance of devices fabricated and tested during this work . in contrast , when a beam with ballistic - dominated thermal transport is integrated into the design , the uniformity in conductance is reduced to a fractional deviation of @xmath28% in devices fabricated across two wafers , and this variability is largely set by the conductance of diffuse beams that support electrical and microwave leads to the device . in addition , we have found no evidence of excess specific heat in single - crystal silicon membranes . hence , the total heat capacity of the detector can be effectively controlled with a normal metal film . for the transition - edge sensors ( tess ) described in this work , the heat capacity was determined by a pd layer 400 nm thick . hence , the strategies outlined for the precision control of thermal conductance and heat capacity are well suited to the fabrication of uniform large - format arrays of tess with sensitivities approaching 10@xmath36 w / hz@xmath160 . we gratefully acknowledge financial support from the nasa roses / apra program . k. rostem was supported by the nasa postdoctoral program at the goddard space flight center during the initial stages of this work . we thank samelys rodriguez for wirebonding and cryo - cable support . we thank our collaborators a. ali , j. appel , and t. a. marriage for their contribution to the data acquisition . we thank the national institute of standards and technology for providing the squid series array and time division multiplexer chips . we start from the usual definition of the infinitesimal view factor @xmath161 @xcite , @xmath162 where @xmath163 is the vector between the infinitesimal surfaces @xmath164 and @xmath165 , and @xmath166 are the angles between the normal vectors to @xmath167 and @xmath163 ( see fig . [ fig : modeldiagram ] ) . we define an equivalent view factor for convenience , @xmath168 and note that the heat exchanged between the infinitesimal surfaces is @xmath169\,f_{da_i - da_j},\end{aligned}\ ] ] where @xmath170 are the temperatures of the surface elements . the total heat flux is the integral of @xmath171 over the beam length , @xmath172 where @xmath173 represents the surfaces of the waveguide , which in our case is a rectangular beam . the equivalent view factor between any two perpendicular surface elements of the beam is @xmath174 ^ 2 } \left [ \,f_m\prod_k ( 1-f_k)^{n_k}\right ] \right\ } \,da_i\,da_j , \\ & = & \mathcal{f}^\perp ( m ; x_i , y_i , y_j , z_j ) \,dx_i\,dy_i\,dy_j\,dz_j , \nonumber \end{aligned}\ ] ] where the indices @xmath96 and @xmath88 sum over all surfaces of the beam . for parallel surface elements , @xmath175 ^ 2 } \left [ \,f_m\prod_k ( 1-f_k)^{n_k}\right ] \right\ } \,da_i\,da_j , \\ & = & \mathcal{f}^\parallel ( f_m ; x_i , y_i , y_j , z_j ) \,dx_i\,dy_i\,dx_j\,dy_j . \nonumber\end{aligned}\ ] ] @xmath99 is the number of specular reflections experienced by phonons reflecting off the element @xmath89 first , and then all other surfaces until a reference plane ( e.g. @xmath94 in fig . [ fig : modeldiagram](a ) ) is reached . @xmath99 can be determined from the geometry of the specular scattering as described in fig . [ fig : modeldiagram](c ) , and is the sum of scattering events from each surface . for example , if the radiating element is on the top surface , @xmath99 can be found for phonons reflecting off the top and bottom surfaces , and reflecting off the sidewalls only . in this way , a different diffuse fraction can be assigned to each surface . it is also useful to note that the linear dependence of the equivalent view factors on @xmath95 can be translated to a linear dependence of the temperature gradient on the diffuse fraction when eq . [ eqn : perp ] and [ eqn : para ] are finally substituted into eq . [ eqn : px01 ] . the integral equations of the equivalent view factors over the surfaces of the beam are @xmath176 where the limits of the integrals depend on the surface @xmath96 from which the phonons were originally radiated . since a temperature gradient is only present in the principal direction of heat flow ( x - axis ) , all surface elements at the plane of constant @xmath177 are at the same temperature and radiate the same flux . hence , the equivalent view factor for the direct ( non - scattered ) heat flux @xmath93 from the surface to the reference plane can be calculated analytically from eq . [ eqn : vf ] by considering the view factor of a infinitesimal strip of width @xmath14 to the reference plane at @xmath94 @xcite .

random matrix theory is of importance for both its conceptual and its predictive powers . at a conceptual level , there is the three fold way classification of global time reversal symmetries of hamiltonians for chaotic quantum systems @xcite , later extended to ten by the inclusion of chiral and topological symmetries @xcite . for each of the ten hamiltonians there is an ensemble of random hermitian matrices , which in turn are the hermitian part of the ten matrix lie algebras associated with the infinite families of symmetric spaces . the symmetric spaces themselves give rise to ten ensembles of random unitary matrices @xcite . to see how the formulation of ensembles leads to a predictive statement , consider for definiteness the three fold way classification of hamiltonians for chaotic quantum systems . the basic hypothesis , initiated by wigner @xcite and dyson @xcite , and refined by bohigas et al . @xcite , is that a quantum mechanical system for which the underlying classical mechanics is chaotic , will , for large energies have the same statistical distribution as the bulk scaled eigenvalues of the relevant ensemble of hermitian matrices . the particular ensemble is determined by the presence or absence of a time reversal symmetry , with the former consisting of two subcases depending on the time reversal operator @xmath2 being such that @xmath3 , or @xmath4 ; see e.g. ( * ? ? ? * ch . 1 ) . this becomes predictive upon the specification of the statistical distributions for the bulk scaled eigenvalues of the random matrix ensemble . thus , according to the hypothesis , these same distributions will be exhibited by the point process formed from the highly excited energy levels of a chaotic quantum system . the point process can be realised in laboratory experiments of nuclear excitations ( see e.g. @xcite ) , and microwave billiards @xcite , amongst other examples . arguably the most spectacular example of both the conceptual and predictive powers of random matrix theory shows itself in relation to the riemann zeta function @xmath5 . the celebrated riemann hypothesis asserts that all the complex zeros of @xmath5 are of the form @xmath6 , with @xmath7 . the montgomery - odlyzko law asserts that for large @xmath8 these zeros referred to as the riemann zeros have the same statistical properties as the bulk eigenvalues of large hermitian random matrices with complex elements . but from the three fold way , the latter is the ensemble of random hermitian matrices giving the statistical properties of the large energy levels of a chaotic quantum mechanical system without a time reversal symmetry . the implied relationship between the riemann zeros and quantum chaos is consistent with and extends the hilbert - plya conjecture @xcite asserting that the riemann zeros correspond to the eigenvalues of some ( unknown ) unbounded self - adjoint operator ; see @xcite for contemporary research on this topic . to realise the predictive consequences of this link between two seemingly unrelated quantities the riemann zeros and chaotic quantum hamiltonians without time reversal symmetry requires a list of the riemann zeros for large @xmath8 . such a list has been provided by odlyzko @xcite , who in a celebrated numerical computation has provided the high precision evaluation of the @xmath9-th riemann zero and over 70 million of its neighbours . from this data set the veracity of the montgomery - odlyzko law can be tested . primary statistical quantities for this purpose include the two - point correlation function @xmath10 which measures the density of zeros a distance @xmath11 from @xmath12 , with @xmath12 itself averaged over a window of zeros which itself is large , but still small relative to @xmath12 and the spacing distribution @xmath13 for the event that the @xmath14-th next ( in consecutive order ) neighbours - th next neighbour of the item @xmath15 is item @xmath16 . ] are a distance @xmath11 apart . as displayed in @xcite , at a graphical level the agreement between the riemann zero data and the predictions of random matrix theory is seemingly exact . although zeros of order @xmath9 are huge on an absolute scale , it turns out that convergence to random matrix predictions occurs on a scale where @xmath17 plays the role of @xmath1 , giving rise to potentially significant finite size effects . such effects were first considered in the work of keating and snaith @xcite in their study of the statistical properties of the value distribution of the logarithm of the riemann zeta function on @xmath18 with @xmath19 large . quite unexpectedly they showed that the finite size corrections could be understood by introducing as a model for the riemann zeros the eigenvalues of haar distributed unitary random matrices from u@xmath20 , instead of complex hermitian random matrices . it had been known since the work of dyson @xcite that the eigenvalues for haar distributed unitary random matrices from u@xmath20 and complex hermitian random matrices with gaussian entries have the same limiting statistical distribution . but for finite @xmath1 they are fundamentally different , with the former being rotationally invariant while the spectral density of the latter is given by the wigner semi - circle law and thus has edge effects . odlyzko s work on the computation of large riemann zeros is ongoing . in @xcite he announced a data set beginning with the @xmath21-nd zero and its next @xmath22 neighbours . taking advantage of the great statistical accuracy offered by this data set , the finite size correction for the deviation of the empirical nearest neighbour spacing distribution and the limiting random matrix prediction was displayed , and shown to have structure . to explain the functional form of this correction , in keeping with the work of keating and snaith , bogomolny and collaborators @xcite were able to present analytic evidence that the correction term could again be understood in terms of a model of eigenvalues from haar distributed u@xmath20 matrices . more explicitly , let @xmath23 refer to the spacing distribution between consecutive eigenvalues for haar distributed unitary random matrices , with the angles of the eigenvalues rescaled to have mean spacing unity . the study @xcite calls for the functional form of @xmath24 in the large @xmath1 expansion @xmath25 here the subscript 2 " on the rhs is the label from dyson s three fold way in the case of no time reversal symmetry . as a straightforward application of the pioneering work of mehta @xcite and gaudin @xcite , dyson @xcite showed that @xmath26 where @xmath27 is the integral operator on @xmath28 with kernel @xmath29 twenty years later this fredholm determinant evaluation was put in the context of painlev theory by the kyoto school of jimbo et al . @xcite , who showed @xmath30 where @xmath31 satisfies the differential equation ( an example of the painlev v equation in sigma form ) @xmath32 with small @xmath33 boundary conditions @xmath34 note that the parameter @xmath35 introduced in ( [ ks ] ) only enters in the characterisation through the boundary condition ; we refer to ( * ? ? ? 8) for background theory relating to the painlev equations as they occur in random matrix theory . in @xcite the leading correction term @xmath24 was estimated using a numerical extrapolation from the tabulation of @xmath36 . in @xcite the task of finding analytic forms for @xmath24 was addressed . one first notes that analogous to the formula ( [ 2 ] ) one has , for finite @xmath1 , @xmath37 where @xmath38 is the integral operator on @xmath28 with kernel @xmath39 expanding now for large @xmath1 gives @xmath40 where the leading correction term in this expansion is given explicitly by @xmath41 it was observed in @xcite ( cf . also lemma [ lem : folklore ] below ) that this expansion lifts to @xmath42 where @xmath43 is as in ( [ 3a ] ) and @xmath44 is the integral operator on @xmath28 with kernel @xmath45 ; the leading correction is now given by the operator expression @xmath46 which is _ linear _ in @xmath47 . substituting in ( [ 1a ] ) and comparing to ( [ 1 ] ) we read off that @xmath48 in section [ sub : numerical ] of the present paper the problem addressed is how to use such formulae to provide a high precision numerical tabulation of @xmath49 . in the case @xmath50 we exhibit the resulting functional form in odlyzko s data set for the riemann zeros ; see section [ sect : riemann ] . an expression for the large @xmath1 expansion of the fredholm determinant in ( [ 1a ] ) involving painlev transcendents was also given in @xcite . thus it was shown that @xmath51 as noted above , @xmath52 satisfies the particular painlev v equation in sigma form ( [ d1 ] ) with boundary condition ( [ d2 ] ) , with only the latter depending on @xmath35 . the transcendent @xmath53 was characterised as the solution of the second order , linear differential equation @xmath54 where , with @xmath55 , @xmath56 and fulfilling the @xmath57 boundary condition @xmath58 substituting ( [ 2.13d ] ) in ( [ 1a ] ) and comparing to ( [ 1 ] ) we read off that @xmath59 by using a nested power series method to solve the differential equation ( [ sig2 ] ) numerically , this expression was used in @xcite to determine the graphical form of @xmath24 , and also to determine its small and large @xmath11 expansions . however , the operator approach advocated in the present paper is simpler and more straightforward to use numerically . random unitary matrices from u@xmath20 chosen with haar measure form one of dyson s three circular ensembles of unitary random matrices coming from the considerations of the three fold way , and in this context is referred to as the circular unitary ensemble ( cue ) . the other two circular ensembles are the circular orthogonal ensemble ( coe ) of symmetric unitary matrices , and the circular symplectic ensemble ( cse ) of self dual unitary matrices . we will show that our methods can be adapted to provide a systematic investigation of the leading correction in the large @xmath1 expansion of spacing distributions for each of these examples . after some preparatory material on the expansion of operator determinants and numerical methods for evaluating the corresponding terms in section [ sect : prep ] , we begin in section [ sect : cue ] by extending the study initiated in @xcite on this problem as it applies to cue matrices . the first question addressed is the computation of the leading order correction term in the large @xmath1 expansion of @xmath60 , or equivalently @xmath61 , in terms of an integral operator formula extending ( [ tq1 ] ) . we then do the same in the circumstance that each eigenvalue has been deleted independently at random with probability @xmath62 . such a thinning procedure , well known in the theory of point processes ( see e.g. @xcite ) , was introduced into random matrix theory by bohigas and pato @xcite , and has been applied to odlyzko s data set for the riemann zeros in @xcite . continuing with the point process perspective , next we consider the setting of a translationally invariant one - dimensional point process , normalised to have average spacing unity , superimposed with a poisson point process . our interest is in the distribution of the minimum distance from each poisson point to a point in the original process . in an ensemble setting , this is equivalent to computing the minimum distance from the origin of a point in the original process . for cue matrices the distribution can be expressed in a form analogous to , which allows the statistic to be expanded for large @xmath1 . in the final subsection of section [ sect : cue ] , the statistic for the nearest neighbor spacing , that is , the minimum of the spacing distance between left and right neighbours in the cue is similarly studied . in section [ sect : riemann ] our results are compared against empirical findings from odlyzko s data set for the riemann zeros . there we will use a unique @xmath63 correction of the cue correlation kernel that was observed in the study @xcite to match the corresponding @xmath63 terms of the two- and three - point correlation functions . including this term in the numerical calculations systematically improves the fit to the empirical data . the study of spacing distributions for matrices from the coe and cse is more involved than for cue matrices . the reason is that the latter is based on the single integral operator @xmath64 with kernel ( [ 3 ] ) , whereas the pathway to tractable expressions in the case of the coe and cse makes use of inter - relations between gap probabilities . the necessary theory is covered in the first part of section [ sect : coecse ] . in the second part of section [ sect : coecse ] we give the leading correction term for the large @xmath1 expansion of @xmath65 where @xmath66 for the coe and @xmath67 for the cse , in terms of a characterisation analogous to ( [ 2.13d ] ) and ( [ sig2 ] ) . the parameter @xmath35 specifies the thinning probability . in this paper we discuss finite size effects up to an error @xmath68 , with @xmath1 in the odlyzko data set of riemann zeros being effectively @xmath69 . we choose the same order of magnitude of @xmath1 for the simulations of the circular ensembles . now , by the law of large numbers , sampling is known to introduce a statistical error of about @xmath70 where @xmath71 denotes the sample size . hence , pushing the sampling error to the same order of magnitude as the remaining finite size error thus requires a sampling size of @xmath72 . this was actually the choice for our simulations and is well matched by the odlyzko data set of a little more than @xmath73 zeros . however , observing structure also in the @xmath68 remainder term would hence require to increase the sampling size by at least two to four orders of magnitude . this is only possible with parallel processing and a subtle memory management of the resulting gigantic data sets ( the raw material provided by odlyzko is already about 22 gb in size ) . the integral operator formulae of this paper are based on the following _ folklore _ lemma , which we prove here for ease of reference . [ lem : folklore ] let @xmath74 be a bounded interval and let @xmath75 , @xmath76 , @xmath77 be continuously differentiable kernels on @xmath78 . if the expansion @xmath79 holds uniformly up to the first derivatives as @xmath80 , then it lifts as an expansion @xmath81 of the induced integral operators on @xmath82 in trace - class norm . if @xmath83 , the operator determinant expands as @xmath84 with the leading correction term given by is the frchet derivative of the map @xmath85 under the identification of the dual of the space of trace - class operators with the space of bounded operators , see ( * ? ? ? the derivative of that map in the direction of the bounded operator @xmath47 is thus given by @xmath86 . by analytic continuation this interpretation of the expression @xmath87 extends well to the case that @xmath88 . ] @xmath89 an expression that is linear in @xmath47 . on _ bounded _ intervals @xmath74 , continuously differentiable kernels @xmath76 induce integral operators @xmath90 on @xmath82 that are trace - class : using the fundamental theorem of calculus one can represent them straightforwardly as the sum of a rank - one operator and a product of two hilbert schmidt operators , see , e.g. , @xcite . by the same construction the expansion lifts from kernels in @xmath91-norm to operators in trace - class norm . now , using the multiplicativity of the operator determinant we get first that @xmath92 truncating the series definition of the operator determinant in terms of exterior products , see , e.g. , ( * ? ? * eq . ( 3.5 ) ) , @xmath93 at the second term we get next that @xmath94 which completes the proof . the computations for all figures of this paper are based on the numerical evaluation of operator terms such as @xmath95 or @xmath87 . in fact , there is an extension of the numerical method introduced in @xcite for the operator determinant that facilitates the efficient highly - accurate evaluation of such terms with exponential convergence such that the approximation error is @xmath96 where @xmath97 denotes the dimension parameter of the method ( e.g. , the number of quadrature points ) . ] if the kernels extend analytically into the complex domain . given a quadrature method @xmath98 with _ positive _ weights @xmath99 , we associate with an integral operator @xmath90 on @xmath82 with the kernel @xmath76 the _ nystrm matrix _ @xmath100 for compact intervals @xmath74 the best choice for smooth @xmath101 is gauss legendre quadrature , for which recently a super - fast algorithm of @xmath102 complexity has been found @xcite ; it converges exponentially fast if @xmath101 extends analytically into the complex domain . as was shown in ( 6.2 ) , these convergence properties are inherited by the determinant of the nystrm matrix as an approximation of the operator determinant , namely @xmath103 where @xmath104 denotes the identity operator in @xmath82 on the left and the @xmath105 identity matrix on the right . in particular , the convergence is exponential if the kernel @xmath106 extends analytically into the complex domain . now , the same method of replacing the integral operator by the associated nystrm matrix applies to the approximation of quite general operator terms . we give four examples : * the trace is approximated straightforwardly by the quadrature formula as @xmath107 * operator products are approximated by matrix products , that is , @xmath108 this follows from @xmath109 * the resolvent kernel @xmath110 of the operator @xmath111 satisfies the integral equation @xmath112 for all @xmath113 . application of the quadrature rule gives with @xmath114 @xmath115 that is @xmath116 and therefore @xmath117 . * by combining all the examples so far , we get @xmath118 in all the examples the convergence properties are inherited from the underlying quadrature formula ; in particular , the convergence is exponential if the kernels extend analytically to the complex domain . the proof is straightforward in the first two examples and follows from the theory of the nystrm method for integral equations in the third one ; the fourth is a combination of all the previous results . the numerical derivatives with respect to the @xmath11 and @xmath119 variables of the generating functions for the spacing distributions are computed in exactly the same fashion as discussed in @xcite , that is , based on chebyshev expansions with respect to @xmath11 and contour integration with respect to @xmath119 . the actual implementation and use of this methodology within the matlab toolbox of @xcite is described in the appendix of the present paper . the sequence of angles , to be referred to as eigen - angles , specifying the eigenvalues of a unitary random matrix from any of the three circular ensembles forms a point process on @xmath120 $ ] with uniform density @xmath121 . due to the rotational symmetry , the spacing distributions of the angles , @xmath122 with @xmath123 for the coe , cue , cse respectively , and @xmath124 , are thus independent of the absolute location of the eigenvalues . if there is an eigen - angle at @xmath12 and at @xmath125 , we may rotate all the angles so that @xmath12 becomes the origin , and speak of a spacing of size @xmath11 . let us do this , and also normalise the angles so that the mean spacing is unity . fundamental to the study of spacing distributions @xmath122 are the so - called gap probabilities @xmath126 . the latter specify the probability that an interval of size @xmath11 contains exactly @xmath127 eigen - angles for @xmath128 matrices . specifically , let us introduce the generating functions @xmath129 of course the naming on the lhs s represent an abuse of notation , due to the same functional forms also appearing on the rhs ; however the appearance of the generating function symbol @xmath119 is enough to distinguish the different quantities . we have the relationship between generating functions ( see e.g. ( * ? ? ? 8.1 ) ) @xmath130 or equivalently the formula @xmath131 also fundamental is the relationship between the generating function ( [ g2 ] ) and the @xmath14-point correlation functions @xmath132 . thus we have ( see e.g. ( * ? ? ? 9.1.1 ) ) @xmath133 the eigenvalues for the cue form a determinantal point process , while the eigenvalues for the coe and cse form pfaffian point processes ; see e.g. ( * ? ? ? * chapters 5 and 6 ) . specifically , in the case of the cue we have @xmath134_{j , l=1,\dots , k},\ ] ] where @xmath135 is given by ( [ 3 ] ) , while the analogous formula for the coe and cse involves a pfaffian of a @xmath136 anti - symmetric kernel function . this structural difference distinguishes the case of the cue and thus makes it simpler . substituting ( [ rk ] ) in ( [ er ] ) , and making use of a standard expansion formula in the theory of fredholm integral operators ( see e.g. ( * ? ? ? * eq . ( 3.14 ) ) ) gives @xmath137 now substituting the expansion ( [ tq0 ] ) in ( [ rk1 ] ) and substituting the result in ( [ g3 ] ) we read off the large @xmath1 expansion to second order . [ p1 ] we have @xmath138 where @xmath139 and @xmath140 in particular @xmath141 and @xmath142 the operation of thinning applied to a point process refers to the procedure of independently deleting each point in a sample with probability @xmath62 , @xmath143 . with the @xmath14-point correlation function before thinning being given by @xmath144 , the @xmath14-point correlation function after thinning is simply @xmath145 , due to the independence . it thus follows from ( [ er ] ) , that the generating function for the gap probability in the presence of thinning , @xmath146 say , is related to the generating function for the gap probability without thinning by @xmath147 specialising further to the cue we can substitute ( [ rk1 ] ) in the rhs of this expression to conclude @xmath148 the analogues of ( [ g3 ] ) and ( [ g4 ] ) for the spacing probabilities are @xmath149 using the first of these , together with ( [ rk1 ] ) and the expansion ( [ tq0 ] ) we can write down the analogue of proposition [ p1 ] . [ p1a ] we have @xmath150 where @xmath151 and @xmath152 in particular @xmath153 and @xmath154 implementation of these formulae according to the method of section [ sub : numerical ] will be carried out in section [ sect : riemann ] , in the context of a comparison with the corresponding statistics for odlyzko s data set of the riemann zeros . ) . left panel : a histogram of empirical data from @xmath155 with @xmath156 scaled to unit mean spacing , computed using a bin size of 0.01 and @xmath157 samples from @xmath155 , for each of which @xmath158 samples of a uniformly distributed random origin were drawn ( blue dots ) ; the large @xmath1 limit @xmath159 ( red solid line ) . right panel : the simulation data minus @xmath159 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath161 ( red solid line).,title="fig:",scaledwidth=55.0% ] ) . left panel : a histogram of empirical data from @xmath155 with @xmath156 scaled to unit mean spacing , computed using a bin size of 0.01 and @xmath157 samples from @xmath155 , for each of which @xmath158 samples of a uniformly distributed random origin were drawn ( blue dots ) ; the large @xmath1 limit @xmath159 ( red solid line ) . right panel : the simulation data minus @xmath159 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath161 ( red solid line).,title="fig:",scaledwidth=55.0% ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] the quantity @xmath163 is the probability that there are no eigenvalues in the interval @xmath164 of the scaled c@xmath165e . differentiating this quantity with respect to @xmath11 gives the probability density function @xmath166 for the scaled eigen - angle closest to the origin , @xmath167 in the case @xmath168 with thinning , it follows from ( [ th1 ] ) , as well as the translation invariance of the sine kernel , that @xmath169 use of ( [ 2.13d ] ) gives a characterisation of the leading two terms of the large @xmath1 expansion of this quantity . we have @xmath170 where in terms of integral operators @xmath171 alternatively , in terms of the transcendents @xmath172 and @xmath173 defined in @xmath174 and @xmath175 to simulate this quantity in translationally invariant empirical data such as the @xmath176 itself , one can actually draw the origin uniformly and independently within the interval for which the data are defined ( that is , one superimposes a poisson point process defining the origin ) . to prepare for corresponding computations within a large set of riemann zeros this was done , as a proof of concept , for cue matrices in figs . [ fig:5 ] and figs . [ fig:6 ] . a variant of the spacing distribution between consecutive eigenvalues is the spacing distribution between nearest neighbour eigenvalues @xcite . for this , at each eigenvalue one measures the smallest of the spacings to the eigenvalue immediately to the left , and the spacing immediately to the right . in a theoretical ensemble formulation , the system is to be conditioned so that there is an eigenvalue at the origin . consider in particular the cue with eigen - angles scaled to have unit spacing . with @xmath177 denoting the @xmath97-point correlation function for the conditioned system in the presence of thinning , we have @xmath178 where the second line follows from ( [ rk ] ) and ( [ 3 ] ) , together with elementary row operations applied to the determinant . let the entry of the determinant in ( [ 2.25 ] ) be denoted by @xmath179 , and denote the corresponding integral operator supported on @xmath180 by @xmath181 . analogous to ( [ 2.20a ] ) , the scaled nearest neighbour spacing in the presence of thinning is then @xmath182 by observing @xmath183 we can read off from ( [ 2.21a ] ) , upon expanding the kernel to order @xmath184 , integral operator formulae for the leading two terms analogous to ( [ 2.25x ] ) . in terms of the notation and , let @xmath185 denote by @xmath186 and @xmath187 the integral operators on @xmath188 with kernels @xmath189 and @xmath190 respectively . we have @xmath191 where @xmath192 in @xcite a painlev transcendent evaluation of @xmath193 has been given ; see also @xcite . -st next neighbour spacing : riemann zeros data vs. random - matrix based prediction with effective dimension @xmath194 , no thinning ( @xmath195 ) . left panel : a histogram of the odlyzko data set with bin size @xmath196 ( blue dots ) ; the large @xmath1 limit @xmath197 ( red solid line ) . right panel : the riemann zero data minus @xmath197 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath198 with interior rescaling ( red solid line ) , exterior rescaling ( thin black line ) and no rescaling ( dashed green line).,title="fig:",scaledwidth=55.0% ] -st next neighbour spacing : riemann zeros data vs. random - matrix based prediction with effective dimension @xmath194 , no thinning ( @xmath195 ) . left panel : a histogram of the odlyzko data set with bin size @xmath196 ( blue dots ) ; the large @xmath1 limit @xmath197 ( red solid line ) . right panel : the riemann zero data minus @xmath197 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath198 with interior rescaling ( red solid line ) , exterior rescaling ( thin black line ) and no rescaling ( dashed green line).,title="fig:",scaledwidth=55.0% ] [ fig:1 ] starting with the 1973 seminal work of montgomery there have been significant developments on a deep ( conjectural ) connection between the statistics of the fluctuation properties of the zeros of the riemann zeta function on the critical axis @xmath18 and those of the eigen - angles of @xmath199 , supported by large - scale numerical calculations based on the extensive tables of riemann zeros provides by odlyzko , see , e.g. , @xcite and the literature cited therein . answering a question of odlyzko @xcite , who had posed the challenge of understanding the structure in the numerical difference graph between the riemann zeta spacing distribution for a set of zeros of large height and the ( conjectured ) asymptotics , finite size effects of this statistic were studied in @xcite . this study gave a precise quantitative association of riemann zeros @xmath200 at height @xmath8 to a ( formal ) size @xmath1 of the corresponding @xmath199 . using the hardy littlewood conjecture of the distribution of prime pairs , bogomonly and keating @xcite had earlier given an analytic expression for the pair correlation of the riemann zeros . the authors of @xcite expanded this for large height @xmath8 , with the local density normalized to unity , to obtain @xmath201 where @xmath202 denotes the smooth asymptotic density of zeros at @xmath8 and @xmath203 , @xmath204 are the following constants ( with @xmath205 the stieltjes constants and @xmath206 summation over the primes ) : ] and the method of @xcite to the sums over the primes . ] @xmath207 on the other hand , the pair correlation function @xmath208 of @xmath199 , normalized to mean spacing unity , expands by as @xmath209 by matching the first correction terms of and the authors of @xcite got as effective dimension @xmath1 of a @xmath199 at height @xmath8 @xmath210 this way one has the ( conjectural ) approximation @xmath211 but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] as noted in @xcite a rescaling of the @xmath11-variable in the leading correction term of the @xmath212-expansion of the two - point functions absorbs the @xmath213 term , respectively the @xmath63 term : @xmath214 where @xmath215 that is , one gets an @xmath216 approximation of @xmath217 by expanding @xmath218 into powers of @xmath219 and rescaling the @xmath11-variable of the leading correction term by @xmath220 . it was suggested in @xcite to apply the same procedure , that is , @xmath220-rescaling the @xmath11-variable of the leading correction term of a @xmath199 based expansion , to improve the fit even for more general statistics , such as the spacing distribution originally considered by odlyzko . in @xcite it was successfully applied to improve the fit of the cue leading correction term to the spacing distribution in the presence of thinning . , no thinning ( @xmath195 ) . left panel : a histogram of the odlyzko data set with bin size @xmath196 ( blue dots ) , where the approximately @xmath22 zeros were broken into 1024 sets of nearly equal size , for each of which @xmath221 samples of a uniformly distributed random origin were drawn ; the large @xmath1 limit @xmath159 ( red solid line ) . right panel : the riemann zero data minus @xmath159 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath161 with interior rescaling ( red solid line ) , exterior rescaling ( thin black line ) and no rescaling ( dashed green line).,title="fig:",scaledwidth=55.0% ] , no thinning ( @xmath195 ) . left panel : a histogram of the odlyzko data set with bin size @xmath196 ( blue dots ) , where the approximately @xmath22 zeros were broken into 1024 sets of nearly equal size , for each of which @xmath221 samples of a uniformly distributed random origin were drawn ; the large @xmath1 limit @xmath159 ( red solid line ) . right panel : the riemann zero data minus @xmath159 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath161 with interior rescaling ( red solid line ) , exterior rescaling ( thin black line ) and no rescaling ( dashed green line).,title="fig:",scaledwidth=55.0% ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] the numerical data , physical arguments and mathematical conjectures surrounding the fluctuations of the riemann zeros indicate that their statistics is asymptotically well approximated , to more than just leading order , by some correlation kernel @xmath222 . consequently , to justify the mechanism of the exterior rescaling introduced in the last paragraph , @xcite suggested to pull - back the absorption of the @xmath213 term from the two - point function into the expansion of such an assumed kernel . in the present paper we will use such an expansion to improve systematically the @xmath199 fit of the fluctuation properties of the riemann zeros statistics from @xmath63 to @xmath68 . by modifying the expansion of the cue correlation kernel @xmath223 to include an @xmath63 term , @xmath224 the task is thus to identify the unknown kernel @xmath225 from the @xmath63 term in . by translation invariance and symmetry of the kernel @xmath222 and by the normalization @xmath226 to mean spacing unity , we get the form @xmath227 of a convolution kernel with a _ symmetric _ function @xmath228 satisfying @xmath229 . a brief calculation results in the corresponding two - point function @xmath230 matching with gives , cf . the first equality in ( * ? ? ? * eq . ( 28 ) ) , @xmath231 in the appendix of @xcite it is shown that this expansion of the kernel allows to reproduce consistently a conjectural expansion of the three - point function ( which was obtained from a calculation published much later in @xcite ) . this is further evidence that the kernel expansion induces , quite generally , the @xmath232 correction terms in the determinantal point process fluctuations statistics . now , the correction kernel @xmath71 can be absorbed easily by an _ interior _ rescaling of the leading correction kernel @xmath45 , namely by introducing @xmath233 which expands as @xmath234 , where would be the same as @xmath220 . ] @xmath235 to summarize , we have @xmath236 and thus get the following recipe to go from the leading correction term of a @xmath199 fluctuation statistics to the corresponding one of the riemann zeros : simply replace the leading correction kernel @xmath45 by its interior rescaling @xmath237 . in our experiments , the most pronounced difference between interior and exterior rescaling can be observed in the statistics shown in the right panel of fig . [ fig:4 ] . the largest data set of riemann zeros currently provided by odlyzko , which was announced already in @xcite , consists of the @xmath238 consecutive zeros starting with zero number @xmath239 the first of them has height @xmath240 while the last one has height @xmath241 because of the logarithmic dependence on the height , even within that large data set the smooth density @xmath212 , the effective dimension @xmath1 and the scaling parameters @xmath220 , @xmath242 remain essentially _ constant _ to 15 digits precision , @xmath243 from this data set we extracted various spacing statistics ,- th next neighbour spacing can be found in ( * ? ? ? * figs . 12 ) ( no thinning and no theoretical prediction of the leading correction term ) , in ( * ? ? ? 13 ) ( no thinning , exterior rescaling only ) and in ( * ? ? ? 910 ) ( exterior rescaling only ) . since in this case there is no difference visible between interior and exterior rescaling we refrained from showing the results once more . ] with ( @xmath162 ) and without ( @xmath195 ) thinning , and compared them up to the leading correction with the theoretical prediction in terms of operator determinants @xmath244 and their corrections @xmath87 as obtained in section [ sect : cue ] for the cue . in figs . [ fig:1][fig:8 ] the riemann zero data are shown as blue dots , the leading order term as a red solid line , and the correction terms * with interior rescaling @xmath245 as a red solid line , * with exterior rescaling @xmath246 as a thin black line , * without rescaling as a green dashed line . in particular , the following statistics of the riemann zeros are shown to be in excellent agreement up to the leading correction term with ( interior ) rescaling : [ cols="^,^,^,^,^",options="header " , ] here we denote by @xmath247 those probability densities for the riemann zeros that are defined analogously to the cue case @xmath248 . , no thinning ( @xmath195 ) . left panel : a histogram of the odlyzko data set with bin size @xmath196 ( blue dots ) ; the large @xmath1 limit @xmath249 ( red solid line ) . right panel : the riemann zero data minus @xmath249 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath250 with interior rescaling ( red solid line ) , exterior rescaling ( thin black line ) and no rescaling ( dashed green line).,title="fig:",scaledwidth=55.0% ] , no thinning ( @xmath195 ) . left panel : a histogram of the odlyzko data set with bin size @xmath196 ( blue dots ) ; the large @xmath1 limit @xmath249 ( red solid line ) . right panel : the riemann zero data minus @xmath249 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath250 with interior rescaling ( red solid line ) , exterior rescaling ( thin black line ) and no rescaling ( dashed green line).,title="fig:",scaledwidth=55.0% ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] as already mentioned , the fact that the eigenvalues of coe matrices form a pfaffian point process tells us that the counterpart to the formula ( [ rk ] ) involves a pfaffian of a @xmath136 anti - symmetric kernel function . substituting this in ( [ er ] ) , the pfaffian analogue of the identity implying ( [ rk1 ] ) ( see e.g. ( * ? ? ? * eq . ( 6.32 ) ) ) allows for a pfaffian analogue in the case of the coe to be given . however , the resulting formula is not of the same practical utility as ( [ rk1 ] ) since the @xmath136 kernel function is not analytic , which in turn means that the numerical quadrature methods of section [ sub : numerical ] have poor convergence properties . fortunately , there is a second option . this presents itself due to fundamental inter - relations between gap probabilities for the coe and cue . these also involve the gap probabilities for haar distributed orthogonal matrices from the classical groups o@xmath251 and o@xmath252 . in reference to the latter , eigenvalues on the real axis and thus corresponding to @xmath253 or @xmath254 are to be disregarded , and only the eigenvalues with angles in @xmath255 are to be considered ( the remaining eigenvalues occur at the negative of these angles , i.e. the complex conjugates ) . the first inter - relation of interest is between the generating functions for the gap probabilities in the cue , and in haar distributed real orthogonal matrices @xcite , ( * ? ? ? * eq . ( 8.127 ) ) @xmath256 with me denoting a particular matrix ensemble , we have used the symbol @xmath257 instead of @xmath258 to indicate that in this expression no scaling of the eigen - angles has been imposed ; this convention will be followed below . to present the next inter - relation , introduce the generating functions @xmath259 with coe@xmath260coe denoting the point process of @xmath261 eigenvalues on the circle that results by superimposing the eigenvalue sequences of two independent coe matrices , a result conjectured by dyson @xcite and proved by gunson @xcite gives that @xmath262 where the operation alt " refers to the operation of observing every second eigenvalue only . as a consequence of this , one has that @xcite , @xcite @xmath263 according to ( [ w1 ] ) , it follows from ( [ w4 ] ) that the gap probabilities for the coe are related to those for o@xmath264 and o@xmath265 with @xmath97 suitably chosen , but this alone does not determine the former . to be able to do this , additional inter - relationships between generating functions are required . in the case @xmath266 and thus @xmath1 even , the additional inter - relationships have been given in @xcite according to the generating function identity @xmath267 note that ( [ w5 ] ) substituted in ( [ w4 ] ) is consistent with ( [ w1 ] ) in the case @xmath1 even . the identity ( [ w5 ] ) has very recently @xcite been shown to be a corollary of the identities between eigenvalue distributions @xmath268 here the notation @xmath269 refers to the eigenvalue distribution of @xmath270 coe matrices , while @xmath271 refers to the distribution in the circumstance that the eigenvalues with angles @xmath272 are reflected in the real axis by @xmath273 , and thus all eigenvalues have angles between 0 and @xmath254 . the operation even ( odd ) refers to observing only those eigenvalues that occur an even ( odd ) number of places from @xmath253 , reading anti - clockwise . substituting ( [ w5 ] ) in ( [ w2 ] ) , and setting @xmath274 so that @xmath275 we obtain , after some minor manipulation , a closed formula for the gap probabilities of @xmath269 in terms of the gap probabilities for @xmath276 ( * ? ? ? * eq . ( 8.150 ) ) @xmath277 the analogue of ( [ w6 ] ) for @xmath278 allows us to deduce the analogue of ( [ w7 ] ) . we have @xmath279 we read off from ( * ? ? ? 7.1 ) that @xmath280 substituting in ( [ w2 ] ) and manipulating as in the derivation of ( [ w7 ] ) gives ( [ w8 ] ) . with @xmath281 the cases and can be combined into a single formula that holds for both parities of @xmath1 : @xmath282 the eigenvalues for @xmath283 in @xmath255 form a determinantal point process with kernel ( see e.g. ( * ? ? ? 5.5.3 ) ) @xmath284 where @xmath285 scaling the eigen - angles of @xmath286 to have unit mean spacing by @xmath287 thus yields the following result : [ cor : coe ] let @xmath288 denote the integral operator on @xmath28 with kernel and denote @xmath274 . we have @xmath289 as in thinning is expressed by @xmath290 we will now turn our attention to spacing probabilities for the cse . these follow from knowledge of the spacing distributions for the coe . thus one has the inter - relation @xcite @xmath291 and it follows from this that @xmath292 as noted in ( * ? ? ? * eq . ( 8.158 ) ) , recalling the definition ( [ w2 ] ) , and making use too of ( [ w5 ] ) , this gives @xmath293 as discussed before , the kernel of the determinantal point process formed by the eigenvalues of @xmath276 is given by @xmath294 scaling the eigen - angles of @xmath295 to have unit mean spacing by @xmath287 thus yields the following result , where we use to express the presence of thinning : [ cor : cse ] let @xmath288 denote the integral operator on @xmath28 with kernel . we have @xmath296 @xmath297-th next neighbour spacing : simulation vs. formulae from theory for finite size coe , no thinning ( @xmath195 ) . left panel : a histogram of empirical data from @xmath298 with @xmath156 scaled to unit mean spacing , computed using a bin size of 0.01 and @xmath157 samples ( blue dots ) ; the large @xmath1 limit @xmath299 ( red solid line ) . right panel : the simulation data minus @xmath299 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath300 ( red solid line).,title="fig:",scaledwidth=55.0% ] @xmath297-th next neighbour spacing : simulation vs. formulae from theory for finite size coe , no thinning ( @xmath195 ) . left panel : a histogram of empirical data from @xmath298 with @xmath156 scaled to unit mean spacing , computed using a bin size of 0.01 and @xmath157 samples ( blue dots ) ; the large @xmath1 limit @xmath299 ( red solid line ) . right panel : the simulation data minus @xmath299 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath300 ( red solid line).,title="fig:",scaledwidth=55.0% ] as in figure [ f8 ] but with thinning @xmath162 . , title="fig:",scaledwidth=55.0% ] as in figure [ f8 ] but with thinning @xmath162 . , title="fig:",scaledwidth=55.0% ] making use of the second equation in ( [ th2 ] ) we get the spacing distribution @xmath301 from which knowledge of the large @xmath1 expansion of the determinants in corollary [ cor : coe ] will allow us to determine the terms in the corresponding expansion of the spacing . [ prop : coe ] let @xmath302 and @xmath303 denote the integral operators on @xmath28 with kernels @xmath304 and @xmath305 respectively . we have @xmath306 where , denoting @xmath307 , @xmath308 and @xmath309 the case @xmath310 of ( [ m1 ] ) is due to gaudin @xcite . as an illustration and test of the above results , we took @xmath157 samples of @xmath286 ( @xmath156 , using the @xmath311 cmv sparse matrix model @xcite ) , and from this made an empirical determination of the spacing distribution scaled to unit mean spacing . this was then subtracted from the large @xmath1 limit @xmath299 and the difference compared against @xmath312 . both @xmath313 ( no thinning ) and @xmath162 were considered ; see figures [ f8 ] and [ f8a ] . the analogue of ( [ sa1 ] ) , @xmath314 allows us to determine the first terms of the large @xmath1 expansion by expanding correspondingly the determinants in proposition [ cor : cse ] . [ prop : cse ] with @xmath302 and @xmath303 as in proposition [ prop : coe ] we have @xmath315 where @xmath316 and @xmath317 the case @xmath310 of ( [ m2 ] ) is due to mehta and dyson @xcite . @xmath297-th next neighbour spacing : simulation vs. formulae from theory for finite size cse , no thinning ( @xmath195 ) . left panel : a histogram of empirical data from @xmath318 with @xmath156 scaled to unit mean spacing , computed using a bin size of @xmath196 and @xmath157 samples ( blue dots ) ; the large @xmath1 limit @xmath319 ( red solid line ) . right panel : the simulation data minus @xmath319 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath320 ( red solid line).,title="fig:",scaledwidth=55.0% ] @xmath297-th next neighbour spacing : simulation vs. formulae from theory for finite size cse , no thinning ( @xmath195 ) . left panel : a histogram of empirical data from @xmath318 with @xmath156 scaled to unit mean spacing , computed using a bin size of @xmath196 and @xmath157 samples ( blue dots ) ; the large @xmath1 limit @xmath319 ( red solid line ) . right panel : the simulation data minus @xmath319 scaled by @xmath160 ( blue dots ) ; the leading correction term @xmath320 ( red solid line).,title="fig:",scaledwidth=55.0% ] as in figure [ f9 ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] as in figure [ f9 ] but with thinning @xmath162.,title="fig:",scaledwidth=55.0% ] as was done for the coe , to illustrate and test the above results , we took @xmath157 samples of @xmath295 ( @xmath156 , again using the @xmath321 cmv sparse matrix model @xcite ) , and from this made an empirical determination of the spacing distribution scaled to unit mean spacing . this was then subtracted from the large @xmath1 limit @xmath319 and the difference compared against @xmath322 . both @xmath313 ( no thinning ) and @xmath162 were considered ; see figures [ f9 ] and [ f9a ] . alternatively to the integral operator characterisation of the expansion terms in propositions [ prop : coe ] and [ prop : cse ] there is one in terms of painlev transcendents . such an expression in the case of the cue has been given in @xcite and restated in above . to simplify we restrict ourselves to @xmath1 being even , writing @xmath323 for definiteness . we begin by noting that the joint eigenvalue pdf for the eigen - angles @xmath324 of @xmath276 matrices in the interval @xmath255 , after changing variables @xmath325 , is proportional to @xmath326 with @xmath327 , @xmath328 , which is an example of the jacobi unitary ensemble ( jue ) ; see e.g. @xcite . hence the corresponding generating functions for the gap probabilities are related by @xmath329 the significance of this result for present purposes is that the rhs can be expressed as a @xmath31pvi transcendent @xcite . specifically , reading off from ( * ? ? ? ( 8.71 ) , ( 8.75 ) , ( 8.76 ) ) we have @xmath330 where @xmath331 satisfies the particular @xmath31pvi equation @xmath332 subject to the boundary condition @xmath333 \displaystyle { 2 n \over \pi } z t^{1/2 } , & a=-1/2 . \end{array } \right.\ ] ] we therefore have @xmath334 to obtain an expansion consistent with ( [ sa ] ) we make the ansatz @xmath335 changing variables in ( [ sa3 ] ) @xmath336 , substituting ( [ ffa ] ) , and equating terms to leading order @xmath337 and to next leading order @xmath338 , we obtain characterisations in terms of differential equations of @xmath339 and @xmath340 . the leading function @xmath339 in satisfies the particular @xmath341 equation ( _ with @xmath342 in the notation of ( * ? ? ? * eq . ( 8.15 ) ) _ ) @xmath343 subject to the boundary condition @xmath344 \displaystyle { z \over \pi } w^{1/2 } , & \nu = - . \end{array } \right.\ ] ] the function @xmath345 in satisfies the second order linear differential equation @xmath346 where the coefficients are given in terms of @xmath347 according to @xmath348 subject to the boundary condition @xmath349 which is thus the same in both cases . in terms of @xmath339 and @xmath340 we have @xmath350 scaling the eigen - angles of @xmath269 to have unit mean spacing by @xmath351 , we can now substitute ( [ 3.31 ] ) into ; then after using to add the presence of thinning substitute the result in ( [ sa1 ] ) and compare with ( [ sa ] ) to obtain painlev transcendent characterisations of the expansion terms @xmath299 and @xmath312 . [ cc1 ] we have @xmath352 and @xmath353 [ r3.6 ] the case @xmath310 of ( [ cc1a ] ) agrees with the @xmath341 formula for @xmath354 ( no thinning ) reported in @xcite and @xcite . the differential equations ( [ ffa1 ] ) and ( [ ffa3 ] ) can be used to successively generate terms in the series expansions of @xmath355 and @xmath356 about the origin , extending the boundary conditions ( [ ffa2 ] ) and ( [ ffa4 ] ) . substituting these into ( [ cc1a ] ) and ( [ cc1b ] ) then provides us with the small @xmath11 expansion of @xmath299 and @xmath312 . [ c3.8 ] we have @xmath357 - \frac{\pi ^8 \xi s^7 } { 90720 } + \frac{\pi ^8 ( \xi -2 ) ( 3 \xi -32)\xi s^8 } { 5292000 } + \frac{\pi ^{10 } \xi s^9 } { 7983360 } + \operatorname{o}(s^{10 } ) \end{gathered}\ ] ] and @xmath358 + \frac{\pi ^8 \xi s^7}{7560}-\frac{\pi ^8 ( \xi -2 ) \left(3 \xi -32 \right)\xi s^8}{352800}-\frac{\pi^{10 } \xi s^9}{435456 } + \operatorname{o}\left(s^{10}\right ) . \end{gathered}\ ] ] we remark that setting @xmath195 in ( [ cc1c ] ) extends the expansion of @xmath359 given in ( * ? ? ? * eq . ( 8.141 ) ) . for the analogous results in the case of the @xmath295 , scaling the eigen - angles to have unit mean spacing by @xmath287 , we substitute ( [ 3.31 ] ) into ( [ w6c ] ) ; then after using to add the presence of thinning substitute the result in ( [ sa16 ] ) and compare with ( [ sa7 ] ) . [ cc2 ] we have @xmath360 and @xmath361 as in remark [ r3.6 ] the case @xmath310 of ( [ cc2a ] ) agrees with the @xmath362 formula for @xmath363 ( no thinning ) reported in @xcite and @xcite . as with corollary [ c3.8 ] we can use these characterisations to generate power series expansions about the origin . [ c3.9 ] we have @xmath364 \nonumber r_{4,\xi}(0;s ) & = -\frac{4}{27 } \pi ^4 \xi s^4+\frac{128 \pi ^6 \xi s^6}{2025}-\frac{128 \pi ^8 \xi s^8}{11025}+\frac{17408 \pi ^{10 } \xi s^{10}}{13395375}+ \operatorname{o}(s^{12}).\end{aligned}\ ] ] the expansions of corollaries [ c3.8 ] and [ c3.9 ] are consistent with the operator theoretic numerical evaluations of figures [ f8][f9a ] for small values of @xmath11 . the method of section [ sub : numerical ] for numerically evaluating general terms of integral operators is most easily added to the matlab toolbox described in @xcite ; the basic code needed to run it is as follows : [ source , numberlines ] ---- function [ val , err , n ] = operatorterm(term , varargin ) % operatorterm evaluates terms of integral operators that have a scalar value . % % operatorterm(term , k1, ... ,km ) returns the value of the % expression ' term(k1, ... ,km ) ' for m discrete integral % operators k1, ... ,km ( that is , for nystrm matrices of % a variable dimension n , which is chosen adaptively ) . % the value to be returned must be scalar . tol = 5e-15 ; n = 7 ; n_max = 1000 ; operators = cell(size(varargin ) ) ; val0 = inf ; while n < n_max n = floor(1.41*n ) ; for k = 1:length(varargin ) operators{k } = varargin{k}(n ) ; end val = term(operators { : } ) ; err = abs(val - val0 ) ; if err < tol ; break ; end val0 = val ; end ---- from the user perspective this command is called with a considerable symbolic look and feel . for example , the leading order term in , that is , @xmath365 evaluates for @xmath366 to the following value : [ source , numberlines ] ---- s = 1 ; k = op(@(x , y ) sinc(pi*(x - y)),[0,s ] ) ; i = @(k ) eye(size(k ) ) ; lead = @(k ) det(i(k ) - k ) ; [ val , err , n ] = operatorterm(lead , k ) ; printcorrectdigits(val , err ) ; 0.170217421379185 ---- and the leading correction term in , that is , @xmath46 evaluates for @xmath366 to the following value : [ source , numberlines ] ---- l = op(@(x , y ) pi*(x - y).*sin(pi*(x - y))/6,[0,s ] ) ; corr = @(k , l ) -det(i(k)-k)*trace((i(k)-k)\l ) ; [ val , err , n ] = operatorterm(corr , k , l ) ; printcorrectdigits(val , err ) ; -0.075241982465122 ---- since the estimated error has been taken into account when printing these values , they are good to about 15 digits ; in both cases the adaptively chosen number of gauss legendre quadrature points ( that is , the dimension of the nystrm matrices representing the integral operators ) was as small as @xmath367 , which is a clear sign of the exponential convergence of the method . cpu time is about a millisecond . we are grateful to a. odlyzko for providing us with the riemann zero data set . the research of f.b . was supported by the dfg - collaborative research center , trr 109 , `` discretization in geometry and dynamics . '' the research of p.j.f . and a.m. is part of the program of study supported by the arc centre of excellence for mathematical & statistical frontiers . to3em , _ a method for calculating spectral statistics based on random - matrix universality with an application to the three - point correlations of the riemann zeros _ , j. phys . a * 46 * ( 2013 ) , 305203 , 17pp . to3em , _ the @xmath371-nd zero of the riemann zeta function _ , dynamical , spectral , and arithmeitc zeta functions ( m. van frankenhuysen and m.l . lapidus , eds . ) , contemporary math . 290bb , amer . soc , providence , ri , 2001 , pp .

thin shell components can be found in many large built - up mechanical structures such as cars , ships , and aeroplanes . the prediction of the mid- and high- frequency vibrational properties of these structures becomes computationally prohibitive for standard element - based methods , such as the finite element method @xcite . the main reasons for this limitation are : firstly , very fine meshes are required for an adequate representation of the highly oscillatory wave solutions and the computational complexity grows with frequency raised to the power of the dimension of the space being modelled . secondly , small uncertainties originating from the manufacturing process lead to a much larger variability in vibro - acoustic responses in the high frequency range @xcite , meaning that the response of any individual manufactured structure is of less interest to computer aided engineering practitioners than the average responses . methods such as statistical energy analysis ( sea ) @xcite and ray tracing @xcite are more commonly applied for modelling wave problems at high frequencies . sea has traditionally proved more popular for structural vibration problems with low damping @xcite , whereas ray tracing has found its niche in applications where relatively few reflections need to be tracked in computer graphics @xcite , room acoustics @xcite and seismology @xcite . ray tracing has also been applied to elastic wave transmission problems on shells and plates @xcite . in this context the ring frequency , that is the frequency above which longitudinal waves in a curved shell behave as they would in a flat plate , provides a useful point of reference . beyond the ring frequency the ray dynamics in a curved shell are relatively simple following geodesic paths , but below the ring frequency one finds that asymptotic ( ray ) theories show richer features . the dispersion relations become highly anisotropic with likewise anisotropic propagation @xcite . an sea treatment would fail to capture the non - trivial way in which the curved shell geometry influences the wave and ray propagation below the ring frequency @xcite , and hence ray methods can provide useful insight @xcite . wave scattering from discontinuous line joints below the ring frequency has been considered in @xcite . however , manufacturers of large built - up mechanical structures are increasingly developing larger and lighter sub - components , whereby large thin shell structures are replacing more traditional plate - beam and multi - plate assemblies . the manufacturing process for such thin shells often entails casting molten metal ( for example aluminium ) , which gives rise to curved components with smooth transitions between flat and curved regions . this raises the question of how the ray and wave scattering , and hence the vibrational properties of the structure , are modified in these smooth designs . in this work we study the case of a singly curved shell , chosen here as an assembly of two plates joined smoothly with a quarter of a cylinder . this simplified assembly represents a typical curved region within one of the larger sub - structures described above . we shall go beyond plane waves and ray tracing calculations by solving the full wave scattering problem numerically . the numerical solution to the full wave scattering problem will be discussed in comparison with the corresponding ray tracing calculations . in both cases we find effective laws for the scattering properties , which may be inserted into ray or wave propagation modelling techniques such as dynamical energy analysis ( dea ) @xcite or the wave and finite element method @xcite . in particular , combining these local scattering models within a larger model of a built - up structure will lead to a hybrid method for structures including curved thin shell components in the mid - frequency regime . here , the natural definition of the mid - frequency regime is given by the range of frequencies that are high enough for a pure fem analysis to be impractical , but low enough so that a simple geodesic description of the trajectory evolution is invalid . in the high - frequency case , dea @xcite can be applied to model the vibrational energy transport of a built - up structure along geodesic paths using the mesh data from a fem analysis @xcite . in fact , dea presents a link between ray tracing and sea by casting the wave or ray problem into the language of evolution operators . we note that an equivalent operator formalism has also been long known in computer graphics @xcite , although the theory underlying dea arose from the more general setting of evolution operators for transporting flows in dynamical systems @xcite . the organisation of the article is as follows : we introduce the necessary shell theory and derive a wave scattering model for a singly curved shell in sect . we then present two approaches for solving the wave scattering model ; a short wavelength asymptotic ray tracing model based on this shell theory is detailed in sect . [ sec : raytrac ] , and a finite difference discretisation of the full wave model is described in sect . [ sec : scatscheme ] . we then discuss and compare numerical results for both the wave and ray scattering models , and the resulting reflection and transmission laws in sect . [ sec : numer ] . the thin shell theory of donnell is one of the simplest and widely adopted models @xcite . in this theory , moments and transverse forces are expressed by the displacement @xmath0 of the middle surface as known from the theory of laterally loaded plates . as with other theories of continuum mechanics , shell theory is formulated in tensor form @xcite . some properties of tensors are summarised in [ sec : tensors ] . we assume an isotropic shell of thickness @xmath1 , young s modulus @xmath2 , density @xmath3 and with poisson ratio @xmath4 . the displacement vector of a point originally on the mid - surface of the shell is decomposed into tangential and normal components thus @xmath5^\text{t}$ ] . the following tensor equation for the normal displacement @xmath0 may be derived @xcite @xmath6 where @xmath7 are the bending and extensional stiffness , respectively . all greek alphabet indices take values from the set @xmath8 . also , the membrane strain is @xmath9 with @xmath10 the second fundamental form and @xmath11 the covariant derivative . these are discussed further in the next section , where they are simplified for a singly curved shell . the tangential displacements @xmath12 in the directions @xmath13 , respectively , satisfy @xcite @xmath14 the term @xmath15 represents the inverse of the first fundamental form as explained in [ sec : tensors ] . note that the above formulation is stated for a shell _ in vacuo _ and , although beyond the scope of the present study , coupling to an acoustic fluid may be included via an additional term in equation ( [ eq : pdesys1 ] ) to express the pressure difference on either side of the shell ; see for example ref . in addition , structural damping may be incorporated in the usual way by replacing the young s modulus @xmath2 in the equations above with @xmath16 , where @xmath17 is the damping loss factor . however , we proceed with the lossless case @xmath18 in order to isolate the effect of curvature on the reflection and transmission properties of the shell , whilst also allowing us to check that our results conserve energy in the correct manner . ) and exterior ( @xmath19 ) regions for the scattering problem . the interfaces between @xmath20 and @xmath19 lie with the flat regions where @xmath21 . for @xmath22 , the curvature increases smoothly to @xmath23 at @xmath24 . the incident , reflected and transmitted waves in @xmath19 are also indicated . ] in this section we simplify the shell theory presented above for the case of a cylindrical ridge as shown in fig . [ fig : ridgegeom ] . the only simplifying assumption is that the shell is not curved with respect to the direction @xmath25 . we define the principal curvature @xmath26 ( in the direction @xmath27 ) as @xmath28 where @xmath29 is a constant corresponding to the maximum curvature in the cylindrical region . a smoothly varying curvature with respect to @xmath27 is obtained via the interpolation function @xmath30 given by @xmath31 where @xmath32 are the centres of the transitions between the two flat regions where @xmath33 and the cylindrical region where @xmath34 , see fig . [ fig : curv ] . also , @xmath35 is the width of the transition region . note that the case of a non - smooth joint between a cylindrical ridge and a pair of connecting plates is obtained in the limit @xmath36 . a cylindrical ridge with a prescribed bending angle @xmath37 can be designed by integrating the angular increment @xmath38 to find the total length @xmath39 of the curved section of the geometry such that @xmath40 for the desired angle . the length @xmath39 is then found by numerically solving the implicit equation ( [ eq : angleinc ] ) . in this work we take a quarter cylinder with @xmath41 as shown in fig . [ fig : ridgegeom ] . in the above described geometrical setting , the shell theory presented in the last section is simplified considerably if we choose an orthonormal basis @xmath42 along the axes @xmath43 . we then measure all tensor fields relative to this basis , that is , we work in physical components . the metric becomes @xmath44 and so the first fundamental form and its dual both reduce to the identity matrix @xmath45 therefore the raising and lowering of indices becomes trivial . the second fundamental form @xmath10 in the basis @xmath42 is simply @xmath46 and zero for all other combinations of @xmath47 and @xmath48 . the covariant derivative ( see [ sec : tensors ] ) , likewise reduces considerably since all derivatives vanish for the metric ( [ eq : metricid ] ) . hence , all christoffel symbols vanish and the covariant derivative is identical to the directional derivative , i.e. @xmath49 . applying the above simplifications in the shell equation for the normal displacement ( [ eq : pdesys1 ] ) and assuming that material constants and the thickness are constant yields the following equation @xmath50 where we write @xmath51 and @xmath52 for simplicity of notation . we adopt this notation for the remainder of the paper since we are no longer considering differential equations written in tensor form . repeating the simplification process for the in - plane wave equation system ( [ eq : pdesys2 ] ) gives @xmath53 \label{eq : pdesys2b}\frac{\rho h}{c } \frac{\partial^2 u^2}{\partial t^2 } = & \:\frac{\partial^2 u^2}{\partial y^2}+\frac{(1+\nu)}{2}\frac{\partial^2u^1}{\partial x \partial y}+\frac{(1-\nu)}{2}\frac{\partial^2u^2}{\partial x^2}+\nu\kappa\frac{\partial w}{\partial y}.\end{aligned}\ ] ] hence , a simplified set of partial differential equations ( pdes ) describe the wave motion in the cylindrical ridge shell geometry under consideration here . in the next section we discuss how under certain modelling assumptions , this system can be reduced further to a set of ordinary differential equations ( odes ) . in this work we are primarily interested in how the scattering properties of a thin shell cylindrical ridge depend on the frequency and the direction of a prescribed incident plane wave originating in the flat region ( where @xmath54 ) . as such it makes sense to consider time - harmonic waves with angular frequency @xmath55 . further , we can extract scattering properties that are independent of the position along the ridge by exploiting the translational symmetry and assuming that the ridge is of infinite extent in the @xmath56direction . under these assumptions we may write @xmath57 where @xmath58 is the component of the wavenumber in the @xmath56direction and @xmath59^\text{t}$ ] are the coefficients of ( [ eq : odeansatz ] ) in the in - plane directions @xmath60 and @xmath61 , and the normal direction , respectively . substitution of the ansatz ( [ eq : odeansatz ] ) into the pde system ( [ eq : pdesys1a ] ) - ( [ eq : pdesys2b ] ) yields the following fourth - order ode system in the variable @xmath60 : @xmath62\label{eq : odesys2 } c_s^2\:\frac{{\mathrm{d}}^2 v}{{\mathrm{d}}x^2}+\left(\omega^2-c_p^2k_y^2\right)v+\mathrm{i } k_y\left((c_p^2-c_s^2)\frac{{\mathrm{d}}u}{{\mathrm{d}}x}+\nu c_p^2\kappa \hat{w}\right)&=0,\\[2pt]\label{eq : odesys3 } \frac{b}{\rho h}\left(\frac{{\mathrm{d}}^4 \hat{w}}{{\mathrm{d}}x^4}-2k_y^2\frac{{\mathrm{d}}^2 \hat{w}}{{\mathrm{d}}x^2}+k_y^4\hat{w}\right)+(c_p^2\kappa^2-\omega^2)\hat{w}+c_p^2\kappa\left(\frac{{\mathrm{d}}u}{{\mathrm{d}}x}+i\nu k_y v\right)&=0.\end{aligned}\ ] ] note that the constants in the above system have been simplified by writing them in terms of the pressure and shear wave velocities , @xmath63 respectively . in particular , we have made use of the following easily verified relations @xmath64 later in the paper we will write the above ode system in the shorthand form @xmath65 for brevity of exposition . we connect the curved region to flat plates on each side using interfaces . imposing a set of conditions at these interfaces will enable us to formulate the set of odes ( [ eq : odesysb ] ) as a boundary value problem , and then by considering the flux of the incoming and outgoing wave fields at these interfaces we are able to formulate a scattering problem . we assume that the interfaces reside in the asymptotically flat regions and as such the interfaces themselves do not give rise to reflection / transmission phenomena , only the interior region between the interfaces governs the scattering properties . each interface is assumed to satisfy the continuity conditions given in ref . @xcite , that is , continuity of displacement , rotation , traction , moment , and normal shear stress . since the material properties are constant throughout the entire geometry , then the interface conditions may be written simply as @xmath66 the superscripts specify the value of the quantity as we approach the interface from either the interior region @xmath67 containing the ridge , or the exterior flat region @xmath68 beyond the interfaces on either side of the ridge . in the sequel we will refer to the interior region as @xmath69 , and the union of the exterior regions as @xmath19 ( see fig . [ fig : curv ] ) . the waves in @xmath19 are precisely those of classical plate theory and therefore the wave modes in @xmath19 that are scattered by @xmath20 will be one ( or more ) of bending , pressure , shear or evanescent bending type . in relation to the vector @xmath70 we have that @xmath71 describes the sum of the bending wave contributions and that the in - plane wave types will each be given by a linear combination of @xmath72 and @xmath73 . in this work we consider only incident bending modes @xmath74 originating in the exterior @xmath19 and being scattered by the ridge in @xmath69 . however , the extension to other incident wave types is straightforward . for the purpose of investigating directional properties we consider plane wave scattering . we introduce the notation that the interface to the left of @xmath69 is located at @xmath75 and the interface to the right is correspondingly at @xmath76 . then an incident wave of unit amplitude travelling in from the left of @xmath20 can be written in the form @xmath77 where @xmath78 is the wave - number associated with the incident bending mode in the @xmath60-direction . in order to write the resulting scattered waves in a concise manner we introduce @xmath79 and @xmath80 . then the scattered waves in @xmath19 may be written @xmath81 where @xmath82 designates the scattered wave type as either bending ( @xmath83 ) , evanescent bending ( @xmath84 ) , pressure ( @xmath85 ) or shear ( @xmath86 ) . the symbol @xmath87 designates whether the scattered wave is reflected ( @xmath88 ) or transmitted ( @xmath89 ) . that is , whether the scattered wave emerges in @xmath19 on the same side of the ridge as the incident wave is sent in , or on the other side . the value of @xmath48 also prescribes the sign in the @xmath90 as negative for @xmath91 and positive for @xmath92 . the coefficient @xmath93 denotes the corresponding wave amplitude , which are related to @xmath70 via the following relations @xmath94 v^{\beta}(x)=&a^p_\beta\exp(\pm \mathrm{i } k^p_x x_\beta)\sin\theta\pm a^s_\beta\exp(\pm \mathrm{i } k^s_x x_\beta ) \cos\theta,\\[2pt]\label{eq : wrt } \hat{w}^{\beta}(x)=&a^b_\beta\exp(\pm \mathrm{i } k^b_x x_\beta)+a^e_\beta\exp(-k^b_x |x_\beta|).\end{aligned}\ ] ] here @xmath95 is the angle between the scattered wave directions and the @xmath96axis ( positive or negatively oriented depending on the direction of propagation ) . hence , the total wave field in @xmath19 to the left of @xmath20 is given by @xmath97^\text{t},\ ] ] and the total wave field to the right of @xmath20 may be written @xmath98^\text{t}.\ ] ] the scattering problem is then formed by connecting the interior problem for @xmath99 in @xmath20 with the plane wave ansatz in the exterior regions for @xmath100 and @xmath101 described above , via the interface conditions ( [ eq : if1 ] ) to ( [ eq : if4 ] ) . here @xmath100 and @xmath101 take the role of @xmath102 on the left and right interfaces , respectively . when this scattering problem is solved , we can extract the scattering solutions @xmath100 and @xmath101 and the scattering coefficients @xmath93 , the latter of which are of primary interest for this study . one obtains a ray tracing model from the pde model ( [ eq : pdesys1a ] ) - ( [ eq : pdesys2b ] ) by moving to short wavelength asymptotics using the ansatz @xmath103 where @xmath104 is a phase function and @xmath105 is a small parameter . the choice of the parameters @xmath106 and @xmath107 determines the wave type , bending or in - plane . then we define the frequency @xmath108 and the wavenumber vector @xmath109 . note that the pre - factors of @xmath110 vanish after applying the asymptotic scaling and setting @xmath111 . assuming that @xmath112 is large in comparison to the curvature , pierce derived a general dispersion relation @xcite that was later presented in a simpler form by norris and rebinsky @xcite . in the next section we apply this dispersion relation to generate the hamiltonian dynamics for our ray tracing model on a singly curved shell . the ring frequency for the cylinder is defined relative to the pressure mode and corresponds to the frequency above which a longitudinal wave can traverse around the cylinder . this extensional motion gives rise to breathing modes in cylinders , which also result in a radial motion . in the vicinity of the ring frequency , approximate forms of the dispersion relations are obtained for the in - plane and bending modes using different scalings . these dispersion relations each reduce to those for a flat plate when the frequency sufficiently exceeds the ring frequency . hence in this study , we refer to the mid - frequency case as the the frequency range in the vicinity of the ring frequency and the high frequency regime as the frequencies that are sufficiently large for the dispersion relations to reduce to those of a flat plate . defining the longitudinal wavenumber to be @xmath113 , with @xmath114 the longitudinal plate wave speed as before , and denoting the corresponding wavenumber at the ring frequency as @xmath115 , we find that @xmath116 for the singly curved shell configuration shown in fig . [ fig : ridgegeom ] . at zeros of the hamiltonian @xmath117 . ( a ) anisotropic dispersion relation for a cylinder . ( b ) isotropic dispersion relation for a flat plate . ] below the ring frequency , the full dispersion relationship is usually considered . for the configuration in fig.[fig : ridgegeom ] with zero curvature in the @xmath56 direction we obtain the following expression from ref . @xcite in physical coordinates : @xmath118 two extreme cases of the dispersion curve are depicted in fig.[fig : disp ] , which shows the respective dispersion relations for a cylinder and for a flat plate using the parameters given in [ sec : parameters ] . the outer part of the cylinder dispersion curve has also been observed experimentally @xcite . the hamiltonian @xmath117 ( [ ham ] ) gives rise to the following ode system , which governs the ray dynamics on a singly curved shell @xmath119 in a full time - domain simulation one would also have two additional equations @xmath120 in this work however , the frequency is constant due to the fact our hamiltonian @xmath117 is time - independent . furthermore , since we also take the material parameters to be constants throughout the shell , equation ( [ eq : hex1 ] ) simply represents a re - parametrisation of time to the fictitious time used here . equations ( [ eq : hex1 ] ) and ( [ eq : hex2 ] ) are therefore not needed in the time - harmonic description . finally , due to the translational invariance in the @xmath61-direction , @xmath58 is constant and it suffices to study the above hamiltonian system in the @xmath121 phase - plane only . in order to solve the full wave scattering problem we discretise the ode system in @xmath20 using finite difference methods and couple this to the scattering of incoming and outgoing waves in @xmath19 using the interface coupling conditions as described below . the differential operator @xmath122 in the ode system ( [ eq : odesysb ] ) includes both bending and in - plane waves , along with the coupling between them . each of the equations in the system is discretised using second order accurate centered finite difference ( fd ) formulae on a set of equi - spaced grid points @xmath123 with @xmath124 . we denote the centered finite difference matrices for approximating derivatives of order @xmath125 by @xmath126 and define @xmath127\hspace{5 mm } \text { with } \:\ : i , j=0,\dots , n \,\ ] ] to be a diagonal matrix containing the curvature values . to create a vectorial ode discretisation we use the kronecker product @xmath128 , see for example ref . @xcite , along with the following @xmath129 projector matrices : @xmath130 \:\text { for } \ : a , b=1,2,3 , \\ { \mathbf{\pi}}_i & = { \mathbf{e}}_{i i } , \\ { \mathbf{h}}_{2 3 } & = \left [ \begin{array}{ccc } 0 & 0 & 0 \\ 0 & 0 & \mathrm{i}\\ 0 & - \mathrm{i } & 0 \end{array } \right ] \,.\end{aligned}\ ] ] using the above notation we can write the various operators of our ode in a compact way . the in - plane equations ( [ eq : odesys1 ] ) and ( [ eq : odesys2 ] ) will have a diagonal ( self - interaction ) and frequency independent part @xmath131 where @xmath132 is the @xmath129 identity matrix . the coupling between the @xmath72 and @xmath73 components is symmetric and is given by @xmath133 the bending wave equation ( [ eq : odesys3 ] ) has a diagonal ( self - interaction ) and frequency independent part @xmath134 and the coupling between the bending and in - plane components is given by @xmath135 the shell operator @xmath122 then has the following discrete representation : @xmath136 the fd operators will extend beyond the grid over @xmath20 as the stencil reaches into the scattering region . this leads to a coupling between the degrees of freedom in the finite difference approximation and the unknown scattering amplitudes @xmath93 . conversely , the incident wave in @xmath19 will produce a forcing term which drives the finite difference calculation in @xmath20 . this coupling of the fd solution in @xmath20 and the plane wave description in @xmath19 is discussed further below . the scattering solution and its derivatives in @xmath19 are matched to the fd solution in @xmath20 at the interfaces between @xmath20 and @xmath19 according to the conditions ( [ eq : if1 ] ) to ( [ eq : if4 ] ) . the incident wave gives rise to inhomogeneous terms for the excitation of the fd model at the left interface . the derivatives appearing in the coupling conditions are implemented in the fd - solution in @xmath20 using one - sided finite difference operators with second order accuracy . forward difference formulae are employed at the left interface and backward difference formulae are employed at the right interface so that the stencils extend only into @xmath20 . the extremal points of these stencils are taken at the scattering boundaries , where the value of the scattering solution is applied as a boundary condition . the interfaces are thus taken to be at positions @xmath137 away from the extremal inner grid points , where @xmath137 is the discretisation step in the fd approximation . as an illustrative example we consider the condition on @xmath138 ( [ eq : if4 ] ) at the interface receiving the incident wave ( [ eq : winc ] ) . the reflected bending modes include both a propagative contribution @xmath139 and an evanescent contribution @xmath140 as shown in equation ( [ eq : wrt ] ) . the forward difference formula for the third derivative takes the form @xmath141 matching the scattering and finite difference solutions using the interface condition ( [ eq : if4 ] ) leads to @xmath142 splitting the discretised boundary operator @xmath143 into a part that acts only on the left - most entry at the interface and one that acts on the remaining nodes in @xmath20 , we obtain @xmath144 at @xmath145 , the left hand side of ( [ eq : ifdisc ] ) evaluates to @xmath146 and using the splitting described above , the right hand side may be written as @xmath147 writing @xmath148 for @xmath149 and using the definition of @xmath143 yields a linear equation @xmath150 in terms of the unknowns @xmath151 , @xmath152 and @xmath153 , @xmath149 . continuing for the remaining boundary conditions at the left interface @xmath154 yields similar relations , all following the same pattern as in ( [ eq : vcond ] ) . the boundary conditions at the right interface give similar results for the transmitted terms , but without the incident wave terms . finally , we combine the discretised interface coupling conditions with the interior finite difference equations . this leads to the matrix problem : @xmath155 \left[\begin{matrix } \mathbf{r}\\ \hat{\mathbf{u}}_\delta\\ \mathbf{t } \end{matrix } \right]= \left[\begin{matrix } * \\ \mathbf{0}\\ \mathbf{0 } \end{matrix } \right]\ ] ] with scattering coefficients @xmath156^\text{t},\ ] ] and @xmath157^\text{t}.\ ] ] here , the finite difference solution in the interior of @xmath20 is represented by @xmath158^\text{t},\:\ : i=1,\dots , n-1.\ ] ] note that the stencil for the second derivative fd operator within @xmath159 takes boundary data from the plane wave terms at the interface points @xmath160 and @xmath161 , whereas the fourth derivative operator extends beyond this and also takes boundary data from within the scattering region at @xmath162 and @xmath163 . in this section we discuss the numerical solution of the full wave scattering problem derived in sect . [ sec2 ] and compare the results to those obtained using an ode time - stepper for the hamiltonian system presented in sect . [ sec : raytrac ] . the parameters chosen for all calculations are summarised in [ sec : parameters ] . for the cylindrical region , these parameters correspond to a circularly cylindrical steel shell as considered in refs . @xcite . note that whilst the width of the cylindrical ridge studied here is on the millimeter scale , it only represents a small region of the larger structures and components that serve to motivate this study . an incident bending mode in the left part of @xmath19 is used to excite the system and is sent in to the interface with @xmath20 at various angles , corresponding to a variation in the trace wave number @xmath164 . the latter is chosen according to the values given by the dispersion curve . in all cases we set @xmath165 rad / s , which is large enough so that waves will always transmit straight through the cylindrical region ( as though it were a flat plate ) when @xmath166 , that is when waves approach the cylindrical region directed parallel to the @xmath60-axis . we note that in the high frequency regime this behaviour will be preserved for almost all incident waves when both @xmath167 and @xmath168 are positive . however , the choice of @xmath55 here corresponds to the more interesting mid - frequency case where reflections are also possible for a range of sufficiently large @xmath169 . phase - plane showing two fixed points at @xmath24 and the separatrix ( solid bold curve ) which connects them as it traverses the cylindrical ridge . ( b ) a close up view of the upper fixed point and separatrix region . the upper fixed point divides the incoming rays from the left hand side ( dashed lines ) into reflecting or transmitting trajectories . the bold dashed line depicts the extremal transmitting trajectory closest to the set of reflecting rays . in both subplots , the solid line trajectories are those which are trapped inside the ridge . ] in this section we consider the reflection/ transmission behaviour of rays corresponding to an incident bending mode ( see fig . [ fig : phasespace4 ] ) . a range of incoming wavenumbers ( directions ) are used corresponding to the strip in the upper left corner of the figure . we find that the rays transmit for sufficiently large @xmath78 , whereas for smaller positive values of @xmath78 the rays reflect . note that the symmetry of the problem means that rays also transmit for @xmath78 negative when the incoming ray is from the right hand side and @xmath170 is large enough . the threshold value of @xmath171 for the change from reflectance to transmittance corresponds to a hyperbolic fixed point and its location can be found conveniently by considering the sizes of the dispersion curves as discussed in @xcite . the upper fixed point shown in both parts of fig . [ fig : phasespace4 ] gives a threshold value of @xmath172 . we will also investigate the existence and location of such a threshold incoming wave direction in the finite difference solution of the full wave problem described in the next section . depending on the angle of incidence , fig . [ fig : phasespace4 ] shows that the point of reflection varies from the centre of the ridge at @xmath24 to slightly towards the flat region to the left . the smooth curvature model here therefore differs from the discontinuous curvature models considered in refs . @xcite , for which the reflection takes place off - centre at a fixed location only . that is , for the models presented in refs . @xcite the reflection would take place at the location of the jump in the curvature between the flat region and the cylindrical region . this would be off - centre in the example here , since the centre point @xmath24 corresponds to the centre of the cylindrical region . in the next section we consider an equivalent scattering problem to the one above , but instead using the finite difference model described in the previous section to numerically solve the full wave problem . the full wave model will include the phase information omitted in the pure ray approximation applied in this section , but at the cost of a computational expense which scales with frequency . such a study is feasible up to reasonably high frequencies due to the one - dimensional setup of the problem derived in sect . however , the high frequency purely transmissive behaviour for the problem here is relatively straightforward to predict , and we find that both methods may be used for the more interesting mid - frequency case studied here . we restrict the study to bending excitations as in the previous section . for smooth joints this means that only the bending mode is active , with negligible conversion to in - plane modes . that is , the scattering probabilities become @xmath173 ) with very small @xmath35 in the interpolation function ( [ eq : interp ] ) , mode conversions appeared in the numerics at almost normal incidence . we defer the study of this case to future work and note the possibility of using discontinuous joints instead for this case . note that we use the term probabilities to describe the coefficients defined in ( [ eq : prob ] ) since for an incoming bending wave of unit amplitude ( [ eq : winc ] ) , conservation of energy gives that @xmath174 . in the full wave picture considered here these coefficients ( [ eq : prob ] ) actually give the proportions of reflected and transmitted wave energy . however , for comparison with the ray tracing results in the last section , they give the probabilities of reflection and transmission for a particular incoming trajectory . plotted on a linear scale . the plot shows anti - resonances close to @xmath175 , @xmath176 and @xmath177 . there is a switch in the dominant behaviour from transmission to reflection close to @xmath178.,scaledwidth=90.0% ] plotted on a logarithmic scale . the logarithmic scale emphasises the anti - resonance reflection peaks and the switch from transmission to reflection that were evident from fig . [ fig : coeff].,scaledwidth=90.0% ] the dependence of the reflection and transmission probabilities on the value of the trace wave number @xmath168 is shown in fig . [ fig : coeff ] . the transmission and reflection probabilities have been computed using the finite difference discretisation described in sect . [ sec : scatscheme ] for both @xmath179 and @xmath180 grid points in @xmath20 , corresponding to @xmath181 m and @xmath182 m , respectively . the ratio of wavelength to @xmath137 is @xmath183 for the case @xmath179 and hence the rule of thumb requiring 6 points per wavelength for reliable results suggests that in both cases our results should be reasonably well converged . this convergence is also evident from fig . [ fig : coeff ] , since it is difficult to detect differences between the plots for @xmath179 and @xmath180 . fig . [ fig : coeff ] shows that the scattering coefficients of the full wave problem include discrete anti - resonance points ( points of perfect reflection ) that were not present in the ray tracing calculation . in addition , the threshold behaviour from transmission to reflection has been smoothed in the full wave calculations , whereas for the ray tracing solution there is a sudden jump from transmission to reflection corresponding to the location of the hyperbolic fixed point . hence , for the calculations in this section there is a region of @xmath168 values where both reflection and transmission take place at the same time . this region lies roughly between @xmath184 and @xmath185 , but one observes an obvious switch from dominant transmission behaviour to dominant reflection behaviour at a threshold value around @xmath186 , which is close to the threshold estimate ( @xmath187 ) from the corresponding ray tracing calculation . it appears therefore that the asymptotic ray model slightly over predicts the transmission / reflection threshold compared with the full wave solution , but still produces a reliable estimate within around @xmath188 of the total @xmath168 range . a more complete picture appears when the scattering coefficients are plotted on a logarithmic scale , see fig . [ fig : logcoeff ] . we can then clearly identify the discrete reflection points in the region of perfect transmission . the dips or peaks can be arbitrarily close to zero , or unity , depending on the resolution of the plot . the discrepancy between the finite difference solution for the reflection coefficient with @xmath179 and @xmath180 is also more evident in this figure since the logarithmic scaling amplifies the differences at very small values of @xmath189 below @xmath190 . however , the positions of the anti - resonances and the switch from transmission to reflection remain in excellent agreement for both @xmath179 and @xmath180 . note also that the values of @xmath191 computed with @xmath179 and @xmath180 are in good agreement with one another , even for very small values . $ ] for @xmath192 ( plots ( a ) to ( c ) ) , @xmath177 ( plots ( d ) to ( f ) ) , @xmath193 ( plots ( g ) to ( i ) ) and @xmath194 ( plots ( j ) to ( l ) ) . each plot shows a displacement in metres against its corresponding @xmath60-coordinate , also in metres . plots ( c ) , ( f ) , ( i ) and ( l ) show the bending mode @xmath71 which exhibits transmission in plot ( c ) , anti - resonance in plot ( f ) , mixed reflection and transmission for the threshold region in plot ( i ) and reflection in plot ( l ) . the remaining plots show the in - plane wave solution , which remains confined inside the cylindrical region in all cases . solid and dashed lines show the real and imaginary parts , respectively . ] figure [ fig : wfmiddle ] shows the behaviour of the wave function solutions @xmath59^\text{t}$ ] for various values of the trace wavenumber @xmath168 . the finite difference computations for these plots were all carried out using @xmath180 grid points . note that there is no mode conversion and so the in - plane contributions @xmath72 and @xmath73 only have support on the curved region and vanish in @xmath19 . in the case of reflection ( @xmath194 ) and transmission ( @xmath192 ) the expected behaviour is observed . that is , for the reflective case the bending wave function @xmath195 localises to the left as shown in plot ( l ) of fig . [ fig : wfmiddle ] . in the ray tracing model one can pinpoint an exact reflection point , but for the wave problem one observes a decay in amplitude as the wave enters the cylindrical region rather than a hard wall reflection . in the case of transmission shown in plot ( c ) of fig . [ fig : wfmiddle ] , the bending wave continues through the ridge without significant deformation from the shape of the initial plane wave excitation . this is equivalent to the corresponding ray tracing result . however , if @xmath168 is chosen coincide with one of the anti - resonance peaks in the transmission region that were not present in the ray tracing result , then the bending wave field is localised in the central cylindrical region ; for example , at @xmath177 as shown in plot ( f ) of fig . [ fig : wfmiddle ] . the bending wave function @xmath195 also localises in the cylindrical region when @xmath168 is chosen in the threshold region as shown in plot ( i ) of fig . [ fig : wfmiddle ] , which shows a plot of @xmath195 at @xmath193 . we have investigated bending wave scattering across a smoothed plate and quarter cylinder configuration in the interesting mid - frequency case , close to the ring frequency for the cylinder . results have been obtained in a simplified and effectively one - dimensional setting using both a high - frequency ray - tracing approximation and a finite difference discretisation of the equivalent full - wave problem . previous studies on plate and cylinder connections have concentrated on joints having discontinuous curvature where reflections can occur , however the smoothed joints considered here also give rise to reflections within the cylindrical part of the structure . furthermore , the ray tracing calculations suggest the existence of a threshold incident wave direction which separates waves or rays that exhibit reflective or transmissive behaviour . this threshold direction is also evident from the smoothed transition between reflection and transmission observed in the full wave calculations . hence , relatively simple scattering laws can be employed to model the propagation of structure - borne noise in shells , and ultimately in built - up structures containing thin shell components . the full wave solution shows that in addition to the switch from transmission to reflection as the incident wave direction becomes increasingly oblique , there are also anti - resonances giving rise to perfect reflection at a discrete set of directions where transmission would typically be expected . in these cases the wave functions appear as trapped modes that localise in the cylindrical part of the configuration . there are several avenues for further research , including extensions to multiply - curved and fluid - loaded shells . in addition , the scattering laws could be incorporated within computed aided engineering simulations of built - up structures via wave methods , such as the wave and finite element method or dynamical energy analysis . support from the eu ( fp7-people-2013-iapp grant no . 612237 ( mhivec ) ) is gratefully acknowledged . we also wish to thank dr gregor tanner for stimulating discussions and dr jonathan crofts for carefully reading the manuscript . 99 o.c . zienkiewicz , r.l . taylor , j.z . zhu , _ the finite element method : its basis and fundamentals _ , 7th edition , butterworth - heinemann , oxford , 2003 . lyon , r.g . dejong , _ theory and application of statistical energy analysis _ , 2nd edition , butterworth - heinemann , boston ma , 1995 . glasser , _ an introduction to ray tracing _ , academic press , new york , 1989 . a. le bot , _ foundation of statistical energy analysis in vibroacoustics _ , oxford university press , oxford , 2015 . h. kuttruff , _ room acoustics _ , 4th edition , spon , london , 2000 . v. cerveny , _ seismic ray theory _ , cambridge university press , cambridge uk , 2001 . norris , d.a . rebinsky , membrane and flexural waves on thin shells . _ journal of vibration and acoustics _ 116 ( 1994 ) 457467 . doi:10.1115/1.2930449 e.g. williams , b.h.houston , j.a . bucaro , experimental investigation of the wave propagation on a point - driven , submerged capped cylinder using k - space analysis . _ journal of the acoustical society of america _ 87(2 ) ( 1990 ) 513522 . doi:10.1121/1.398922 r.s . langley , wave motion and energy flow in cylindrical shells . _ journal of sound and vibration _ 169(1 ) ( 1994 ) 2942 . doi:10.1006/jsvi.1994.1004 a.n . norris , reflection and transmission of structural waves at an interface between doubly curved shells . _ acta acustica united with acustica _ 84(6 ) ( 1998 ) 10661076 . g. tanner , dynamical energy analysis - determining wave energy distributions in vibro - acoustical structures in the high - frequency regime . _ journal of sound and vibration _ 320 ( 2009 ) 10231038 . doi:10.1016/j.jsv.2008.08.032 d.j . chappell , d. lchel n. sndergaard , g. tanner , dynamical energy analysis on mesh grids : a new tool for describing the vibro - acoustic response of engineering structures . _ wave motion _ 51(4 ) ( 2014 ) 589597 . doi:10.1016/j.wavemoti.2014.01.004 e. manconi , b.r . mace , wave characterization of cylindrical and curved panels using a finite element method . _ journal of the acoustical society of america _ 125(1 ) ( 2009 ) 154163 . doi:10.1121/1.3021418 j.m . renno , b.r . mace , calculation of reflection and transmission coefficients of joints using a hybrid finite element / wave and finite element approach . _ journal of sound and vibration _ 332(9 ) ( 2013 ) 21492164 . doi:10.1016/j.jsv.2012.04.029 j.t . kayija , the rendering equation . _ acm siggraph computer graphics _ 20(4 ) ( 1986 ) 143150 . doi:10.1145/15886.15902 p. gaspard . _ chaos , scattering and statistical mechanics _ , cambridge university press , cambridge uk , 1998 . p. cvitanovi , r. artuso , r. mainieri , g. tanner , g. vattay _ chaos : classical and quantum _ , chaosbook.org , niels bohr institute , copenhagen , denmark , 2012 . pierce , waves on fluid - loaded inhomogeneous elastic shells of arbitrary shape . _ journal of vibration and acoustics _ 115(4 ) ( 1993 ) 384390 . doi:10.1115/1.2930361 w. flgge , _ tensor analysis and continuum mechanics _ , springer , berlin , 1972 . van loan , the ubiquitous kronecker product . _ journal of computational and applied mathematics _ 123 ( 2000 ) 85100 . doi:10.1016/s0377 - 0427(00)00393 - 9 d.a . rebinsky , a.n . norris , dispersion of flexural waves on shells . _ journal of vibration and acoustics _ 118(3 ) ( 1996 ) 526529 . doi:10.1115/1.2888218 n. sndergaard , d.j . chappell , g. tanner , tracking vibrational energy on curved shell structures in the mid - high frequency limit - a ray - tracing approach . _ proceedings of noise and vibration - emerging technologies ( novem ) 2015 _ , dubrovnik , croatia , 2015 . m. do carmo , _ differential geometry of curves and surfaces _ , prentice hall , new jersey , 1976 . t. frankel , _ the geometry of physics _ , cambridge university press , new york , 2004 . several physical theories are concisely written in tensor form @xcite . the tensor formalism has also proved to be useful in continuum mechanics . the components of tensors are with respect to a given choice of coordinates @xmath196 . a change of coordinates from @xmath196 to @xmath197 leads to expressions for components in the new coordinates in relation to those of the old . for example , given a tensor of type @xmath198 ( that is , with one superscript and two subscript indices ) the transformation takes the form @xmath199 where jacobians and inverse jacobians have been used for the change of coordinates . we have adopted the important summation convention for repeated indices ; this indicates that a summation is to take place over the repeated index ( unless otherwise stated ) . the metric tensor @xmath200 , known as the first fundamental form , measures distances via @xmath201 the inverse of the metric tensor @xmath15 obeys @xmath202 and is used to alter the type of a tensor . for example , we can introduce a new tensor of type ( 1,1 ) @xmath203 from a tensor of type ( 0,2 ) by raising the indices of the latter tensor . likewise @xmath200 lowers indices . the directional derivative in arbitrary coordinates is generalized to the covariant derivative @xmath204 . this derivative may be introduced from an embedding using projection of gradients @xcite or intrinsically with quantities only related to the curved space itself @xcite . for example , in a coordinate basis , the covariant derivative of a tensor of type ( 1,1 ) becomes @xmath205 with @xmath206 the christoffel symbols . the computations throughout this work were done using the following parameter choices @xcite :

it is well known that any link can be represented by the closure of a braid . the minimum number of strands needed in a braid whose closure represents a given link is called the _ braid index _ of the link . defined as the extreme value of a quantity over an infinite family of links that are topologically equivalent like other link invariants ( such as the minimum crossing number ) , the braid index of a link is hard to compute @xcite in general . in the case of the minimum crossing number , there was a long standing conjecture which states that for a reduced alternating link diagram , the number of crossings in the diagram is equal to the minimum crossing number of the link . it is known that the span of the jones polynomial @xcite of a link gives a lower bound on the crossing number of the link . in the case of a reduced alternating link diagram , it was shown that the span of the jones polynomial of a link equals the number of crossings in the diagram , which leads to the proof of the conjecture @xcite . in the case of braid index , there is a similar inequality relating the braid index of a link to the @xmath1-span of its homfly polynomial ( which is a polynomial of two variables @xmath2 and @xmath1 to be defined in the next section ) . s. yamada proved that any link diagram of a given link @xmath3 with @xmath4 _ seifert circles _ can be realized as the closure of a braid on @xmath4 strands , which implies that the braid index of an oriented link @xmath3 equals the minimum number of seifert circles of all link diagrams of @xmath3 @xcite . in @xcite , h. morton showed that the number of seifert circles of a link @xmath3 , hence the braid index of @xmath3 ( in light of yamada s result ) , is bounded from below by @xmath1-span@xmath5 ( which is called the morton - frank - williams inequality , or mfw inequality for short ) . in analogy to the crossing number conjecture for a reduced alternating link diagram , k. murasugi conjectured that the number of seifert circles , hence the braid index , in such a diagram equals @xmath1-span@xmath5 ( the murasugi conjecture ) @xcite . although this conjecture turned out to be false in general @xcite ( for example the knot @xmath6 has 4 seifert circles , but the @xmath1-span of its homfly polynomial is 4 so @xmath1-span@xmath7 and the braid index of @xmath6 is also 3 ) , researchers had shown that the mfw inequality is sharp for many classes of links ( including some non - alternating ones ) hence the @xmath1-span of the homfly polynomial for these links can be used to determine their braid index . examples include the closed positive braids with a full twist ( in particular the torus links ) @xcite , the 2-bridge links and fibered alternating links @xcite , and a new class of links discussed in a more recent paper by s. y. lee and m. seo @xcite . for more readings on related topics , interested readers can refer to j.s . birman and w.w . menasco @xcite , p.r . cromwell @xcite , e.a . elrifai @xcite , h. morton , h. b. short @xcite , t. nakamura @xcite and a. stoimenow @xcite . motivated by this question , in this paper the authors seek a special and explicit formulation of the homfly polynomial for certain classes of links where the explicit forms of the homfly polynomial would allow us to analyze and derive the @xmath1-spans of the homfly polynomials of these links . our main result expresses the homfly polynomial of a link presented in a closed braid diagram form in two explicit formulas . we show that one of our formulas of the homfly polynomial is equivalent to the expansion derived by f. jaeger @xcite . however , our approach is combinatorial in nature and the proof of the formula is shorter than the proof in @xcite . we show that the morton - frank - williams inequality is an immediate consequence of these two homfly polynomial formulas . as a more significant application of our result , we use it to show that if a link has an @xmath0-string closed braid diagram that is also reduced and alternating , then the braid index of the link is exactly @xmath0 . the proof of this is direct and short . our approach is very different from the proof of murasugi s more general results @xcite on oriented alternating fibered links , which is inductive in nature . this paper is structured as follows . in section [ s2 ] , we introduce basic definitions and terminology about link diagrams , braids and the homfly polynomial . we introduce two special classes of resolving trees for closed braids , in which every vertex is a closed braid and the leaf vertices are closed braids that represent trivial links . we call these resolving trees descending trees and ascending trees respectively . in section [ s3 ] , we state and prove our main result , namely two formulas expressing the homfly polynomial of a closed braid as a total contribution of all leaf vertices in our trees and show that the morton - frank - williams inequality @xcite is a direct consequence of these two formulas . in section [ s4 ] , we show that the @xmath1-span of the homfly polynomial of a reduced alternating braid on @xmath0 strands is exactly @xmath8 and an application of the the morton - frank - williams inequality shows that the braid index of a reduced alternating braid equals the number of strands in the braid . as another application of our main result from section [ s3 ] , we also give a short proof that the leading coefficient of the alexander polynomial of such a closed braid equals @xmath9 . finally , in section [ acp ] , we show that one of our formulas is equivalent to the expansion of the homfly polynomial derived by f. jaeger @xcite based on the concept of admissible circuit partitions . we assume that the reader has the basic knowledge about the definition of a link and refer a reader without such knowledge to a textbook such as @xcite . for the convenience of the reader , however , we will review one important result that is needed in section [ s3 ] . figure [ r - moves ] defines three moves one can make on a link diagram without changing its topology , and these are called reidemeister moves of type i , ii and iii . in 1926 , k. reidemeister @xcite proved that two link diagrams represent the same link if and only if one diagram can be changed to the other through a finite sequence of reidemeister moves . for a given oriented link diagram @xmath3 , we assign each crossing @xmath10 or @xmath11 according to its sign as defined in figure [ fig : cross ] . the writhe of @xmath3 , written as @xmath12 , is the sum of these @xmath10 s and @xmath11 s over all crossings of @xmath3 . however , if we only sum the @xmath10 s and @xmath11 s over crossings between two different components @xmath13 and @xmath14 of the link , then this number is always even and half of it is called the _ linking number _ between @xmath13 and @xmath14 . using the reidemeister moves , it is easy to show that the linking number is a link invariant . in particular , if @xmath13 and @xmath14 are separated by a topological plane ( in which case @xmath13 and @xmath14 are said to be _ splittable _ ) , then the linking number between @xmath13 and @xmath14 is zero . we make the following remark for future reference . its proof is trivial and is left to the reader . [ r_remark ] _ reidemeister moves of type ii and iii do not change the writhe of a link diagram hence link diagrams related by a finite sequence of reidemeister moves ii and iii have the same writhe . in particular , if a link diagram @xmath15 is obtained through @xmath16 by a finite sequence of moves that involve only deformation of segments of the link within planes that are parallel to the projection plane of the link diagram , then @xmath17 since such moves do not introduce reidemeister moves of type i. also , if @xmath3 is a link with splittable components @xmath13 , @xmath14 , , @xmath18 , then @xmath19 . _ a crossing in a link diagram is called _ nugatory _ if there is a simple closed curve such that the link diagram intersects the curve only at the crossing and the link diagram intersects both components of the complement of the curve . a link diagram is said to be _ reduced _ if no crossing in the diagram is nugatory . a link diagram is _ alternating _ if as we travel through the link diagram by any given orientation , the strands go through the crossings alternately between overpasses and underpasses . for example , the closure of the braid in figure [ fig : bd ] is an alternating link diagram . as we mentioned in the introduction , reduced alternating link diagrams are special since the number of crossings in a reduced alternating link diagram is the minimum crossing number of the link @xcite . consider @xmath20 as the standard euclidean @xmath21-plane . a _ braid diagram _ ( or just a _ braid _ ) on @xmath0 strands is a set @xmath22 $ ] consisting of @xmath0 curves called @xmath23 of @xmath24 such that the following four conditions are met . first , each strand is a monotonic curve in the @xmath25 direction . second , every point of @xmath26 is a starting point of a unique strand and every point of @xmath27 is an end point of a unique strand . third , every point of @xmath28 belongs to at most two strands . a point that belongs to two strands is called a _ crossing_. at each crossing , one strand is identified as an _ overpass _ and the other is as an _ underpass _ @xcite . fourth , there is at most one crossing in @xmath29 for each @xmath30 $ ] . note that the second condition gives the braid diagram a downward orientation , so the closure of a braid diagram is an oriented link diagram . treated as topological objects , one can speak of topological equivalence of braid diagrams , and such equivalence relations provide the foundations for one to treat the braids as elements in the algebraic objects called the _ braid groups_. not to deviate from our main task , we will only point out that a braid group @xmath31 is a group with @xmath32 generators @xmath33 satisfying certain relations . an element of @xmath31 is a word of these generators and each letter in the word corresponds to a crossing in the braid . an example of a braid on 5 strands and its counterpart in @xmath34 is given in figure [ fig : braid ] . please refer to @xcite for more details on braid groups . in this paper we are only interested in closed braids as topological objects , not the braids in the algebraic sense . for this reason , _ we will not distinguish a braid and its closure _ , that is , the word braid ( or a symbol of it ) can either represent the braid itself or its closure . the reader should rely on the context to determine its meaning ( in many cases it really does not matter ) . we define the _ label _ of a strand by the @xmath35 coordinate of its starting point and we denote the corresponding mapping by @xmath36 , that is , if a strand @xmath37 starts at @xmath38 , then @xmath39 . on the other hand , the mapping @xmath40 that takes the label of a strand to the @xmath35 coordinate of its ending point defines a permutation of the labels ( namely the integers from @xmath41 to @xmath0 ) . denote this permutation by @xmath42 and write it as a product of disjoint cycles , we have @xmath43 where @xmath44 is the label of the first strand in the @xmath45-th cycle , @xmath46 is the label of the second strand in the @xmath45-th cycle , ... , and so on , @xmath47 is the number of cycles in the permutation . furthermore , we can re - arrange the orders of the cycles and the numbers in each cycle so that @xmath44 is the smallest integer in each cycle for each @xmath45 , and @xmath48 if @xmath49 . we call this special form of @xmath42 the _ standard form_. from now on , @xmath42 will always be expressed in its standard form . note that the standard form of @xmath42 defines a total order among the strand labels , namely @xmath50 we call this order the _ return order _ of the strands in the braid diagram @xmath24 . we call each @xmath44 in @xmath42 the _ pivot label _ within its corresponding cycle and @xmath51 the _ pivot point _ of the cycle when @xmath42 is expressed in its standard form . note that each cycle in @xmath42 corresponds to a connected component in the closed braid diagram and we can travel through @xmath24 by traveling through each such component . we say that we travel through @xmath24 _ naturally _ if we travel along the strands of @xmath24 in its return order , starting from the pivot point at each component and follow the orientation of the braid @xmath24 . on 5 strands that has 8 crossings . ] _ consider the braid diagram @xmath24 shown in figure [ fig : braid ] whose corresponding word in @xmath34 is @xmath52 . the dotted curves show how the braid is closed to form a link diagram . the circled points on top are pivot points of the three components . we have @xmath53 ( in its standard form ) , so its return order is @xmath54 . one can see that if we close the braid using disjoint curves starting and ending at the points @xmath55 , @xmath56 for each @xmath57 ( shown in figure [ fig : braid ] with dotted curves ) , strands @xmath41 and @xmath58 are in the same component , strand @xmath59 is a component by itself and strands @xmath60 and @xmath61 are in the same component . the pivot labels of these components are @xmath41 , @xmath59 and @xmath60 respectively . _ let @xmath62 be a crossing in a braid diagram @xmath24 , @xmath63 be the overpassing strand at @xmath62 and @xmath64 be the underpassing strand at @xmath62 . we say that @xmath62 is _ descending _ if @xmath65 and _ ascending _ if @xmath66 . if all crossings in @xmath24 are descending ( ascending ) , then we say that @xmath24 is a descending ( ascending ) braid diagram . for example , all crossings except the circled one in figure [ fig : braid ] are descending crossings . switching the circled crossing to a descending crossing will then results in a descending braid . [ descending_remark ] _ descending ( ascending ) braid diagrams have the easy and well known property that the closure of any such braid diagram is a trivial link , that is , a link topologically equivalent ( ambient isotopic ) to the disjoint union of several circles contained in the same plane . take the braid given in figure [ fig : braid ] as an example by flipping the circled crossing to make the braid descending . using the order @xmath54 , we can actually move strands 1 , 4 , 2 , 3 and 5 into planes @xmath67 , 4 , 3 , 2 and 1 respectively so that their projections would still give the descending braid , and that this move will not change the topology of the corresponding link ( since there are no crossing changes in the process ) . it is then easy to see that each component corresponding to a cycle @xmath68 , @xmath69 or @xmath70 is an unknot and these knots are splittable since they are separated by planes . _ let @xmath71 , @xmath72 , and @xmath73 be oriented link diagrams that coincide except at a small neighborhood of a crossing where the diagrams are presented as in figure [ fig : cross ] : the crossing in @xmath71 ( @xmath72 ) is positive ( negative ) and is assigned @xmath10 ( @xmath11 ) in the calculation of the writhe of the link diagram . we say the crossing presented in @xmath71 is of a _ positive _ sign and the crossing presented in @xmath72 is of a _ negative _ sign . the following result appears in @xcite . [ ho ] there is a unique function that maps each oriented link diagram @xmath3 to a two - variable laurent polynomial with integer coefficients @xmath74 such that 1 . if @xmath16 and @xmath15 are ambient isotopic , then @xmath75 . 2 . @xmath76 . 3 . if l is an unknot , then @xmath77 . the laurent polynomial @xmath74 is called the _ homfly polynomial _ of the oriented link @xmath3 . the second condition in the proposition is called the _ skein relation _ of the homfly polynomial . with conditions ( 2 ) and ( 3 ) above , one can easily show that if @xmath3 is a trivial link with @xmath0 connected components , then @xmath78 ( by applying these two conditions repeatedly to a simple closed curve with @xmath32 twists in its projection ) . for our purposes , we will actually be using the following two equivalent forms of the skein relation : @xmath79 a rooted and edge - weighted binary tree @xmath80 is called a _ resolving tree _ of an oriented link diagram @xmath3 ( for the homfly polynomial ) if the following conditions hold . first , every vertex of @xmath80 corresponds to an oriented link diagram . second , the root vertex of @xmath80 corresponds to the original link diagram @xmath3 . third , each leaf vertex of @xmath80 corresponds to a trivial link . fourth , if we direct @xmath80 using the directions of the paths from the root vertex to the leaf vertices , then under this direction any internal vertex has exactly two children vertices and the corresponding link diagrams of these three vertices are identical except at one crossing and they are related by one of the two possible relations at that crossing as shown in figure [ fig : rtree ] , where the edges are weighted and the directions of the edges coincide with the direction of @xmath80 . [ tree_formula_remark ] _ if @xmath3 admits a resolving tree @xmath80 , then one can easily show that @xmath74 is a summation in which each leaf vertex of @xmath80 contributes exactly one term in the following way . let @xmath81 be the trivial link corresponding to a leaf vertex in @xmath80 and let @xmath82 be the unique path from the root ( @xmath3 ) to the leaf vertex ( @xmath81 ) . then the contribution of the leaf vertex is simply @xmath83 multiplied by the weights of the edges in @xmath82 , where @xmath84 is the number of components in @xmath81 . it is known ( and not hard to prove ) that resolving trees exist for any given oriented link diagram @xmath3 , and that they are not unique in general . if @xmath85 is the mirror image of @xmath3 , a resolving tree for @xmath85 can be obtained from a resolving tree of @xmath3 by taking mirror images of all link diagrams in it and replacing @xmath1 by @xmath86 and @xmath2 by @xmath87 in the edge weights . it follows the well known fact that @xmath88 . _ ) and ( [ skein2 ] ) . ] let @xmath89 be a braid on @xmath0 strands and let us define an algorithm ( let us call it algorithm d ) that operates on @xmath89 as follows . if @xmath89 is descending , then the algorithm does not do anything and simply returns @xmath89 . if @xmath89 is not descending , then it contains at least one ascending crossing . in this case , let us travel through @xmath89 naturally , until we run into the first ascending crossing @xmath62 , which can be a positive or a negative crossing as shown in either the left hand side or the right hand side of figure [ fig : rtree ] . the algorithm then returns the two corresponding braids split from the original one as shown in figure [ fig : cpd ] , which we will name as @xmath90 and @xmath91 ( @xmath92 for flipping " the crossing and @xmath37 for smoothing " the crossing ) . notice that @xmath90 has one less ascending crossing than @xmath89 does since @xmath90 and @xmath89 are identical ( including the return order of the strands ) except at the crossing @xmath62 . but one can not say the same for @xmath91 since it does not share the same strands ( and the return order of the strands ) with @xmath89 . however , @xmath89 and @xmath91 share the same strands and crossings up to crossing @xmath62 ( which are all descending ) while travel through them naturally , and @xmath91 has one less crossing than @xmath89 . so if algorithm d is repeatedly applied to a braid and the resulting braids of this operation , then this process will end after a finite number of repeats ( in fact this number is bounded above by the number of crossings in @xmath24 ) . it follows that we can construct a special resolving tree @xmath80 for @xmath93 ( as a link diagram ) as follows . we apply algorithm d to @xmath24 first if it is not descending , and then apply the algorithm again to the two resulting braids ( if they are not descending ) , and so on , until all the leaf vertices are descending braids . the closures of the braids involved in the process are the vertices of @xmath80 . in particular , the closures of the resulting descending braids form the leaf vertices of @xmath80 since they are all trivial links . by assigning appropriate weights to the edges of @xmath80 , one can easily verify that @xmath80 is indeed a resolving tree of @xmath93 . by the way it is constructed , @xmath80 is unique . in a similar fashion , we can also construct another ( unique ) resolving tree of @xmath93 by replacing descending " with ascending " in the above ( the algorithm corresponding to algorithm d would be called algorithm a ) . to distinguish the two , we will call the first the descending tree " and the later the ascending tree " of @xmath93 , and denote them by @xmath94 and @xmath95 respectively . an example of a descending tree is shown in the right hand side of figure [ fig : cpd ] . ; right : the descending tree of ( the closure of ) the braid on the top left ( only the braids are shown ) . ] [ r_cor ] _ for a given braid @xmath81 obtained from @xmath24 by flipping and smoothing some of its crossings , there is a simple way to check whether @xmath96 by the way @xmath97 is constructed . we travel through @xmath81 naturally and visit each crossing of @xmath24 exactly twice . for each crossing of @xmath24 that we encounter in this process for the first time ( including the smoothed ones , these are marked by small circles in figure [ fig : cpd ] ) we perform the following test . if this crossing is smoothed in @xmath81 , we check whether we would be approaching it from its underpass if the crossing ( in its original form in @xmath24 ) were not smoothed . on the other hand , if the crossing is also a crossing in @xmath81 ( which may or may not have been flipped from its original form in @xmath24 ) , we check whether we are approaching it from its overpass . if all crossings pass this check , then @xmath98 , otherwise @xmath99 . _ in this section , we derive the main result of this paper , namely the two formulas of the homfly polynomial of a braid @xmath93 based on @xmath94 and @xmath95 respectively , given in theorem [ p3 ] . let @xmath89 be a vertex in @xmath94 ( or @xmath95 ) , @xmath100 be the writhe of @xmath89 , @xmath101 be the number of components in @xmath102 . note that @xmath89 is obtained from @xmath24 by applying algorithm d ( or algorithm a ) repeatedly , and in this process some crossings of @xmath24 may have been smoothed . let @xmath103 be the number of smoothed crossings and @xmath104 be the number of smoothed crossings that are negative . it is trivial to note that @xmath103 is simply the difference between the number of crossings in @xmath24 and the number of crossings in @xmath89 . [ l3 ] if @xmath81 is a descending braid on @xmath0 strands , then @xmath105 . on the other hand , if @xmath106 is an ascending braid on @xmath0 strands , then @xmath107 . since @xmath81 is descending , by remark [ descending_remark ] , @xmath108 can be realized by such space curves that its components are separated by planes that are parallel to the @xmath21-plane , hence the writhe contribution of crossings whose strands belong to different components is zero . let @xmath109 , @xmath13 , @xmath14 , , @xmath18 be the cycles of @xmath110 and @xmath111 be the length of the cycle @xmath112 . we have @xmath113 and @xmath114 . we claim that for each component @xmath115 of @xmath108 , @xmath116 . for simplicity let @xmath117 and let the labels of the strands in @xmath115 be @xmath118 , , @xmath119 , ordered by their return order in @xmath81 . without loss of generality , assume that strand @xmath120 is in the plane @xmath121 , strand @xmath122 is in the plane @xmath123 , ... and strand @xmath119 is in the plane @xmath124 . furthermore , the curve connecting the ending point of @xmath125 to the starting point of strand @xmath126 is bounded between the two planes @xmath127 and @xmath128 for @xmath129 . we first observe that each strand and part of the curves connecting to its ends can be deformed to straight line segments ( within the plane that it is in ) , in a form as shown in the left hand side of figure [ fig : layers ] , where each strand resides in a plane parallel to @xmath130 ( and the equation of the plane is marked in the figure ) . the part of the connecting curve attached to the strand residing in the same plane is marked by solid lines and the dotted curves have their end points on different planes parallel to the @xmath21-plane whose @xmath2 coordinates differ by exactly one , with the exception of the dotted curve on the far left ( which is between the planes @xmath124 and @xmath121 ) . with its strands and the connecting curves straightened . right : a special re - arrangement of the strands of @xmath115 through a finite sequence of reidemeister moves of type ii and iii . ] by remark [ r_remark ] , this deformation does not change the writhe . for the same reason , these straight line segments can freely slide within the plane they reside in since there are no other curves in that plane . see figure [ slide ] for one such move . in particular , nothing can preventing us from sliding them into the position where the strands @xmath131 are parallel lines arranged in this order from left to right , as shown in the right hand side of figure [ fig : layers ] . since all moves are made within the planes where these curves reside in , the writhe does not change by remark [ r_remark ] . we see that @xmath132 from the right hand side of figure [ fig : layers ] since there are exactly @xmath133 crossings in the projection and all of them are negative . thus @xmath134 , _ i.e. _ , @xmath105 . an ascending braid diagram @xmath106 is the mirror image of a descending braid diagram @xmath81 . it is known that @xmath135 and it follows that @xmath107 . is the only strand in the top plane and can be slid freely within the plane without causing self intersection of the link ; right : the same strand after a sliding move . ] [ p3 ] let @xmath24 be a braid on @xmath0 strands , @xmath136 and @xmath137 be the set of leaf vertices of @xmath94 and @xmath95 respectively , then @xmath138 @xmath139 where @xmath140 is the number of crossings in @xmath24 that are smoothed in obtaining @xmath81 and @xmath141 is the number of negative crossings among these smoothed crossings . let us consider the descending tree first . by remark [ tree_formula_remark ] , the contribution of @xmath142 to @xmath143 is @xmath83 multiplied by the weights of the edges on the unique path of @xmath94 from @xmath93 to @xmath108 . as shown in figure [ fig : rtree ] , the degree of @xmath1 in the weight of an edge is exactly the change of writhe from the starting vertex of the edge ( remember that it is directed from the root to the leaf ) to the ending vertex of the edge , whereas a @xmath2 term in the weight of the edge indicates that the ending vertex is obtained from the starting vertex by a crossing smoothing and a negative sign in the weight indicates that the smoothed crossing is a negative crossing . it follows that the total contribution of @xmath81 is @xmath144\end{aligned}\ ] ] by lemma [ l3 ] . this proves ( [ e1 ] ) . equation ( [ e2 ] ) can be proved in a similar fashion and is left to the reader . _ let @xmath3 be a link with braid index @xmath0 and @xmath93 a braid representation of @xmath3 where @xmath24 is a braid of @xmath0 strands . let @xmath145 and @xmath146 be the maximum and minimum degrees of @xmath1 in @xmath74 . the @xmath1-span of @xmath74 is defined as @xmath147 . since @xmath148 for any @xmath142 , formulas ( [ e1 ] ) and ( [ e2 ] ) imply that @xmath149 and @xmath150 . it follows that @xmath1-span@xmath151 , that is , the morton - frank - williams inequality is a direct consequence of theorem [ p3 ] . _ a link is _ splittable _ if there exist components of the link that lie on different sides of a topological plane in @xmath152 and a link is _ non - splittable _ if it is not splittable @xcite . a braid diagram is _ reduced _ if its closure is a reduced link diagram . a braid diagram is _ alternating _ if its closure is an alternating link diagram . an oriented link is a _ reduced alternating braid _ on @xmath0 strands if it can be represented by a reduced alternating braid diagram on @xmath0 strands . a reduced alternating braid is an alternating link but the converse is not true in general . figure [ fig : bd ] is an example a non - splittable , reduced alternating braid diagram . in this section , we prove the following theorem with a simple and direct proof based on our results from the last section . @xcite[theorem4.1 ] let @xmath24 be the closure of a reduced alternating braid on @xmath0 strands , then the braid index of @xmath24 is @xmath0 . _ the above theorem is a special case of a more general theorem on a class of oriented alternating fibered links due to murasugi @xcite where the proof relies on using a sequence of lemmas shown by induction . the alternating links in this class are the @xmath153-products of @xmath154 torus links ( which include the alternating closed braids ) @xcite and are also known to be fibered @xcite . the fact that the reduced alternating closed braids are fibered can be established directly by using another result due to murasugi , which states that an alternating link is fibered if the leading coefficient of its alexander polynomial is @xmath9 @xcite , and by proving that the leading coefficient of the alexander polynomial of a reduced alternating closed braid is indeed @xmath9 . this latter fact seems to be known but we failed to find a direct proof of it in the literature so we will provide a short one at the end of this section , which also serves as another application of theorem [ p3 ] and our method . we also note that fact that the reduced alternating closed braids are fibered is a direct consequence of a result due to stallings in which he proved that all homogeneous closed braids ( which include the alternating ones ) are fibered @xcite . _ before we proceed to the proof of the theorem , we note that it suffices for us to prove this result for reduced alternating braids that are non - splittable . since if not , say @xmath155 with the @xmath156 s being the non - splittable components of @xmath24 , we can simply apply our result to each @xmath156 . since the braid index of a link equals the sum of the braid indices of its non - splittable components , and @xmath1-span@xmath5 of @xmath157 is the sum of the ( @xmath1-span@xmath5 ) s of the @xmath158 s as one can easily check , the general result then follows . let @xmath24 be a braid diagram on @xmath159 strands . consider @xmath24 as a word of generators of the braid group @xmath31 . note that each crossing in the braid diagram corresponds to a letter @xmath160 or @xmath161 in the word @xmath24 . in a standard drawing of @xmath24 as a braid , a crossing corresponding to @xmath160 or @xmath161 is drawn in the space between the vertical line @xmath162 and @xmath163 in the @xmath21-plane . we call this space a _ gap_. a gap is _ odd _ if @xmath45 is odd , and _ even _ if @xmath45 is even . thus , a crossing corresponds to @xmath160 or @xmath161 is in an odd gap ( even gap ) if @xmath45 is odd ( even ) . if @xmath24 is non - splittable , then each gap in the braid diagram contains at least one crossing . if @xmath24 is also reduced , then each gap in the diagram contains at least two crossings . suppose @xmath24 is alternating , the overpass ( underpass ) of a crossing @xmath62 must be the underpass ( overpass ) at next crossing @xmath164 . so if @xmath164 is in the same gap , @xmath164 must have the same sign as @xmath62 and if @xmath164 is in an adjacent gap , @xmath164 must have the opposite sign . therefore , if @xmath24 is alternating , all odd gaps contain crossings of the same sign and all even gaps contain crossings of the opposite sign . if the first gap contains positive crossings , we say the alternating braid diagram @xmath24 is _ positive - leading _ and if the first gap contains negative crossings , we say the alternating braid diagram @xmath24 is _ negative - leading_. [ l4 ] let @xmath24 be a non - splittable reduced alternating braid diagram on @xmath0 strands , @xmath145 be the highest degree of @xmath1 in @xmath157 and @xmath146 the lowest degree of @xmath1 in @xmath157 . then @xmath165 and @xmath166 . it follows that @xmath1-span@xmath167 . a simple application of the morton - frank - williams inequality and lemma [ l4 ] , together with the note in the second paragraph of this section , immediately shows that if @xmath24 is a reduced alternating braid on @xmath0 strands , then the braid index of @xmath24 is exactly @xmath0 . we now proceed to prove lemma [ l4 ] . we will consider the positive - leading alternating braid diagrams first . bear in mind that for a positive - leading alternating braid diagram , all the odd gaps contain only positive crossings and all the even gaps contain only negative crossings . consider @xmath157 as a laurent polynomial of @xmath1 with coefficients in the ring of laurent polynomials in the variable @xmath2 . by ( [ e1 ] ) , the highest possible degree of @xmath1 is @xmath168 and the only leaf vertices @xmath98 that can make contributions to the term of @xmath157 with this @xmath1 degree must satisfy the condition @xmath169 . let us consider the braid @xmath170 obtained from @xmath24 by first smoothing all crossings in the odd gaps , then smoothing all crossings in even gaps except the first one and the last one and finally flip the sign of the last crossings in the even gaps . this is obvious by the checking method in remark [ r_cor ] . part of the @xmath170 contribution to @xmath172 is a term of the form @xmath173 . proof of claim 2 . by ( [ e1 ] ) , the contribution of @xmath170 to @xmath172 is @xmath174 we have @xmath175 and the result follows . claim 3 . for any @xmath98 , @xmath176 , the contribution of @xmath81 to @xmath172 either has its maximum degree in the variable @xmath1 less than @xmath168 , or has its degree in @xmath2 less than @xmath177 . proof of claim 3 . if @xmath178 then there is nothing to prove . assuming that @xmath169 , then the contribution of @xmath81 to @xmath157 is @xmath179 so the degree of @xmath2 is @xmath180 . we need to show that @xmath181 , that is , @xmath182 . since @xmath169 , the return order of @xmath81 is @xmath183 . it follows that each gap contains either no crossings or at least two crossings of @xmath81 since each strand , say it has label @xmath45 with starting point @xmath55 , has to go through a gap an even number of times in order for it to end at the point @xmath56 . we claim that @xmath81 contains the first crossing of @xmath24 in each even gap in its original sign . if this is not the case , let @xmath62 be the first such crossing that has been changed ( so it is either smoothed or flipped in @xmath81 ) . assume that @xmath62 is in the @xmath45-th gap ( with @xmath45 being even ) . notice that any strand with label @xmath184 less than @xmath45 can only cross the @xmath45-th gap into gaps to the right of the @xmath45-th gap and must return to the point @xmath185 at the end because the return order of @xmath81 is @xmath183 and the strands can only travel downward . it follows that the strand entering @xmath62 from its right side must have label greater than @xmath45 . now consider the strand of @xmath81 that first enters @xmath62 as we travel through @xmath81 naturally . this strand has to come from the left side of @xmath62 by the above observation . but then it fails the test given in remark [ r_cor ] since @xmath62 is descending in @xmath24 and can not be changed in @xmath81 . this gives us the needed contradiction . it now follows that @xmath81 has at least two crossings in each even gap . if @xmath81 also contains some crossings in the odd gaps , or contains more than two crossings in some even gaps , then we already have @xmath182 and there is nothing left to prove . so the only case left is the case when @xmath81 contains no crossings in the odd gaps and exactly two crossings in each even gap . claim 4 . if @xmath98 and it contains no crossings in the odd gaps and exactly two crossings in each even gap , then @xmath186 , that is , @xmath170 is the only element in @xmath187 with this property . proof of claim 4 . by the proof of claim 3 , if @xmath188 , then there exists an even gap such that @xmath81 contains the first crossing of @xmath24 in this gap in its original sign , and exactly one other crossing @xmath62 of @xmath24 in this gap which is not the last crossing in this gap . by remark [ r_cor ] again , the sign of @xmath62 in @xmath24 has to be changed to make it descending in @xmath81 . but as we travel through @xmath81 naturally past @xmath62 and encounter the first crossing of @xmath24 in the same gap below @xmath62 ( keep in mind that this crossing exists because @xmath62 is not the last crossing of @xmath24 in this gap , and the other crossings of @xmath24 in the adjacent odd gap have all been smoothed in @xmath81 ) . this crossing has been smoothed in @xmath81 but we are now approaching it from its overpass , so it fails the check in remark [ r_cor ] and @xmath189 , contradicting to @xmath190 . the consequence of claims 1 to 4 is that if we write @xmath157 as a laurent polynomial of @xmath1 whose coefficients are laurent polynomials of @xmath2 , then it contains a nontrivial term of the form @xmath191 and all other terms have degrees less than @xmath168 , that is , @xmath165 . to obtain @xmath146 , we will use @xmath192 and ( [ e2 ] ) , where @xmath193 is obtained from @xmath24 by keeping the first crossing and flipping the last crossing in each odd gap , and smoothing all other crossings . the details are left to the reader . finally , if @xmath24 is a non - splittable reduced negative - leading alternating braid diagram , then its mirror image @xmath194 is a non - splittable reduced positive - leading alternating braid diagram and we have @xmath195 . let @xmath196 and @xmath197 be the highest and lowest degrees of @xmath1 in @xmath198 . then by the first part of the lemma , we have @xmath199 and @xmath200 . therefore by remark [ tree_formula_remark ] we have @xmath201 and @xmath202 . [ remark4.4 ] _ another way to handle the case when @xmath24 is a non - splittable reduced negative - leading alternating braid diagram is to take its connected sum with a positive hopf link , thus creating a new first gap with a single pair of positive crossings . the statement now follows the positive leading case and from the fact that the braid index is additive under taking the connected sum . _ finally , as another application of theorem [ p3 ] , we provide a short proof to the following theorem . this result seems to be known @xcite , although we failed to find a specific proof in the literature . [ t4.4 ] let @xmath24 be the closure of a reduced alternating braid on @xmath0 strands , then the leading coefficient in the alexander polynomial of @xmath24 is @xmath9 . let @xmath203 be the alexander polynomial of @xmath24 , then @xmath204 @xcite . assume that @xmath24 is positive leading , substituting @xmath205 , @xmath206 in ( [ e1 ] ) leads to @xmath207 construct @xmath208 by the following procedure : ( i ) for the first gap of @xmath24 , smooth all crossings ( which are positive ) except the last one , which we will flip and cross it into the second gap . once there , we have no choice but to keep the first ( negative ) crossing we encounter which will lead us into the third gap . ( ii ) now we will smooth all crossings we encounter ( negative and positive ones in the second and the third gaps ) until we encounter the last positive crossing in the third gap . notice that in doing so we may have reached the bottom of the braid at @xmath209 and returned to the top of the braid at @xmath210 , so this last positive crossing we encounter may not be the last positive crossing in the third gap . at this positive crossing we flip it and cross it into the next gap and repeat ( ii ) . this process is repeated until we reach the last gap and we can and will smooth all crossings in the last gap except one , no matter it is even or odd . at this point , we have exactly one positive crossing in each odd gap but may have negative crossings in the even gaps that we have not visited . however in the last gap we only have one crossing left and we will be traveling back through this crossing . as one can easily verify , each time we travel back to an even gap it is through the only crossing left in the odd gap to its right , so if there are any negative crossings that we have not visited before , these crossings will be between the two positive crossings in the two odd gaps to the left and right of the said even gap . by the descending rule , we can and will smooth all these crossings . thus we have shown that @xmath208 , @xmath211 and @xmath212 where @xmath213 is the total number of crossings in @xmath24 . furthermore , it is also rather easy to see that @xmath214 implies @xmath215 . finally we leave it to our reader to verify that @xmath216 is the only element in @xmath136 with the property @xmath217 . the result of the theorem then follows . if @xmath24 is negative leading , we will simply apply the above argument to ( [ e2 ] ) instead . in this section we show that the homfly polynomial formation ( [ e1 ] ) given in theorem [ p3 ] is equivalent to the expansion derived by f. jaeger for braids in @xcite . thus our approach used to prove proposition [ p3 ] provides an alternative ( and in fact shorter ) proof of his expansion . unlike our approach ( which is combinatorial in nature ) , jaeger stated his formula using a concept called the _ admissible circuit partitions _ of a braid diagram . we establish this equivalence by showing that for any braid @xmath24 , there is a bijection between the leaf vertices in the descending resolving tree @xmath94 and the admissible circuit partitions of @xmath24 , such that each leaf vertex and its corresponding admissible partition under this bijection make the same contributions to their respective homfly polynomial formulas . we begin with a brief review of the essential concepts introduced by f. jaeger with some small modification of the terminology to make the terms more consistent with our current paper . interested readers please refer to @xcite for the details and his original terminologies . given a braid diagram @xmath24 , a circuit partition @xmath218 of @xmath24 is a braid diagram obtained from @xmath24 by smoothing every crossing in a subset @xmath219 of its crossings , while leaving the other crossings unchanged . the smoothed crossings are denoted by small dotted circles in figure 5 . we can identify a circuit partition @xmath218 by the ordered pair @xmath220 where @xmath24 is the original braid diagram and @xmath219 the set of crossings of @xmath24 to be smoothed . two circuit partitions @xmath220 and @xmath221 are defined to be equal if and only if @xmath222 and @xmath223 . let @xmath62 be a crossing in @xmath24 . if @xmath62 is smoothed , the upper left part of the strand entering @xmath62 is connected to the lower left part of the strand exiting @xmath62 and the resulting curve is called a _ left tangence _ at @xmath62 . similarly one can define the _ right tangence _ at @xmath62 . while traveling through @xmath218 naturally , we will meet each crossing in @xmath24 exactly twice ( including the crossings that have been smoothed ) . a smoothed crossing @xmath62 in @xmath218 is said to be _ admissible _ if it has the following property . if @xmath62 is a positive crossing , then the first passage at @xmath62 is a left tangence , if @xmath62 is negative in @xmath24 , then the first passage at @xmath62 is a right tangence . a circuit partition is _ admissible _ if all the smoothed crossings in @xmath218 are admissible . in particular , @xmath24 itself is an admissible circuit partition , in which no crossing is smoothed . we denote the set of admissible circuit partitions of @xmath24 by @xmath224 . then in @xcite jaeger showed that for any braid diagram @xmath24 on @xmath0 strands , @xmath225 where @xmath226 , @xmath227 is the number of crossings in @xmath219 , @xmath228 is the number of negative crossings in @xmath219 and @xmath229 is the number of components in the closure of @xmath218 . [ thm : da ] for any braid diagram @xmath24 , there exists a bijection @xmath230 such that for each @xmath142 , @xmath81 and @xmath231 are both obtained from @xmath24 by smoothing crossings from the same set . define a mapping @xmath232 from @xmath136 to the set of all circuit partitions of @xmath24 as follows . for any @xmath233 , let @xmath234 be the circuit partition @xmath220 where @xmath219 is the set of crossings in @xmath24 that are smoothed in the process of obtaining @xmath81 . we claim that @xmath232 is a bijection between @xmath136 and @xmath224 . we proceed by induction on @xmath4 , the number of vertices in the descending tree @xmath94 . if @xmath235 , then the tree consists of only the root vertex , that is , @xmath24 is descending . so @xmath236 . assume that there is an admissible circuit partition @xmath237 , and let @xmath62 be the first crossing encountered as we travel through @xmath218 naturally . up to the crossing @xmath62 , traveling through @xmath218 naturally is the same as traveling through @xmath24 naturally since nothing has been changed up to that point . since @xmath62 is descending , the first strand entering @xmath62 is the top strand at @xmath62 , and smoothing @xmath62 results in a right tangence if @xmath62 is positive and in a left tangence if @xmath62 is negative . this contradicts the definition of an admissible circuit partition . thus the only admissible circuit partition of @xmath24 is itself . so @xmath232 is a bijection . assume that the statement is true for all @xmath24 whose descending tree has at most @xmath238 vertices and let @xmath24 be such that @xmath94 has at most @xmath239 vertices . since @xmath240 , @xmath24 contains at least one ascending crossing . let @xmath62 be the first ascending crossing of @xmath24 encountered when we travel through @xmath24 naturally . all crossings preceding @xmath62 in the return order being descending , they are not switched or smoothed in any vertex ( braid ) of @xmath94 . let @xmath241 and @xmath242 be the children of the root vertex of @xmath94 , which are obtained by flipping and smoothing @xmath62 respectively . the descendants of @xmath241 and @xmath242 respectively form the rooted trees @xmath243 and @xmath244 respectively . the set of leaves @xmath136 is the union of the sets @xmath245 and @xmath246 . note that this union is disjoint , since @xmath62 is not smoothed in elements of @xmath245 but is smoothed in elements of @xmath246 . using an argument similar to the one used in the case @xmath235 , we see that for every admissible circuit partition @xmath247 , @xmath219 does not contain any crossing preceding @xmath62 when we travel along the strands of @xmath24 naturally thus each @xmath248 can be identified with @xmath249 if @xmath250 , and with @xmath251 if @xmath252 . let @xmath253 be the inverse of this identifying map ( which is a bijection , of course ) . since @xmath241 and @xmath242 are the children of the root vertex of @xmath94 , @xmath254 and @xmath255 both have at least one less vertex than @xmath94 does . by the induction hypothesis , @xmath256 and @xmath257 are bijections that map @xmath258 to @xmath259 and @xmath260 to @xmath261 respectively . it follows that the mapping @xmath262 defined by @xmath263 if @xmath264 and @xmath265 if @xmath266 is a bijection . since @xmath267 , @xmath232 is a bijection between @xmath136 and @xmath268 . this concludes our proof . _ for the descending tree represented in figure [ fig : cpd ] , the admissible circuit partitions corresponding to the leaf vertices are depicted on the left hand side of the picture . for each leaf the corresponding admissible circuit partition is at the same level . note that for each corresponding pair the same crossings of the braid at the root are smoothed . _ the bijection introduced in proposition [ thm : da ] has the property that @xmath142 and @xmath269 are identical except the signs at some crossings . hence we have @xmath270 , @xmath271 and @xmath272 . this leads to the following result . f. jaeger s expansion of the homfly polynomial ( [ j_equation ] ) ( * ? ? ? * proposition 3 ) is an immediate consequence of ( [ e1 ] ) . as a final note to this section , we point out that if we change the definition of admissible circuit partition by defining a smoothed crossing @xmath62 in @xmath218 to be admissible if the first passage at @xmath62 is a left tangence when @xmath62 is negative , and the first passage at @xmath62 is a right tangence if @xmath62 is positive , then we can use ( [ e2 ] ) to show that f. jaeger s expansion of the homfly polynomial ( [ j_equation ] ) becomes @xmath273 where @xmath274 is the set of admissible circuit partitions under this new definition . we end this paper by noting the potential application of theorem [ p3 ] to other classes of links that are presented in a closed braid form . it is also an interesting yet challenging question to explore whether there exist other classes of links that allow formulations similar to the ones in theorem [ p3 ] . these will be the future research directions of the authors . the research of the third author was partially supported by a grant from the simons foundation ( # 245153 to gbor hetyei ) . the authors thank alex stoimenow for his insightful comments . they are also indebted to an anonymous referee for the careful reading of this manuscript and for suggesting several substantial improvements . the alternative argument given in remark [ remark4.4 ] is also due to the referee .

groups generated by finite automata ( groups of automata or automaton groups ) were formally introduced at the beginning of 1960 s @xcite , but more substantial work on this remarkable class of groups started only in 1970 s after aleshin @xcite confirmed a conjecture by glushkov @xcite that these groups could be used to study problems of burnside type ( note that groups of automata should not be confused with automatic groups as described in @xcite ) . it was observed in 1960 s and 1970 s that groups of automata are closely related to iterated wreath products ( pioneering work in this direction is due to kaloujnin @xcite ) and that the theory of such groups could be studied by using the language of tables developed by kaloujnin @xcite and sushchanskii @xcite . even more intensive study of groups of finite automata started in the beginning of 1980 s after the development of some new ideas such as self - similarity , contracting properties , and a geometric realization as groups acting on rooted trees . these developments allowed for elegant constructions of burnside groups @xcite and pushed the study of groups of automata in many directions : analysis @xcite , geometry @xcite , probability @xcite , dynamics @xcite , formal languages @xcite , etc . two well known and important problems were solved using groups of automata in the early 1980 s , namely milnor problem @xcite on intermediate growth and day problem @xcite on amenability . a 5-state automaton constructed in @xcite ( on the right in figure [ 3automata ] ) generates a 2-group , denoted @xmath0 . it was shown in @xcite that @xmath0 has intermediate growth ( between polynomial and exponential ) . this led to construction of other examples of this type @xcite and also made important contribution to and impact on the theory of invariant means on groups @xcite initiated by von neumann @xcite by providing an example of amenable , but not elementary amenable group ( in the sense of day @xcite ) . among the most interesting newer developments is the spectral theory of groups generated by finite automata and graphs associated to such groups @xcite . for instance , automaton groups provided first examples of regular graphs realized as schreier graphs of groups for which the spectrum of the combinatorial laplacian is a cantor set @xcite . further , the realization of the lamplighter group @xmath1 as automaton group ( bottom left in figure [ 3automata ] ) was crucial in the proof that this group has a pure point spectrum ( with respect to a system of generators related to the states of the automaton ) and thus has discrete spectral measure , which was completely described @xcite . this , in turn , led to a construction @xcite of a 7-dimensional closed manifold with non - integer third @xmath2-betti number providing a counterexample to the strong atiyah conjecture @xcite . another fundamental recent discovery is the relation of groups of automata to holomorphic dynamics @xcite . namely , it is shown that to every rational map @xmath3 on the riemann sphere with finite postcritical set one can associate a finite automaton generating a group , denoted @xmath4 and called iterated monodromy group of @xmath5 . the geometry and the topology of the schreier graphs of @xmath4 is closely related to the geometry of the julia set of @xmath5 . figure [ b - schreier5 ] depicts a schreier graph associated to an automaton group , denoted @xmath6 and called basilica group . its reminiscence to the julia set of the map @xmath7 is related to the fact that @xmath6 is the iterated monodromy group @xmath8 of the holomorphic map @xmath9 @xcite ) . groups of automata represent the basis of the theory of self - similar groups and actions @xcite and are related to the study of belyi polynomials and dessins denfants of grothendieck @xcite . the use of iterated monodromy groups was crucial in the recent solution of hubbard s twisted rabbit problem in @xcite . one of the most important developments in the theory of automaton groups is the introduction of groups with branch structure , providing a link to just - infinite groups @xcite and groups of finite width @xcite . in particular , a problem suggested by zelmanov was solved by using the profinite completion of @xmath0 @xcite . the problem of gromov on uniformly exponential growth was solved recently by using branch automaton groups @xcite . an unexpected link of groups of automata and their profinite completions to galois theory was found by r. pink ( private communication ) and aitken , hajir and maire @xcite , while n. boston @xcite related branch groups of automata to fontaine - mazur conjecture and other problems in number theory . the class of branch groups is also a new source for infinitely presented groups , for which the presentation can be written in a recursive form ( see @xcite ) . a recent observation @xcite is that automaton groups and their schreier graphs stand behind the famous hanoi towers problem ( see @xcite ) and some of its generalizations @xcite . there are indications that spectral properties of groups generated by finite automata could be used in the study of kaplansky conjecture on idempotents ( and thus also baum - connes conjecture @xcite and novikov conjecture @xcite ) , dixmier unitarizability problem @xcite , and for construction of new families of expanders , and perhaps even ramanujan , graphs @xcite . in this article we are going to describe some progress which was achieved during the last few years in the problem of classification of automaton groups . two important characteristics of an automaton are the cardinalities @xmath10 and @xmath11 of the set of states and the alphabet , respectively , and the pair @xmath12 is a natural measure of complexity of an automaton and of the group it generates . the groups of complexity @xmath13 are classified @xcite and there are only @xmath14 such groups ( see theorem [ thm_class2states ] in section [ sec_proofs ] here ) . the problem of classification of @xmath15 groups or @xmath16 groups is much harder . the current text represents the progress being made by the research group at texas a&m university over the last few years toward classification of @xmath15 groups . the total number of invertible automata of complexity @xmath15 is @xmath17 . however , the number of non - isomorphic groups generated by these automata is much smaller . there are no more than @xmath18 pairwise non - isomorphic groups of complexity @xmath15 . the proof of this theorem is too long to be presented here ( even the list of all groups takes a lot of space ) . instead , we have chosen for this article a set of @xmath19 automata generating @xmath20 groups ( among the most interesting in this class , in our opinion ) , which we list in the form of a table . the table provided here is a part of the table listing the whole set of @xmath18 groups . we keep the numeration system from the whole table ( the rule for numeration is explained in section [ sec_approach ] ) . major results obtained for the whole family are the following theorems . the numbers in the brackets indicate the numbers of corresponding automata in the class . there are @xmath14 finite groups in the class : @xmath21 [ 1 ] , @xmath22 [ 1090 ] , @xmath23 [ 730 ] , @xmath24 [ 847 ] , @xmath25 [ 802 ] and @xmath26 [ 748 ] . there are @xmath14 abelian groups in the class : @xmath21 [ 1 ] , @xmath22 [ 1090 ] , @xmath23 [ 730 ] , @xmath25 [ 802 ] , @xmath27 [ 731 ] and @xmath28 [ 771 ] . note that there are also virtually abelian groups in this class ( having @xmath29 , @xmath30 [ 2212 ] , @xmath31 [ 752 ] or @xmath32 [ 968 ] as subgroups of finite index ) . [ thm : class_free ] the only free non - abelian group in the class is the free group of rank 3 generated by the aleshin - vorobets automaton [ 2240 ] . moreover , the isomorphism class of this automaton group coincides with its equivalence class under symmetry . the definition of symmetric automata is given in section [ sec_approach ] . there are no infinite torsion groups in the class . we do not provide the complete proofs of these theorems ( by the reason explained above ) . instead , we give here some information about each of the chosen groups and include proofs of most facts . properties that are in our focus are the contracting property , self - replication , torsion , relations ( we list the relators up to length @xmath33 ) , rank of quotients of the stabilizers series , shape of the related schreier graphs . the article is organized as follows . we start with a general information about rooted trees and their automorphisms . then we provide quick introduction to the theory of automaton groups . we continue with the definition of schreier graphs and explain how they naturally appear for the actions on rooted trees . then we list @xmath19 automata generating @xmath20 groups together with some of their properties . in the last section we give proofs of many facts related to the groups in the list . the last part also contains some more general results ( such as an algorithm detecting transitivity of an element and a criterion for group transitivity on the binary tree ) . we recommend the articles @xcite and the book @xcite to the reader who is interested in becoming more familiar with automaton groups . let @xmath34 be fixed and let @xmath35 be the alphabet @xmath36 . the set of words @xmath37 over @xmath35 ( the free monoid over @xmath35 ) can be given the structure of a _ regular rooted labeled @xmath38-ary tree _ @xmath39 in which the empty word @xmath40 is the _ root _ , the _ level _ @xmath11 in @xmath39 consists of the words of length @xmath11 over @xmath35 and every vertex @xmath41 has @xmath38 children , labeled by @xmath42 , for @xmath43 . denote by @xmath44 the group of automorphisms of @xmath39 . let @xmath5 be an automorphism in @xmath44 . any such automorphism can be decomposed as @xmath45 where @xmath46 , for @xmath43 , are automorphisms of @xmath39 and @xmath47 is a permutation of @xmath35 . the automorphisms @xmath46 ( also denoted by @xmath48 ) , @xmath43 , are called ( the first level ) _ sections _ of @xmath5 and each one acts as an automorphism on the subtree @xmath49 hanging below the vertex @xmath50 in @xmath39 consisting of the words in @xmath37 that start with @xmath50 ( any such subtree is canonically isomorphic to the whole tree ) . the action of @xmath5 is decomposed in two steps . first the @xmath38-tuple @xmath51 acts on the @xmath38 subtrees hanging below the root , and then the permutation @xmath47 , called the _ root permutation _ of @xmath5 , permutes these @xmath38 subtrees . thus the action of @xmath5 from on @xmath37 is given by @xmath52 for @xmath50 a letter in @xmath35 and @xmath53 a word over @xmath35 . further iterations of the decomposition ( [ decomposition ] ) yield the second level sections @xmath54 , @xmath55 , and so on . algebraically , we have @xmath56 where @xmath57 is the _ permutational wreath product _ in which the coordinates of @xmath58 are permuted by @xmath59 . iterations of the decomposition show that @xmath44 has the structure of an iterated wreath product @xmath60 . thus @xmath44 is a pro - finite group and in particular , all of its subgroups are residually finite . an obvious and natural sequence of normal subgroups of finite index intersecting trivially is the sequence of level stabilizers . the @xmath11-th _ level stabilizer _ @xmath61 of a group @xmath62 consists of those tree automorphisms in @xmath63 that fix the vertices in @xmath39 up to level @xmath11 . the group @xmath44 is obviously an uncountable object . we are interested in finitely generated subgroups of @xmath44 that exhibit some important features of @xmath44 . one such feature is self - similarity . a group @xmath63 of tree automorphisms is _ self - similar _ if , for every @xmath64 in @xmath63 and a letter @xmath50 in @xmath35 there exists a letter @xmath65 in @xmath35 and an element @xmath66 in @xmath63 such that @xmath67 for all words @xmath53 over @xmath35 . another way to express self - similarity of a group @xmath63 of tree automorphisms is to say that every section @xmath68 of every element @xmath64 in @xmath63 is again an element of @xmath63 . the full tree automorphism group @xmath44 is clearly self - similar ( see ( [ fxw ] ) ) . a self - similar group @xmath63 embeds in the permutational wreath product @xmath69 by @xmath70 consider a finite system of recursive relations @xmath71 where each symbol @xmath72 , @xmath73 , @xmath74 , is equal to one of the symbols @xmath75 and @xmath76 . the system ( [ f^0 ] ) has a unique solution in @xmath44 . the action of @xmath77 on @xmath39 is given recursively by @xmath78 . the group generated by the automorphisms @xmath75 is finitely generated self - similar group of automorphisms of @xmath39 . this group can be described by a finite invertible automaton ( just called automaton in the rest of the article ) . [ automaton ] a _ finite invertible automaton _ @xmath79 is a @xmath80-tuple @xmath81 where @xmath82 is a finite set of _ states _ , @xmath35 is a finite _ alphabet _ of cardinality @xmath34 , @xmath83 is a map , called _ output map _ , @xmath84 is a map , called _ transition map _ , and for each state @xmath85 in @xmath82 , the restriction @xmath86 given by @xmath87 is a permutation , i.e. @xmath88 . the automaton @xmath81 reads words from @xmath37 and provides output words that are also in @xmath37 . the behavior is encoded in the output and transition maps . an _ initial automaton _ @xmath89 is just an automaton @xmath79 with a distinguished state @xmath90 selected as an initial state . we first informally describe the action of the initial automaton @xmath89 on @xmath37 . the automaton starts at the state @xmath85 , reads the first input letter @xmath91 , outputs the letter @xmath92 and changes its state to @xmath93 . the rest of the input word is handled by the new state @xmath94 in the same fashion ( in fact it is handled by the initial automaton @xmath95 ) . formally , the action of the states of the automaton @xmath79 on @xmath37 can be described by extending the output function @xmath96 to a function @xmath97 recursively by @xmath98 for all states @xmath85 in @xmath82 , letters @xmath43 and words @xmath53 over @xmath35 . then the action of the initial automaton @xmath89 is defined by @xmath99 , for words @xmath100 over @xmath35 . in fact , ( [ rhoext ] ) shows that each initial automaton @xmath89 , @xmath90 , defines a tree automorphism , denoted by @xmath85 , defined by @xmath101 where the section @xmath102 is the state @xmath103 and the root permutation @xmath104 is the permutation @xmath105 . given an automaton @xmath81 , the group of tree automorphisms generated by the states of @xmath79 is denoted by @xmath106 and called the _ automaton group _ defined by @xmath79 . the generating set @xmath82 is called the _ standard _ generating set of @xmath106 . _ boundary _ of the tree @xmath39 , denoted @xmath107 , is the set @xmath108 of words over @xmath35 that are infinite to the right ( infinite geodesic rays in @xmath39 starting at the root ) . it has a natural metric ( infinite words are close if they agree on long finite prefixes ) and the group of isometries @xmath109 is canonically isomorphic to @xmath44 . thus the action of the automaton group @xmath106 on @xmath39 can be extended to an isometric action on @xmath107 . in fact , and are valid for infinite words @xmath53 as well . an automaton @xmath79 can be represented by a labeled directed graph , called moore diagram , in which the vertices are the states of the automaton , each state @xmath85 is labeled by its own root permutation @xmath104 and , for each pair @xmath110 , there is an edge from @xmath85 to @xmath111 labeled by @xmath50 . for example , the 5-state automaton in the right half of figure [ 3automata ] generates the group @xmath0 mentioned in the introduction ( @xmath112 denotes the permutation exchanging @xmath113 and @xmath114 ) . the two 2-state automata given on the left of figure [ 3automata ] are the so called _ adding machine _ ( top ) , which generates the infinite cyclic group @xmath29 and the _ lamplighter automaton _ ( bottom ) generating @xmath115 . recursion relations of type ( [ f^0 ] ) for the adding machine and the lamplighter automaton are given by @xmath116 respectively . various classes of automaton groups deserve special attention . an automaton group @xmath117 is _ contracting _ if there exist constants @xmath118 , @xmath119 , and @xmath120 , with @xmath121 , such that @xmath122 , for all vertices @xmath41 of length at least @xmath120 and @xmath123 ( the length is measured with respect to the standard generating set @xmath82 ) . for sufficiently long elements @xmath64 this means that the length of its sections at vertices on levels deeper than @xmath120 is strictly shorter than the length of @xmath64 . this length shortening leads to an equivalent definition of a contracting group . namely , a group @xmath63 of tree automorphisms is contracting if there exists a finite set @xmath124 , such that for every @xmath125 , there exists @xmath126 , such that @xmath127 for all vertices @xmath128 of length not shorter than @xmath120 . the minimal set @xmath129 with this property is called the _ nucleus _ of @xmath63 . the contraction property is a key feature of various inductive arguments and algorithms involving the decomposition @xmath130 . another important class is the class of automaton groups of _ branch type_. branch groups arise as one of the three @xcite possible types of just - infinite groups ( infinite groups for which all proper homomorphic images are finite ) . every infinite , finitely generated group has a just - infinite image . thus if a class of groups @xmath131 is closed under homomorphic images and if it contains infinite , finitely generated examples then it contains just - infinite examples . such examples are , in a sense , minimal infinite examples in @xmath131 . for example , @xmath0 is a branch automaton group that is a just - infinite 2-group . i.e. , it is an infinite , finitely generated , torsion group that has no proper infinite quotients . also , the hanoi towers group @xcite and the iterated monodromy group @xmath132 @xcite are branch groups , while @xmath133 is not a branch group , but only weakly branch ( for definitions see @xcite ) . the class of _ polynomially growing automata _ was introduced by sidki in @xcite , where it is proved that no group @xmath106 defined by such an automaton contains free subgroups of rank 2 . moreover , for a subclass of so called _ bounded automata _ it is known that the corresponding groups are amenable @xcite ( this class of automata , for instance , includes the automata generating @xmath0 , @xmath6 and hanoi towers group on @xmath134 pegs , but not for more pegs ) . finally , self - replicating groups play an important role . a self - similar group @xmath63 is called _ self - replicating _ if , for every vertex @xmath100 , the homomorphism @xmath135 from the stabilizer @xmath100 in @xmath63 to @xmath63 , given by @xmath136 , is surjective . this condition is usually easy to check and , together with transitivity of the action on level 1 , it implies transitivity of the action on all levels . another way to show that a group of automorphisms of the binary tree is level transitive is to use proposition [ prop : transitivity ] . let us fix some self - similar contracting group acting on @xmath137 by automorphisms . denote by @xmath138 the space of left infinite sequences over @xmath35 . two elements @xmath139 are said to be _ asymptotically equivalent _ with respect to the action of the group @xmath63 , if there exist a finite set @xmath140 and a sequence @xmath141 of elements in @xmath142 such that @xmath143 for every @xmath144 . the asymptotic equivalence is an equivalence relation . moreover , sequences @xmath145 are asymptotically equivalent if and only if there exists a sequence @xmath146 of the elements in the nucleus of @xmath63 such that @xmath147 and @xmath148 , for all @xmath144 . the quotient space @xmath149 of the topological space @xmath138 by the asymptotic equivalence relation is called the _ limit space _ of the self - similar action of @xmath63 . the limit space @xmath149 is metrizable and finite - dimensional . if the group @xmath63 is finitely - generated and level - transitive , then the limit space @xmath149 is connected . the last decade witnessed a shift in the attention payed to the study of schreier graphs . let @xmath63 be a group generated by a finite set @xmath150 and let @xmath63 act on a set @xmath151 . the _ schreier graph _ of the action @xmath152 is the graph @xmath153 with set of vertices @xmath151 and set of edges @xmath154 , where the arrow @xmath155 starts in @xmath65 and ends in @xmath156 . if @xmath157 then the schreier graph @xmath158 of the action of @xmath63 on the @xmath63-orbit of @xmath65 is called _ orbital schreier graph_. let @xmath63 be a subgroup of @xmath44 generated by a finite set @xmath150 ( not necessary self - similar ) . the levels @xmath159 , @xmath160 , are invariant under the action of @xmath63 and we can consider the schreier graphs @xmath161 . let @xmath162 . then the pointed schreier graphs @xmath163 converge in the local topology ( topology defined in @xcite ) to the pointed orbital schreier graph @xmath164 . the limit space of a finitely generated contracting self - similar group @xmath63 can be viewed as a hyperbolic boundary in the following way . for any given finite generating system @xmath150 of @xmath63 define the self - similarity graph @xmath165 as the graph with set of vertices @xmath137 in which two vertices @xmath166 are connected by an edge if and only if either @xmath167 , for some @xmath168 ( vertical edges ) , or @xmath169 for some @xmath170 ( horizontal edges ) . if the group is contracting then the self - similarity graph @xmath165 is gromov - hyperbolic and its hyperbolic boundary is homeomorphic to the limit space @xmath149 . the set of horizontal edges of @xmath165 spans the disjoint union of all schreier graphs @xmath171.thus , the schreier graphs @xmath171 in some sense approximate the limit space @xmath149 of the group @xmath63 . moreover , for many examples of self - similar contracting groups there exists a sequence of numbers @xmath172 such that the metric spaces @xmath173 , where @xmath38 is the combinatorial metric on the graph , converge in the gromov - hausdorff metric to the limit space of the group . we recall the definition and basic properties of iterated monodromy groups ( img ) . let @xmath174 be a path connected and locally path connected topological space and let @xmath175 be its open path connected subset . let @xmath176 be a @xmath38-fold covering . by @xmath177 we denote the @xmath11-th iteration of the map @xmath5 . the map @xmath178 , where @xmath179 , is a @xmath180-fold covering . choose an arbitrary base point @xmath181 . let @xmath182 be the disjoint union of the sets @xmath183 ( these sets are note necessarily disjoint by themselves ) . the set of pre - images @xmath182 has a natural structure of a rooted @xmath38-ary tree with root @xmath184 in which every vertex @xmath185 is connected to the vertex @xmath186 , @xmath187 . the fundamental group @xmath188 acts naturally on every set @xmath189 and , in fact , acts by automorphisms on @xmath182 . _ iterated monodromy group _ @xmath4 of the covering @xmath5 is the quotient of the fundamental group @xmath188 by the kernel of its action on @xmath182 . it is proved in @xcite that all iterated monodromy groups are self - similar . this fact provides a connection between holomorphic dynamics and groups generated by automata . the iterated monodromy group of a sub - hyperbolic rational function is contracting and its limit space is homeomorphic to the julia set of the rational function . in particular , the sequence of schreier graphs @xmath190 of the iterated monodromy group of a sub - hyperbolic rational function can be drawn on the riemann sphere in such a way that they converge in the hausdorff metric to the julia set of the function . schreier graphs also play a role in computing the spectrum of the markov operator @xmath191 on the group . namely , given a group @xmath63 generated by a finite set @xmath192 , acting on a tree @xmath37 there is a natural unitary representation of @xmath63 in the space of bounded linear operators @xmath193 given by @xmath194 . the markov operator @xmath195 corresponding to this unitary representation plays an important role . the spectrum of @xmath191 for a self - similar group @xmath63 is approximated by the spectra of finite dimensional operators arising from the action of @xmath63 on the levels of the tree @xmath37 . for more on this see @xcite . let @xmath196 be a subspace of @xmath197 spanned by the @xmath198 characteristic functions @xmath199 , of the cylindrical sets corresponding to the @xmath198 vertices on level @xmath11 . then @xmath196 is invariant under the action of @xmath63 and @xmath200 . denote by @xmath201 the restriction of @xmath202 on @xmath196 . then @xmath203 are finite dimensional operators , whose spectra converge to the spectrum of @xmath191 in the sense @xmath204 if @xmath205 is the stabilizer of an infinite word from @xmath108 , then one can consider the markov operator @xmath206 on the schreier graph of @xmath63 with respect to @xmath205 . the following fact is observed in @xcite and can be applied to compute the spectrum of markov operator on the cayley graph of a group in case if @xmath205 is small . if @xmath63 is amenable or the schreier graph @xmath207 ( the schreier graph of the action of @xmath63 on the cosets of @xmath205 ) is amenable then @xmath208 . the next three sections are devoted to the groups generated by 3-state automata over the 2-letter alphabet @xmath209 . fix @xmath210 as the set of states . every @xmath15 automaton is given by @xmath211 where @xmath212 , for @xmath213 , @xmath214 , @xmath215 , and @xmath216 . a number is assigned to the automaton above by the following formula @xmath217 thus every @xmath15 automaton obtains a unique number in the range from @xmath114 to @xmath218 . the numbering of the automata is induced by the lexicographic ordering of all automata in the class . the automata numbered @xmath114 through @xmath219 act trivially on the tree and generate the trivial group . the automata numbered @xmath220 through @xmath218 generate the group @xmath22 of order @xmath221 , because every element in any of these groups is either trivial , or changes all letters in any word over @xmath35 . therefore the `` interesting '' automata have numbers @xmath222 through @xmath223 . denote by @xmath224 the automaton numbered by @xmath11 and by @xmath225 the corresponding group of tree automorphisms . sometimes , when the context is clear , we use just the number to refer to the corresponding automaton or group . the following operations on automata change neither the group generated by this automaton , nor , essentially , the action of the group on the tree . 1 . passing to inverses of all generators 2 . permuting the states of the automaton 3 . permuting the letters of the alphabet two automata @xmath226 and @xmath227 that can be obtained from one another using a composition of the operations ( @xmath228)(@xmath229 ) , are called _ symmetric_. if the minimization of an automaton @xmath226 is symmetric to the minimization of an automaton @xmath227 , we say that the automata @xmath226 and @xmath227 are _ minimally symmetric _ and write @xmath230 . another equivalence relation we consider is the isomorphism of the groups generated by the automata . the minimal symmetry relation is a refinement of the isomorphism relation , since the same abstract group may have different actions on the binary tree . there are @xmath231 classes of automata , which are pairwise not minimally symmetric , @xmath33 of which are minimally symmetric to automata with fewer than @xmath134 states . these @xmath33 classes of automata are subject of theorem [ thm_class2states ] , which states that they generate @xmath14 different groups . at present , it is known that there are at most @xmath18 non - isomorphic groups in the considered class . in this section we provide information about selected groups in the class of all groups generated by @xmath15 automata . the groups are selected in such a way that the corresponding proofs in section [ sec_proofs ] show most of the main methods and ideas that were used for the whole class . the following notation is used : * rels - this is a list of some relators in the group . all independent relators up to length @xmath20 are included . on some situations additional longer relators are included . for @xmath232 and @xmath233 there are no relators of length up to @xmath20 and the relators provided in the table are not necessarily the shortest . in many cases , the given relations are not sufficient ( for example , some of the groups are not finitely presented ) . * sf - these numbers represent the size of the factors @xmath234 , for @xmath235 . * gr - these numbers represent the values of the growth function @xmath236 , for @xmath235 , and generating system @xmath237 , @xmath238 , @xmath239 . finally , for each automaton in the list a histogram for the spectral density of the operator @xmath240 acting on level @xmath241 of the tree is shown . in some cases , in order to show the main ways to prove the group isomorphism , we provide several different automata generating the same group . @p172ptp174pt@ * automaton number @xmath242 * @p48ptp200pt @xmath243 @xmath244 @xmath245 & group : @xmath246 contracting : _ yes _ self - replicating : _ no _ + rels : @xmath247 , @xmath248 , @xmath249 , @xmath250 , @xmath251 , @xmath252 , + @xmath253 , @xmath254 , @xmath255 , + @xmath256 , @xmath257 , @xmath258 , + @xmath259 , @xmath260 , + @xmath261 , @xmath262 , + @xmath263 + sf : @xmath264,@xmath265,@xmath266,@xmath267,@xmath268,@xmath269,@xmath270,@xmath271,@xmath272 + gr : 1,4,9,17,30,47,68,93,122,155,192 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 190,10)@xmath273 ( 1155,10)@xmath274 ( 680,77)@xmath275 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath276 * @p48ptp200pt @xmath277 @xmath244 @xmath245 & group : contracting : _ no _ self - replicating : _ yes _ + rels : @xmath278 , @xmath279 , @xmath280 , @xmath281 + sf : @xmath264,@xmath265,@xmath266,@xmath267,@xmath270,@xmath282,@xmath283,@xmath284,@xmath285 + gr : 1,7,37,187,937,4687 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 230,700)@xmath274 ( 680,240)@xmath275 ( 1155,10)@xmath274 ( 680,77)@xmath275 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath286 * @p48ptp200pt @xmath243 @xmath287 @xmath245 & group : @xmath26 contracting : _ yes _ self - replicating : _ no _ + rels : @xmath247 , @xmath248 , @xmath249 , @xmath250 , @xmath288 , @xmath289 + sf : @xmath264,@xmath265,@xmath266,@xmath290,@xmath290,@xmath290,@xmath290,@xmath290,@xmath290 + gr : 1,4,8,12,15,16,16,16,16,16,16 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 190,10)@xmath273 ( 890,585)@xmath274 ( 680,77)@xmath275 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath291 * @p48ptp200pt @xmath277 @xmath287 @xmath245 & group : contracting : _ no _ self - replicating : _ yes _ + rels : @xmath292 + @xmath293 , @xmath294 + @xmath295 , @xmath296 + @xmath297 + sf : @xmath264,@xmath265,@xmath266,@xmath267,@xmath270,@xmath282,@xmath283,@xmath284,@xmath285 + gr : 1,7,37,187,937,4687 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 230,700)@xmath274 ( 680,240)@xmath275 ( 890,585)@xmath274 ( 680,77)@xmath275 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath298 * @p48ptp200pt @xmath277 @xmath299 @xmath245 & group : @xmath28 contracting : _ yes _ self - replicating : _ yes _ + rels : @xmath238 , @xmath300 + sf : @xmath264,@xmath265,@xmath301,@xmath266,@xmath290,@xmath302,@xmath267,@xmath303,@xmath268 + gr : 1,5,13,25,41,61,85,113,145,181,221 + limit space : @xmath221-dimensional torus @xmath304 & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 230,700)@xmath274 ( 680,240)@xmath275 ( 1080,10)@xmath273 ( 455,585)@xmath273 + @p172ptp174pt@ * automata number @xmath305 and @xmath306 * @p48ptp12ptp48ptp200pt @xmath243 @xmath307 @xmath245 & 783 : & @xmath308 @xmath307 @xmath245&group : @xmath309 contracting : _ yes _ self - replicating : _ yes _ + rels : @xmath247 , @xmath248 , @xmath249 , @xmath250 , @xmath310 , + @xmath311 , @xmath312 , + @xmath313 , @xmath314 + sf : @xmath264,@xmath265,@xmath301,@xmath290,@xmath267,@xmath315,@xmath316,@xmath317,@xmath318 + gr : 1,4,9,17,30,51,85,140,229,367,579 + limit space : & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 190,10)@xmath273 ( 890,585)@xmath274 ( 1160,10)@xmath275 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath319 * @p48ptp200pt @xmath320 @xmath321 @xmath245 & group : @xmath28 contracting : _ yes _ self - replicating : _ yes _ + rels : @xmath322 , @xmath300 + sf : @xmath264,@xmath265,@xmath301,@xmath266,@xmath290,@xmath302,@xmath267,@xmath303,@xmath268 + gr : 1,7,21,43,73,111,157,211,273,343,421 + limit space : @xmath221-dimensional torus @xmath304 & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 680,240)@xmath274 ( 193,10)@xmath275 ( 820,585)@xmath273 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath323 * @p48ptp200pt @xmath308 @xmath324 @xmath325 & group : @xmath326 contracting : _ no _ self - replicating : _ no _ + rels : @xmath247 , @xmath248 , @xmath249 + sf : @xmath264,@xmath265,@xmath266,@xmath302,@xmath303,@xmath269,@xmath327,@xmath272,@xmath328 + gr : 1,4,10,22,46,94,190,382,766,1534 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 1200,300)(1123,733)(787,1020 ) ( 1216,354)(1200,300)(1167,345 ) ( 150,700)@xmath273 ( 680,77)@xmath274 ( 1160,10)@xmath275 ( 1115,700)@xmath274(460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath329 * @p48ptp200pt @xmath277 @xmath299 @xmath325 & group : _ basilica group _ contracting : _ yes _ self - replicating : _ yes _ + rels : @xmath238 , @xmath330 , @xmath331 , + @xmath332 , @xmath333 , + @xmath334 , @xmath335 , + @xmath336 + sf : @xmath264,@xmath265,@xmath266,@xmath267,@xmath270,@xmath282,@xmath283,@xmath284,@xmath285 + gr : 1,5,17,53,153,421,1125,2945,7545 + limit space : & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 1200,300)(1123,733)(787,1020 ) ( 1216,354)(1200,300)(1167,345 ) ( 230,700)@xmath274 ( 680,240)@xmath275 ( 1080,10)@xmath273 ( 1115,700)@xmath274(460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath337 * @p48ptp200pt @xmath320 @xmath307 @xmath325 & group : contracting : _ no _ self - replicating : _ yes _ + rels : @xmath338 , @xmath339 , + @xmath340 , @xmath341 , @xmath342 , @xmath343 , @xmath344 , + @xmath345 , @xmath346 + sf : @xmath264,@xmath265,@xmath266,@xmath303,@xmath327,@xmath347,@xmath348,@xmath349,@xmath350 + gr : 1,7,35,165,758,3460 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 1200,300)(1123,733)(787,1020 ) ( 1216,354)(1200,300)(1167,345 ) ( 680,240)@xmath274 ( 193,10)@xmath275 ( 890,585)@xmath274 ( 1160,10)@xmath275 ( 1115,700)@xmath274(460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath351 * @p48ptp200pt @xmath352 @xmath307 @xmath325 & group : contracting : _ no _ self - replicating : _ yes _ + rels : @xmath353 + @xmath354 , @xmath355 + @xmath356 + sf : @xmath264,@xmath265,@xmath266,@xmath303,@xmath327,@xmath357,@xmath358,@xmath349,@xmath350 + gr : 1,7,37,187,937,4687 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 1200,300)(1123,733)(787,1020 ) ( 1216,354)(1200,300)(1167,345 ) ( 230,700)@xmath274 ( 193,10)@xmath275 ( 890,585)@xmath274 ( 1160,10)@xmath275 ( 1115,700)@xmath274(460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath359 * @p48ptp200pt @xmath277 @xmath360 @xmath325 & group : _ baumslag - solitar group @xmath361 _ contracting : _ no _ self - replicating : _ yes _ + rels : @xmath362 , @xmath363 + sf : @xmath264,@xmath265,@xmath266,@xmath290,@xmath267,@xmath268,@xmath269,@xmath270,@xmath271 + gr : 1,7,33,127,433,1415 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 1200,300)(1123,733)(787,1020 ) ( 1216,354)(1200,300)(1167,345 ) ( 230,700)@xmath274 ( 680,240)@xmath275 ( 680,77)@xmath274 ( 890,585)@xmath275 ( 1115,700)@xmath274(460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath364 * @p48ptp200pt @xmath365 @xmath366 @xmath325 & group : @xmath367 contracting : _ yes _ self - replicating : _ yes _ + rels : @xmath247 , @xmath248 , @xmath249 , @xmath368 , @xmath369 + sf : @xmath264,@xmath265,@xmath266,@xmath303,@xmath327,@xmath357,@xmath358,@xmath370,@xmath371 + gr : 1,4,10,22,46,94,184,352,664,1244,2296,4198,7612 + limit space : & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 1200,300)(1123,733)(787,1020 ) ( 1216,354)(1200,300)(1167,345 ) ( 650,250)@xmath273 ( 1155,10)@xmath274 ( 890,585)@xmath275 ( 1115,700)@xmath274(460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath372 * @p48ptp200pt @xmath320 @xmath299 @xmath373 & group : contracting : _ no _ self - replicating : _ yes _ + rels : @xmath238 , @xmath374 + sf : @xmath264,@xmath265,@xmath266,@xmath267,@xmath270,@xmath282,@xmath283,@xmath284,@xmath285 + gr : 1,5,17,53,161,475,1387 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 700,1211 ) ( 820,1168)(771,1141)(779,1196 ) ( 680,240)@xmath274 ( 193,10)@xmath275 ( 1080,10)@xmath273 ( 545,1261)@xmath274 ( 460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath375 * @p48ptp200pt @xmath277 @xmath307 @xmath373 & group : contains the lamplighter group contracting : _ no _ self - replicating : _ yes _ + rels : @xmath376 , @xmath377 , @xmath378 , + @xmath379 , @xmath380 , @xmath381 , + @xmath382 , @xmath383 , @xmath384 , + @xmath385 , @xmath386 , + @xmath387 , @xmath388 , + @xmath389 , @xmath390 , + @xmath391 , @xmath392 , + @xmath393 , + @xmath394 , + @xmath395 + sf : @xmath264,@xmath265,@xmath266,@xmath303,@xmath327,@xmath347,@xmath348,@xmath349,@xmath350 + gr : 1,7,33,143,597,2465 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 700,1211 ) ( 820,1168)(771,1141)(779,1196 ) ( 230,700)@xmath274 ( 680,240)@xmath275 ( 890,585)@xmath274 ( 1160,10)@xmath275 ( 545,1261)@xmath274 ( 460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath396 * @p48ptp200pt @xmath365 @xmath321 @xmath373 & group : contains @xmath32 as a subgroup of index @xmath397 contracting : _ yes _ self - replicating : _ no _ + rels : @xmath247 , @xmath248 , @xmath249 , @xmath368 , @xmath256 , + @xmath398 , @xmath369 , @xmath399 , + @xmath400 , @xmath401 , + @xmath311 , @xmath402 + sf : @xmath264,@xmath265,@xmath266,@xmath267,@xmath315,@xmath327,@xmath403,@xmath404,@xmath347 + gr : 1,4,10,22,46,94,184,338,600,1022 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,152)@xmath114 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 700,1211 ) ( 820,1168)(771,1141)(779,1196 ) ( 650,250)@xmath273 ( 820,585)@xmath273 ( 545,1261)@xmath274 ( 460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath405 * @p48ptp200pt @xmath308 @xmath406 @xmath245 & group : @xmath309 contracting : _ yes _ self - replicating : _ yes _ + rels : @xmath247 , @xmath249 , @xmath250 , @xmath407 , @xmath408 , + @xmath409 , @xmath410 , @xmath411 , @xmath412 , + @xmath413 , @xmath414 , @xmath415 , @xmath416 , + @xmath417 , @xmath418 + sf : @xmath264,@xmath265,@xmath301,@xmath290,@xmath267,@xmath315,@xmath316,@xmath317,@xmath318,@xmath419 + gr : 1,5,16,40,88,184,376,746,1458 + limit space : & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,165)@xmath112 ( 664,1022)@xmath114 ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 150,700)@xmath273 ( 1155,10)@xmath274 ( 680,77)@xmath275 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath420 * @p48ptp200pt @xmath421 @xmath422 @xmath245 & group : _ klein bottle group _ , virtually @xmath30 contracting : _ yes _ self - replicating : _ no _ + rels : @xmath423 , @xmath424 + sf : @xmath264,@xmath265,@xmath301,@xmath290,@xmath267,@xmath268,@xmath269,@xmath270,@xmath271 + gr : 1,7,19,37,61,91,127,169,217,271,331 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,165)@xmath112 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 190,10)@xmath274 ( 230,700)@xmath275 ( 890,585)@xmath274 ( 680,77)@xmath275 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath425 * @p48ptp200pt @xmath426 @xmath427 @xmath245 & group : @xmath428 free of rank @xmath134 ( aleshin - vorobets ) contracting : _ no _ self - replicating : _ no _ + rels : + sf : @xmath264,@xmath265,@xmath301,@xmath290,@xmath303,@xmath269,@xmath271,@xmath404,@xmath429 + gr : 1,7,37,187,937,4687 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,165)@xmath112 ( 664,1022)@xmath114 ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 680,240)@xmath274 ( 230,700)@xmath275 ( 890,585)@xmath274 ( 1160,10)@xmath275 ( 455,585)@xmath273 + @p172ptp174pt@ * automaton number @xmath430 * @p48ptp200pt @xmath308 @xmath431 @xmath325 & group : @xmath432 contracting : _ yes _ self - replicating : _ yes _ + rels : @xmath247 , @xmath248 , @xmath249 , @xmath408 , @xmath433 , + @xmath434 , @xmath435 , @xmath436 , + @xmath437 , @xmath438 , + @xmath439 , @xmath440 , + @xmath441 + sf : @xmath264,@xmath265,@xmath301,@xmath290,@xmath302,@xmath267,@xmath303,@xmath268,@xmath315 + gr : 1,4,10,19,31,46,64,85,109,136,166 + limit space : @xmath221-dimensional sphere @xmath442 & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,165)@xmath112 ( 664,1022)@xmath114 ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 1200,300)(1123,733)(787,1020 ) ( 1216,354)(1200,300)(1167,345 ) ( 150,700)@xmath273 ( 650,87)@xmath273 ( 1115,700)@xmath274(460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath443 * @p48ptp200pt @xmath320 @xmath422 @xmath373 & group : contracting : _ no _ self - replicating : _ yes _ + rels : @xmath376 , @xmath377 , @xmath378 , + @xmath379 , @xmath380 , @xmath383 , + @xmath384 , @xmath385 , + @xmath387 , @xmath388 , + @xmath389 , @xmath390 , + @xmath392 , @xmath393 , + @xmath394 + sf : @xmath264,@xmath265,@xmath266,@xmath303,@xmath327,@xmath347,@xmath348,@xmath349,@xmath350 + gr : 1,7,33,143,602,2514 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,165)@xmath112 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 300,200)(1,0)800 ( 1050,225)(1100,200)(1050,175 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 750,983)(1150,287 ) ( 753,927)(750,983)(797,952 ) ( 650,983)(250,287 ) ( 297,318)(250,287)(253,343 ) ( 700,1211 ) ( 820,1168)(771,1141)(779,1196 ) ( 680,240)@xmath274 ( 193,10)@xmath275 ( 890,585)@xmath274 ( 680,77)@xmath275 ( 545,1261)@xmath274 ( 460,585)@xmath275 + @p172ptp174pt@ * automaton number @xmath444 * @p48ptp200pt @xmath421 @xmath406 @xmath445 & group : isomorphic to @xmath446 contracting : _ no _ self - replicating : _ yes _ + rels : @xmath239 , @xmath447 + sf : @xmath264,@xmath265,@xmath266,@xmath267,@xmath270,@xmath282,@xmath283,@xmath284,@xmath285,@xmath448 + gr : 1,5,17,53,161,485,1445 + & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,165)@xmath112 ( 664,1022)@xmath114 ( 100,100 ) ( 46,216)(100,200)(55,167 ) ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 700,1211 ) ( 820,1168)(771,1141)(779,1196 ) ( 190,10)@xmath274 ( 230,700)@xmath275 ( 1155,10)@xmath274 ( 680,77)@xmath275 ( 465,1261)@xmath273 + @p172ptp174pt@ * automaton number @xmath449 * @p48ptp200pt @xmath308 @xmath406 @xmath445 & group : @xmath450 contracting : _ yes _ self - replicating : _ yes _ + rels : @xmath239 , @xmath247 , @xmath451 + sf : @xmath264,@xmath265,@xmath301,@xmath266,@xmath302,@xmath268,@xmath271,@xmath347,@xmath348 + gr : 1,4,10,22,46,94,190,375,731,1422,2752,5246,9908 + limit space : & ( 1450,1090)(0,130 ) ( 200,200 ) ( 1200,200)(700,1070 ) ( 45,280)@xmath237 ( 1280,280)@xmath238 ( 820,1035)@xmath239 ( 164,165)@xmath112 ( 1164,165)@xmath112 ( 664,1022)@xmath114 ( 200,300)(277,733)(613,1020 ) ( 559,1007)(613,1020)(591,969 ) ( 287,150)(700,0)(1113,150 ) ( 325,109)(287,150)(343,156 ) ( 1300,100 ) ( 1345,167)(1300,200)(1354,216 ) ( 700,1211 ) ( 820,1168)(771,1141)(779,1196 ) ( 150,700)@xmath273 ( 1155,10)@xmath274 ( 680,77)@xmath275 ( 465,1261)@xmath273 + we start this section with a few useful observations , which simplify computations and were used often in the classification process . [ thm_class2states ] there are , up to isomorphism , @xmath14 different groups generated by @xmath221-state automata over a @xmath221-letter alphabet ( automata of complexity @xmath13 ) . namely , trivial group , @xmath22 , @xmath23 , infinite cyclic group @xmath29 , infinite dihedral group @xmath452 and the lamplighter group @xmath1 . [ nontors ] let @xmath63 be a group generated by an automaton @xmath79 over the @xmath221-letter alphabet @xmath209 that has the following property . the set of states of @xmath79 splits into two nonempty parts @xmath205 and @xmath82 such that 1 . one of the parts contains all active states and the other contains all inactive states ; 2 . for each state in part @xmath205 , both arrows go to states in the same part ( either both to @xmath205 pr both to @xmath82 ) ; 3 . for each state in part @xmath82 , one arrow goes to part @xmath205 and the other to part @xmath82 then any element of the group that can be written as a product of odd number of active generators and odd number of inactive generators in any order , has infinite order . in particular , @xmath63 is not a torsion group . let @xmath64 be such an element . let us prove by induction on @xmath11 that , for each @xmath160 , there exists a vertex @xmath41 fixed by @xmath453 such that the section of @xmath453 at @xmath41 has the same form ( i.e. is a product of odd number of active generators and odd number of inactive generators ) . for @xmath454 this is true . suppose it is true for @xmath455 and let @xmath66 be a section of @xmath456 at some vertex @xmath128 satisfying the conditions of the assumption . since @xmath66 is a product of odd number of active states and odd number of inactive states we can write @xmath457 . since @xmath41 is fixed under @xmath458 we have that @xmath459 is fixed under @xmath460 and @xmath461 . the element @xmath462 is product ( in some order ) of the first level sections of the generators ( and/or their inverses ) used to express @xmath66 . by assumption , among these generators , there are odd number of active and odd number of inactive . the generators from part @xmath205 , by condition ( @xmath463 ) , will produce even number of active and even number of inactive generators , while the generators from part @xmath82 , by condition @xmath464 , will give odd number of generators from both categories , which proves the induction step . thus @xmath465 for all @xmath11 . the abelianization of @xmath466 is isomorphic to the infinite cartesian product @xmath467 . the canonical isomorphism sends @xmath125 to @xmath468 , where @xmath469 is the number of vertices @xmath470 , such that @xmath471 acts nontrivially on the first level ( i.e. @xmath469 is the number of active sections of @xmath64 at level @xmath228 ) [ alg_trans ] let @xmath63 be a self - similar group generated by a finite automaton over the @xmath221-letter alphabet @xmath209 , and let @xmath125 be given as a product of states and their inverses . denote by @xmath475 the image of @xmath64 in @xmath472 $ ] . the element @xmath64 acts level transitively on @xmath37 if and only if @xmath476 . suppose @xmath477 , @xmath478 . then @xmath479 we can produce similar equations for the sections @xmath480 , @xmath481 and so on . since @xmath63 is generated by a finite automaton the number of different sections of @xmath64 is finite . therefore we get finite linear system of equations over the variables @xmath482 , whose solution will express @xmath475 as a rational function @xmath483 , where @xmath484 are polynomials with degrees not higher than @xmath485 . expanding this rational function as a power series will produce a preperiodic sequence of coefficients from @xmath22 with period and preperiod no longer than @xmath486 . in particular , @xmath64 acts level transitively if and only if all @xmath469 , @xmath487 are equal to @xmath114 . for the converse , let us first prove that all level stabilizers @xmath488 are different . for this it suffices to show that for every @xmath489 @xmath490 . since all stabilizers have finite indices in @xmath63 and @xmath63 is infinite we get that all of them are infinite . let @xmath491 be an arbitrary nontrivial element and @xmath492 be a word of shortest length ( one of them ) such that @xmath493 ( in other words , @xmath494 and such @xmath495 is maximal ) . clearly @xmath496 and we can consider the section @xmath497 which is an element of @xmath63 because of self - similarity . the fact that @xmath498 implies @xmath499 . on other hand @xmath500 . therefore @xmath501 , thus @xmath502 and we found the desired element . suppose @xmath63 acts transitively on level @xmath11 . let @xmath503 be an arbitrary element and let @xmath504 be one of the words such that @xmath505 , where @xmath506 . for @xmath507 , let us find an element @xmath125 such that @xmath508 . this will prove the induction step . note , that the last proof works also for self - similar subgroups of the infinitely iterated permutational wreath product @xmath516 ( the subgroup of @xmath44 consisting of those automorphisms of the @xmath38-ary tree for which the activity at every vertex is a power of some fixed cycle of length @xmath38 . also , certain generalizations of this method could be used in more complex situations ( see , for example , the construction of @xmath517 in @xcite , or proof of transitivity of sushchansky groups on a subtree in @xcite ) . all generators have order @xmath221 . the elements @xmath520 and @xmath521 generate @xmath30 because @xmath522 is the adding machine and has infinite order . also we have @xmath523 and @xmath524 is normal in @xmath525 because @xmath526 and @xmath527 . in other words , @xmath528 . furthermore , the element @xmath541 has infinite order , fixes the vertex @xmath542 and its section at this vertex is equal to @xmath541 . therefore @xmath535 is not contracting ( all powers of @xmath541 would have to belong to the nucleus ) . it follows from the relations @xmath544 that @xmath543 is a homomorphic image of @xmath26 . since @xmath545 , @xmath546 and @xmath547 , it follows that @xmath548 . one can verify directly that @xmath239 is not equal to any of the four elements in @xmath549 that stabilize level 1 ( namely @xmath114 , @xmath238 , @xmath550 and @xmath551 ) . thus @xmath552 . for a word @xmath53 in @xmath554 , let @xmath555 , @xmath556 and @xmath557 denote the sum of the exponents of @xmath237 , @xmath238 and @xmath239 in @xmath53 . let @xmath53 represents the element @xmath123 . if @xmath555 and @xmath556 are odd , then @xmath64 acts transitively on the first level , and @xmath558 is represented by a word @xmath559 , which is the product ( in some order ) of all first level sections of all generators appearing in @xmath53 . hence , @xmath560 and @xmath561 are odd again . therefore , similarly to proposition [ nontors ] , any such element has infinite order . in particular @xmath562 has infinite order . since @xmath563 and @xmath564 , the element @xmath565 has infinite order ( and so does @xmath237 ) . since @xmath565 fixes the vertex @xmath539 and its section at that vertex is equal to @xmath565 , the group @xmath232 is not contracting . since @xmath566 is finitely generated , abelian , and self - replicating ( easy to check ) , it follows from @xcite that it is free abelian . there are two options : either it has rank @xmath114 or rank @xmath221 ( since @xmath567 ) . let us prove that the rank is @xmath221 . for this it is sufficient to show that @xmath568 in @xmath63 . assume on the contrary that @xmath569 for some integer @xmath11 and @xmath10 and choose such integers with minimal @xmath570 . since @xmath571 stabilizes level 1 , @xmath10 must be even and we have @xmath572 . but then @xmath573 and by the minimality assumption @xmath10 must be @xmath113 , implying @xmath574 . the last equality can only be true for @xmath454 since @xmath566 is torsion free ( free abelian ) and @xmath575 . thus @xmath576 . we have @xmath578 , @xmath579 and @xmath580 . hence , for the subgroup @xmath581 , we have @xmath582 . on the other hand @xmath530 is normal in @xmath63 since @xmath583 and @xmath584 . thus @xmath585 , where @xmath22 is generated by @xmath237 and the action of @xmath237 on @xmath530 is given above . it is proved below ( see @xmath586 ) , that @xmath587 . therefore @xmath577 also contains a torsion free subgroup of index @xmath80 . all generators have order @xmath221 and @xmath237 commutes with @xmath239 . conjugating this group by the automorphism @xmath589 yields an isomorphic group generated by the 4-state automaton defined by the recursive relations @xmath590 , @xmath591 and @xmath592 . on the other hand , we obtain the same automaton after conjugating @xmath577 by @xmath593 ( here @xmath237 denotes the generator of @xmath577 ) . since @xmath566 is finitely generated , abelian , and self - replicating , it follows from @xcite that it is free abelian . consider the @xmath601-endomorphism @xmath602 associated to the vetrex @xmath113 , given by @xmath603 for @xmath604 , provided @xmath605 . consider also the linear map @xmath606 induced by @xmath607 . it has the following matrix representation with respect to the basis corresponding to the triple @xmath608 : @xmath609 its characteristic polynomial @xmath610 has three distinct complex roots @xmath611 , @xmath612 and @xmath613 . choose an eigenvector @xmath614 associated to the eigenvalue @xmath615 , @xmath215 . in particular , we may choose @xmath616 , which shows that @xmath617 in @xmath600 . in order to show that @xmath618 ( except when @xmath619 ) we will prove that the vector @xmath620 is eventually pushed out from the domain corresponding to the first level stabilizer , i.e. from the set @xmath621 , by iterations of the action of @xmath79 . consider the expansion of @xmath41 in the basis @xmath622 : @xmath623 . since @xmath624 or @xmath625 , @xmath41 is not a scalar multiple of @xmath626 . we have @xmath627 , as @xmath628 , since @xmath629 . we can choose a neighborhood of @xmath630 that does not contain points from @xmath631 , except maybe @xmath630 . eventually @xmath632 will be in this neighborhood and , since @xmath633 for all @xmath634 , @xmath632 will be outside of @xmath631 . this implies that the word @xmath635 represents a nontrivial element in @xmath600 . thus @xmath636 . * 846 * : @xmath326 . this is a result of muntyan . see the proof in @xcite . in particular , @xmath637 contains a self - similar free group of rank @xmath221 generated by a 6-state automaton . the automaton 846 is sometimes called bellaterra automaton . * 852 * : basilica group @xmath638 . first studied in @xcite , where it is proved that @xmath6 is not in the class @xmath639 of sub - exponentially amenable groups , does not contain a free subgroup of rank @xmath221 , and the nontrivial generators @xmath237 and @xmath238 generate a free subsemigroup . spectral properties are considered in @xcite . it is proved in @xcite that @xmath6 is amenable , providing the first example of an amenable group not in the class @xmath639 . let us prove that @xmath238 has infinite order . for any @xmath641 @xmath642 . since @xmath643 equals either @xmath238 or @xmath239 and @xmath644 , we have @xmath645 . therefore all elements in the forward orbit of @xmath646 under the action of @xmath238 end in @xmath647 . the length of the non - zero prefix of any infinite word ending in @xmath647 can not decrease under the action of @xmath238 . indeed , for any @xmath641 @xmath648 . the section @xmath649 is one of the three generators , for which we have @xmath650 , @xmath651 and @xmath652 . on the other hand , the length of the non - zero prefix along the orbit can not stabilize , because in this case the orbit must be finite and we must have @xmath653 , for some @xmath654 . but this is impossible since @xmath655 and thus the length of the non - zero prefix of @xmath656 must be at least @xmath134 . thus the orbit is infinite and @xmath238 has infinite order . using the same approach as for @xmath640 one can show that @xmath239 has infinite order . namely the length of the non - zero prefix of the forward orbit of @xmath658 under @xmath239 is nondecreasing , which then implies that this orbit is not finite . the automaton satisfies the conditions of proposition [ nontors ] . thus , @xmath660 has infinite order in @xmath63 , which implies that @xmath661 , @xmath662 also have infinite order . hence , we can claim the same for @xmath237 and @xmath360 . furthermore , the element @xmath663 also has infinite order ( it is conjugate of the adding machine ) . since @xmath664 we have @xmath665 and @xmath666 . let us check that @xmath667 . since @xmath668 and @xmath669 all we need to check is that @xmath670 , i.e. @xmath671 . the last is correct since @xmath672 and @xmath673 is a conjugate of the inverse of @xmath674 . thus , since @xmath238 and @xmath675 have infinite order , @xmath676 . denote @xmath678 and @xmath679 . all generators have order @xmath221 , and therefore the subgroup @xmath680 is a normal subgroup of index @xmath221 in @xmath677 . moreover @xmath681 , where @xmath22 is generated by @xmath239 and the action of @xmath22 on @xmath530 is given by @xmath682 and @xmath683 . we have @xmath684 and @xmath685 . exchanging the letters @xmath113 and @xmath114 leads to an isomorphic copy of @xmath530 defined by @xmath686 and @xmath687 , which is the iterated monodromy group @xmath688 , according to @xcite . thus , @xmath689 . it is known @xcite that the group @xmath692 defined by @xmath693 is the lamplighter group @xmath1 ( compare to ) . consider the subtree @xmath694 of @xmath37 consisting of all words over the alphabet @xmath695 . the element @xmath237 swaps the letters of @xmath151 and @xmath238 fixes them . since @xmath696 , @xmath697 , the tree @xmath694 is invariant under the action of @xmath698 and the action of @xmath530 on @xmath694 coincides with the action of the lamplighter group @xmath699 on @xmath37 ( with the identification @xmath700 ) . therefore the map @xmath701 given by @xmath702 , @xmath703 extends to a homomorphism . we claim that this homomorphism has trivial kernel . indeed , let @xmath704 be a group word representing an element of the kernel of @xmath607 . since the word @xmath705 represents the identity in @xmath706 the total exponent of @xmath237 in @xmath53 must be even and the total exponent @xmath707 of both @xmath237 and @xmath238 in @xmath53 must be 0 . but in that case the element @xmath708 fixes the top two levels of the tree @xmath37 and has decomposition @xmath709 where the @xmath710 s denote words over @xmath237 and @xmath238 representing the identity in @xmath530 ( these words correspond to the first level sections of @xmath705 in @xmath706 ) . therefore @xmath711 and the kernel of @xmath607 is trivial . this group contains @xmath32 as a subgroup of index @xmath397 . it is contracting with nucleus consisting of @xmath713 elements , whose self - similar closure consists of @xmath714 elements . all generators have order @xmath221 . let @xmath715 , @xmath716 , and consider the subgroup @xmath717 . direct computations show that @xmath50 and @xmath65 commute ( @xmath718 and @xmath719 ) . conjugating by @xmath720 leads to the self - similar copy @xmath721 of @xmath142 generated by @xmath722 and @xmath723 , where @xmath724 and @xmath725 . since @xmath726 , the virtual endomorphism of @xmath721 is given by the matrix @xmath727 the eigenvalues @xmath728 of this matrix are not algebraic integers , hence , according to @xcite , the group @xmath721 is free abelian of rank @xmath221 , and so is @xmath142 . since all generators have order @xmath221 , the subgroup @xmath729 has index @xmath221 in @xmath712 . the stabilizer @xmath730 of the second level has index @xmath731 in @xmath530 . moreover , the quotient group is isomorphic to the dihedral group @xmath24 ( since @xmath660 acts on the second level by permuting @xmath732 and @xmath733 , while @xmath734 acts by permuting @xmath735 ) . the stabilizer @xmath730 , conjugated by the element @xmath736 , is generated by @xmath737 therefore , all @xmath738 commute and @xmath739 . if @xmath740 , then all sections must be trivial , hence , @xmath741 . but @xmath142 is free abelian , whence @xmath742 , @xmath743 . thus , @xmath730 is a free abelian group of rank @xmath744 . since @xmath748 and @xmath749 , the generators @xmath237 and @xmath239 commute further , @xmath750 , which shows that @xmath751 , and therefore @xmath752 and @xmath753 . since @xmath754 , the element @xmath237 has infinite order and so does @xmath755 . finally , since @xmath756 , we have @xmath757 and @xmath747 is the klein bottle group . going back to the generating set consisting of @xmath237 and @xmath238 , we get the presentation @xmath758 . * 2240 * : free group of rank @xmath134 . the automaton generating this group first appeared in @xcite . it is proved in @xcite that @xmath759 is a free group of rank 3 with basis @xmath608 . this is the smallest example of a free nonabelian group among all automata over a @xmath221-letter alphabet ( see theorem [ thm : class_free ] ) . all generators have order @xmath221 . let @xmath762 and @xmath763 and let @xmath764 . then @xmath765 and @xmath766 . it is easy to check that @xmath50 and @xmath65 commute and that @xmath530 is self - replicating . the matrix of the associated virtual endomorphism is given by @xmath767 since the eigenvalues @xmath768 are not algebraic integers , according to @xcite @xmath530 is free abelian of rank 2 . the subgroup @xmath530 is normal of index 2 in @xmath761 because the generators of @xmath761 are of order @xmath221 . thus @xmath769 , where the action of @xmath770 on @xmath530 is by inversion of the generators . for any vertex @xmath128 , we have @xmath772 , @xmath773 and @xmath774 , for @xmath775 . therefore , for any vertex @xmath776 , @xmath777 and the forward orbit of @xmath658 under @xmath237 is infinite , because the length of the non - zero prefix grows by 2 with each application of @xmath237 . thus @xmath237 has infinite order . the subsemigroup @xmath787 is free . indeed , let @xmath53 be a nonempty word in @xmath788 . if @xmath789 in @xmath690 , then @xmath53 contains both @xmath237 and @xmath238 , because they both have infinite order . suppose the length of @xmath53 is minimal among all nonempty words over @xmath790 representing the identity element in @xmath690 . then one of the projections of @xmath53 will be shorter than @xmath53 , nonempty , and will represent the identity in @xmath690 , which contradicts the minimality assumption . thus @xmath791 in @xmath690 , for any nonempty word in @xmath788 . 1 . if @xmath53 ends in @xmath794 then @xmath559 is a word that is shorter than @xmath53 ending in @xmath237 , while @xmath795 is a word not longer than @xmath41 ending in @xmath238 . since @xmath796 in @xmath690 and @xmath797 , we have a contradiction . if @xmath53 ends in @xmath798 then @xmath799 is a word shorter than @xmath53 ending in @xmath238 , while @xmath626 is a word not longer than @xmath41 ending in @xmath237 . since @xmath800 in @xmath690 and @xmath801 , we have a contradiction . 3 . if @xmath802 then @xmath803 in @xmath63 and @xmath626 is a nonempty word , which is impossible , as already proved above . it is proved in @xcite that @xmath450 is generated by @xmath808 and @xmath809 . we have then @xmath810 . conjugate the right hand side of the wreath recursion by @xmath811 to obtain a copy of @xmath450 given by @xmath812 , @xmath813 and @xmath814 ( this is equivalent to conjugating by @xmath815 in @xmath816 ) . this shows that @xmath807 is isomorphic to @xmath450 via the isomorphism @xmath817 and @xmath818 . moreover , they are conjugate by the element @xmath819 , where @xmath820 ( this is the automorphism of the tree changing all letters which stand on even places ) . the group @xmath807 is contained in @xmath577 as a subgroup of index @xmath221 ( see @xmath577 ) . it contains the torsion free subgroup @xmath530 mentioned in the discussion of @xmath577 as subgroup of index 2 and is a weakly branch group over @xmath596 . all schreier graphs on the boundary of the tree have polynomial growth of degree @xmath221 . diameters of schreier graphs on the levels grow as @xmath822 ( see @xcite for details ) . laurent bartholdi and rostislav i. grigorchuk . lie methods in growth of groups and groups of finite width . in michael atkinson et al . , editor , _ computational and geometric aspects of modern algebra _ , volume 275 of _ london math . _ , pages 127 . cambridge univ . press , cambridge , 2000 . laurent i. bartholdi and zoran unik . some solvable automaton groups . in _ topological and asymptotic aspects of group theory _ , volume 394 of _ contemp . _ , pages 1129 . soc . , providence , ri , 2006 . steven c. ferry , andrew ranicki , and jonathan rosenberg . a history and survey of the novikov conjecture . in _ novikov conjectures , index theorems and rigidity , vol . 1 ( oberwolfach , 1993 ) _ , volume 226 of _ london math . lecture note ser . _ , pages 766 . cambridge univ . press , cambridge , 1995 . rostislav i. grigorchuk and andrzej uk . spectral properties of a torsion - free weakly branch group defined by a three state automaton . in _ computational and statistical group theory ( las vegas , nv / hoboken , nj , 2001 ) _ , volume 298 of _ contemp . _ , pages 5782 . soc . , providence , ri , 2002 . wolfgang lck . , volume 44 of _ ergebnisse der mathematik und ihrer grenzgebiete . 3 . folge . a series of modern surveys in mathematics [ results in mathematics and related areas . 3rd series . a series of modern surveys in mathematics]_. springer - verlag , berlin , 2002 . v. nekrashevych and s. sidki . automorphisms of the binary tree : state - closed subgroups and dynamics of @xmath825-endomorphisms . volume 311 of _ london math . _ , pages 375404 . cambridge univ . press , 2004 .

single - transverse spin asymmetries ( ssas ) in high - energy collisions with one transversely polarized hadron are important phenomena that have been observed for more than three decades in various physical processes @xcite . in these processes , the observed final - state hadrons show an asymmetric distribution in a plane perpendicular to the beam direction when the transversely polarized hadron scatters off an unpolarized hadron ( or a virtual photon ) . the ssa is defined as @xmath2 , the ratio of the difference and the sum of ( differential ) cross sections when the hadron s spin vector , @xmath3 , is flipped . recent experimental measurements of ssas in polarized semi - inclusive lepton - nucleon deep inelastic scattering ( sidis ) @xcite and in hadronic collisions @xcite have renewed the theoretical interest in ssas and in understanding their roles in hadron structure and quantum chromodynamics ( qcd ) . although it was realized a long time ago @xcite that perturbative qcd can be used to study the effects of transverse spin , the size of the observed asymmetries came as a surprise and has posed a challenge for researchers in this field @xcite . within a model calculation @xcite , brodsky , hwang , and schmidt showed that the final state interaction in deep inelastic scattering ( dis ) can generate a phase required for a nonzero ssa in sidis . it was later realized that this final state interaction can be factorized into the gauge link of the gauge invariant transverse momentum dependent ( tmd ) quark distributions @xcite . the nonvanishing ssa obtained in ref . @xcite is a consequence of the existence of a naively time - reversal - odd tmd quark distribution , the so - called sivers function @xcite . when applying the same calculation to the drell - yan production of lepton pairs at hadron colliders , the final state interaction in sidis becomes an initial state interaction in drell - yan , and the phase changes sign , which leads to a prediction of a sign change in the ssas between these two processes @xcite . this nontrivial `` universality '' property associated with the tmd parton distributions is the consequence of gauge interactions in qcd @xcite , and of the qcd factorization theorems for these two processes @xcite . experimental tests of this prediction will be crucial for our understanding of the origin of ssas in qcd @xcite . in ref . @xcite , it was proposed to study the sivers functions by means of a ssa in azimuthal - angular correlations of two jets produced nearly back - to - back at hadron colliders . measurements of this ssa for dijet production have begun at rhic @xcite , complementing the measurements in sidis . unlike the sidis or drell - yan process , dijet production at hadron colliders involves both initial and final state interactions that may produce the phase needed for a nonvanishing ssa . consequently , the sign and the size of the asymmetry will depend on the relative strength of these interactions . following the previous works on sidis and drell - yan , the authors of ref . @xcite developed a systematic approach to describe the role of initial- and final - state interactions in generating ssas in hadronic collisions , and they found that summing all initial / final state interactions into the gauge link of the tmd parton distributions leads to a very complicated functional form of the gauge link . in particular , the tmd parton distributions studied in dijet correlations in hadronic scattering will have no connection to those in the sidis and drell - yan processes , because their definitions are completely different @xcite . one then has to question the universality of the tmd parton distributions , and the predictive power of perturbative qcd calculations , which relies on comparing physical observables with the same factorized long - distance physics while having different perturbatively calculable short - distance dynamics . in ref . @xcite , we briefly reported a new result for the ssa in dijet production in hadronic collisions in the twist-3 efremov - teryaev - qiu - sterman ( etqs ) approach @xcite . we considered the spin - dependent cross section , @xmath4 , for the process @xmath5 with the jet momenta @xmath6 and @xmath7 . when both @xmath8 and @xmath9 , the transverse components of the momenta @xmath10 and @xmath11 , respectively , are much larger than @xmath12 , a nonvanishing @xmath13 can be generated by the etqs mechanism in the collinear factorization approach . in this framework , the ssas are attributed to the spin - dependent twist - three quark - gluon correlation functions , which correspond to a quantum interference between different partonic scattering amplitudes . since the incoming partons are approximated to be collinear to the corresponding incoming hadrons in this approach , the momentum imbalance of the two jets is generated by producing a three - parton final - state . @xcite , we calculated the contribution from initial - state gluon radiation to @xmath13 in the kinematic region where @xmath14 . we presented the final result for the leading contributions to @xmath13 in the expansion of the partonic scattering in @xmath15 involving a hard @xmath16 subprocess . in this paper , we will provide the detailed derivations of this result , and we will also present the full contributions from all other partonic subprocesses at the same order . we find that although both initial - state and final - state interaction lead to a nonvanishing ssa , the final - state interactions give the dominant contributions to @xmath13 . we therefore expect that the ssa in dijet production will have the same sign as the sivers asymmetry in sidis . we find that at leading order in the @xmath15 expansion , the perturbatively calculated partonic parts can be further factorized into a single - scale ( @xmath8 ) hard part and perturbatively generated tmd parton distributions with transverse momenta @xmath17 . we also find that our perturbatively calculated result is equal to the leading - order term in the @xmath18 expansion of the following generalized tmd factorization formula @xcite : @xmath19_c\ , \delta^{(2)}(\vec{k}_{1\perp}+\vec{k}_{2\perp}+ \vec{\lambda}_\perp-\vec{q}_\perp ) \ , , \nonumber\end{aligned}\ ] ] where @xmath20 runs over all parton flavors , @xmath21 and @xmath22 are partonic hard and soft factors , respectively , and the @xmath23_c$ ] represents a trace in color space between the hard and soft factors due to the color flow into the jets @xcite . the hard factor in eq . ( [ e4 ] ) only depends on the single hard scale @xmath8 in terms of partonic mandelstam variables of the reaction @xmath24 : @xmath25 with @xmath26 , @xmath27 and @xmath28 and @xmath29 the rapidities of the two jets . in eq . ( [ e4 ] ) , @xmath30 and @xmath31 denote the transverse - spin dependent tmd quark distributions ( known as the sivers function ) and the unpolarized tmd quark distribution , respectively ; these tmd parton distributions were chosen to follow their definitions in the semi - inclusive dis process . for example , for a polarized proton with momentum @xmath32 with @xmath33 and transverse spin vector @xmath34 , the tmd distribution for quark flavor @xmath35 can be defined through the decomposition of the following matrix element @xcite , @xmath36 \ , \label{e3}\end{aligned}\ ] ] where @xmath37 is a hadronic mass scale introduced to keep the tmd distributions @xmath38 and @xmath39 at the same mass dimension , and the gauge link @xmath40 is defined in a covariant gauge as @xmath41 with the path link extended to @xmath42 . @xmath43 is a vector conjugate to the momentum vector @xmath10 . since we will work in a covariant gauge throughout this paper , the vector @xmath43 could be chosen to be a light - cone vector with @xmath44 and @xmath45 . an off - light - cone vector ( @xmath46 ) will have to be used when high order corrections are taken into account @xcite . if we work in a singular gauge , like the light - cone gauge , an additional gauge link at the spatial infinity ( @xmath47 ) will have to be included in order to ensure the gauge invariance of the above definitions @xcite . we have chosen tmd parton distributions defined in sidis because of the dominance of final - state interactions . choosing tmd parton distributions defined according to the drell - yan process would change the sign of the partonic hard factors , but not affect the overall sign of the physical cross section . we note that the various factors in eq . ( [ e3 ] ) , apart from being functions of transverse momentum , also depend on the renormalization and factorization scales , and especially on the gluon rapidity cut - off when higher - order corrections are taken into account . the latter dependence is governed by the collins - soper evolution equation @xcite , which leads to a resummation of large logarithms of the form @xmath48 $ ] in the perturbative series @xcite . in order to evaluate the ssa for dijet production , we also calculate in this paper the contributions from initial - state gluon radiation to the spin - averaged dijet cross section , @xmath49 , in the same kinematic region where @xmath50 . using the collinear factorization approach , we find that like the spin - dependent case , the leading contribution in the @xmath15 expansion of the perturbatively calculated partonic scatterings can be further factorized into a hard part at a single - scale @xmath8 and perturbatively generated unpolarized tmd parton distributions with transverse momenta @xmath17 . our leading contribution in the @xmath15 expansion is equal to the leading - order term of the following tmd factorization formula @xcite , @xmath51_c \ , \delta^{(2)}(\vec{k}_{a\perp}+\vec{k}_{b\perp } + \lambda_\perp-\vec{q}_\perp ) \ ; , \label{e2}\end{aligned}\ ] ] where the superscript `` @xmath52 '' indicates the scattering of an unpolarized beam off an unpolarized target / beam . for consistency , we again express the factorization formula in terms of tmd parton distributions defined in the sidis process . actually , the unpolarized tmd parton distributions defined in sidis and the drell - yan process are identical , because the unpolarized parton distributions are invariant under the naive - time - reversal transformation . a key feature of any qcd factorization is that the perturbatively calculated short - distance hard factors should not be sensitive to details of the _ factorized _ long distance physics . in this paper , we will derive all short - distance hard factors in eqs . ( [ e4]),([e2 ] ) also by using the brodsky - hwang - schmidt model for ssas @xcite . we find that the hard - scattering factors derived in this model are in fact the same as those derived in the collinear factorization approach , despite of the clear difference in the treatment of the nucleon . both the spin - averaged and the spin - dependent cross sections calculated in this model are therefore consistent with the generalized factorization formulas in eqs . ( [ e4]),([e2 ] ) . to fully investigate the above factorization formalism , we would have to consider the contributions from gluon interactions and radiations in all possible regions of the phase space , and to all orders in perturbation theory , which is not what we are trying to do in this paper . instead , as a step to test the above factorization formalism , we will study the contribution at the first non - trivial order with a single gluon radiated nearly parallel to one of the incident nucleons . from this study , we will be able to show how to factorize this gluon into the tmd parton distributions of the incoming nucleons , and to verify the definition of the tmd parton distributions used in the generalized factorization formulas . our first order calculation at @xmath53 clearly shows the factorization of a hard part at @xmath54 and that the radiation of a collinear gluon can be absorbed into the relevant tmd parton distributions . however , our work does not address the factorization between the various tmd parton distributions at the scale @xmath55 . our calculation should be regarded merely as a `` necessary condition '' for such a factorization to hold . since the predictive power of perturbative - qcd calculations involving measured hadrons relies on factorization , a full proof or disproof of this factorization remains an important challenge @xcite . the rest of this paper is organized as follows . in sec . ii , we calculate the partonic hard factors in a simple model - inspired approach . using the quark - diquark model for the proton and the brodsky - hwang - schmidt model for ssas , we extract the hard factors by the factorization of the tmd parton distributions . we find that the hard factors for the unpolarized scattering in eq . ( [ e4 ] ) are identical to the partonic differential cross sections @xmath56 , and the ones for the transverse - spin dependent cross section are different due to the initial- and final - state color interactions . in sec . iii , we calculate the ssa in the collinear factorization approach , and show that the collinear gluons parallel to the polarized proton can be factorized into the sivers function , while those parallel to the unpolarized proton can be factorized into the unpolarized tmd parton distribution . we also demonstrate that the hard factors calculated in both approaches are the same and independent of the details of the factorized long - distance physics . in sec . iv , we extend our formulas to the @xmath9-weighted ssa , and compare with previous results in the literature . finally , we summarize our paper in section v. in this section , we will calculate the hard factors at @xmath57 in the generalized factorization formulas in eqs . ( [ e4]),([e2 ] ) , based on a model in which an energetic parton scatters off a nucleon that is made of a quark and a scalar diquark @xcite . we will follow the approach of brodsky et al . @xcite and incorporate the correct color interactions between the gauge boson and the scalar diquark in the proton wave function . we extract the hard factor by factorizing the model - sensitive long - distance physics from the differential cross section into the parton distributions . because the model - dependence was factorized into the parton distributions , the extracted hard factors resulting from our calculations do not depend on how the nucleon couples to the partons , and therefore , should be model - independent . " represents the color - index for the gauge boson , and @xmath58 the color - indices for the quark and/or the di - quark . _ , width=340 ] -0.4 cm in this subsection , we derive the hard factors @xmath59 for the unpolarized differential cross section in eq . ( [ e2 ] ) . we will first give a simple example of how the hard factors may be extracted in the context of the quark - diquark model of @xcite . we will also give a more general approach based on power counting techniques to calculate the hard factors directly from partonic diagrams . to specify the quark - diquark model of the proton , we give in fig . [ f2 ] graphical rules for the vertex coupling the quark - diquark to the proton , and the vertex coupling the diquark to a gauge boson . in order to correctly take into account the color degree of freedom of the strong interactions , we included proper color factors for these vertices . using this quark - diquark model to extract the hard factors not only verifies the fact that the hard factors are independent of the dynamics at the scale of the proton , but also provides a relatively simpler demonstration of how to factorize the cross sections into the tmd quark distributions and corresponding hard factors at the leading order in the @xmath15 expansion . in this particular model , we calculate the cross section for an energetic parton of momentum @xmath60 scattering off a parton of momentum @xmath61 from the ( modeled ) nucleon , producing two jets of momenta @xmath62 and @xmath63 plus an unobserved particle with momentum @xmath64 in the final state , as shown in fig . [ f3 ] , where the feynman diagram represents the contribution from a quark - quark scattering channel : @xmath16 with the quark @xmath11 from the nucleon . we extract the hard factors by comparing the calculated cross section with the factorized formulas in either eq . ( [ e4 ] ) or ( [ e2 ] ) . the diquark in the final state has a transverse momentum @xmath65 to balance the small momentum imbalance of the two jets , @xmath66 . channel , for the unpolarized cross section . _ , width=207 ] -0.4 cm according to the factorization formulas in eqs . ( [ e4]),([e2 ] ) , the transverse momentum imbalance @xmath9 of the two jets is a consequence of adding the transverse momenta of the two active partons and from the soft factor . since the incoming quark in this model calculation has only a longitudinal momentum component , its transverse as well as longitudinal momentum distributions should be given by delta functions , @xmath67 similarly , at this order , the soft factor is also a delta function of the transverse momentum , i.e. , @xmath68 . therefore , the only contribution to the transverse momentum imbalance of the two jets comes from the transverse momentum of the quark emerging from the nucleon , and the generalized factorization formulas can be reduced to @xmath69 for the unpolarized cross section , and @xmath70 for the single transverse - spin dependent cross section . in order to extract the hard factor from the differential cross sections in eqs . ( [ e7]),([e8 ] ) , we will need to calculate the tmd quark ( unpolarized and sivers ) distributions within the same model . they are given by the diagrams shown in fig . [ f4 ] , and are available @xcite . we can summarize their results as follows : @xmath71 \ , \label{e9}\\ q_t^{\rm sidis}(x , k_\perp)&=&c_f\frac{g_n^2\alpha_s}{16\pi^3}\frac{(1-x)(m_q+xm_p ) } { \lambda^2(x , k_\perp^2)}\frac{m_p}{k_\perp^2 } \ln\frac{\lambda^2(x , k_\perp^2)}{\lambda^2(x,0 ) } \ , \label{e10}\end{aligned}\ ] ] where @xmath72 and @xmath73 , and @xmath74 , @xmath75 and @xmath76 are masses for the proton , the quark and the gauge boson , respectively . in the above formulas , we have included the color - factors arising from the vertices described in fig . [ f2 ] . -0.4 cm we now present details of the calculation of the contribution of the quark - quark scattering channel @xmath77 to the unpolarized scattering cross section . we also provide full results for all other channels , which can be derived similarly . the feynman diagram for @xmath77 has been shown in fig . [ f3 ] . in the limit @xmath78 , the two final - state jets have approximately the same transverse momentum , i.e. , @xmath79 . we define the following kinematic variables : @xmath80 where @xmath81 . from these , we immediately find the incident partons momentum fractions @xmath82 and @xmath83 , and the scalar diquark s momentum can be written as @xmath84 in the evaluation of the cross sections , we use the power counting analysis @xcite , keep only the leading power contributions and neglect all higher order corrections in @xmath85 , where @xmath86 represents any lower mass scale like @xmath9 , @xmath74 , @xmath75 and @xmath76 . in this limit , the differential cross section for fig . [ f3 ] can be written as @xmath87 is the scattering amplitude for the diagram . expanding the above delta function for the phase space integral of @xmath64 , we find that in the leading power contributions @xcite , @xmath88 where the `` plus '' distribution follows the usual definition @xcite . this delta function will help to simplify our calculations for contributions from different kinematic regions of @xmath64 . for example , if @xmath64 is parallel to @xmath89 , which means @xmath90 , we only have a contribution from the first term of the above expansion , which also implies @xmath91 in this limit . furthermore , the quark propagator in fig . [ f3 ] reads @xmath92 where @xmath93 is defined above after eq . ( [ e10 ] ) , and is order of @xmath94 . combining this propagator with the delta function expansion , we find that indeed only the first term in the delta function contributes . neglecting all higher order terms in @xmath85 , we obtain the amplitude squared for fig . [ f3 ] as @xmath95 \ , \ ] ] where @xmath96 is the color - factor , and the hadronic mandelstam variables @xmath97 , @xmath98 , and @xmath99 are defined as @xmath81 , @xmath100 , @xmath101 . in the limit @xmath91 , these variables can be related to the partonic mandelstam variables as : @xmath102 , @xmath103 , @xmath104 at @xmath105 , and we will have @xmath106 . substituting the above results into the differential cross section , we obtain , @xmath107 \frac{2\pi\alpha_s^2}{\hat s^2}\ , \frac{\hat s^2+\hat u^2}{\hat t^2}\ , \delta(x_b-1 ) \ .\ ] ] comparing this result with eq . ( [ e7 ] ) , and identifying the tmd parton distribution @xmath108 defined in eq . ( [ e9 ] ) , we extract the partonic hard factor for the differential cross section from the diagram in fig . 3 as@xmath109 where @xmath110 are the partonic mandelstam variables defined before . this hard factor is just the well - known differential cross section @xmath111 for the partonic process @xmath16 . the hard factors for all other partonic channels can be calculated similarly , and they are all identical to the partonic differential cross section @xmath112 , and collected in appendix a. instead of calculating the differential cross section directly , we now show how to use the power counting techniques @xcite to analyze the contribution from all partonic channels and to demonstrate the factorization of the hard factors . we then apply this approach to calculate the contributions to the spin dependent differential cross section . we again first consider the simple example of the @xmath16 channel . the jet transverse momentum @xmath8 is the only hard scale , while all other momentum scales , such as the imbalance @xmath9 of the two jets and all other mass scales @xmath74 , @xmath75 and @xmath76 , are relatively soft and much smaller than the hard scale . we expand the scattering amplitude squared in the small parameter @xmath15 . using power counting techniques , we identify the leading power contributions and match them to the momentum regions of the unobserved ( or integrated ) parton of momentum @xmath64 . in our current model , where a parton of momentum @xmath113 scatters off a nucleon of momentum @xmath89 , the leading power contribution is from the region when @xmath64 is parallel to @xmath89 . if @xmath64 is parallel to @xmath113 , the contribution is power suppressed , because the denominator of the quark propagator @xmath114 will be of order @xmath115 , instead of @xmath116 as is the case when @xmath64 is parallel to @xmath89 . into the tmd quark distribution of the proton and the hard factor represented by the partonic scattering feynman diagram.__,width=491 ] -0.4 cm when the momentum @xmath64 is parallel to @xmath89 , the quark propagator in fig . [ f3 ] has a virtuality @xmath117 , which is much smaller than the hard scale @xmath118 . compared to the short distance physics in the partonic scattering taking place at the time scale @xmath119 , the incident quark of momentum @xmath120 is long - lived , and the contribution from this subprocess can be separated into two parts : the parton distribution part relevant at the soft scale @xmath121 and the hard partonic part at the scale @xmath122 , representing the long distance and short distance physics , respectively . we demonstrate this factorization in fig . [ f5 ] , where the contribution to the differential cross section from the process in fig . [ f3 ] is factorized into a parton distribution multiplied by a hard factor . we can further separate the hard factor into a partonic scattering amplitude squared stripped of any color , and the color - factor @xmath123 . in this way we can take into account the color decomposition of the subprocess in the case of unpolarized scattering . this is trivial for this particularly simple partonic channel , but becomes a convenience for the more complicated ones . after subtraction of the parton distribution from the differential cross section , the hard factor in fig . [ f5 ] can then be written as @xmath124 where @xmath125 represents a common factor and is introduced to simplify the notation , and @xmath126 . this of course agrees with the above result in eq . ( [ hqqp ] ) . most importantly , it demonstrates that the hard factor is model independent . it only depends on the partonic scattering cross section and the color - factor associated with the partonic diagram . it does not depend on the model we used to describe how the nucleon couples to the quark and diquark . -0.4 cm a non - vanishing single transverse - spin asymmetry requires initial / final state interactions generating a phase . because all initial and final partonic states relevant for dijet production are colored , both initial and final state interactions have to be taken into account for a complete result . in fig . [ f6 ] , we show all possible diagrams with gluon exchange between the diquark and the hard scattering part which may contribute to the ssa for the dijet - correlation , again for the case of an underlying @xmath16 process . the initial and final state interaction diagrams in fig . [ f6](a - c ) contribute to a soft gluon pole , which corresponds to the quark sivers function contribution we are considering in this paper . however , the gluon interaction with the internal gluon propagator shown in fig . [ f6](d ) does not contribute to a soft gluon pole , but to a soft fermion pole @xcite . this soft fermion pole contribution is not related to the quark sivers function and will not be discussed in this paper . therefore , in the following calculations , we will only consider the contributions from the diagrams ( a)-(c ) . we will follow the same factorization approach that was used in last subsection for calculating the hard factors for the unpolarized cross section . we use the power counting method to factorize the additional gluon interaction between the hard scattering and the diquark into the sivers parton distribution for every partonic channel . as discussed above , the leading power contribution is from the region of phase space where the outgoing diquark momentum @xmath64 is parallel to the polarized proton @xmath89 : @xmath127 , @xmath128 and @xmath129 . to obtain a leading power contribution , the gluon momentum @xmath130 also needs to be parallel to @xmath89 . this is so because the diquark propagator in fig . [ f6 ] reads @xmath131 which will be power suppressed unless @xmath132 . this result can also be derived from the existence of a `` pinch '' singularity in the @xmath133 integral over the quark propagator of momentum @xmath134 and the diquark propagator of momentum @xmath135 @xcite . since the momentum @xmath130 of the gluon is parallel to @xmath89 , its polarization will also be proportional to @xmath89 , because the gauge - boson - scalar vertex is proportional to @xmath136 , as shown in fig . therefore , the gluon attaching to the initial- or final - state quark is longitudinally polarized , i.e. , its polarization is along its momentum . because of this , we can further decouple the gluon interaction with the external particles by using the eikonal approximation . for example , for the diagram with final - state interaction in fig . [ f6](b ) , the gluon interaction part can be reduced to @xmath137 where the gamma matrix @xmath138 appears because of the interaction with a longitudinally polarized gluon , and where @xmath35 is the color - index for this gluon . in the above derivations , we have only kept the leading power contributions . for ( 1,2,4,5 ) where @xmath35 is the color index for the gluon with momentum @xmath130 and @xmath58 are indices for the quark lines ; @xmath139 for ( 3,6 ) where @xmath140 are indices for incoming and outgoing gluons.__,width=529 ] -0.4 cm similarly , the initial state interaction in fig . [ f6](a ) and the other final state interaction in fig . [ f6](c ) can be simplified following the same eikonal approximation . the eikonal approximation will give rise to eikonal propagators and vertices . in fig . [ f7 ] , we summarize these eikonalized expressions for all the initial and final state interactions , including the attachments of the longitudinally polarized gluon to a quark or anti - quark line , or to a gluon line as well . the eikonal approximation for quark- and antiquark - gluon interactions with incoming or outgoing momentum can be derived following the above example . for the three - gluon interactions , we can derive the eikonal vertex and propagator by choosing physical polarizations for the external gluons . with this choice , we find the vertex and gluon propagator can be reduced to , for example for fig . [ f7](6 ) : @xmath141\frac{i(-g^{\nu\nu'})}{(p_b+l)^2+i\epsilon}g^{\rho -}\nonumber\\&&\approx \frac{(2p_b+l)^- g_{\mu\nu}}{(p_b+l)^2+i\epsilon}(-if_{abc})=\frac{g}{-l^+-i\epsilon } g_{\mu\nu}(-if_{abc } ) \ , \end{aligned}\ ] ] where we have used the fact that the incoming gluon is transversely polarized , so that the polarization tensor can be chosen as @xmath142 similarly , we can simplify the longitudinally polarized gluon attachment to the outgoing gluon line as shown in fig . [ f7](3 ) . these eikonal propagators and the associated vertices can be absorbed into the parton distribution for the polarized proton , after factorizing the diagram . the only difference between this factorization and the one for the unpolarized scattering is the color - flow , because of the additional gluon exchange between the diquark and the hard part . the color - flow is sensitive to the difference between initial- and final - state interactions and depends on the specific partonic subprocess . therefore , the generalized factorization would immediately fail if the eikonal propagators and associated vertices , which are absorbed into the tmd distributions , were sensitive to the process dependent color flows . however , at the leading order , one can always attribute the color - flow to the partonic hard part and keep the corresponding tmd distribution insensitive to it . therefore , as compared to the unpolarized case ( see fig . [ f5 ] ) , for the single - spin - dependent cross section we shall obtain a more complicated factorization . for the case of the three diagrams of fig . [ f6 ] the factorization is shown in fig . the contribution from each diagram can be factorized into the sivers function and the hard partonic scattering amplitude squared , multiplied by a color - factor that takes into account the color - flow of the initial / final state interaction . however , in order to verify the generalized factorization beyond the leading order , one would have to show that additional initial - state as well as final - state gluon interactions could also be factorized into a sum of terms , each of which is factorized into a product of higher order tmd distributions and corresponding color factors and partonic hard parts . for example , to verify the factorization at the next - to - leading order with two eikonal gluon interactions , one would have to show that the initial- and final - state interactions can be factorized into the same color factors and the lowest order hard parts as in fig . 8 , multiplied by the corresponding second order tmd distribution . we will leave the detailed study of the generalized factorization at higher orders to a future publication . . _ , width=529 ] -0.4 cm at least to the leading order , the extracted color factors @xmath143 with @xmath144 in fig . [ f8 ] are insensitive to the long - distance physics . that is , they do not depend on how the quarks and the gluon couple to the nucleon . this is because the factorization of the color - flow between the hard partonic part and the tmd part results in a color configuration connecting an incoming and an outgoing quark ( with color indices @xmath145 ) and a gluon ( with color @xmath35 ) . the only possible color structure to describe this connection is @xmath146 , due to @xmath147 invariance and the wigner - eckart theorem . therefore , one can reduce all the color matrices from the hard part to a simple color matrix @xmath146 , and the coefficient will be exactly the color - factor in the factorization formula . in this way , the extracted color - factor can not depend on the bottom part of the diagram describing the coupling of the quarks and the gluon to the nucleon , but will only depend on the color - flow in the hard partonic diagrams with an additional gluon insertion . for example , the color factors appearing in fig . [ f8 ] can be derived in the following way , @xmath148 where the color - matrices on the left side of the equations come from the partonic diagrams with one gluon attachment to the initial- or final - state quark line . the extracted color - factor @xmath149 is for the initial - state interaction , and @xmath150 and @xmath151 are for the final - state interactions on the lines with the momentum @xmath62 and @xmath63 , respectively . in the factorization shown in fig . [ f8 ] , the leading order sivers functions for all three diagrams are the same . this is because we normalize all eikonal vertices to the one used in the calculations of the sivers function in sidis ( shown in fig . the difference from the initial or final state interaction effects is summarized into the relevant color - factors . for other partonic channels , it is in some cases necessary to introduce an extra sign when the eikonal propagator contributes with an opposite phase compared to that in the sivers function in sidis . for example , the eikonal propagators in the diagrams ( 2,5,6 ) of fig . [ f7 ] contribute an opposite sign , whereas those of ( 1,3,4 ) contribute the same sign as in sidis . because all the initial and final state interactions in fig . [ f8 ] contribute the same sign , there is no sign change for their color - factors . another important point is that the partonic scattering amplitude squared is the same for all the three diagrams in fig . it can be calculated directly from the relevant feynman rules for the partonic diagrams and is also the same as the one for unpolarized scattering given above . after factorizing out the sivers function from the cross section contributions for the diagrams in fig . [ f8 ] , the color factors for the three diagrams can then be summed to give the final result for the hard factor for the @xmath16 subprocess : @xmath152h_{qq'\to qq'}= \frac{\alpha_s^2\pi}{\hat s^2}\frac{n_c^2 - 5}{4n_c^2}\frac{2(\hat s^2+\hat u^2)}{\hat t^2 } \ , \label{hqq}\ ] ] where @xmath153 has been defined above after eq . ( [ hqu ] ) . the above derivations can be extended to all other partonic channels involving the quark sivers function . they can also be extended to the case of sub - processes initiated by the gluon sivers function for which , however , more color - structures ( like @xmath154 and @xmath155 for the three gluon coupling to the nucleon state ) will emerge @xcite . in this paper , we will focus on the processes with a quark sivers function , which include all quark - quark and quark - antiquark scattering processes , such as @xmath16 calculated above , @xmath156 , @xmath157 , @xmath158 and @xmath159 , as well as the quark - gluon process @xmath160 and quark - antiquark annihilation into a gluon pair , @xmath161 . following the above derivations for the @xmath16 process , the hard factors for the unpolarized and single - spin - dependent cross sections for these channels can be summarized by the following master formulas : @xmath162 where @xmath163 labels an individual feynman diagram ( meant as the square of an amplitude or an interference between two amplitudes , see below ) with @xmath164 the associated expression for the hard factor . @xmath165 is the color - factor for the unpolarized cross section , @xmath166 for the initial state interaction for the single - spin dependent cross section , and @xmath167 and @xmath168 for the final state interactions when the gluon attachment is to the lines of momentum @xmath62 and @xmath63 respectively . from the above analysis , we can easily see that the hard factors in eqs . ( [ e29]),([e30 ] ) are model - independent . the model - dependence is removed once we factorize out the unpolarized quark distribution or the sivers function from the differential cross sections . from the above master formulas , the hard factors depend on the partonic scattering amplitude squared , and the color - factors , derived from partonic diagrams with or without the additional gluon attachment . both the color and partonic factors , and therefore the hard factors , are independent of the model for the proton used in our calculation . -0.4 cm .the color - factors and hard cross sections of individual diagrams of fig . [ f9 ] for the quark - quark scattering channels . [ cols="<,^,^,^,^,^,^",options="header " , ] in summary , in the intermediate transverse momentum region @xmath169 , we will only have leading - power contributions to the derivative terms from the expansion of the delta function for the on - shell condition of the radiated gluon @xmath64 , and when @xmath64 is parallel to @xmath89 . we have shown these diagrams in fig . this part of the contribution is relatively easy to derive , as we mentioned above , by multiplying a factor @xmath170 to the unpolarized cross sections . the only additional involvement is the color factor . we list the color - factors for all diagrams of fig . [ f15 ] in table iv . each of the color factors has a factor @xmath171 . this can be easily seen for the color factors for diagrams @xmath172 , @xmath173 and @xmath174 , for which we have the following simplifications , @xmath175 where @xmath149 , @xmath150 and @xmath151 have been defined above in eqs . ( [ e25]),([e26]),([e27 ] ) . combining these color factors with the squared amplitude calculated in the last subsection for the unpolarized cross section , we find that the contributions to the derivative terms can be factorized into the hard factor given in eq . ( [ hqq ] ) times the derivative terms of the perturbatively generated sivers function for the polarized nucleon in eq . ( [ e38 ] ) . we illustrate this factorization in fig . for example , the derivative contribution from diagram ( @xmath172 ) is factorized into a soft pole diagram contribution to the derivative term in the sivers function where the radiated gluon attaches to the quark line shown in fig . [ f17 ] , multiplied by the color - factor @xmath149 and the partonic amplitude squared for @xmath16 . @xmath173 and @xmath174 will be factorized into the same diagram contribution to the quark sivers function , but with different color - factors @xmath150 and @xmath151 , respectively . the same conclusion holds for all other diagrams . for example , the sum of diagrams ( @xmath176-@xmath177 ) is factorized into a soft pole diagram contribution to the sivers function where the radiated gluon attaches to the gauge link , multiplied by the color factor @xmath149 and the partonic amplitude squared for @xmath16 . is parallel to @xmath89 . this part can be factorized into a soft pole contribution to the perturbatively generated sivers function in polarized hadron @xmath89 . this factorization applies both to the derivative and the non - derivative contributions from the soft poles . the circle with the attached gluon in the feynman diagrams represents all possible gluon attachments . likewise , the gluon attachment to the circle in the sivers function diagram includes the diagrams with gluon attachments to the gauge link and to the quark line.__,width=453 ] -0.4 cm finally , we present the total result for the derivative contribution by the @xmath178 process to the single - spin asymmetry for the dijet - correlation , @xmath179 \nonumber\\ & \times & \frac{\alpha_s}{2\pi^2}\ , \left[-\frac{1}{2n_c}\right ] \int dx\ , \xi \left[x\frac{\partial}{\partial x}t_f(x , x)\right ] \left[1+\xi^2\right ] \ , , \label{x - ssa - d}\end{aligned}\ ] ] where the superscript `` d '' indicates the derivative term , and where the single - scale partonic hard part @xmath180 has been defined in eq . ( [ hqq ] ) . the result in eq . ( [ x - ssa - d ] ) indeed reproduces the derivative terms in the first order tmd factorization formula in eq . ( [ e39 ] ) . in addition to the derivative contributions , the soft poles also contribute to the non - derivative terms . one way of calculating the non - derivative terms is to follow the @xmath181-flow in the scattering amplitude squared , summing up all contributions . keeping in mind that @xmath182 , we have to keep the full dependence on @xmath183 and @xmath184 in the calculations . it is easy to keep track of the explicit @xmath185 dependence coming from momentum conservation . in addition , we also need to include the `` indirect '' @xmath185 dependence resulting from kinematic constrains : the on - shell condition of the radiated gluon and the double poles in the final state interactions . for example , the on - shell delta function for the momentum @xmath64 of the radiated gluon requires that the longitudinal momentum component depends on @xmath181 , which will lead to a @xmath181 dependence for the longitudinal momentum fractions of @xmath186 and @xmath187 as well . as mentioned above , if the polarized gluon attaches to the left side of the cut , the radiated gluon will have momentum @xmath188 . since @xmath187 has transverse momentum @xmath183 , the transverse component of @xmath189 will be equal to @xmath190 . furthermore , the minus component of @xmath189 is fixed by momentum conservation to @xmath191 , where the contribution by @xmath192 has been neglected , because it is of order @xmath193 . therefore , @xmath194 does not depend linearly on @xmath181 . the plus component of @xmath189 , on the other hand , will have linear @xmath181 dependence because of the on - shell condition . we therefore parameterize @xmath189 as @xmath195 where we have indicated the term linear in @xmath183 in the plus component of @xmath189 . as mentioned above , the minus component @xmath196 does not depend linearly on @xmath181 and can be determined to be @xmath197 . by using the on - shell condition @xmath198 , we can then obtain the @xmath181-dependent term in the plus component of @xmath189 , and the final parameterization result for @xmath189 is @xmath199 it is easy to check that @xmath198 up to terms quadratic in @xmath183 . from this result , we can also determine the @xmath183-dependence of the plus components of @xmath186 and @xmath187 . for example , @xmath187 can be directly calculated from momentum conservation , and we find @xmath200p_a^\mu+k_{2\perp}^\mu \ , \label{e76}\ ] ] where the minus component of @xmath187 has again been neglected . we next obtain the parameterization for @xmath186 by @xmath201 , where @xmath202 is determined by the pole and will be different for different poles . for the initial state interaction , the pole @xmath203 gives @xmath204 , and the longitudinal momentum fraction does not depend on the linear term in @xmath181 . however , for the final state interaction with gluon attachment to the line with momentum @xmath62 , the pole will be given by @xmath205 , and the plus component of @xmath202 will depend on @xmath181 through @xmath206 . similarly , the final state interaction with @xmath63 will lead to @xmath207 . in this way , the parameterizations for @xmath186 and @xmath202 will vary for different poles , whereas that for @xmath187 will always remain the same . this is because the parameterization of @xmath187 comes from the on - shell condition for the radiated gluon , which is not sensitive to the location of the poles . when the polarized gluon attaches to the right of the cut lines , on the other hand , it is @xmath186 that is determined by the on - shell condition for the radiated gluon , @xmath208 . following the same method as above , we find @xmath209p_a^\mu+k_{1\perp}^\mu \ .\end{aligned}\ ] ] furthermore , in this case @xmath210 , with @xmath202 determined from the position of the poles ; @xmath187 will be different for the different poles . with the above parameterizations for all the external particles momenta , the calculations for the linear expansion in @xmath181 are straightforward . the final result for the soft pole contributions to the non - derivative terms when @xmath64 is nearly parallel to @xmath89 have the following form : @xmath211 \nonumber\\ & \times & \frac{\alpha_s}{2\pi^2}\ , \left[-\frac{1}{2n_c}\right ] \int dx\ , \xi \ , t_f(x , x ) \left[\frac{2\xi^3 - 3\xi^2 - 1}{1-\xi}\right]\ , , \label{x - ssa - nd - sa}\end{aligned}\ ] ] where @xmath180 is given in eq . ( [ hqq ] ) . the separate factorization for initial and final state interactions diagrams in fig . [ f17 ] , which we used to describe the factorization of the derivative contributions , also applies to the non - derivative ones . for example , the contributions from diagrams ( @xmath172-@xmath177 ) of fig . [ f15 ] can be factorized into the sivers function shown in fig . [ f17 ] , multiplied by @xmath149 and the partonic scattering function , including both the derivative and non - derivative contributions . the same happens for the final state interaction diagrams ( @xmath173-@xmath212 ) and ( @xmath174-@xmath213 ) . there are also hard pole contributions when @xmath64 is parallel to @xmath89 . we have shown these diagrams in fig . [ f18 ] . as discussed above , the hard pole diagrams only contribute to the non - derivative terms . to calculate their contributions , we follow the above method as for the soft pole contributions . because the hard poles appear in the internal propagator , we have to take _ all _ possible gluon attachments into account . for example , diagrams @xmath172-@xmath214 of fig . [ f18 ] display all contributions with the same hard pole ( indicated by the short bars in the propagators ) , with three different attachments of the polarized gluon . another important feature for the hard pole contributions is that , after taking the pole , the polarized gluon will have nonzero longitudinal momentum fraction , i.e. , @xmath215 as we mentioned above . this is in fact the reason for referring to these poles as `` hard '' poles . for example , the hard poles in the diagrams @xmath172-@xmath214 have the momentum @xmath216 with @xmath217 , and their pole arises at the following value for @xmath218 : @xmath219 where the last equality comes from the fact that @xmath64 is parallel to @xmath89 and is dominated by its plus component as shown in eq . ( [ e66 ] ) . we have also neglected all higher power corrections in @xmath220 in the above derivations . from this result , we find the associated twist - three quark - gluon correlation function will take the form @xmath221 , that is , its two arguments are not at the same . this is different from the soft pole contributions where @xmath222 and @xmath223 are equal in the twist-3 quark - gluon correlation matrix element . is parallel to @xmath89 . this part can be factorized into a hard pole contribution to the sivers function in the polarized hadron @xmath224 . the little circle in the diagrams represents all possible gluon attachments , as required by gauge - invariance . for example , the first one in the left panel represents all three diagrams @xmath172-@xmath214 of fig . [ f18].__,width=453 ] -0.4 cm in our calculations , we further choose physical polarization for the radiated gluon , with the following polarization tensor : @xmath225 with this choice , we only have contributions from the three - gluon vertex diagrams @xmath176 , @xmath226 and @xmath227 and , apart from the color factor , each of them contributes a term @xmath228 . their color factors are straightforward to decompose : @xmath229 we then directly find that the contributions of these diagrams can be factorized into the hard pole contributions to the sivers function , multiplied by @xmath149 ( or @xmath150 and @xmath151 ) and the partonic hard - scattering function . we display this factorization graphically in fig . the total contribution is given by @xmath230 \nonumber\\ & \times & \frac{\alpha_s}{2\pi^2}\ , \left[-\left(\frac{1}{2n_c } + c_f\right)\right ] \int dx\ , \xi \ , t_f(x , x-\hat x_g ) \left[\frac{1+\xi}{1-\xi}\right ] \ , \label{x - ssa - nd - ha}\end{aligned}\ ] ] where @xmath231 and @xmath180 is given in eq . ( [ hqq ] ) . indeed , the above result reproduces the contribution to eq . ( [ e39 ] ) that arises from the hard pole part of the quark sivers function . in summary , by combining all the soft - pole and hard - pole contributions when the radiated gluon is nearly parallel to the polarized nucleon , we can reproduce the contribution from the perturbatively generated sivers function to the ssa in dijet - correlations in eq . ( [ e39 ] ) , derived from the tmd factorization formula in the regime @xmath169 . with these explicit calculations , we have demonstrated that at this perturbative order the contribution to the ssa when @xmath64 is parallel to @xmath89 can be factorized into the quark sivers function , multiplied by the initial / final state interaction color - factors ( @xmath149 , @xmath150 , and @xmath151 ) , and the partonic scattering amplitude squared . the calculation for the contributions when @xmath64 is parallel to @xmath113 follows the same procedure as above . as we have shown earlier , there is no contribution to the derivative terms when @xmath64 is parallel to @xmath113 , and so we only need to consider the non - derivative contributions . we will also have both soft and hard poles contributions . in fig . [ f20 ] , we have shown the soft pole diagrams , and in figs . [ f21],[f22 ] the hard pole ones . again we have to keep track of the @xmath181-flow in the scattering amplitudes . first , we need to parameterize the momenta of all the external particles including their linear dependence on @xmath181 . for example , when the polarized gluon attaches to the left of the cut line in the diagrams of figs . [ f20]-[f22 ] , the momentum @xmath189 of the radiated gluon and the quark momentum @xmath187 can be parameterized as @xmath232 and @xmath233 with @xmath202 determined by the pole . note that these parameterizations are different from those in eqs . ( [ e75]),([e76 ] ) . this is because , when @xmath64 is parallel to @xmath113 , its minus component is dominant and is parameterized by @xmath234 . again , this minus component does not depend linearly on @xmath181 . the plus component of @xmath189 does have linear dependence , which can be calculated from the on - shell condition @xmath198 . the linear @xmath181-dependence for the plus component of @xmath187 can be calculated as well . likewise , when the gluon attaches to the right of the cut , we will have @xmath235 and @xmath210 with @xmath202 determined from the position of the poles . again , different poles will give different results for @xmath202 , following the above discussions for the soft pole contributions . is parallel to @xmath113 . this contribution is factorized into the tmd quark distribution in the unpolarized nucleon @xmath236 . the little circle in the diagrams represents all possible gluon attachments . the bars indicate the poles . the second and third diagrams in the left panel represent the soft pole diagrams @xmath173-@xmath212 and @xmath174-@xmath213 of fig . [ f20 ] respectively . the first one contains both soft poles ( represented by the bar on the circle ( diagrams @xmath172-@xmath177 of fig . [ f20 ] ) ) and hard poles ( indicated by the bar on the straight line ( all diagrams in fig . [ f21])).__,width=453 ] -0.4 cm by using the above parameterizations for the momenta of the external particles , we can calculate the contributions from the soft pole diagrams in fig . [ f20 ] , and the hard pole diagrams in figs . [ f21 ] and [ f22 ] . we further find that the contribution from the hard pole diagrams in fig . [ f22 ] is power suppressed by @xmath15 . so , for the hard pole contributions , we only need to consider the diagrams shown in fig . [ f21 ] , which all have the same hard pole indicated by the bars on the propagators . as discussed in the last subsection , the hard pole contributions may in general lead to the polarized gluon having nonzero longitudinal momentum fraction , @xmath215 . however , when the radiated gluon s momentum @xmath64 is nearly parallel to @xmath113 , the hard poles lead to @xmath237 at leading power , as was the case for the soft pole contributions . this is very different from what we found for the hard pole contributions in the last subsection in the case @xmath64 parallel to @xmath89 . the momentum giving the hard pole , @xmath238 with @xmath217 , is the same as in the last subsection . however , taking the pole , we now find @xmath239 where the last equation holds in the limit @xmath240 . the reason for this is that , when @xmath64 becomes parallel to @xmath113 , it will be dominated by its minus component and can be parameterized as @xmath241 . substituting this into the above equation ( [ xdef ] ) , we find that the numerator is of order @xmath242 , whereas the denominator is of order @xmath118 . so , in the leading power approximation , @xmath243 for the hard pole contributions when @xmath64 is nearly parallel to @xmath113 . the contributions from the soft poles in fig . [ f20 ] and the hard poles in fig . [ f21 ] is added to obtain the final results . we emphasize again that our power - counting based analysis is only valid for the leading power terms in the @xmath15 expansion . the final result for the contributions from figs . [ f20 ] and [ f21 ] , plus their mirror diagrams , is @xmath244 \nonumber\\ & \times & \frac{\alpha_s}{2\pi^2}\ , c_f \ , \int dx'\ , \xi ' \ , q'(x ' ) \left[\frac{1+\xi^{\prime 2}}{1-\xi'}\right ] \ , \label{x - ssa - nd - b}\end{aligned}\ ] ] which is again proportional to the same hard factor @xmath245 , and reproduces the last term in the tmd factorization formula eq . ( [ e39 ] ) . it is important to note that a part of the hard pole contributions from fig . [ f21 ] cancels out the soft pole contributions from diagrams @xmath172-@xmath177 of fig . the remainder of the hard pole contribution is factorized into the quark distribution multiplied by @xmath149 and the partonic scattering function . the soft pole diagrams @xmath173-@xmath212 and @xmath174-@xmath213 are separately factorized into the same quark distribution for the unpolarized nucleon multiplied by @xmath150 and @xmath151 , respectively , and the partonic scattering function . we illustrate this factorization in fig . [ f23 ] . after summing up all contributions from eqs . ( [ x - ssa - d ] ) , ( [ x - ssa - nd - sa ] ) , ( [ x - ssa - nd - ha ] ) , and ( [ x - ssa - nd - b ] ) , we obtain the total leading - power contribution to @xmath13 from the @xmath246 partonic subprocess in the @xmath15 expansion . it reproduces the factorized formula in eq . ( [ e39 ] ) , which is the leading term in the corresponding tmd factorization formula , eq . ( [ e4 ] ) . in summary , we have calculated the single transverse - spin asymmetry in dijet - correlation in the twist - three approach when the radiated gluon is parallel to either the polarized nucleon or the unpolarized nucleon . by using the power counting technique , we have shown that the contribution to the ssa can be factorized into the perturbatively generated quark sivers function when the radiated gluon is parallel to the polarized nucleon , and into the unpolarized quark distribution when it is parallel to the unpolarized nucleon . we have demonstrated that the result reproduces the leading order terms in the corresponding tmd factorization formalism in the regime @xmath247 . this is a nontrivial result , especially for the single spin asymmetry , because the calculations for the quark sivers function at this order have to take into account the perturbative expansion of the gauge link at next - to - leading order ( @xmath248 ) . the single spin dependent differential cross section term in eq . ( [ e4 ] ) can be further simplified by taking a moment in @xmath9 . such a moment was also considered in @xcite . we can rewrite @xmath249 where @xmath250 is the so - called bi - sector angle of the two jets : @xmath251 with @xmath252 and @xmath253 the azimuthal angles of the two jets relative to the polarization vector @xmath34 . the angle @xmath254 measures how far the two jets are away from the back - to - back configuration . all these azimuthal angles are defined in a frame in which the polarized proton is moving in the @xmath255 direction . since @xmath256 , the @xmath9-moment of the asymmetry is also related to the @xmath257-moment . one interesting moment is the following : @xmath258 where @xmath259 is the twist-3 matrix element of the quark - gluon correlation function defined in eq . ( [ tf ] ) @xcite . it is also related to the @xmath260-moment of the tmd quark sivers function , see eq . ( [ e34 ] ) . an advantage of taking this moment of the asymmetry is that the transverse momentum integrals of the various factors in the factorization formula decouple from each other , without further assumptions for the @xmath260-dependence of the tmd distributions . in fact , the ssa in the @xmath9 moment of the dijet momentum imbalance is effectively a physical quantity with only a _ single _ large observed scale , @xmath8 . for such an observable , the conventional collinear factorization is more appropriate . in this case , all partons transverse momentum dependence is integrated into the usual collinear parton distributions . the hard factors we obtained above for all the partonic channels agree with those given in @xcite , although a very different approach was adopted there . in @xcite , in order to investigate the @xmath9-moment of the single spin asymmetry , the appropriate gauge links were derived and expanded to first order in @xmath261 . since both approaches have included the initial / final state interactions in their formalisms , they should agree with each other at this order . in this paper , we have studied the asymmetric production of two jets in hadronic collisions . using the collinear factorization approach in perturbative qcd , we calculated both the spin - averaged and the spin - dependent differential cross section for the hadronic production of dijets with momentum imbalance @xmath9 , in a kinematic region where @xmath262 . at the first nonvanishing order , the momentum imbalance is generated by radiating a gluon with transverse momentum equal to the imbalance . in the limit when the imbalance @xmath9 is much less than the averaged jet momentum @xmath8 , we derived the leading - power contribution to the cross sections in the @xmath15 expansion when the radiated gluon is nearly parallel to one of the incoming hadrons . we found that the perturbatively calculated leading contributions to both the spin - averaged and the spin - dependent cross sections can be factorized into a partonic hard part which is a function of @xmath8 , and into perturbatively generated tmd parton distributions at @xmath248 . our results derived in the collinear factorization approach in the limit @xmath263 reproduce the same factorized expressions in eqs . ( [ e37]),([e39 ] ) that were derived as the leading order terms of the generalized tmd factorization formulas in eqs . ( [ e4]),([e2 ] ) when the observed @xmath9 is solely due to the perturbatively generated tmd parton distribution of one of the incoming hadrons . the consistency between the results derived in the collinear factorization approach and those derived by expanding the generalized tmd factorization formulas in eqs . ( [ e4]),([e2 ] ) to the same order naturally leads us to ask if the generalized tmd factorization formulas in eqs . ( [ e4]),([e2 ] ) are actually valid for the hadronic dijet production to all orders in perturbative qcd when @xmath264 , like the tmd factorization formulas for drell - yan and sidis . _ if this were true _ , the consistency that we demonstrated in this paper in the overlap region where @xmath50 would effectively provide a unified picture and treatment for the hadronic dijet production over the full kinematic range of jet momenta : using the tmd factorization formalism when the dijet momentum imbalance is small and the collinear factorization approach when it is large . this is exactly what was achieved in ref . @xcite for the drell - yan and sidis processes . however , as we mentioned at the beginning of sec . iii , showing the consistency between two factorized formalisms at the leading order when @xmath265 is not sufficient to prove the tmd factorization formula over its full kinematic region . when we started with the collinear factorization approach to calculate the momentum imbalance of the dijet cross section , we already took as a fact that the long - distance interaction between the hadrons / jets at their mass scales , @xmath266 , are factorized for the given @xmath267 . that is , our results derived in the limit @xmath265 are valid when @xmath268 ; but , our derivation does not address the issues concerning the factorization of the interactions between the hadrons / jets when @xmath269 or the mass scale of the hadrons / jets involved . the consistency that we demonstrated here is certainly a necessary condition that needs to be satisfied should the tmd factorization formulas be valid . for the tmd factorization to be true , we need to establish two key facts : the partonic hard part at @xmath54 must be insensitive to any physics at the scale @xmath78 , and the tmd parton distributions at the scale @xmath9 should be process independent and universal , or at least , `` quasi''-universal up to a sign change @xcite . in this paper , we also calculated in sec . ii the partonic hard parts at @xmath57 by using the brodsky - hwang - schmidt model with the proper color factor for the interaction vertices . we derived the partonic hard parts from the lowest order scattering diagrams and showed that the dynamics of partonic scattering at scale @xmath57 is independent of the model of nucleon that we used in our derivation . since the ssa can be generated by either initial- or final - state interactions , the color flow of the partonic scattering in the dijet momentum imbalance is very different from that of drell - yan and sidis . since all poles from the initial- and final - state one gluon interactions contribute to the ssa , and the color factor for each individual scattering diagram is insensitive to the long - distance details of the nucleon because there is only one unique color structure for the active quark - gluon combination at this order , we were able to factor the process - dependent as well as the diagram - specific color factors along with the partonic hard parts , and leave the physics at scale @xmath270 in the tmd parton distributions defined in sidis . by doing that , we found that the resulting full partonic hard parts are the same as those derived in the collinear factorization approach . however , for a full tmd factorization , we will have to show that the poles from multiple gluon interactions can be exponentiated into the gauge link that defines the tmd parton distributions , while leaving the same partonic hard parts with the same color factors . due to the mixture of initial- and final - state interactions and the complications of color flows , this task is very non - trivial and is beyond the scope of this paper . to better understand the dijet momentum imbalance in hadronic collisions , a number of extensions can be performed based on our results . first , in this paper , we have only studied the contribution of one gluon radiation collinear to one of the incident hadrons . this should be extended to other kinematic regions of the gluon . for example , when the gluon is radiated parallel to the final state jets or hadrons , it is important to see how the corresponding contribution can be accommodated by the jet definition , or be factorized into the fragmentation functions for the hadron , especially in the spin - dependent case . another important kinematic regime is when the radiated gluon becomes soft . it is very important to see if the resulting contribution can be factorized into a soft factor , as predicted by the tmd factorization formalism . in addition , in this paper we have only discussed the ssa for dijet correlations , which has sivers - type contributions . for the related di - hadron correlations , one expects also the collins mechanism to be important , and it will be interesting to explore it and the factorization properties of the corresponding spin - dependent cross section . the method and procedure used in this paper should provide useful guidance in all these related studies . we finally note that if the tmd factorization indeed fails , a careful comparison between our results and experimental data as the momentum imbalance @xmath15 decreases could furnish an ideal test of the breaking of factorization and provide new opportunities to explore qcd dynamics . we are grateful to cedran bomhof , john collins , andreas metz , and piet mulders for useful discussions . we thank xiangdong ji for his general remarks and his reminding us the wigner - eckart theorem for the color - factor decomposition . we especially thank george sterman for continued discussions on the factorization issues in hadronic reactions . is supported in part by the u. s. department of energy under grant no . de - fg02 - 87er-40371 . w.v . are finally grateful to riken , brookhaven national laboratory and the u.s . department of energy ( contract number de - ac02 - 98ch10886 ) for providing the facilities essential for the completion of their work . thanks the high energy theory group at argonne national laboratory for its hospitality during the writing of this work . in this appendix , we list the hard factors for the dijet - correlation at hadron colliders , for the unpolarized and the single - transverse spin dependent cases . for unpolarized scattering , the hard factors are well - known in the literature . they may also be obtained from the results shown in tables i - iii , using eq . ( [ e29 ] ) . for completeness , we list them here : @xmath271-\frac{n_c^2 - 1}{4n_c^3}\frac{4\hat s^2}{\hat t\hat u}\right\ } \ , \nonumber\\ h_{q\bar q\to q\bar q}^{uu}(\hat s,\hat t,\hat u)&=&\frac{\alpha_s^2\pi}{\hat s^2}\left\{\frac{n_c^2 - 1}{4n_c^2}\left[\frac{2(\hat s^2+\hat u^2)}{\hat t^2}+\frac{2(\hat u^2+\hat t^2)}{\hat s^2}\right]-\frac{n_c^2 - 1}{4n_c^3}\frac{4\hat u^2}{\hat s\hat t}\right\ } \ , \nonumber\\ h_{qg\to qg}^{uu}(\hat s,\hat t,\hat u)&=&\frac{\alpha_s^2\pi}{\hat s^2}\left\{\frac{1}{2}\frac{2(\hat s^2+\hat u^2)}{\hat t^2}-\frac{c_f}{2n_c}\frac{2(\hat s^2+\hat u^2)}{\hat s\hat t}\right\ } \ , \nonumber\\ h_{q\bar q\to gg}^{uu}(\hat s,\hat t,\hat u)&=&\frac{\alpha_s^2\pi}{\hat s^2}\left\{\frac{c_f^2}{n_c}\frac{2(\hat t^2+\hat u^2)}{\hat t\hat u}-c_f\frac{2(\hat t^2+\hat u^2)}{\hat s^2}\right\ } \ , \nonumber\end{aligned}\ ] ] @xmath272 for the single - transverse spin dependent case , using eq.([e30 ] ) , we obtain @xmath273+\frac{n_c^2 + 3}{4n_c^3}\frac{4\hat s^2}{\hat t\hat u}\right\ } \ , \nonumber\\ h_{q\bar q\to q\bar q}^{\rm sivers}(\hat s,\hat t,\hat u)&=&\frac{\alpha_s^2\pi}{\hat s^2}\left\{-\frac{n_c^2 - 3}{4n_c^2}\frac{2(\hat s^2+\hat u^2)}{\hat t^2}+\frac{n_c^2 + 1}{4n_c^2}\frac{2(\hat u^2+\hat t^2)}{\hat s^2}-\frac{n_c^2 + 1}{4n_c^3}\frac{4\hat u^2}{\hat s\hat t}\right\ } \ , \nonumber\\ h_{qg\to qg}^{\rm sivers}(\hat s,\hat t,\hat u)&=&\frac{\alpha_s^2\pi}{\hat s^2}\left\{-\frac{n_c^2}{4(n_c^2 - 1)}\frac{2(\hat s^2+\hat u^2)}{\hat t^2}\left[\frac{\hat s}{\hat u}-\frac{\hat u}{\hat s}\right]-\frac{1}{2(n_c^2 - 1)}\frac{2(\hat s^2+\hat u^2)}{\hat t^2}\right.\nonumber\\ & & ~~~~~~~\left.-\frac{1}{4n_c^2(n_c^2 - 1)}\frac{2(\hat s^2+\hat u^2)}{\hat s\hat u}\right\ } \ , \nonumber\\ h_{q\bar q\to gg}^{\rm sivers}(\hat s,\hat t,\hat u)&=&\frac{\alpha_s^2\pi}{\hat s^2}\left\{-\frac{1}{2n_c}\frac{2(\hat t^2+\hat u^2)}{\hat s^2}+\frac{n_c}{4}\frac{2(\hat t^2+\hat u^2)}{\hat s^2}\left[\frac{\hat t}{\hat u}+\frac{\hat u}{\hat t}\right]\right.\nonumber\\ & & ~~~~~~~\left.-\frac{2n_c^2 + 1}{4n_c^3}\frac{2(\hat t^2+\hat u^2)}{\hat t\hat u}\right\ } \ , \nonumber\\ h_{qg\to q\gamma}^{\rm sivers}(\hat s,\hat t,\hat u)&=&\frac{\alpha_s\alpha_e e_q^2\pi}{\hat s^2}\left(-\frac{n_c^2 + 1}{2n_c(n_c^2 - 1)}\right)\frac{2(\hat s^2+\hat u^2)}{-\hat s\hat u } \ , \nonumber\\ h_{q\bar q\to g\gamma}^{\rm sivers}(\hat s,\hat t,\hat u)&=&\frac{\alpha_s\alpha_e e_q^2\pi}{\hat s^2}\left(\frac{n_c^2 + 1}{2n_c^2}\right)\frac{2(\hat t^2+\hat u^2)}{\hat t\hat u } \ .\end{aligned}\ ] ] the above results are for @xmath98-channel scattering . in dijet - production , both @xmath98 and @xmath99 channel processes will contribute . the hard factors for the @xmath99-channel scattering can be obtained from the results above by exchanging @xmath274 . g. bunce _ et al . _ , lett . * 36 * , 1113 ( 1976 ) . a. airapetian _ et al . _ [ hermes collaboration ] , phys . lett . * 84 * , 4047 ( 2000 ) ; phys . lett . * 94 * , 012002 ( 2005 ) . a. bravar [ spin muon collaboration ] , nucl . a * 666 * , 314 ( 2000 ) ; h. avakian [ clas collaboration ] , talk presented at the rbrc workshop `` single - spin asymmetries '' , brookhaven national laboratory , upton , new york , june 1 - 3 , 2005 , rbrc proceedings volume 75 ( bnl-74717 - 2005 ) ; v. y. alexakhin _ et al . _ [ compass collaboration ] , phys . lett . * 94 * , 202002 ( 2005 ) ; e. s. ageev _ et al . _ [ compass collaboration ] , arxiv : hep - ex/0610068 ; f. bradamante [ compass collaboration ] , talk presented at the `` 17th international spin physics symposium ( spin 2006 ) '' , kyoto , japan , october 2 - 7 , 2006 . j. adams _ et al . _ [ star collaboration ] , phys . rev . lett . * 92 * , 171801 ( 2004 ) ; l. nogach , talk presented at the `` 17th international spin physics symposium ( spin 2006 ) '' , kyoto , japan , october 2 - 7 , 2006 , arxiv : hep - ex/0612030 . s. s. adler [ phenix collaboration ] , phys . * 95 * , 202001 ( 2005 ) ; m. chiu , talk presented at the `` 17th international spin physics symposium ( spin 2006 ) '' , kyoto , japan , october 2 - 7 , 2006 ; c. a. aidala , ph.d . thesis , columbia u. , arxiv : hep - ex/0601009 . f. videbaek [ brahms collaboration ] , aip conf . * 792 * , 993 ( 2005 ) ; j. h. lee [ brahms collaboration ] , talk presented at the `` 17th international spin physics symposium ( spin 2006 ) '' , kyoto , japan , october 2 - 7 , 2006 . g. l. kane , j. pumplin and w. repko , phys . lett . * 41 * , 1689 ( 1978 ) . for reviews , see : m. anselmino , a. efremov and e. leader , phys . * 261 * , 1 ( 1995 ) [ erratum - ibid . * 281 * , 399 ( 1997 ) ] ; z. t. liang and c. boros , int . j. mod . a * 15 * , 927 ( 2000 ) ; v. barone , a. drago and p. g. ratcliffe , phys . * 359 * , 1 ( 2002 ) . s. j. brodsky , d. s. hwang and i. schmidt , phys . b * 530 * , 99 ( 2002 ) ; nucl . b * 642 * , 344 ( 2002 ) . j. c. collins , phys . b * 536 * , 43 ( 2002 ) . d. w. sivers , phys . d * 41 * , 83 ( 1990 ) ; phys . d * 43 * , 261 ( 1991 ) . x. ji and f. yuan , phys . b * 543 * , 66 ( 2002 ) ; a. v. belitsky , x. ji and f. yuan , nucl . phys . b * 656 * , 165 ( 2003 ) . d. boer , p. j. mulders and f. pijlman , nucl . phys . b * 667 * , 201 ( 2003 ) . j. c. collins and d. e. soper , nucl . b * 193 * , 381 ( 1981 ) [ erratum - ibid . b * 213 * , 545 ( 1983 ) ] ; nucl . b * 197 * , 446 ( 1982 ) . j. c. collins , d. e. soper and g. sterman , nucl . b * 250 * , 199 ( 1985 ) . d. boer and w. vogelsang , phys . rev . d * 69 * , 094025 ( 2004 ) . b. i. abelev , _ et al . _ , [ star collaboration ] arxiv:0705.4629 [ hep - ex ] ; j. balewski , talk presented at the spin 2006 symposium , kyoto , japan , october 2 - 7 , 2006 , arxiv : hep - ex/0612036 . c. j. bomhof , p. j. mulders and f. pijlman , phys . b * 596 * , 277 ( 2004 ) ; eur . j. c * 47 * , 147 ( 2006 ) ; a. bacchetta , c. j. bomhof , p. j. mulders and f. pijlman , phys . d * 72 * , 034030 ( 2005 ) ; c. j. bomhof and p. j. mulders , jhep * 0702 * , 029 ( 2007 ) . j. w. qiu , w. vogelsang and f. yuan , arxiv:0704.1153 [ hep - ph ] , phys . b in press . a. v. efremov and o. v. teryaev , sov . j. nucl . * 36 * , 140 ( 1982 ) [ yad . 36 * , 242 ( 1982 ) ] ; a. v. efremov and o. v. teryaev , phys . b * 150 * , 383 ( 1985 ) . j. qiu and g. sterman , phys . lett . * 67 * , 2264 ( 1991 ) ; nucl . b * 378 * , 52 ( 1992 ) ; phys . d * 59 * , 014004 ( 1999 ) . j. botts and g. sterman , nucl . b * 325 * , 62 ( 1989 ) . n. kidonakis and g. sterman , nucl . b * 505 * , 321 ( 1997 ) ; nucl . b * 525 * , 299 ( 1998 ) ; n. kidonakis , g. oderda and g. sterman , nucl . b * 531 * , 365 ( 1998 ) . x. ji , j. w. qiu , w. vogelsang and f. yuan , phys . rev . lett . * 97 * , 082002 ( 2006 ) ; phys . d * 73 * , 094017 ( 2006 ) ; phys . b * 638 * , 178 ( 2006 ) . w. vogelsang and f. yuan , phys . d * 72 * , 054028 ( 2005 ) . j. collins and j. w. qiu , arxiv:0705.2141 [ hep - ph ] . g. altarelli and g. parisi , nucl . b * 126 * , 298 ( 1977 ) . g. t. bodwin , phys . d * 31 * , 2616 ( 1985 ) [ erratum - ibid . d * 34 * , 3932 ( 1986 ) ] . x. ji , phys . b * 289 * , 137 ( 1992 ) . a. bacchetta , c. bomhof , u. dalesio , p. j. mulders and f. murgia , arxiv : hep - ph/0703153 . c. j. bomhof , p. j. mulders , w. vogelsang and f. yuan , phys . rev . d * 75 * , 074019 ( 2007 ) . j. w. qiu and g. sterman , aip conf . proc . * 223 * , 249 ( 1991 ) ; nucl . b * 353 * , 137 ( 1991 ) . c. kouvaris , j. w. qiu , w. vogelsang and f. yuan , phys . d * 74 * , 114013 ( 2006 ) . h. eguchi , y. koike and k. tanaka , nucl . b * 752 * , 1 ( 2006 ) ; nucl . b * 763 * , 198 ( 2007 ) ; y. koike and k. tanaka , phys . b * 646 * , 232 ( 2007 ) ; arxiv : hep - ph/0703169 . g. ratcliffe and o. v. teryaev , arxiv : hep - ph/0703293 .

while accretion of gas onto compact objects has been studied for over fifty years , much uncertain remains about how angular momentum is transported . there is increasing evidence that the interplay between the local angular momentum transport from magnetohydrodynamic ( mhd ) turbulence and the more global accretion flow shapes the behavior of the mass accretion ( e.g. * ? ? ? the primary goal of this paper is to connect mhd accretion disk theory with one of the most widely discussed phenomenological models for disk variability , the propagating fluctuations model . we explore how a stochastically varying effective viscosity leads to the growth of so - called propagating fluctuations " in our long , global accretion disk simulation and relate the behavior of our simulation to observable properties from accreting black holes . the idealized @xmath1-disk from the seminal work of has long served as the canonical disk model . assuming turbulence was present in the disk , they argued that an effective kinematic viscosity originating from turbulent eddies spanning a radial range provided the internal viscous stress required to transport angular momentum and allow for accretion . in their model , the turbulent viscosity takes the form @xmath2 , where @xmath3 is the turbulent velocity and @xmath4 is a characteristic length scale of the turbulent eddies . this viscosity can be parameterized with an efficiency " factor in terms of the local sound speed , @xmath5 , and the local scale height of the disk , h , @xmath6 . with this prescription , restrict the @xmath1 parameter to a range of @xmath7 . if the turbulent velocity is greater than the sound speed , the gas will shock , convert excess kinetic energy to thermal energy , and return to the subsonic domain , ensuring @xmath8 . likewise , they postulated @xmath9 because the differential shear in the disk will quickly break - up large turbulent eddies leading to localized turbulence . an alternative but equivalent view , motivated by dimensional analysis , is to say that the relevant viscous stresses in the disk are proportional to the pressure , @xmath10 , where @xmath11 , where @xmath11 is the off - diagonal component of the stress energy tensor and @xmath12 is pressure . @xmath1 and @xmath13 are related to each other by a factor of order unity that depends upon the equation of state . even in the early work of , the influence of magnetic stress was intuitively assumed to be an essential part of the turbulent injection mechanism , though a viable driving mechanism was absent until the magneto - rotational instability ( mri ) was put forth as a candidate to drive the disk turbulence . the instability of a differentially rotating , magnetized plasma was originally discovered by @xcite and @xcite , but its astrophysical significance was not realized until @xcite studied it in the context of an ionized accretion disk . they showed that radial perturbations of gas elements are linearly unstable if @xmath14 , where @xmath15 is the wavenumber , @xmath16 is the alfvn speed , and @xmath17 is the orbital angular frequency . as a fluid element is perturbed , the displacement acts to enhance a weak seed field until it goes nonlinear , which subsequent numerical mhd simulations have verified drives and sustains turbulence @xcite . to date , a large and rich body of simulation work has been dedicated to studying how mri driven turbulence behaves and transports angular momentum including studies using unstratified shearing boxes ( e.g. * ? ? ? * ; * ? ? ? * ) , stratified shearing boxes ( e.g. * ? ? ? * ; * ? ? ? * ) , unstratified global models ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , stratified global disk models ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , relativistic global models ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , and global models with radiative transfer ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) . the viscous behavior of the disk ultimately arises because of net internal stress due to correlated fluctuations in the magnetic field and correlated fluctuations in gas velocity @xcite . these correlated fluctuations lead to the maxwell stress ( @xmath18 ) and the reynolds stress ( @xmath19 ) , respectively . following the @xmath1-prescription , the effective @xmath1 from the net stress is thus the ratio of the volume averaged stress to the volume averaged pressure , @xmath20 this serves to parameterize the effectiveness of angular momentum loss by the stresses acting on the gas in the disk . in this paper we use angled brackets to denote a volume average for a given quantity , @xmath21 , where @xmath22 turbulent fluctuations in the disk will naturally lead to stochastic variability in the effective @xmath1 and provide an intuitive explanation for the existence of broadband photometric variability observed in both galactic black hole binaries ( gbhbs ) and active galactic nuclei ( agns ) . the photometric variability is observed across several decades in frequency , typically in the form of aperiodic flicker " type noise . employing a stochastic @xmath1 , @xcite developed a model for this flicker noise by linearizing the standard geometrically thin , optically thick accretion disk evolution equations . the model was limited to small perturbations in @xmath1 , but it reproduced the power spectrum of the fluctuations and laid the foundation for the subsequent development of the propagating fluctuations model . the propagating fluctuations model has since taken a more generalized form and served as a phenomenological description for the structure and organization of the broadband variability for gbhbs and agns . the evolution of a geometrically thin , optically thick accretion disk ( in units where @xmath23 ) is governed by a nonlinear diffusion equation @xmath24\ ] ] where r is the distance from the central engine , @xmath25 is the local effective viscosity in the disk , and @xmath26 is the local surface density . in this equation , the diffusive redistribution of angular momentum depends on the product of the effective viscosity and the local surface density . if the effective viscosity fluctuates , the instantaneous mass accretion rate ( @xmath27 ) at each point in the disk is thus the product of the many prior stochastic events . a consequence of the multiplicative combination is that fluctuations in @xmath27 will tend to be preserved as they accrete as higher density regions will typically have higher @xmath27 and lower density regions will typically have lower @xmath27 . additionally , since the viscous time increases with radii , higher frequency fluctuations will be imposed on the accretion flow as material moves inwards . this will lead to the development of a hierarchical structure in the fluctuations with the smaller , high frequency fluctuations imposed upon the larger , low frequency fluctuations . the appeal of the propagating fluctuations model is that it explains three properties of black hole variability that have been difficult to reproduce with other variability models . first , gbhb and agn light curves are log - normally distributed @xcite which is accounted for through the multiplicative combination of many independent fluctutations . second , and related to the log - normal distribution , the root - mean square ( rms ) of the flux variability in the light curves of both gbhbs and agns depends linearly on the flux level @xcite , the so - called rms - flux relationship . this indicates the mass accretion rate variation has a constant fractional amplitude and that the mechanism driving variability is , therefore , independent of the accretion rate @xcite . finally , coherent fluctuations in the emission at different wavebands from gbhbs have frequency dependent time lags with variability in higher energy bands lagging variability in lower energy bands @xcite . this hard lag " is believed to be a signature of fluctuations in the accretion flow moving to smaller radii in the disk where the emission temperature is higher . despite the success the propagating fluctuations model has had in explaining the phenomenology of gbhb and agn variability , the model has yet to be examined in the context of modern mhd theory . to date , the model has been examined in the context of geometrically - thin ( vertically - integrated ) @xmath1-disks in which some stochasticity is imposed upon @xmath1 . how well this viscous behavior translates into an mhd disk is unknown . a particularly challenge is the timescales on which @xmath1 fluctuates and drives variability within the accretion flow . @xcite prescribes @xmath1 fluctuations to be slow , of order the viscous time . however , if fluctuations in @xmath1 are solely due to turbulent fluctuations , we might expect them to be on the dynamical timescale . @xcite show , using a semi - analytic model , that propagating fluctuations only exist if fluctuations in @xmath1 occur at sufficiently low frequencies . if the frequency of @xmath1 fluctuations is too high , they find that @xmath27 fluctuations are damped out and the nonlinear features of the variability do not develop . therefore , fluctuations in the effective-@xmath1 on timescales longer than the dynamical timescale are needed to reproduce the phenomenology seen in the variability . in this paper , we perform the first detailed analysis of propagating fluctuations in the context of a long - duration , mhd accretion disk simulation . we show that the well - known quasi - periodic dynamo induces slow fluctuations in the effective @xmath1 that then drive propagating fluctuations . in section [ sec - model_conv ] we introduce the simulation and perform a convergence study . section [ sec - overview ] presents a brief discussion of the nature of the turbulence . in section [ sec - dynamo ] , we investigate the relationship of the quasi - periodic disk dynamo and angular momentum transport , finding that it imposes an intermediate - timescale modulation of the effective @xmath1 . we then examine the variability of both the mass accretion rate ( section [ sec - prop_flucs ] ) and a proxy for bolometric radiative luminosity ( section [ sec - proxy_emission ] ) finding that both quantities display the non - linear variability characteristics of real sources and that the coherence of these quantities across the disk is exactly as expected in the propagating fluctuations picture . in section [ sec - discussion ] , we place our results into a broader context , and then conclude in section [ sec - summary ] . in this section , we describe the construction and convergence properties of our numerical model . in brief , we evolve a three - dimensional ideal mhd simulation in order to study the local and global dynamics of a geometrically - thin ( constant opening angle @xmath28 ) accretion disk . given our focus on the time - variability of the accretion flow , the crux of our problem is the description of the mhd turbulent dynamics and how it influences the long - timescale behavior of the disk . thus , we concentrate our computational resources into the high - spatial resolution and the long duration nature of the simulation . we demonstrate below that the numerical grid is of high enough resolution that the macroscopic behavior of the disk turbulence is converged and insensitive to increases in resolution . as a penalty , we simplify the physics to the bare minimum . we employ non - relativistic ideal mhd , and a pseudo - newtonian gravitational potential to emulate the dynamical effects of a general relativistic potential , including the presence of an inner - most circular orbit ( isco ) . we use a simple cooling function to keep the disk thin , but otherwise neglect all radiation physics . renderings of density and magnetic turbulent structure of our fiducial disk simulation are shown in figure [ fig - diskimg ] . our simulation employs the finite - difference mhd code _ zeus - mp _ _ _ zeus - mp _ solves the differential equations of ideal compressible mhd , @xmath29 @xmath30 @xmath31 @xmath32 where @xmath33 to second order accuracy in space . the method of constrained transport is used to maintain zero divergence in the magnetic field to machine precision . the explicit integration time step is set by the usual courant conditions , and is first - order accurate in time . we begin by defining our fiducial simulation . the simulation domain is covered by spherical coordinates ( @xmath34 ) and spans @xmath35 , \theta \in [ \pi/2 -0.5,\pi/2 + 0.5 ] , \phi \in [ 0,\pi/3)$ ] , shown in figure [ fig - diskimg ] . the zone aspect ratio is approximately @xmath36 and each scale height is resolved with @xmath37-zones . outflowing boundary conditions were used for the @xmath38 and @xmath39 boundaries and periodic boundary conditions were used in the @xmath40 direction . we modified the standard zeus - mp boundaries to conserve @xmath41 across the @xmath42- and @xmath43-boundaries . the simulation domain was broken into a well - resolved inner region ( @xmath44 ) , used for our analysis , and an outer region with poorer resolution to act as a gas reservoir . the gas reservoir mitigates the effect of secular decay from the draining of material from the disk over the course of the simulation . the radial spacing in the inner region logarithmically increases in r from 4 - 45 r@xmath45 . in all , the simulation has @xmath46 zones in this region . the @xmath47 spacing in the outer region logarithmically increases from @xmath48 and a total of @xmath49 zones in this region . we approximate the gravitational field around a nonrotating black hole with a psuedo - newtonian gravitational potential of the form : @xmath50 with this potential , several important aspects of a general relativistic gravity field are captured including an isco ( at @xmath51 ) and the qualitative change of radial shear in the disk . a @xmath52 adiabatic equation of state is used for the gas . an initially axisymmetric thin disk was initialized with constant midplane density and radially decreasing pressure corresponding to an isothermal vertical profile : @xmath53 and @xmath54 the disk is initialized to a scale height of @xmath55 and was maintained by an _ ad hoc _ , optically thin cooling function similar to @xcite . the cooling function plays the role of the real physical processes and removes energy at a rate equivalent to the thermal time of the disk ( @xmath56 ) , as was done in @xcite . if the gas energy in a cell is above the target energy , @xmath57 , it is cooled over the cooling time . the cooling function used is @xmath58 , where @xmath59 $ ] . the switch function , @xmath60 , acts as a threshold function and is zero when @xmath61 . a protection routine is implemented to ensure the density and pressure values do not become artificially small and/or negative . the gas was initialized with a purely azimuthal velocity field and a weak magnetic field . the local velocity is set so that the effective centripetal force balances the gravitational force . a series of weak , large - scale poloidal magnetic field loops are set for the initial magnetic field configuration ( e.g. @xcite ) to perturb the disk and allow for the growth of the mri . the loops were set from the vector potential @xmath62 to ensure the initial field was divergence free and had the form : @xmath63 where @xmath64 is a normalization constant , @xmath65 is an envelope function that is 1 in the body of the disk and smoothly goes to 0 at @xmath66 and 3 scale heights above the disk . the alternating field polarity is set by the final multiplicative term such that loops have a radial wavelength of @xmath67 . the @xmath64 constant was set to initialize the disk to an average ratio of gas- to magnetic - pressure of @xmath68 . the following analysis treats the simulation after the mri has saturated and transient behavior from initialization has died away . @xmath69 and @xmath70 reach a quasi - steady state in the well - resolved region of the disk and the mri saturates , as measured by the convergence metrics discussed in section [ sec - conv ] , after approximately 200 isco orbits ( @xmath71 ) . we take this point to be @xmath72 . the simulation was then run for 1410 isco orbits ( @xmath73 ) . in the interest of data management , the values of @xmath69 , @xmath74 , @xmath70 , and @xmath75 were output every two isco orbits ( @xmath76 ) . the issue of convergence is forefront when considering the reliability of results from any _ ab initio _ accretion disk simulation because the mhd turbulence that mediates angular momentum transport has structure on the smallest accessible scales . with this type of simulation , the macroscopic behavior can have a resolution dependence if the turbulent scales dominating the shear stresses are not adequately resolved @xcite . as a preface to our study of variability , here we assess the degree to which we are adequately resolving the turbulent dynamics using two measures . firstly , we examine the resolvability of the fastest growing linear mri mode through the commonly employed quality factors introduced by @xcite . secondly , to characterize the non - linear saturation of the turbulence , we use the magnetic tilt angle . using these diagnostics , we compare our fiducial ( long ) simulation with higher resolution comparison simulations . the first set of diagnostics used to measure resolvability were the vertical and azimuthal quality factors @xcite , @xmath77 and @xmath78 , respectively . @xmath79 and @xmath80 measure resolvability through the number of grid cells per fastest - growing mri wavelength , @xmath81 , and critical toroidal field length , @xmath82 where @xmath83 using the quality factors , we characterized the global resolvability of the disk . the local values of @xmath79 and @xmath80 were calculated in every voxel of the well - resolved region ( @xmath84 ) for every data dump after the transient behavior in the disk had died away ( @xmath72 onwards ) . the volume - weighted average was calculated for each quality factor by averaging over in the region within one scale height above and below the disk mid - plane to get the instantaneous value and then over the entire duration of the simulation to get a temporal average . the averaging was restricted to the body of the disk because we want the most conservative estimate of the resolvability of the simulation . in the highly magnetized , low - density coronal regions the quality factors are very high and , if included , the quality factors would appear artificially high , overestimating how well we resolve the mri . we find @xmath85 and @xmath86 , a bit below the nominal resolvability criteria of @xmath87 @xcite . in addition to the global averages , we are interested in how the resolvability depends on the position within the grid . shown in figure [ fig - qzqp ] is the time and azimuthally averaged profile of @xmath79 , as well as the time and polar average profile of @xmath80 . the average of @xmath79 was calculated as a function of @xmath88 and @xmath43 by averaging over entire @xmath40 domain . as we did in our global average , the vertical averaging over the polar angle in our calculation of @xmath80 is only done in the region within one scale height above and below the disk to exclude the coronal region . the spatial variation can be seen in both quantities , but no apparent dead zones " of gross under - resolution are present . cccccccccccc fiducial & @xmath89 & @xmath90 & @xmath91 & @xmath92 & @xmath93 & @xmath94 & @xmath95 & @xmath96 & @xmath97 & @xmath98 & @xmath99 + hireswedge & @xmath89 & @xmath100 & @xmath101 & @xmath102 & @xmath103 & @xmath104 & @xmath105 & @xmath106 & @xmath107 & @xmath108 & @xmath109 + hires2@xmath110 & @xmath111 & @xmath112 & @xmath101 & @xmath102 & @xmath103 & @xmath104 & @xmath113 & @xmath114 & @xmath115 & @xmath116 & @xmath117 + these quality factors probe the ability of the simulation to resolve linear mri modes but do not address the structure of the ( non - linear ) saturated turbulence . thus , we additionally measure the average in - plane magnetic tilt angle , @xmath118 the value of which is closely connected to the processes by which the mhd turbulence saturates @xcite . we note that , given the dominance of @xmath119 over the other field components , the tilt angle is approximately given by @xmath120 where @xmath121 is the maxwell stress normalized by the magnetic pressure , @xmath122 and @xmath123 is the ratio of gas pressure to magnetic pressure , @xmath124 thus , the magnetic tilt angle is directly tied to the effectiveness of angular momentum transport at a given field strength and can be used to probe whether a given simulation has correctly captured the non - linear saturation of the mhd turbulence . we calculate @xmath125 within the mid - plane regions of the disk , restricting to half a scale height above and below the disk midplane in the well - resolved region . the temporally averaged magnetic tilt ( neglecting the first 200 isco orbits ) is @xmath126 ( fig . [ fig - mag_tilt ] ) , comparable to estimates from both analytic theory @xcite and previous high - resolution local @xcite and global @xcite simulations . we further assess the fidelity of our fiducial simulation by performing two additional high - resolution comparison runs . the first comparison simulation had the exact same geometry as the fiducial simulation , but twice the resolution in each of the three dimensions ( hence voxels of @xmath127 the volume ) . the second comparison simulation also doubles the number of zones in each of the three dimensions , but then extends the domain to include the full 2@xmath110 in azimuth . both of these comparison simulations possess @xmath128 total zones with a vertical resolution of @xmath129-zones per disk scale - height . of course , the @xmath130 ratio was larger in the @xmath111 higher resolution simulation due to the extension of the @xmath40 domain . the evolution of @xmath125 for the simulation used in our analysis and the two test simulations is shown in figure [ fig - mag_tilt ] . in the higher resolution simulations the evolution of @xmath125 is slightly faster than our fiducial run , but settles at similar levels as the mri - driven turbulence saturates . one difference between the development of the turbulence in the comparison simulations and our fiducial run is that our fiducial run has a transient spike in magnetic tilt angle @xmath131 ( 94 isco orbits ) into the initialization . from visual inspection , we see the growth and break - up of a large channel flow in the disk body in the fiducial run . this does not develop in the higher resolution comparison simulations because the non - axisymmetric parasitic instabilities are more effective at breaking up nascent channel flows . with this minor difference , the similarity in the evolution and saturation of the turbulence between the three simulations indicates that our fiducial run is adequately resolved . , for long , fiducial disk simulation ( black line ) , high resolution wedge simulation ( red line ) , and high resolution @xmath111 simulation ( blue line ) . the t=0 time we set use in our analysis is marked by the black dashed line . the evolution of the long disk simulation used for the science run in our paper has been truncated to only show the initial saturation of the turbulence and a short time period short after , but continues for @xmath73 . @xmath125 grows similarly in all three simulations and saturates at the same value , @xmath132 . the black dashed line indicates t=0 used in our analysis . [ fig - mag_tilt],scaledwidth=50.0% ] in all , we can be confident in the convergence of our model . by standard measures the simulation is adequately resolved and has reached a level of saturated turbulence . the quality factors slightly are lower than what has been deemed well - resolved by prior convergence studies ( i.e. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , but the similarities between @xmath125 in our long science run and those of our two shorter , high - resolution comparison simulations demonstrate the turbulence has reached a saturated level . we will now look at the behavior of the effective @xmath1 in the disk . of particular interest is the variable behavior of the effective @xmath1 as this is required to drive fluctuations in @xmath27 . using equation [ eqn - alpha ] , we calculate the effective @xmath1 by restricting the domain over which we average to the disk midplane , @xmath133 $ ] , where h is one disk scale height . for each data output from our simulation ( @xmath134 ) we calculate the average stress and pressure within one scale height across the entire azimuthal domain for each radial bin . the ratio of the stress to pressure was then taken to provide the effective @xmath1 as a function of radius and time . figure [ fig - alpha_var ] shows the time variability of the effective @xmath1 parameter in the disk after averaging over radius . large variability occurs on an intermediate timescale , longer than a dynamical time and shorter than a viscous time . the value of the effective @xmath1 in the disk typically fluctuates between @xmath135 and @xmath136 in a quasi - periodic manner . on the lower end of the fluctuating range , @xmath135 seemingly acts as a floor where the minima consistently returns . the upper range on the effective @xmath1 is less well - defined with maxima found at different levels as low as @xmath137 and as high as @xmath138 . parameter in the disk after averaging over radius . the @xmath1 parameter is highly variable and fluctuates between @xmath139 and @xmath136 in a quasi - periodic manner . [ fig - alpha_var],scaledwidth=50.0% ] fit with a normal distribution ( red line ) and log - normal distribution ( blue - line ) . the normal distribution provides a better fit to the distribution . the distribution is best fit by a log - normal distribution . [ fig - alpha_hist],scaledwidth=50.0% ] figure [ fig - alpha_hist ] shows the histogram of effective @xmath1 with normal ( gaussian ) and log - normal fits to the probability density function ( pdf ) . statistically , the pdf of the effective @xmath1 is better fit by a log - normal distribution . the best fit with a normal distribution has @xmath140 and the best fit with a log - normal distribution has @xmath141 . here , we have a brief digression in order to characterize the velocity and length scales of the turbulence . given the range of time- and length - scales involved in a global disk simulation , we choose to perform this exercise over a limited range of radii ; we select the region @xmath142 , \theta \in [ \pi/2 -0.35,\pi/2 + 0.35 ] , \phi \in [ 0,\pi/3)$ ] . in physical coordinates , this domain spans @xmath143 . the size of the subdomain was chosen such that any localized turbulent structure would be well contained within the box , but also so that any medium - scale structure would be captured by our analysis if it is present . within this region , we first calculated the ratio of the turbulent speed to the sound speed within the disk midplane as a function of time . the turbulent velocity is defined as @xmath144 and the sound is defined in its usual form @xmath145 for each data dump , we calculated the volume weighted spatial average of @xmath146 to serve as a diagnostic for the average velocity of the turbulence within the disk . figure [ fig - turb_variability ] shows the time variability of @xmath147 . this ratio is highly variable and fluctuates within a range of @xmath148 to @xmath149 . comparing with the time variability of @xmath1 in figure [ fig - alpha_var ] , the correlation between @xmath1 and @xmath146 can be readily observed . the average value is @xmath150 . measuring the spatial scales of the turbulence is significantly more involved than measuring the turbulent velocity , but , nevertheless , provides a method for characterizing the turbulence in simple terms . the spatial scales of the density , magnetic field components and maxwell stress of the turbulence were measured in our @xmath151 subdomain using the autocorrelation the method of @xcite & @xcite with some modifications . for each data dump after the initialization of the simulation , the two - point autocorrelation of a given quantity of interest , @xmath152 , was calculated within this subdomain in the following way . first , large - scale structure of @xmath152 was removed . for a given time step , t , the average over @xmath153 gm / c@xmath154 ( 11 data dumps ) for each cell in the subdomain was calculated . the time average was centered on the timestep of interest such that the average for a single cell included its values 5 data dumps before and after . this was then subtracted off of the present cell value : @xmath155 leaving only the high - frequency perturbation . since the average was determined for an individual cell , this also removes vertical gradients . ccccc & & a - 0.32 & a - 0.41 & a - 0.66 + @xmath75 & 16.7@xmath156 & b - 1.17 & b - 1.67 & b - 3.49 + & & c - 0.24 & c - 0.42 & c - 1.08 + & & a - 0.29 & a - 0.38 & a - 0.60 + b@xmath157 & 30.2@xmath156 & b - 0.75 & b - 0.97 & b - 1.79 + & & c - 0.07 & c - 0.10 & c - 0.17 + & & a - 0.20 & a - 0.24 & a - 0.36 + b@xmath158 & 14.8@xmath156 & b - 0.56 & b - 0.74 & b - 1.20 + & & c - 0.14 & c - 0.21 & c - 0.38 + & & a - 0.27 & a - 0.35 & a - 0.58 + b@xmath159 & 22.2@xmath156 & b - 0.97 & b -1.31 & b - 2.52 + & & c - 0.10 & c - 0.14 & c - 0.24 + & & a - 0.20 & a - 0.28 & a - 0.47 + b@xmath160 & 22.2@xmath156 & b - 0.78 & b - 1.03 & b - 2.03 + & & c - 0.07 & c - 0.10 & c - 0.21 + & & a - 0.23 & a - 0.30 & a - 0.53 + b@xmath157b@xmath161 & 26.4@xmath156 & b - 0.70 & b - 0.92 & b - 1.64 + & & c - 0.07 & c - 0.10 & c - 0.17 + next , the physical coordinates were normalized by the local scale height to express the turbulent scales in terms of @xmath162 . this converts the wedge geometry into a rectangular geometry and allows for easier comparison with prior work . in this normalized geometry , @xmath88 maps to @xmath163 , @xmath40 maps to @xmath164 , and @xmath43 maps to @xmath165 . after the normalization of the coordinates , the three - dimensional fourier transform was then taken of @xmath166 : @xmath167from this , the autocorrelation , @xmath168 , was calculated according to : @xmath169 for the first part of our analysis into the spatial scales , we fit the time average three - dimensional autocorrelations . two - dimensional slices through the origin along the x - y , x - z , and y - z planes are shown in figure [ fig - auto_corr ] . to measure the shape of the three - dimensional autocorrelations , we defined three surfaces corresponding to where the value of the autocorrelation falls to @xmath168 = 0.5 , @xmath170 , and @xmath171 . the standard definition of the correlation length is the full - width , half - max of the autocorrelation which is where the autocorrelation falls to 0.5 . however , other works use different definitions of the correlation length , so we present those values to aid in comparing the turbulent scales between simulations . a least - squares minimization was used to fit these surfaces with an ellipsoid of the form : @xmath172 ^ 2}{a^2 } + \frac{[y\cos(\theta)-x\sin(\theta)]^2}{b^2 } + \frac{z^2}{c^2}. \end{split}\ ] ] where @xmath43 is the tilt angle introduced by the disk shear and @xmath173 , @xmath174 , and @xmath175 are the axes corresponding to the @xmath163 , @xmath164 , and @xmath165 directions , respectively . the tilt angle and best fits for the three different correlation lengths are given in table [ tab - ellips_fit ] . the radial correlation length of the gas density , @xmath75 , provides an estimate for the characteristic length scale , @xmath4 , of the turbulence . in the shakura & sunyaev @xmath1-prescription the turbulence is less than the disk scale height . this requirement does not strictly hold for the y- and z - directions , but it does hold for the x - direction , which is the important direction for the angular momentum transport . figure [ fig - turb_variability ] shows the time variability of @xmath4 calculated at individual data dumps . unlike @xmath146 , the variability of @xmath4 is stochastic and lacks the structure seen in the time variability of @xmath176 . instead of varying in lock step with @xmath1 , the time trace of @xmath4 is flat and simply fluctuates around @xmath177 . this means that increases in stress act to drive the turbulence more quickly , rather than injecting energy at a larger scale and allowing it to cascade down to smaller scales and that the turbulent eddy scales are preserved regardless of the magnitude of the stress in the disk . at @xmath178 @xmath179 . within the disk midplane the fluctuations in @xmath180 are largely random . large , stratified regions of coherent magnetic field develop in the corona as buoyancy lifts highly magnetized gas vertically and regions of similar polarity grow through magnetic reconnection . the coronal regions are symmetric across the midplane , but have opposite polarity . [ fig - bfield_strat],scaledwidth=50.0% ] we will now study the disk dynamo and its influence on the variability of the effective @xmath1 in the simulation . the cyclical magnetic dynamo is a well - established phenomenon in accretion disks and is seen in both local @xcite and global @xcite mhd simulations . it is characterized by rising bundles of toroidal magnetic field in which the toroidal field flips polarity between two such events . the dynamo is quasi - periodic in nature with @xmath181 , where @xmath182 is the local dynamical frequency . the influence the dynamo has on the disk is not fully understood . since the maxwell stress and , thus , turbulent injection from the mri is proportional to @xmath183 , we might imagine that the stress would be modulated by the large - scale toroidal magnetic field generated by the dynamo . indeed , a correlation between stress and toroidal magnetic field has been found @xcite , however , the coupling of maxwell stress to the low - frequency , periodic fluctuations of the dynamo has not been well - investigated . of additional interest is how the dynamo could provide a local , net poloidal magnetic field , thereby increasing the turbulent transport of angular momentum . local , shearing box simulations with a net poloidal magnetic field typically find higher @xmath1-values than similar simulations with zero net magnetic field @xcite . when considering the impact of the dynamo on the global disk , we can think of the global domain as a collection of many local regions . the large - scale magnetic field generated by the dynamo could produce a net local poloidal field through these regions , even if the overall magnetic field in the global domain is zero . we begin with an overview of the spatio - temporal properties of the magnetic field in the disk . after the intialization of the simulation , a stratified magnetic field geometry developed and was sustained . figure [ fig - bfield_strat ] shows a snapshot of the azimuthally average of @xmath180 at an arbitrary time ( @xmath178 @xmath179 ) . correlated regions of field in the corona are seen spanning large portions of the disk . like other simulations , the large - scale magnetic field is generated in the disk and is then slowly lifted by magnetic buoyancy @xcite . this growth can be seen in the psds of the azimuthally averaged @xmath180 . shown in figure [ fig - dynamo_ffts ] are the psds of @xmath180 as a function of radius for the midplane and 2h , 3h , and 4h above the disk midplane . the psds were calculated similarly to @xcite and are defined as @xmath184 where @xmath185 is the fourier transform of the time sequence of the variable of interest , @xmath186 in our case , the time sequence is the time variability of the azimuthal average of @xmath180 in the disk mid - plane . the fourier transform was calculated for each point on the radial grid . the large gas reservoir supplies additional material to the disk during the simulation which protected from secular changes due to the draining from accretion . therefore , there was no need pre - whiten " before taking fourier transforms of the basic fluid variability , as @xcite discuss the need for in the case of significant secular evolution . in the disk midplane the dynamo behaves similarly to the dynamo in @xcite . rather than having a single characteristic frequency , a band of enhanced power parallels the orbital frequency at frequencies @xmath187 . for a given frequency , a band of power extends 10 - 15 r@xmath45 , highlighting the radial coherence of the dynamo action pointed out by @xcite . for a given radius , the band of power spans approximately a factor of three . the psds of the azimuthal magnetic field above the disk show a larger radial range of enhanced power . as the the height above the disk midplane increases , the vertical bands of power at a given frequency stretches to smaller radii . at @xmath188 , power at the lowest frequencies is enhanced along the entire radial domain of the simulation . all the while , the high frequency envelope seen in the psd of the disk midplane is preserved . figure [ fig - dynamo_vs_brbp ] shows the spatio - temporal evolution of the dynamo ( probed by the azimuthal average of @xmath180 ) compared with the maxwell stress for three distinct radii ( 15 r@xmath45 , 20 r@xmath45 , & 25 r@xmath45 ) . we will refer to these three radii in order to provide a comparison of different properties related to the magnetic field throughout the disk . as we saw in the previous section , the dynamo plays a central role in the disk s magnetic field evolution and , as we can see from the butterfly diagrams , this translates into modulation of the maxwell stress when the magnetic field is stronger at the maxima and minima of the dynamo cycle , the stress is largest . this behavior is quite similar to that seen in the stratified shearing boxes of @xcite . since the polarity of the global field flips during the dynamo cycle the stress is always positive . it is also worth noting that while the dynamo defines the trend within a given cycle , the fluctuations in the stress are dominated by high - frequency variability . and @xmath183 for 15 , 20 , and 25 r@xmath45 . the characteristic butterfly " pattern is seen in the spacetime diagrams of @xmath180 . when @xmath180 is stronger , @xmath183 is stronger , although the variability is mostly dominated by high - frequency variability . [ fig - dynamo_vs_brbp],scaledwidth=50.0% ] ) of @xmath180 ( left panel ) and @xmath183 ( right panel ) for 15 ( top ) , 20 ( middle ) , and 25 ( bottom ) r@xmath45 . the strong , multi - peaked bands of power correspond to the oscillations seen in figure [ fig - dynamo_vs_brbp ] . a noticeable difference between the frequency - weighted psds of @xmath180 and @xmath183 is the strong power at high - frequencies in @xmath183 . [ fig - alpha_2dfft],scaledwidth=50.0% ] shown in figure [ fig - alpha_2dfft ] are the frequency - weighted psds of @xmath180 and @xmath183 for 15 , 20 , and 25 r@xmath45 . the modulation of the maxwell stress by the dynamo is more clearly seen in the strong , multi - peaked bands of power in the psds of both @xmath180 and @xmath183 corresponding to oscillations of the low - frequency dynamo . however , the psds of @xmath183 also have a high - frequency component to its variability from fluctuations due to mri - driven turbulence . this component is much more significant than the high - frequency variability in @xmath180 . interestingly , the bands of low - frequency variability in the frequency - weighted psds of @xmath183 are found around the local dynamo frequency , but the frequency of the variability is constant across the radii we probe . the strong band of power in the psds of @xmath180 show the expected shift to lower frequency with increasing radii due to the decrease in orbital frequency . this discrepancy seems to indicate that while the dynamo does introduce the low frequency variability in @xmath1 needed to drive the propagating fluctuations , the frequency at which @xmath1 is modulated is subject to a more global interaction of the magnetic field . in the analysis of propagating fluctuations we are interested in the behavior of the instantaneous mass accretion rate , @xmath189 and how it is related to the local disk variables , specifically the surface density and the effective @xmath1-parameter . additionally , we are interested in how evolution of the accretion disk might appear if we could observe our simulated disk . we therefore employ a dissipation proxy to generate synthetic time - dependent luminosity . in accretion disks , the outward transport of angular momentum liberates gravitational potential energy from the disk gas . this is then converted to thermal energy through the turbulent cascade of mri driven turbulence and removed from the system through radiative processes . assuming the disk is radiatively - efficient , the magnetic stress can be used as a tracer of energy dissipation since the energy injected by the stress is quickly radiated away . adopting eqn 9 from @xcite , the local flux at the photosphere of the disk ( at one scaleheight ) due to dissipation is given by @xmath190[eqn - proxy ] where a and b are relativistic correction factors given by @xmath191 and @xmath192 this scheme has been used to produce synthetic light curves in other studies of global accretion disks , e.g. @xcite and @xcite . shown in figure [ fig - globalplot ] are the spacetime diagrams of @xmath193 , @xmath27 and the emission proxy . as can be seen in the figure , periods and regions of enhanced accretion ( higher @xmath27 ) are coincident with higher @xmath193 . the correlation of @xmath27 with surface density acts to ( partially ) preserve the pattern of fluctuations in surface density as disk material loses angular momentum and moves to smaller radii . additionally , when @xmath27 is greater at a given radii , there is a corresponding increase in the emission proxy . it is worth noting that the name propagating fluctuations " is a slight misnomer as there is no real propagation mechanism at play in the disk . rather , the mass accretion is a diffusive process that scales with @xmath193 , providing the qualitative appearance of propagating " to smaller radii . figure [ fig - mdottime ] shows the time trace and psd of @xmath27 at the isco ( 6 r@xmath45 ) . as expected , there are stochastic accretion events when @xmath27 dramatically increases ( by a factor of 3 - 5 ) , and then returns to a stationary baseline value , a value of @xmath194 . the psd is featureless and is best fit with a powerlaw with index of @xmath195 , corresponding to flicker - type noise . the histogram of the instantaneous mass accretion rate at the isco in our simulation is shown in figure [ fig - dist ] . the shape of the distribution resembles the distribution of gbhb systems like cygnus x-1 @xcite and is , similarly , very well fit by a log - normal distribution of the form @xmath196,\ ] ] where @xmath197 is the normalization , @xmath198 is the log of the mean , and @xmath199 is the standard deviation . the best - fitting model parameters for the probability density function ( pdf ) of the mass accretion rate at @xmath200 are @xmath201 and @xmath202 and gives a @xmath203 of @xmath204 . for comparison , the best fit with a normal distribution has a @xmath203 of @xmath205 . statistically , the log - normal distribution is overwhelmingly favored ( @xmath206 ) and , visually , we can see the normal distribution fails to reproduce the fast rise of the distribution at low @xmath27 values and shallower tail of the distribution at higher @xmath27 values . the implication of this distribution is , hence , that fluctuations in @xmath27 combine multiplicatively , rather than additively in our mhd disk . at the isco fit with a normal ( red line ) and log - normal ( blue line ) distribution . the log - normal distribution provides a superior fit . [ fig - dist],scaledwidth=50.0% ] we can also quantify the asymmetry of the distribution by measuring skewness . skewness is the third momentum of the distribution and is given by , @xmath207 where @xmath198 is the mean , @xmath199 is the standard deviation , and @xmath208 is the number of data points . the distribution of @xmath27 at the isco has a value of @xmath209 , indicating a strongly positively skewed distribution . we will now look at the correlations in @xmath1 , @xmath27 , @xmath210 @xmath193 to verify that @xmath27 does indeed scale with both @xmath1 and @xmath193 . it is typically assumed @xmath1 is independent of @xmath193 thus , according the canonical disk equation , @xmath27 is expected to be higher when @xmath1 and @xmath193 are larger . however , there is a degeneracy between these three parameters and we must remove the stochasticity of the third in order to tease out relationships between any two quantities . the easiest way to find the underlying trend is to simply bin the data , which essentially averages out the stochasticity of the variable we are not interested in . since there are radial gradients in the disk , we focused on the behavior at @xmath211 @xmath212 . ten bins were used with 70 data points per bin . the bars indicate the standard deviation of the bin . figure [ fig - ams ] shows the @xmath213 , @xmath214 , and @xmath215 correlations .. we find that , indeed , the mass accretion rate scales with both @xmath1 and @xmath193 , validating the _ a priori _ assumptions that went into the phenomenological motivation of the propagating fluctuations model . the pearson correlation coefficient between @xmath193 and @xmath27 is @xmath216 and between @xmath1 and @xmath27 it is @xmath217 , indicating these are both statistically strong correlations . additionally , the correlation coefficient of @xmath1 with @xmath193 is @xmath218 , which is considered a moderately strong anti - correlation . while the strong positive correlations of @xmath27 with @xmath1 and @xmath193 confirm the underlying assumptions of the propagating fluctuations model , the negative @xmath1-@xmath193 correlation is a slightly different than expected . we believe the anti - correlation between @xmath1 and @xmath193 can be attributed to the magnetic buoyancy in the disk . in the higher density regions , pressure balance in the disk displaces magnetically dominated ( lower @xmath123 ) gas upwards , decreasing the local field strength , thereby decreasing @xmath1 and causing the negative trend . the strongest evidence for propagating fluctuations in our simulation comes from radial coherence and frequency dependent phase shifts of the @xmath27 variability . at the heart of the propagating fluctuations model is the predication that modulations in the accretion rate at larger radii will be seen at the inner radii with a time - lag set by the viscous inflow time . the coherence function provides the cleanest way to assess the causal connection in the disk . adopting the convention of @xcite , we will consider @xmath27 at two radii , @xmath219 and @xmath220 , with fourier transforms @xmath221 and @xmath222 , respectively . the coherence of these two signals , given by @xmath223 is a real - valued positive function that has a maximum of unity if and only if @xmath220 is related to @xmath219 via simple a linear transfer function , @xmath224 where @xmath225 is some transfer function . in the other extreme , @xmath226 implies no linear relationship between @xmath219 and @xmath220 . in general , @xmath227 can be considered a measure of the fraction of variability in @xmath220 that is coherently related to @xmath219 and has values between @xmath228 ( completely incoherent ) and @xmath229 ( perfectly coherent ) . for non - zero coherence , we can compute the cross - spectrum @xmath230 . the complex phase @xmath231 of the cross spectrum gives the frequency - dependent phase shift of the coherent parts of @xmath219 and @xmath220 . the coherence function and phase shifts are shown in figure [ fig - coh_phase ] . the coherence function was calculated with the isco taken to be the inner , reference radii . the @xmath27 signal " was broken into three segments for averaging and the coherence function was calculated with each radial bin . similar to @xcite , we find the coherent regions are found at frequencies below that of the local viscous frequency . the viscous time can hence be interpreted as a minimum timescale on which inward communication of flow fluctuations can occur . from the average phases , time - lags can be determined . the frequency - dependent time lag at a given radii is @xmath232 . figure [ fig - phase_lag_slice ] shows the coherence function calculated between @xmath27 at the isco and 10 @xmath212 . given the length of our simulation , only the inner regions have been run for a viscous time , so we only show the structure of the phases and time lags calculated at these two radii . from the two plots we can see fluctuations in @xmath27 have phase shifts that increase with frequency , but a constant time lag up until the fluctuations become incoherent at @xmath233 . the constant time lags of the fluctuation in @xmath27 at different radii was seen in the semi - analytic model of @xcite . this implies that the behavior of the propagating fluctuations in our mhd simulation is consistent with those of the one - dimensional viscous disk . we finish our analysis of @xmath27 in by looking at the rms - flux relationship . while the rms - flux relationship is a staple of broadband variability @xcite , its physical origins remain uncertain . this relationship is of great interest because it appears to be an intrinsic property of the accretion process and , possibly , even more fundamental than the psd of the variability . @xcite originally hypothesized the rms - flux relationship is evidence of @xcite type propagating fluctuations in @xmath27 since the rms - flux relationship is a ubiquitous property from all black holes , regardless of mass . if the rms - flux relationship in the photometric variability is truly due to the propagating fluctuations in the accretion disks of these sources , it should be seen in the @xmath27 of our simulation . to test this we calculated the mean , @xmath234 and rms @xmath235 of the mass accretion rate at the isco over a sliding window of 40 isco orbits ( n=20 data dumps ) . shown in figure [ fig - rms_flux ] is a plot of the raw @xmath236 vs @xmath237 scatter plot . the relationship was fit using a bayesian , mcmc method with a linear function of the form @xmath238 with @xmath15 = @xmath239 and @xmath240 = @xmath241 . the ( almost ) one - to - one correspondence of the rms to the average mass accretion rate is expected , as well as the intercept passing through the origin , i.e. when there is no flux there should be no variability . however , the observed rms - flux relationships are typically offset and best described by a model with a constant baseline flux , @xmath242 , that varies with black hole spectral state . our results suggest the variable component is related to the underlying mhd turbulence of the disk and the flux offset can be attributed to disk processes not captured by our simulation . we now discuss the global behavior of our emission proxy and how the propagating fluctuations in @xmath27 might appear as photometric variability . the full details of the radiation produced through the interaction of the accretion flow with the various emission mechanisms can only be captured by including more realistic radiation physics , which our simulation lacks . nevertheless , we can use our emission proxy to explore the emission variability in a broad sense . in particular , we want to determine if the structure of the @xmath27 variability can be easily observed when the emission is integrated over the entire disk . the flux we observe from gbhbs and agns is , in effect , the integrated emission from a range of radii that evolve on timescales proportional to their dynamical times . consequently , there is the possibility some of the variability could be smoothed out as lower - frequency variability from larger radii can wash - out higher - frequency variability . given how clear the nonlinear signal is in astrophysical sources , we want to check it is reproduced by the variability of the emission proxy from our simulation . vs its average value at the isco calculated over 40 isco orbit periods of the simulation . a strong linear dependence is present in the variability , consistent with the rms - flux relationship observed from astrophysical black hole systems . [ fig - rms_flux],scaledwidth=50.0% ] figure [ fig - lc_plots ] shows the synthetic light curve and radial profile of the average emission for our disk . we created the light curve by integrating the dissipation from the emission proxy over the well - resolved region of the simulation ( @xmath84 ) . given our radial range , the synthetic light curve is composed of signals from radii with dynamical times that vary by a factor of 37.5 , imitating observation with a broad - band filter . several aspects of the disk emission are apparent . first , the light curve is similar to the time trace of @xmath27 at the isco . the signal is aperiodic and flares in the light curve correspond to large accretion events . second , any flares in the light curve quickly decay down to a well - defined , stationary average . the property of mean - reversion is of statistical interest as a way to model its behavior . finally , and worth the most discussion , is the dominate role emission from the inner region plays in driving variability this is more clearly seen in the radial profile of the emission per unit area of the disk , but can also be inferred by the high - frequency flares . in figure [ fig - lc_plots ] the disk profile of is shown for comparison . the two profiles trace each other well at larger radii , but diverge in the innermost regions . in the canonical @xmath1-disk the energy flux goes to zero at the isco because it was believed at the time that no torques could be felt from the gas in the inner plunging region . mhd simulations have shown that , in fact , torques can be felt across the isco and radiation can originate from closer to the black hole than the isco @xcite . in addition to the physical effect of torques spanning the isco region , there may also be a nonphysical contribution to our profile because we assume any energy from stresses is immediately converted into radiated energy . in reality , there is a time delay between the injection and the dissipation as the energy is transported through the turbulent cascade . in the inner regions near the black hole the material could be rapidly swept into the hole before it has had time to completely radiate the injected energy . this would lead to an overestimation of the emission from the inner regions by our emission proxy due to this advection of energy . in our analysis of the mass accretion rate in the simulation we found the variability of @xmath27 had near - identical structure to the universally observed flux variability from black hole systems . we will now focus on assessing how well @xmath27 fluctuations translate into fluctuations in disk luminosity using the diagnostics used in section [ sec - prop_flucs ] . figure [ fig - lc_diagnostics ] ( upper and middle panels ) shows the flux histogram and rms - flux relationship for the total ( `` broad - band '' ) proxy light curve shown in fig . [ fig - lc_plots ] . the preferred fit to the flux histogram is a log - normal distribution with parameters @xmath243 and @xmath244 . the @xmath203 of this fit is @xmath245 which is statistically well - fit and much better than that of a normal distribution with @xmath203 of this fit is @xmath246 . the skewness of this distribution is @xmath247 , indicating the histogram of the light curve is more highly skewed than the histogram of the @xmath27 and the log - normal distribution is more pronounced . naively , the increased skewness is surprising given that the light curve is integrated over the disk and presumably additively combining many random processes . according to the central limit theorem , this should drive the distribution to be more gaussian . however , since @xmath27 is radially coherent , the fluctuations of the emission proxy are not independent , as is assumed in a gaussian process . the full synthetic light curve also has a linear relationship between the rms of the variability and the average flux value . as we did with @xmath27 , the light curve was broken into 20 sections and the standard deviation and mean values were calculated for each . the distribution is well fit by a linear function with @xmath15 = @xmath248 and @xmath240 = @xmath249 . the rms - flux relationship of the synthetic emission is flatter than the linear relationship between the rms of @xmath27 and the average value . in fact , the slope of our linear fit is remarkably similar to those seen in gbhbs , specifically cyg x-1 and sax j1808.4 - 3658 @xcite . while it is tempting to make a direct comparison , the statistics on our measurement prevent drawing any deeper conclusions . however , given how much the difference in the slopes of the rms - flux relationship of @xmath27 compared to the rms - flux relationship of the light curve , it does appear that integrating over the disk plays a significant role . by integrating over the entire disk , the low - frequency emission adds a slowly varying component to the flux , thereby decreasing the percentage of the flux from the fast varying inner regions and decreasing the slope . an examination of coherence requires us to create two light curves to imitate observation with two bands ( for instance hard and soft x - rays ) . we formed these light curves by integrating the emission proxy in the regions spanning @xmath250 and @xmath251 . then , as we did with @xmath27 , the coherence function was calculated by dividing the light curves into three sections , resulting in the coherence function shown in fig . [ fig - lc_diagnostics ] ( bottom panel ) . the coherence shows the same behavior as found for the mass accretion rate , with highly - coherent signals at low - frequencies , and a falling coherence at higher - frequencies . de - coherence is seen at @xmath252 which , while a relatively high - frequency , agrees with the coherence and phase plots of @xmath27 fluctuations with the isco in figure [ fig - coh_phase ] . having established the signatures of propagating fluctuations in @xmath27 are present in the variability of our emission proxy , we will now look at the psd and the statistical behavior of the variability by fitting a simple ornstein - uhlenbeck ( ou ) process . shown in figure [ fig - plaw_fit ] is the psd of the synthetic light curve fit with a broken power law . the last 512 data points of our light curve were used to calculate the psd which ensures no phase dependencies were present . following the method described in @xcite , the psd was fit with a power law plus white - noise component and broken power law plus white - noise component of the form @xmath253 where k is a normalization constant , @xmath254 is the low frequency slope , @xmath255 is the high frequency slope , and @xmath256 is the break frequency , to determine which model provides the best description of the psd . the power density was assumed to have an exponential probability distribution with the probability of measuring a given power between @xmath257 and @xmath258 expressed by , @xmath259 where p@xmath260 is the mean power . for an individual bin , the likelihood of a measuring power @xmath261 is@xmath262 for the entire psd , the likelihood is the product of the individual likelihoods . since this can often be very large or very small , the standard statistic used is the log likelihood , @xmath263.\ ] ] a markov chain monte carlo ( mcmc ) algorithm was used to conduct the fitting and find the parameters that maximize the log likelihood of each psd model . , scaledwidth=50.0% ] the psd is well fit by both a broken power law with @xmath264 , @xmath265 , and @xmath266 and a single power law with @xmath267 . however , the broken power law provides a much better statistical fit with @xmath268 at the expense of two degrees of freedom . this translates into a @xmath269 . hence , while the underlying psd of @xmath27 at the isco in our simulation has @xmath270 , the integrated emission proxy has a steeper psd . this is simply due to the integrated contributions from across the accretion disk , with additional low frequency power from large radii being added to the psd . it is not immediately clear what physics sets the break frequency @xmath256 , corresponding to a timescale of @xmath271 , although it approximately corresponds to the cooling / thermal timescale of the inner regions of the disk . in order to ensure that the low - frequency break is not imposed by the truncation of our emission proxy at @xmath272 , we generated light curves and the corresponding pdss truncating at @xmath273 . in all cases , the break persisted and the frequency was unchanged . in addition to the psd , the statistical properties of the variability of the synthetic light curve are of interest and provide an additional method of characterizing the synthetic light curve . recent work has shown a damped random - walk , and more specifically an ornstein - uhlenbeck ( ou ) process , provides a very good statistical model of gbhb and agn variability @xcite . the ou process is part of a class of statistical models called _ continuous time , first order autoregressive _ , car(1 ) , processes . car(1 ) processes can take many forms and behave very different depending on their model parameters . the ou process describes a noisy relaxation process where a system is stochastically perturbed from a stationary mean and decays back to the mean with a relaxation time @xmath274 @xcite . the stochastic differential equation describing such a process is : @xmath275 where @xmath276 , @xmath198 is the mean and @xmath199 is a measure of the average magnitude of volatility per @xmath277 . this equation has the following direct solution : @xmath278 where @xmath279 is the time step and @xmath280 is a zero - centered , gaussian distribution with a variance of 1 . the even sampling of our light curve provides some simplicity in fitting the ou process . @xmath281 is linearly dependent on @xmath282 and takes the form : @xmath283 where @xmath284 is a normal - random term that is independent and identically distributed . from a simple linear fit , we can solve for @xmath285 , @xmath198 , and @xmath199 where @xmath286 @xmath287 @xmath288 since the ou process is inherently a gaussian process we must therefore perform the fit to the synthetic light curve in log - space since our flux " distribution is log - normally distributed . thinking in terms of real observations , this is equivalent to fitting in magnitude space . for our light curve , the best fit parameters are @xmath289 , @xmath290 and @xmath291 . the @xmath285 parameter corresponds to a decay time of @xmath292 . in our simulation we find that @xcite type propagating fluctuations " quickly develop and are sustained throughout the entire simulation . the multiplicative manner by which fluctuations in @xmath27 combine and grow defines the behavior and subsequent evolution of the disk . the most significant consequence from the multiplicative combination of @xmath27 fluctuations is that the accretion history of a given parcel of gas is retained , allowing for the apparent propagation of fluctuations on a viscous time . the @xmath27 variability at the isco in our simulation bears striking resemblance to variability observed from astrophysical black holes and shares the same phenomenological properties . in particular , it reproduces the log - normal flux distributions , rms - flux relationship , and radial coherence ubiquitously observed from gbhbs and agns . these nonlinear features have typically been difficult for variability models to account for , but are a consequence of a scenario where the mass accretion rate is proportional to the multiplicative product of a stochastically varying viscosity and the local surface density . in our simulated disk they develop naturally from the mri driven turbulence . additionally , we use an emission proxy to generate a disk - integrated synthetic light curve to verify that the nonlinear characteristics of the @xmath27 variability will actually translate into observable features in the disk emission . while rudimentary , our emission proxy follows the turbulent energy injected to the gas , which we assume is quickly radiated away since we are considering a radiatively efficient disk . we find that , indeed , the nonlinear features of @xmath27 are present when the entire disk integrated emission is considered and are not lost with the addition of low - frequency variability originating from gas at large radii . our work was able to address several issues raised by @xcite regarding the timescale on which the effective @xmath1 parameter of the disk fluctuates and how well an mhd disk can really be described by a strictly viscous disk on long time scales . the previous efforts to model propagating fluctuations in mass accretion rate of @xcite using a linearized model and @xcite using a one dimensional viscous disk model relied on the _ a priori _ assumption that the low - frequency fluctuations in effective @xmath1 occurred on a viscous time . however , this is very long for the magnetic field to vary . when @xcite allowed @xmath1 to vary on a dynamical time , the typically assumed evolutionary timescale for the mri , they found that the fluctuations were too rapid and propagating fluctuations did not develop . we find the disk dynamo can provide the necessary intermediate timescale modulation of the effective @xmath1 to drive propagating fluctuations in mass accretion rate . dynamo action is universally observed in modern accretion disk simulations and seems to be intimately tied to the mhd physics of a differentially rotating plasma . thus , the dynamo can provide a natural mechanism to introduce variability in the effective @xmath1 . while the ubiquity of the dynamo has been established , its presence is typically neglected when considering accretion disk evolution . our results highlight the significant impact the dynamo can have on the disk and that it is a fundamental ingredient in the driving of propagating fluctuations . the properties of the @xmath27 and light curve variability from our global , _ ab initio _ mhd simulation are very similar to those from the simple viscous disk with stochastic viscosity parameter of @xcite . from this similarity , we can conclude that on long timescales our mhd disk acts in a viscous manner . this is not surprising , but there is not a trivial connection between the internal mhd stresses that transport angular momentum and the expected viscous behavior of the disk , even though it has largely been intuited from empirical results . our mhd simulation captures gross aspects of real disk variability , but it is not without limitations . in particular , we use simple prescriptions to approximate the affects of general relativity and radiative physics . as a first effort at modeling propagating fluctuations in mass accretion rate in an mhd disk we are mainly concerned with capturing the dynamics of the disk and the interplay of the accretion flow with the stochastically fluctuating effective @xmath1 . to do this , a long , high - resolution simulation is required . therefore , the priority of where computational resources are used shifts away from more complex physics and into the duration and resolution of the simulation . while the focus of our analysis has been on understanding the signatures of propagating fluctuations around black holes , we do not expect the effects of general relativity and radiation physics to have a significant impact on the characteristics of the disk evolution are primarily concerned with . in fact , the phenomenology we seek to understand is also readily observed in cataclysmic variable ( cv ) systems @xcite . these systems are non - relativistic and not hot enough to be largely affected by radiation pressure in the disk . the similarity of the broad - band variability between cvs , gbhbs , and agns implies that propagating fluctuations are a generic feature of highly ionized accretion disks . the fluctuations resulting from the underlying accretion processes behave consistently across many orders of magnitude of mass , temperature and evolutionary time implying that the growth of propagating fluctuations is not unique to an specific class of system , but rather intrinsically tied to the angular momentum transport resulting from mhd turbulence . therefore , we believe that the results from this simulation are robust enough to provide a qualitative description the accretion mechanism for highly ionized accretion disks . we have performed the first examination of the propagating fluctuation picture using _ ab initio _ mhd turbulent models of disk accretion . the propagating fluctuations model has served to explain the universally observed nonlinear features of photometric variability from accreting black holes of all masses . in our simulation , propagating fluctuations in @xmath27 form and behave in a manner consistent with those of the viscous disk model of @xcite . our main results are summarized here : 1 . the effective @xmath1-parameter in our simulated disk is highly variable . the variability has two components : a high - frequency component due to local fluctuations from the mri and a lower frequency component due to modulation of the stress by the disk dynamo . the lower frequency variability in the effective @xmath1-parameter from the disk dynamo drives fluctuations in @xmath27 by modulating the internal disk stress on an intermediate timescale . as was shown in @xcite , if fluctuations in @xmath1 are too rapid , fluctuations will be damped out and propagating fluctuations will not grow . the dynamo frequency is low enough that propagating fluctuations can easily grow . previous efforts directed at understanding the role of propagating fluctuations in @xmath27 relied on _ ad hoc _ prescriptions for the variability timescale . we show , for the first time , that the dynamo fills this role and provides a natural driver for fluctuations in @xmath27 . the @xmath27 in our simulation is positively correlated with @xmath193 and the effective @xmath1 . in the analytic disk equations the mass accretion rate is proportional to the product of the local surface density and the local viscosity . these correlations occur in our _ ab initio _ simulation , suggesting that they naturally result from the interaction of the mri with accretion flow . the @xmath27 variability at the isco of our simulation resembles the photometric variability observed from gbhbs and agns . the @xmath27 variability is log - normally distributed , has a linear rms - flux relationship , and displays radial coherence . these features are empirically observed and have been attributed to propagating fluctuations in @xmath27 . we find that they are , in fact , properties of accretion flow when propagating fluctuations in @xmath27 are present . 5 . using an emission proxy to generate a disk integrated , synthetic light curve , we find that the qualitative properties of the accretion flow are reflected in the observable emission . while the nonlinear signatures are preserved in the light curve , integrating over the disk adds additional low - frequency power from more distant radii . however , since inner region of the disk is brightest , the high frequency variability is preferentially weighted and the nonlinear properties of @xmath27 are preserved . we would like to thank mark avara , m. coleman miller , phil uttley , and christian knigge for the many insightful discussions throughout the course of this work . jdh is grateful for the financial support from the umd - goddard joint space science institute ( jsi ) through his graduate fellowship . this research was also supported by nasa through grants nnx15ac40 g and nnx14aj04 g .

quantum information processing seeks to perform tasks which are impossible or not efficient with the use of conventional classical information processing , by using systems described by quantum mechanics . we can consider two kinds of quantum information tasks based on the types input states : the classical input tasks and the quantum input tasks . whereas input states of the classical input tasks are quantum states but encode just classical information , input states of the quantum input tasks encodes quantum information , of which unit is described by a qubit @xmath0 where @xmath1 and @xmath2 are unknown complex parameters satisfying @xmath3 and @xmath4 and @xmath5 are a fixed basis of a qubit . for example , quantum algorithms @xcite are classical input tasks and quantum error correcting codes @xcite and quantum universal optimal cloning @xcite are quantum input tasks . to investigate yet unveiled full quantum potential of quantum information processing , it is necessary to understand properties of quantum input tasks . in many quantum input tasks , how quantum information is encoded in the larger hilbert space of composite systems determines the main functionality of the tasks . for example , in quantum error correcting codes , qubit information is encoded in a subspace of a larger hilbert space such that it can be still recovered after being influenced by certain errors ( or noises ) which map input qubit information into the whole hilbert space . the encoding process can be described by a transformation of a computational basis @xmath6 where the original quantum information is given into a set of orthogonal states in the larger hilbert space @xmath7 . in this picture , the properties of encoding for a task are captured by the choice of a set of states @xmath7 , which represents _ how _ original quantum information is spread across the hilbert spaces of subsystems . entanglement , or a non - local quantum correlation , of an _ individual _ state is an essential resource for performing quantum input tasks such as quantum teleportation @xcite , namely , the existence of entanglement is necessary for performing teleportation beyond the classical limit . to analyze _ non - local _ properties of spread quantum information described by the set of states is a way to characterize how quantum information is spread by encoding . here , we use the word non - local to represent properties which are not fully accessible by just using local operations on the subspaces and classical communications ( locc ) but global operations on the whole systems . for individual states , the existence of this kind of non - locality is accompanied by the existence of entanglement . however , it is also known that such non - local properties of a set of states can be essentially different from non - locality of individual states . an important example is a set of nine product states which can not be locally discriminated by using locc presented in @xcite ( the `` non - locality without entanglement '' paper by bennet et al . ) . in this example , there is no entanglement in the quantum states where classical information is encoded , therefore , no entanglement resource required for encoding classical information , but it is not possible to decode ( i.e. , identify encoded classical information ) deterministically locc , without using entanglement resources . entanglement properties of each encoded state does not fully capture the non - local property appearing in the decoding process . as it had been also pointed out in the context of local copy and local state discrimination in @xcite , impossibility of tasks involving locc transformation of a set of states implies _ non - locality beyond individual entanglement_. for characterizing non - locality of the spread of quantum information , non - local resources required for decoding quantum information should be considered as well as for encoding quantum information , since required minimum resources for encoding process and decoding process are not necessary the same in locc transformations . in this paper , we focus on non - local resources required for decoding processes and investigate properties of an extreme case of spreading quantum information that does not consume non - local resources for decoding . we study a simple but fundamental case of spreading qubit information into two - party states . we present necessary and sufficient conditions for spreading qubit information into bipartite states such that qubit information can be extracted by only using locc between the two parties . we call this task _ remote extraction_. in this task , since we have to investigate simultaneous transformations of _ two _ states @xmath8 under all possible locc , unlike the case of a single known pure bipartite state where lo - popescu theorem @xcite is applicable , the proof of necessity is involved and it consists of seven steps . we also present the explicit form of the locc for achieving the perfect remote extraction . interestingly , the obtained conditions indicate that qubit information can be asymmetrically spread into a bipartite state . [ fig : asymmetricloccextraction ] ) . there is a way of encoding to spread qubit information where it can be remotely extracted with unit probability at one of the parties but not at the other as long as they are using locc . thus , we can introduce a non - local property of a set of states in an asymmetric manner for two parties , whereas entanglement properties of an _ individual _ bipartite pure state are always symmetric between the two parties due to schmidt decomposition . this fact is another indication of the difference of non - locality of a set of states and an individual state . , width=283 ] the asymmetric property of spreading quantum information can be used for controlling the transmission of quantum information . namely , we can distribute quantum information such that one of the party , say alice , can act as a controller of quantum information to support other party to extract full quantum information , but she can not obtain full quantum information for herself , as long as they are acting by locc . therefore we can consider a simple two qubit device which globally encodes qubit information and also acts as a local `` switch '' for transmission of qubit information to one of the qubits . from the viewpoint of controlling transmission of quantum information , it is also useful to spread quantum information into two parties such that alice s local operation can _ irreversibly destroy _ quantum information such that the state of bob s qubit after alice s operation is set to a _ pure _ state which does not contain quantum information , i.e. the parameters @xmath1 and @xmath2 . note that this process is not a randomizing process to transform bob s qubit to be in a completely mixed state . in this task , a part of classical information of the parameter @xmath1 and @xmath2 can be retrieved . we call this task _ remote destruction_. ( fig . [ fig : loccdestruction ] ) such a way for spreading quantum information can be used another kind of `` switch '' for controlling quantum information transmission . using a similar technique for proving conditions of remote extraction , we also present necessary and sufficient conditions for this task . spreading quantum information for remote destruction can also be asymmetric between two parties , but we show that the conditions for remote destruction are incompatible for those of remote extraction . and @xmath2 ) . note that some classical information represented by @xmath9 can be retrieved from the outcome of alice s measurement @xmath10 . [ fig : loccdestruction],width=283 ] this paper is organized as the following : in section [ statements ] , the definitions and precise statements of remote extraction and destruction are given . in section [ ext - sufficiency ] , we show the proof of sufficiency for remote extraction . the preparations and outline of the proof of necessity are given in section [ ext - outline ] . the proof of necessity consists of seven steps and they are presented in section [ ext - necessity ] . the proof of conditions for remote destruction is presented in section [ des ] , and the summary and discussions are given in section [ summary ] . we take two orthonormal vectors @xmath11 in two qubit hilbert space @xmath12 , which we will call _ basis states _ , and encode qubit information into a two - qubit state represented by @xmath13 . we consider two qubits are spatially separated and one of the qubit is at alice s side and the other qubit is at bob s side . the task of remote extraction is to extract qubit information at bob s side from the two - qubit state @xmath14 by using finite rounds of locc . that is , we look for a finite round locc procedure @xmath15 such that @xmath16 for arbitrary @xmath17 satisfying @xmath18 is a fixed orthonormal basis in @xmath19 and @xmath20 is an arbitrary vector in @xmath21 . throughout this paper , we use a notation @xmath22 , and @xmath23 where @xmath24 and @xmath25 are complex conjugates of @xmath1 and @xmath2 , respectively . we also denote the conjugation of a single qubit state @xmath26 with respect to an orthonormal basis @xmath27 of the qubit by @xmath28 . the finite round locc procedure @xmath15 is given by a sequence of alice s measurements @xmath29 and bob s @xmath30 , where @xmath31 is an index for alice s @xmath32-th round measurement and @xmath33 is an index for bob s @xmath32the round measurement ( @xmath34 ) , satisfying the normalization conditions @xmath35 for each @xmath32 . we use the notation @xmath36 . it is easy to see that ( [ task ] ) is equivalent to @xmath37 for @xmath38 and all @xmath39 . when there exists a locc procedure satisfying the condition of eq.([eq ] ) , we say that qubit information on the basis states @xmath11 can be _ deterministically extracted _ by locc at bob s side . we call the pair of the state given by the right hand side of eq.([eq ] ) @xmath40 as the _ extracted form _ and @xmath41 with some unitary operator @xmath42 is said to be _ locally equivalent to the extracted form to bob_. we use these notations in our proof of remote extraction . now our problem is to find the condition of the basis states @xmath43 satisfying eq.([eq ] ) . in this paper , we prove the following theorem : [ one ] qubit information spread between alice and bob @xmath44 can be deterministically extracted using only locc at bob s side @xmath45 if and only if the schmidt decompositions of the basis states @xmath46 are given by @xmath47 where @xmath48 are the schmidt coefficients satisfying @xmath49 and @xmath50 , and @xmath51 and @xmath52 are the schmidt basis of alice s qubit and @xmath53 is the schmidt basis of bob s qubit . if the conditions of eq.([cond1 ] ) and ( [ cond2 ] ) are satisfied , then @xmath54 are of the form @xmath55 using three real parameters @xmath56 , @xmath57 and @xmath58 . we define a family of orthonormal basis @xmath59 labeled by positive real number @xmath60 , by @xmath61 where @xmath62 and @xmath63 . then if the pair of vectors @xmath64 is not locally unitary equivalent to the extracted form , all the measurements by alice on it are of the form @xmath65 where @xmath66 is an unitary operator , @xmath67 and @xmath68 . on the other hand , all the measurements that bob carries out @xmath69 are scalar multiplications of unitary operators . from the theorem , we see that the schmidt coefficients of the basis states have to be identical for remote extraction , therefore the basis states should have , at least , same entanglement for remote extraction . on the other hand , the schmidt base of alice s qubit of the basis states are not necessary to be same . although the orthogonality condition of the basis states @xmath70 fixes one of the parameters to be @xmath71 , @xmath72 or @xmath73 for @xmath74 , we can choose @xmath57 and one of @xmath56 and @xmath58 in eq . ( [ ov ] ) arbitrary . this property allows asymmetry of remote extraction ; we can encode qubit information such that the conditions for remote extraction at bob are satisfied but the conditions for remote extraction at alice are not satisfied . we can also obtain necessary and sufficient conditions of the basis states for symmetric remote extraction , where the deterministic remote extraction at either alice or bob is possible depending on the choice of locc procedures from the theorem 1 . [ both ] extraction to either alice or bob is possible if and only if the schmidt decompositions of the basis states are given by @xmath75 thus , if a set of the basis states @xmath43 satisfies the conditions of remote extraction eqs . ( [ cond1 ] ) and ( [ cond2 ] ) , but does not satisfy the conditions for the symmetric ones eqs . ( [ sym1 ] ) and ( [ sym2 ] ) , it gives an asymmetric way of spreading qubit information , where deterministic remote extraction is only possible at bob , not at alice . for @xmath76 , any choice of two orthogonal ( maximally entangled ) states can be transformed into the forms of eqs . ( [ cond1 ] ) and ( [ cond2 ] ) , therefore there is no asymmetric remote extraction . however , for @xmath77 where the schmidt base are determined uniquely , asymmetry of ( perfect ) remote extraction is guaranteed as long as @xmath78 in eq . ( [ ov ] ) . the case of @xmath79 presents an interesting picture how qubit information is spread between two parties in terms of symmetry and asymmetry ; for @xmath72 , qubit information is already extracted at bob from the beginning , and no qubit information can be extracted at alice by locc , for @xmath80 , qubit information is symmetrically shared between alice and bob , and for @xmath81 , qubit information is shared but asymmetrically . we note that bob s operation is restricted to scalar multiplication of unitary operators for extracting qubit information at bob . therefore , once one of the party performs an extraction measurement of the form of eq . ( [ measurement ] ) , qubit information can be only extracted to the party who has not performed the extraction measurement , even with the basis states allowing symmetric remote extraction . the measurement condition also implies that one - way locc , where alice performs a projective measurement on her qubit in the @xmath82 basis and bob performs a conditional unitary operation depending on alice s measurement outcome is sufficient for remote extraction of qubit information . the task of remote destruction is to _ irreversibly _ destroy spread qubit information by acting one of the party ( alice ) and to prevent extracting quantum information at the other party ( bob ) . we assume that bob would not cooperate to destroy information , and also we would like to prevent recovery of quantum information even if classical information about alice measurement is known . we look for alice s measurement @xmath83 such that for arbitrary @xmath17 satisfying @xmath84 , @xmath85 for each @xmath10 . here , @xmath86 is some scalar which depends on @xmath87 , and @xmath88 are vectors that do not depend on @xmath87 . in this paper , we show the following : [ distract ] deterministic remote destruction by alice is possible if and only if and only if the schmidt decompositions of the basis states are given by @xmath89 where @xmath90 , @xmath91 and @xmath51 is the schmidt base of alice s qubit and @xmath53 and @xmath92 are the schmidt basis of bob s qubit . if the schmidt rank of @xmath93 ( resp . @xmath94 ) is @xmath95 , then the measurement operators for deterministic remote destruction @xmath83 are of the form @xmath96 where @xmath97 is an arbitrary vector , @xmath98 , and @xmath99 is an orthonormal basis diagonalizing a matrix @xmath100 if the schmidt rank of both of @xmath93 and @xmath94 are @xmath101 , then the measurement operators @xmath83 are of the form @xmath102 here , the vector @xmath103 have to be @xmath104 or @xmath105 if @xmath106 and @xmath107 are not parallel to each other , while it can be an arbitrary vector if @xmath106 and @xmath107 are parallel to each other . we see that the conditions given by eq . ( [ apple ] ) is identical for the conditions for deterministic remote extraction at _ alice _ , instead of bob . therefore , the conditions for remote destruction by alice s measurement are incompatible for those of remote extraction by alice s measurement . the conditions for symmetric remote destruction are also given by eqs . ( [ sym1 ] ) and ( [ sym2 ] ) , therefore , in the symmetric case , alice can determine whether destructing qubit information or letting bob to extract full qubit information by the choice of her measurement , but bob is also in the same position . we first observe that if the conditions given by eqs . ( [ cond1 ] ) and ( [ cond2 ] ) are satisfied , then the representation of the base of alice s qubit ( eq . ( [ ov ] ) ) is obtained . the case of @xmath108 or @xmath109 is trivial . let us assume @xmath110 . since @xmath93 and @xmath94 are orthogonal , the two base of alice s qubit appearing in eqs . ( [ cond1 ] ) and ( [ cond2 ] ) have to satisfy @xmath111 if we represent the basis state @xmath112 by @xmath113 and another basis state @xmath114 by @xmath115 with complex parameters @xmath116 and @xmath117 satisfying @xmath118 , and a real parameter @xmath56 , the condition of eq.([iti ] ) implies @xmath119 . by introducing two real parameters @xmath58 and @xmath57 , @xmath116 and @xmath117 are represented by @xmath120 and @xmath121 , respectively . thus , we obtain the representation of the basis of alice s qubit in eq . ( [ ov ] ) . now we choose another basis of alice s qubit @xmath122 given by @xmath123 we will check that qubit information can be extracted to bob s qubit by alice s projective measurement described by @xmath124 followed by an appropriate unitary operation performed by bob depending on the measurement outcome of alice . if alice obtains the measurement result corresponding to @xmath125 , the basis states are transformed to @xmath126 @xmath127 if alice obtains the measurement result @xmath128 , the basis states are transformed to @xmath129 @xmath130 note that the resulting pairs are locally equivalent to the extracted form to bob . hence , by choosing a suitable unitary operation transforming the basis of bob s qubit back to @xmath131 , spread qubit information can be faithfully extracted to bob s side by only using locc . in our proof , we employ matrix representations of states . in this section , we first describe the matrix representation , and then introduce the key notion in our proof : _ extraction measurements ( e - measurements)_. we also present the outline of our proof of necessity for remote extraction consisting of seven steps . let @xmath132 be a @xmath133-dimensional hilbert space , and let @xmath134 be an orthonormal basis of @xmath132 . we consider a bi - partite system @xmath135 . let @xmath136 be a maximal entangled state in @xmath135 . the conjugation of a state @xmath137 with respect to @xmath138 is represented by @xmath139 . the conjugation of an operator @xmath140 with respect to a basis @xmath141 is denoted by @xmath142 , i.e. , @xmath143 one can easily check that the useful relations @xmath144 , @xmath145 and @xmath146 . these relations are extensively used in our proof . by straight forward calculation , we can check the following properties : [ form ] + 1 . for all @xmath147 , there exists unique @xmath140 such that @xmath148 . 2 . @xmath149 3 . @xmath150 an e - measurement performed by alice on a pair of orthonormal states @xmath43 is a measurement described by a set of measurement operators @xmath151 satisfying @xmath152 , which preserve orthogonality of the states @xmath153 for all @xmath10 and also equi - probability , namely , equal probability for measuring each basis state @xmath154 , while there exists @xmath10 such that @xmath155 . an e - measurement performed by bob is defined in the same manner . an e - measurement is not always possible and the existence of the e - measurement restricts the form of @xmath43 . note that the final pair of extraction @xmath40 is measurable by e - measurement ( e - measurable ) of alice given by @xmath156 . on the other hand , we call another type of measurement such that @xmath157 for all @xmath10 , a _ c - measurement_. note that if extraction to bob is possible , alice should be able to perform the e - measurement on the last pair , otherwise extraction to bob at the next round is not possible . now , we introduce a set of orthonormal base of alice s qubit , @xmath158 . we define @xmath158 by a set of all orthogonal basis @xmath159 such that the decompositions @xmath160 satisfy @xmath161 of course , @xmath158 can be an empty set , depending on @xmath162 . we call an element in @xmath158 , an orthonormal basis on which alice can perform an e - measurement . in fact , we will see that if alice can operate an e - measurement @xmath83 on @xmath43 , then each @xmath163 have to be of the form @xmath164 where @xmath165 is a single qubit unitary , @xmath166 , and @xmath167 . we also define @xmath168 in the same manner . then it is obvious that for arbitrary single qubit unitary operators @xmath169 and a complex number @xmath170 , we have @xmath171 we prove the necessary conditions for remote extraction in the following seven steps : _ step 1 : _ we prove that the orthogonality and equi - probability conditions should be satisfied for all rounds of locc . from this , we show that the local operations in the locc procedure have to be e - measurements or c - measurements . _ step 2 : _ we show that if alice can perform an e - measurement @xmath83 on a pair @xmath43 , then @xmath158 is non - empty . furthermore , we see that each @xmath163 has to be of the form given by eq . ( [ mform ] ) . _ step 3 : _ we derive the explicit form of the set @xmath158 when it is not empty . we see that it is parameterized by a positive scalar @xmath172 . _ step 4 : _ we derive the necessity conditions for both of alice and bob to be able to perform an e - measurement on a pair of states @xmath43 . _ step 5 : _ using the result of _ step 4 _ , we prove that the following situation is impossible : alice performs some e - measurement @xmath83 on @xmath43 , and for all the results of her measurement @xmath173 , bob can sequently perform an e - measurement . _ step 6 : _ we show that if deterministic remote extraction is possible , @xmath158 is not empty . _ step 7 : _ we show that eqs . ( [ bunk ] ) and ( [ deco ] ) imply that the schmidt forms of @xmath11 to be given by eqs . ( [ cond1 ] ) and ( [ cond2 ] ) . we show that deterministic remote extraction requires that the two vectors ( unnormalized basis states ) have to be orthogonal to each other and have the same norm at every step in locc . let us consider a locc described by a sequence of conditional local measurements of alice @xmath174 and bob @xmath30 , for @xmath34 . each set of measurement operators @xmath175 , @xmath176 satisfies @xmath177 we use a notation @xmath178 , as introduced in section [ statements ] , and denote the vectors at each step by @xmath179 where @xmath180 or @xmath181 . as seen in section [ statements ] , at the last turn ( @xmath182 ) , the two vectors are orthogonal @xmath183 and they have the same length @xmath184 . by summing up with respect to @xmath185 , using the relation ( [ sum ] ) , we have @xmath186 and @xmath187 . repeating this summation procedure , we obtain @xmath188 and @xmath189 for all @xmath190 and @xmath191 , i.e. , the orthogonality and equi - probability conditions should be satisfied for all rounds in locc . therefore , the local operations in the locc procedure have to be e - measurements or c - measurements . as the c - measurements can not extract information on its own , we need the e - measurements . we derive the necessity and sufficient conditions for alice to be able to carry out the e - measurement . [ on ] if alice can carry out an e - measurement on a pair of orthonormal states @xmath43 in @xmath192 , then , @xmath158 is not empty . furthermore , the e - measurement have to be of the form ( [ mform ] ) . _ proof : _ let @xmath193 be an e - measurement by alice on @xmath43 . as it is the e - measurement , there exists @xmath10 such that @xmath194 . as @xmath195 is positive , it can be diagonalized in a suitable basis @xmath159 . we will show that @xmath166 . in the basis @xmath159 , we have @xmath196 where @xmath197 . we define two matrices @xmath198 and @xmath199 for the matrix representation of the basis states @xmath200 and @xmath201 . let us represent @xmath202 in this basis @xmath203 as @xmath204 as @xmath11 are orthogonal unit vectors satisfying @xmath205 and @xmath206 , we have @xmath207 the condition of the proposition is represented in this basis as @xmath208 and @xmath209 . hence we have @xmath210 these conditions are rewritten in term of @xmath159 as follows : we have @xmath211 where @xmath212 it is easy to check that eq . ( [ alph ] ) is equivalent to @xmath213 and we conclude the basis @xmath159 is in @xmath158 . as @xmath159 was defined as a basis that diagonalizes @xmath195 , @xmath163 has to be of the form given by eq . ( [ mform ] ) , i.e. , the e - measurement have to be of the form of eq . ( [ mform ] ) . @xmath214 we derive the explicit form of vectors in the set @xmath158 . suppose that @xmath215 is not local unitary equivalent to the extracted form , and @xmath158 is not empty . let us fix one element @xmath159 in @xmath158 . then @xmath216 where @xmath217 is an orthonormal basis of @xmath218 , labeled by a positive real number @xmath60 : @xmath219 here , the phase factor @xmath220 is determined as follows : if the schmidt rank of @xmath93 is @xmath95 , then we have @xmath221 , @xmath222 , and we define a phase factor @xmath223 by @xmath224 . if the schmidt rank of @xmath93 is @xmath101 and @xmath43 is not local unitary equivalent to the extracted form , then we have @xmath225 , and we define the phase @xmath226 . _ proof _ + from the equivalence of eqs . ( [ alph ] ) and ( [ vr ] ) , the matrix forms of the basis states , @xmath198 and @xmath199 , are represented by @xmath227 in the @xmath159 basis . a general orthonormal basis can be written as @xmath228 and @xmath229 in terms of @xmath230 using a parameter @xmath231 . if @xmath232 , then the two basis states have to satisfy the condition @xmath233 and @xmath234 , for @xmath38 . by eq . ( [ mr ] ) , this condition is equivalent to @xmath235 these conditions are also equivalent to the following conditions @xmath236 if @xmath237 or @xmath238 we have @xmath239 , hence we obtain @xmath240 where @xmath241 . however , it is easy to see @xmath242 for @xmath243 . therefore , if ( [ net1 ] ) or ( [ net2 ] ) holds , we can parameterize @xmath158 with a positive scalar @xmath244 . it is also easy to see that if the schmidt rank of @xmath93 is @xmath95 , ( [ net1 ] ) holds . furthermore , in this case , @xmath223 is given by @xmath224 . on the other hand , if the schmidt rank of @xmath93 is @xmath101 and it is not local unitary equivalent to the extracted form , then eq . ( [ net2 ] ) holds . in this case , the phase is given by @xmath245 . hence , in both cases , @xmath158 is @xmath60-parameterized . @xmath214 by a direct calculation , complex conjugation of the vectors @xmath246 in @xmath247 are given by @xmath248 in the @xmath249 basis . suppose that both of alice and bob can perform an e - measurement on a pair of basis states @xmath43 . this assumption excludes the possibility that @xmath43 is local unitary equivalent to the extracted form at alice or bob from the beginning . since alice can perform an e - measurement , @xmath158 is non - empty and its elements are @xmath60-parameterized @xmath250 as we have shown in _ step 3_. the vectors @xmath43 can be decomposed with respect to the elements of @xmath158 @xmath251 so that @xmath252 and @xmath253 . furthermore , every e - measurement by alice on @xmath43 is of the form @xmath254 where @xmath165 is a unitary operator , @xmath255 and @xmath256 . in _ step 4 _ , we show that under the assumption that both alice and bob can perform an e - measurement on @xmath43 , the vectors @xmath257 and @xmath258 satisfy @xmath259 and @xmath260 for all @xmath172 . to prove this , note that if bob can perform an e - measurement , there exists a basis set @xmath261 satisfying @xmath262 for @xmath38 , from _ step 2_. note that @xmath202 can be represented as @xmath263 because of eq . ( [ tb ] ) . let us represent the @xmath60-parameterized vectors in the @xmath264 basis : @xmath265 then @xmath266 are represented in the @xmath267 basis as @xmath268 where @xmath269 represents irrelevant elements for our evaluation . hence eq . ( [ ao ] ) implies @xmath270 it is easy to derive the relations @xmath271 and @xmath272 which imply @xmath273 for all @xmath274 . representing @xmath202 in the @xmath159 basis , from the representation in eq . ( [ xi ] ) , we see @xmath275 for all @xmath60 . this implies @xmath276 . but as we have a freedom about the choice of the fixed basis @xmath159 ( we could take @xmath277 ) , we obtain @xmath278 for all @xmath274 . in the matrix representation in eq . ( [ xi ] ) , we have @xmath279 let us consider the following situation : alice performs some e - measurement @xmath83 on @xmath43 , and for all the results of the measurements @xmath280 , bob can sequently perform another e - measurement . can this situation occur ? in this _ step 5 _ , we show this is not possible . by symmetry , the situation that interchanging alice s and bob s roles is also impossible . if this situation occurs , the pair @xmath43 is not local unitary equivalent to the extracted form at bob . therefore , @xmath158 should be @xmath60-parameterized and @xmath43 is decomposed as in eq . ( [ tb ] ) . each @xmath163 is of the form of eq . ( [ mf ] ) . as it is an e - measurement , there exists @xmath10 such that @xmath281 . after alice s e - measurement , the two basis states are transformed as @xmath282 and @xmath283 . note that alice still can perform an e - measurement on this pair , ( with @xmath284 , @xmath285 , for example . ) now assume that bob can perform an e - measurement on @xmath286 , for all @xmath10 . then , we have @xmath287 , and the pair of basis states @xmath286 have to satisfy @xmath288 by _ step 4_. as @xmath83 is an e - measurement , there exists @xmath10 such that @xmath289 . in the following , we see that if there exists @xmath10 such that @xmath281 , eqs . ( [ af ] ) and ( [ ef ] ) have at most one solution @xmath290 in @xmath291 . note that eq . ( [ ef ] ) is equivalent to eq . ( [ to ] ) . if @xmath292 , and @xmath293 are satisfied , we have @xmath294 for all @xmath60 from eq . ( [ te ] ) . from eq . ( [ af ] ) , this implies @xmath295 for all @xmath10 , which contradicts our assumption . therefore , we have @xmath296 or @xmath297 . if @xmath298 and @xmath299 are satisfied , then eq . ( [ ef ] ) has the only solution @xmath300 . if @xmath301 , we have @xmath302 . this equation has one negative solution and one positive solution . hence in any case , eqs . ( [ af ] ) and ( [ ef ] ) have at most one solution in@xmath303 . if eqs . ( [ af ] ) and ( [ ef ] ) have no solution , we can conclude that the situation is impossible . let us consider the case that there exists a unique solution @xmath304 . as each @xmath10 has to satisfy eqs . ( [ af ] ) and ( [ ef ] ) , we have @xmath305 for all @xmath10 . then by eq . ( [ af ] ) , we obtain @xmath306 for all @xmath10 , i.e. , the ratio of @xmath307 and @xmath308 is independent of @xmath10 . furthermore , as there exists @xmath10 such that @xmath281 , the ratio @xmath309 is not @xmath101 . therefore , we have @xmath310 as for @xmath311 , this is not equal to @xmath312 . this contradicts the normalization condition of the measurement operator @xmath313 therefore , the situation we have considered can not occur . suppose that remote extraction to bob s qubit is possible somehow . if @xmath43 is not local unitary equivalent to the extracted form , it should be possible to carry out the first e - measurement , for either alice or bob . if alice can carry it out , @xmath158 is not empty . on the other hand , if alice can not perform the first e - measurement , bob have to be able to do it . however , it is impossible from the following reason : recall that the final pair eq . ( [ eq ] ) is a pair of the basis states that alice can carry out an e - measurement . as we consider only finite rounds of locc , this means at some point of locc , bob performs an e - measurement and after any result of bob s e - measurement , alice should be able to perform an e - measurement i.e. , the situation considered in _ step 5 _ should occur . however , we have shown that it is impossible in _ step 5_. therefore , the extraction to bob s qubit is only available for the case that alice can perform the first e - measurement , i.e. , the case that @xmath158 is not empty . furthermore , as we have seen above , bob can carry out only c - measurements in the extraction procedure . in this _ step 7 _ , we represent the conditions of eq . ( [ bunk ] ) for @xmath158 in the schmidt form given by eqs . ( [ cond1 ] ) and ( [ cond2 ] ) . suppose that eq . ( [ bunk ] ) is satisfied and let @xmath314 be the schmidt decompositions of @xmath11 , respectively . then we have @xmath315 by the relations @xmath316 and @xmath317 , @xmath318 , the vectors @xmath319 are represented in the basis parallel to @xmath320 as @xmath321 for @xmath322 and @xmath323 . therefore , we have @xmath324 from these relations , we derive that @xmath325 , @xmath326 is parallel to @xmath327 , and @xmath328 is parallel to @xmath329 . by modifying the phase of @xmath330 and @xmath331 appropriately , we can take @xmath332 and @xmath333 . finally , we present the explicit form of locc . by section [ ext - sufficiency ] , the basis @xmath159 represented by @xmath334 and @xmath335 is an element of @xmath158 . therefore , if @xmath43 is not local unitary equivalent to the extracted form , all the elements in @xmath158 are @xmath60-parameterized by _ step 3 _ , with respect to @xmath336 , which is equal to the form of eq . ( [ labelt ] ) in the theorem . by _ step 5 _ , bob can carry out only a c - measurement for remote extraction at bob , therefore , all @xmath337 are scalar multiplications of unitary operators . on the other hand , the form of alice s measurement on the pair of the states @xmath43 have to be @xmath338 , with a unitary operation @xmath165 , @xmath339 and @xmath340 . for this @xmath163 and a nonzero scalar multiplication of a unitary operator @xmath341 , we can see @xmath342 if @xmath287 , from the argument in _ step 2_. note that if @xmath343 , the pair @xmath344 is locally unitary equivalent to the extracted form at bob . thus , we obtain the form of alice s measurement given in the theorem [ one ] , inductively . let @xmath345 , @xmath346 . we consider the basis states in the matrix representation : @xmath347 . first we show the following lemma . [ odis ] suppose that there exists a matrix @xmath348 and vectors @xmath349 , @xmath350 satisfying the following property : for all @xmath351 , there exists @xmath352 such that @xmath353 then if the schmidt rank of @xmath93 ( resp . @xmath94 ) is @xmath95 , @xmath354 where @xmath355 is an eigenvector of a matrix @xmath356 ( resp . @xmath357 ) . if the schmidt rank of both of @xmath93 and @xmath94 are @xmath101 and the basis states are represented by @xmath358 then one of the followings occurs : 1 . @xmath359 2 . @xmath360 3 . @xmath320 and @xmath361 are parallel to each other and @xmath362 for an arbitrary vector @xmath355 . _ proof _ + if the schmidt rank of @xmath363 is @xmath95 , and @xmath364 , @xmath365 has to be of rank @xmath101 . we represent it as @xmath362 by introducing a vector @xmath355 . then we have @xmath366 hence we obtain @xmath367 as the schmidt rank of @xmath93 is @xmath95 , the rank of @xmath368 is @xmath95 , i.e. , @xmath368 is invertible . therefore , @xmath369 , and @xmath370 then we obtain @xmath371 as @xmath368 is invertible , we have @xmath372 hence @xmath355 is an eigenvector of @xmath356 . the result is also unchanged for the case that schmidt rank of @xmath373 is @xmath101 . on the other hand , suppose that the schmidt rank of both basis states @xmath93 and @xmath94 are @xmath101 . we write @xmath374 . by the assumption , @xmath375 and @xmath376 have to be parallel . from this fact , the statement holds . @xmath214 + + next we present the following lemma for the cases where the schmidt the schmidt rank of @xmath93 ( resp.@xmath94 ) is @xmath95 . [ matrixform1 ] if the schmidt rank of @xmath93 ( resp.@xmath94 ) is @xmath95 , deterministic remote destruction is possible if and only if the matrix @xmath377 ( resp.@xmath378 ) is diagonalized by some orthonormal basis @xmath99 with eigenvalues @xmath379 . furthermore , the measurement operators for deterministic remote destruction @xmath83 are of the form @xmath380 where @xmath97 is an arbitrary vector and @xmath98 . _ proof : _ note that @xmath381 if @xmath356 is diagonalized by an orthonormal basis @xmath382 with corresponding eigenvalues @xmath383 , then we have @xmath384 for @xmath385 . we set @xmath386 . by defining @xmath387 , we obtain @xmath388 hence we have @xmath389 and remote destruction is possible . conversely , suppose that a measurement represented by @xmath390 of deterministic remote destruction is possible . then , from lemma [ odis ] , we have @xmath391 , where @xmath103 is an eigenvector of @xmath356 . if @xmath356 has only one eigenvector , then it is impossible to obtain @xmath392 . therefore , @xmath356 has two eigenvectors . however , if we can not take them orthogonal to each other , it is again impossible to have @xmath392 . therefore , @xmath356 has two orthogonal eigenvectors , which means that it is diagonalized by the orthonormal basis , and @xmath163 has to be given as in eq.([m2 ] ) . the proof for the case that the schmidt rank of @xmath94 is @xmath95 is identical . @xmath214 now we present a lemma for the cases where both of the basis states @xmath162 have the schmidt rank @xmath101 . if the schmidt rank of both of @xmath93 and @xmath94 are @xmath101 , then deterministic remote destruction is possible if and only if @xmath393 are of the form @xmath394 where @xmath395 are orthogonal . the measurement operators @xmath390 are of the form @xmath102 here , the vector @xmath103 have to be @xmath104 or @xmath105 if @xmath396 are not parallel to each other , while it can be an arbitrary vector if @xmath396 are parallel to each other . _ proof : _ if @xmath162 are of the forms given by eq . ( [ s11 ] ) , then we can take @xmath397 for destruction . conversely , suppose that by performing a measurement represented by @xmath390 , deterministic remote destruction is possible . we represent the basis states as @xmath398 note that either @xmath399 or @xmath400 should hold . by lemma [ odis ] , @xmath163 has to be either of the form _ 1 - 3 _ in the lemma . if @xmath401 , we have @xmath400 and the situation _ 3 _ can not occur . therefore , @xmath365 has to of the form @xmath402 or @xmath403 . however , as we have @xmath401 , we also have @xmath404 and @xmath163s can not satisfy @xmath405 . hence we obtain @xmath399 . again by lemma [ odis ] , if @xmath320 and @xmath361 are not parallel , each @xmath163 has to be of the form @xmath406 , @xmath407 while if they are parallel , then @xmath391 for an arbitrary @xmath103 . @xmath214 now , let us prove theorem [ distract ] for remote destruction . from lemma [ matrixform1 ] , if @xmath408 has schmidt rank @xmath95 , destruction is possible if and only if @xmath409 for some orthogonal basis @xmath410 . this is equivalent to @xmath411 using orthogonality condition @xmath412 , one can easily check that eq . ( [ apple ] ) implies eq . ( [ xzx ] ) . below , we show that eq . ( [ xzx ] ) implies eq . ( [ apple ] ) . + now , using schmidt decomposition , we have @xmath413 representing @xmath414 in this basis as @xmath415 eq . ( [ xzx ] ) can be written as @xmath416 using the conditions @xmath417 we obtain @xmath418 and @xmath419 on the other hand , eq . ( [ xzx ] ) requires @xmath420 combining these relations , we see that @xmath421 have to be of the form @xmath422 substituting these , @xmath199 is given by @xmath423 note that @xmath424 and @xmath425 are orthogonal . hence this decomposition in the matrix form gives the schmidt decomposition of the vector @xmath426 , and we obtain eq . ( [ apple ] ) . in this paper , we present necessary and sufficient conditions for remote extraction and destruction of single - qubit information encoded and spread in two - qubit states . the conditions show that there are ways to spread qubit information into a two - qubit hilbert space in a less `` non - local '' manner , namely , the recovery ( extraction ) of qubit information at one of the qubits and irreversible erasure ( destruction ) of qubit information can be achieved by measurements on only one of the qubits and classical communications , even though the encoded two qubit states are entangled in general . the main part of the paper is devoted to the proof of necessity for remote extraction shown in seven steps . by introducing the notation of extraction measurements ( e - measurements ) , which are local measurements preserving orthogonality of the basis states for encoding and equal probability for measuring each basis state , we evaluate the conditions of the basis states for the existence of e - measurements . we have derived necessary and sufficient conditions for basis states for encoding and the explicit form of measurements . necessary and sufficient conditions for remote extraction indicate the possibility of sharing qubit information between two parties in an asymmetric manner , namely , we can spread qubit information such that it can be remotely extracted to only one of the parties by using locc but not to the other party . since impossibility of locc tasks implies the existence of a kind of non - locality , the possibility of asymmetric qubit information sharing indicates that such non - locality can be asymmetric between two parties . the obtained necessary and sufficient conditions for remote destruction indicate that such asymmetric non - locality for irreversibly destroying qubit information can also be introduced , but the conditions for asymmetric cases are not compatible with those for remote extraction . remote destruction by one of the parties is possible if and only if remote extraction at that party is possible . if we consider alice s system in our model as an environment , it is a similar situation to _ quantum lost and found _ considered by gregoretti and werner in @xcite . in their paper , they have obtained necessary and sufficient conditions for extracting quantum information spread over the system and environment due to the system - environment coupling , by measuring the environment and performing conditional operations on the system . by using quantum information theoretical analysis , they have proven that extraction is possible if and only if when a map @xmath15 for the system state @xmath427 can be represented by the random unitary channel @xmath428 , where @xmath429 is a set of random unitary operators acting on the system qubit , and @xmath430 is probability satisfying @xmath431 . we note that our remote extraction can be regarded as a special case where both the system and environment consist of a single qubit , however we consider the conditions for general locc whereas they consider only one - way ( from the environment to the system ) locc for quantum information extraction . in the case of the single qubit system - environment , a purification of the map @xmath15 gives a corresponding transformation describing the spread of qubit information to the two parties , the system and environment ; @xmath432 for arbitrary @xmath1 and @xmath2 , where @xmath433 is an orthonormal basis of the environment . they are equivalent to our locc extraction conditions for basis states given by eqs . ( [ cond1 ] ) and ( [ cond2 ] ) . therefore , the encoding conditions for remote extraction of qubit information spread into two - qubit states are same for both one - way locc and general locc , in spite of the inapplicability of the lo - popescu theorem to this task . the authors are grateful to d. markham for discussions . this work was supported by the sumitomo foundation , the asahi glass foundation , and the japan society of the promotion of science . y. o. was supported by the coe21 program at graduate school of mathematical sciences , the university of tokyo . m. m. was partially supported by the special coordination funds for promoting science and technology . shor , in _ proc . , 35th annnual symposium of foundation of computer science _ , ieee press ( 1994 ) ; l. grover , in _ proc . 28th anual acm symposium on the theory of compuation _ , 212 , acm press ( 1996 ) . shor , phys . a * 52 * , r2493 ( 1995 ) ; a.r . calderbank and p.w . shor , phys . a * 54 * , 1098 ( 1996 ) ; a.m. steane , phys . 77 * , 793 ( 1996 ) . v. buek and m. hillery , phys . a * 54 * , 1844 ( 1996 ) . bennett et.al . lett . * 70 * , 1895 ( 1993 ) . c. h. bennett , d. p. divincenzo , c. a. fuchs , t. mor , e. rains , p. w. shor , j. a. smolin , and w. k. wootters , phys . a * 59 * , 1070 ( 1999 ) . m. owari and m. hayashi , phys . rev . a * 74 * , 032108 ( 2006 ) h .- k . lo and s. popescu , a * 63 * , 022301 ( 2001 ) . m. gregoretti and r. f. werner , j. mod . opt . * 50 * , 915 ( 2003 ) .

in this paper we set out to study the functions in multivariate complex variables which however always take real values . such functions are frequently encountered in engineering applications arising from signal processing @xcite , electrical engineering , and control theory @xcite . it is interesting to note that such complex functions are usually not studied by conventional complex analysis , since they are typically not even analytic because the cauchy - riemann conditions will never be satisfied unless the function in question is trivial . there has been a surge of research attention to solve optimization models related to such kind of complex functions @xcite . sorber et al . @xcite developed a matlab toolbox for optimization problems in complex variables , where the complex function in question is either _ pre - assumed _ to be always real - valued @xcite , or it is the modulus / norm of a complex function @xcite . an interesting question thus arises : _ can such real - valued complex functions be characterized ? _ indeed there does exist a class of special complex functions that always take real values : the hermitian quadratic form @xmath0 where @xmath1 is a hermitian matrix . in this case , the quadratic structure plays a key role . this motivates us to search for more general complex polynomial functions with the same property . interestingly , such complex polynomials can be completely characterized , as we will present in this paper . as is well - known , polynomials can be represented by tensors . the same question can be asked about complex tensors . in fact , there is a considerable amount of recent research attention on the applications of complex tensor optimization . for instance , hilling and sudberythe @xcite formulated a quantum entanglement problem as a complex multilinear form optimization under the spherical constraint , and zhang and qi @xcite and ni et al . @xcite discussed quantum eigenvalue problems , which arised from the geometric measure of entanglement of a multipartite symmetric pure state in the complex tensor space . examples of complex polynomial optimization include aittomaki and koivunen @xcite who formulated the problem of beam - pattern synthesis in array signal processing as complex quartic polynomial minimization , and aubry et al . @xcite who modeled a radar signal processing problem by complex polynomial optimization . solution methods for complex polynomial optimization can be found in , e.g. , @xcite . as mentioned before , polynomials and tensors are known to be related . in particular in the real domain , homogeneous polynomials ( or forms ) are bijectively related to _ symmetric _ tensors ; i.e. , the components of the tensor is invariant under the permutation of its indices . this important class of tensors generalizes the concept of symmetric matrices . as the role played by symmetric matrices in matrix theory and quadratic optimization , symmetric tensors have a profound role to play in tensor eigenvalue problems and polynomial optimization . a natural question can be asked about complex tensors : _ what is the higher order complex tensor generalization of the hermitian matrix ? _ in this paper , we manage to identify two classes of symmetric complex tensors , both of which include hermitian matrices as a special case when the order of the tensor is two . in recent years , the eigenvalue of tensor has become a topic of intensive research interest . perhaps a first attempt to generalize eigenvalue decomposition of matrices can be traced back to 2000 when de lathauwer et al . @xcite introduced the so - called higher - order eigenvalue decomposition . shortly after that , kofidis and regalia @xcite showed that blind deconvolution can be formulated as a nonlinear eigenproblem . a systematic study of eigenvalues of tensors was pioneered by lim @xcite and qi @xcite independently in 2005 . various applications of tensor eigenvalues and the connections to polynomial optimization problems have been proposed ; cf . @xcite and the references therein . we refer the interested readers to the survey papers @xcite for more details on the spectral theory of tensors and various applications of tensors . computation of tensor eigenvalues is an important source for polynomial optimization @xcite . essentially the problem is to maximize or minimize a homogeneous polynomial under the spherical constraint , which can also be used to test the ( semi)-definiteness of a symmetric tensor . in this paper we are primarily interested in complex polynomials / tensors that arise in the context of optimization . by nature of optimization , we are interested in the complex polynomials that always take real values . however , it is easy to see that if no _ conjugate _ term is involved , then the only class of real - valued complex polynomials is the set of real constant functions . ] . therefore , the conjugate terms are necessary for a complex polynomial to be real - valued . hermitian quadratic forms mentioned earlier belong to this category , which is an active area of research in optimization ; see e.g. @xcite . in the aforementioned papers @xcite on eigenvalues of complex tensors , the associated complex polynomials however are not real - valued . the aim of this paper is different . we target for a systematic study on the nature of symmetricity for higher order complex tensors which will lead to the property that the associated polynomials always take real values . the main contribution of this paper is to give a full characterization for the real - valued conjugate complex polynomials and to identify two classes of symmetric complex tensors , which have already shown potentials in the algorithms design @xcite . we also proposed two new types of tensor eigenvalues / eigenvectors for the new classes of complex tensors . this paper is organized as follows . we start with the preparation of various notations and terminologies in section [ sec : preparation ] . in particular , two types of conjugate complex polynomials are defined and their symmetric tensor representations are discussed . section [ sec : condition ] presents the necessary and sufficient condition for real - valued conjugate complex polynomials , based on which two types of symmetric complex tensors are defined , corresponding to the two types of real - valued conjugate complex polynomials . as an important result in this paper , we then present the definitions and properties of eigenvalues and eigenvectors for two types of symmetric complex tensors in section [ sec : eigenvalue ] . in section [ sec : banach ] , we discuss banach s theorem , which states that the largest eigenvalue ( in the absolute value sense ) of a symmetric real tensor is equal to its largest singular value , and extend it to the two new types of symmetric complex tensors . some application examples are discussed in section [ sec : application ] to show the significance in practice of the theoretical results in this paper . finally , we conclude this paper by summarizing our main findings and outlining possible future work in section [ sec : conclusion ] . throughout this paper we use usual lowercase letters , boldface lowercase letters , capital letters , and calligraphic letters to denote scalars , vectors , matrices , and tensors , respectively . for example , a scalar @xmath2 , a vector @xmath3 , a matrix @xmath4 , and a tensor @xmath5 . we use subscripts to denote their components , e.g. @xmath6 being the @xmath7-th entry of a vector @xmath3 , @xmath8 being the @xmath9-th entry of a matrix @xmath4 and @xmath10 being the @xmath11-th entry of a third order tensor @xmath5 . as usual , the field of real numbers and the field of complex numbers are denoted by @xmath12 and @xmath13 , respectively . for any complex number @xmath14 with @xmath15 , its real part and imaginary part are denoted by @xmath16 and @xmath17 , respectively . its modulus is denoted by @xmath18 , where @xmath19 denotes the conjugate of @xmath20 . for any vector @xmath21 , we let @xmath22 be the transpose of its conjugate , and we define it analogously for matrices . throughout this paper we uniformly use the 2-norm for vectors , matrices and tensors in general , which is the usual euclidean norm . for example , the norm of a vector @xmath21 is defined as @xmath23 , and the norm of a @xmath24-th order tensor @xmath25 is defined as @xmath26 a multivariate complex polynomial @xmath27 is a polynomial function of variable @xmath21 whose coefficients are complex , e.g. @xmath28 . a multivariate _ conjugate _ complex polynomial ( sometimes abbreviated by conjugate polynomial in this paper ) @xmath29 is a polynomial function of variables @xmath30 , which is differentiated by the subscript @xmath31 , standing for ` conjugate ' , e.g. @xmath32 . in particular , a general @xmath33-dimensional @xmath24-th degree conjugate complex polynomial can be explicitly written as summation of monomials @xmath34 in the above notation for a monomial @xmath35 , the indices of the coefficient @xmath36 are always partitioned by a ` , ' to separate that of conjugate variables and that of regular variables . in particular , the coefficient of a monomial that only has conjugate variables such as @xmath37 will be written as @xmath38 . in this definition , it is obvious that complex polynomials are a subclass of conjugate complex polynomials . remark that a pure complex polynomial can never only take real values unless it is a constant . this observation follows trivially from the basic theorem of algebra . given a @xmath24-th order complex tensor @xmath25 , its associated multilinear form is defined as @xmath39 where @xmath40 for @xmath41 . a complex tensor @xmath42 is called _ symmetric _ if @xmath43 and every component @xmath44 are invariant under all permutations of the indices @xmath45 . we remark that conjugation is not involved here when speaking of symmetricity for complex tensors . closely related to a symmetric tensor @xmath46 is a general @xmath24-th degree complex homogeneous polynomial function @xmath27 ( or complex form ) of variable @xmath47 , i.e. , @xmath48 in fact , symmetric tensors ( either in the real domain or in the complex domain ) are bijectively related to homogeneous polynomials ; see @xcite . in particular , for any @xmath33-dimensional @xmath24-th degree complex form @xmath49 there is a uniquely defined @xmath33-dimensional @xmath24-th order symmetric complex tensor @xmath46 with @xmath50 satisfying , where @xmath51 is the set of all distinct permutations of the indices @xmath52 . on the other hand , in light of formula , a complex form @xmath27 is easily obtained from the symmetric multilinear form @xmath53 by letting @xmath54 . to discuss higher order conjugate complex forms and complex tensors , let us start with the well - established properties of hermitian matrices . let @xmath55 with @xmath56 , which is not symmetric in the usual sense because @xmath57 in general . the following conjugate quadratic form @xmath58 always takes real values for any @xmath21 . in particular , we notice that each monomial in the above form is the product of one ` conjugate ' variable @xmath59 and one usual ( non - conjugate ) variable @xmath60 . to extend the above form to higher degrees , let us consider the following special class of conjugate polynomials : [ def : sform ] a symmetric conjugate form of the variable @xmath21 is defined as @xmath61 essentially , @xmath62 is the summation of all the possible @xmath63-th degree monomials that consist of @xmath24 conjugate variables and @xmath24 usual variables . here the subscript ` @xmath64 ' stands for ` symmetric ' . the following example is a special case of . given a @xmath24-th degree complex form @xmath65 , the function @xmath66 is a @xmath63-th degree symmetric conjugate form . notice that @xmath67 is actually a real - valued conjugate polynomial . later in section [ sec : condition ] we shall show that a symmetric conjugate form @xmath62 in always takes real values if and only if the coefficients of any pair of conjugate monomials @xmath68 and @xmath69 are conjugate to each other , i.e. , @xmath70 as any complex form uniquely defines a symmetric complex tensor and vice versa , we observe a class of tensors representable for symmetric conjugate forms . [ def : partial - symmetric ] an even order tensor @xmath71 is called partial - symmetric if for every @xmath72 @xmath73 we remark that the so - called partial - symmetricity was studied earlier in algebraic geometry by carlini and chipalkatti @xcite , and was also studied in polynomial optimization @xcite in the framework of mixed polynomial forms , i.e. , for any fixed first @xmath24 indices of the tensor , it is symmetric with respect to its last @xmath24 indices , and vise versa . it is clear that partial - symmetricity is weaker than the usual symmetricity for tensors . let us formally define the bijection @xmath74 ( taking the first initial of symmetric conjugate forms ) between symmetric conjugate forms and partial - symmetric complex tensors , as follows : + ( i ) @xmath75 : given a partial - symmetric tensor @xmath71 with its associated multilinear form @xmath76 , the symmetric conjugate form is defined as @xmath77 ( ii ) @xmath78 : given a symmetric conjugate form @xmath79 , the components of the partial - symmetric tensor @xmath71 are defined by @xmath80 for all @xmath81 and @xmath82 . given a bivariate fourth degree symmetric conjugate form @xmath83 , the corresponding partial - symmetric tensor @xmath84 satisfies that @xmath85 and other entries are zeros . conversely , @xmath62 can be obtained from @xmath86 . according to the mappings defined previously , the following result readily follows . [ thm : tensors ] the bijection @xmath74 is well - defined , i.e. , any @xmath33-dimensional @xmath63-th order partial - symmetric tensor @xmath71 uniquely defines an @xmath33-dimensional @xmath63-th degree symmetric conjugate form , and vice versa . in , for each monomial the numbers of conjugate variables and the usual variables are always equal . this restriction can be relaxed further . [ def : gform ] a general conjugate form of the variable @xmath21 is defined as @xmath87 essentially , @xmath88 is the summation of all the possible @xmath24-th degree monomials , allowing any number of conjugate variables as well as the usual variables in each monomial . here the subscript ` @xmath89 ' stands for ` general ' . obviously @xmath62 is a special case of @xmath88 , and @xmath88 is a special case of @xmath29 . in section [ sec : condition ] we shall show that a general conjugate form @xmath88 will always take real values for all @xmath3 if and only if the coefficients of each pair of conjugate monomials are conjugate to each other . to this end , below we shall explicitly treat the conjugate variables as new variables : + ( i ) @xmath90 : given a symmetric tensor @xmath91 with its associated multilinear form @xmath53 , the general conjugate form of @xmath21 is defined as @xmath92 ( ii ) @xmath93 : given a general conjugate form @xmath94 of @xmath21 as , the components of the symmetric tensor @xmath91 are defined as follows : for any @xmath95 , sort these @xmath96 s in a nondecreasing order as @xmath97 and let @xmath98 , then @xmath99 given a symmetric second order tensor ( matrix ) @xmath100 , the corresponding general conjugate form is @xmath101 conversely , @xmath102 can obtained component - wisely by . similar to lemma [ thm : tensors ] , the following is easily verified ; we leave its proof to the interested readers . [ thm : tensorg ] the bijection @xmath103 is well - defined , i.e. , any @xmath104-dimensional @xmath24-th order symmetric tensor @xmath91 uniquely defines an @xmath33-dimensional @xmath24-th degree general conjugate form , and vice versa . to conclude this section we remark that a partial - symmetric tensor ( representation for a symmetric conjugate form ) is less restrictive than a symmetric tensor ( representation for a general conjugate form ) , while a symmetric conjugate form is a special case of a general conjugate form . one should note that the dimensions of these two tensor representations are actually different . in this section , we study the two types of conjugate complex forms introduced in section [ sec : preparation ] : symmetric conjugate forms and general conjugate forms . let us first focus on polynomials , and present the following general characterization of real - valued conjugate complex polynomials . [ thm : realvalue ] a conjugate complex polynomial function is real - valued if and only if the coefficients of any pair of its conjugate monomials are conjugate to each other , i.e. , any two monomials @xmath105 and @xmath106 with @xmath2 and @xmath107 being their coefficients satisfying @xmath108 must have that @xmath109 . the above condition actually implies that the coefficient of any self - conjugate monomial must be real . applying theorem [ thm : realvalue ] to the two classes of conjugate forms that we just introduced , the conditions for them to always take real values can now be characterized : [ thm : condition ] a symmetric conjugate form @xmath110is real - valued if and only if @xmath111 a general conjugate form @xmath112is real - valued if and only if @xmath113 let us now prove theorem [ thm : realvalue ] . we first show the ` if ' part of the theorem , which is quite straightforward . to see this , for any pair of conjugate monomials ( including self - conjugate monomial as a special case ) of a conjugate complex polynomial : @xmath105 and @xmath114 with @xmath115 being their coefficients , if @xmath109 , then @xmath116 implying that @xmath117 is real - valued . since all the conjugate monomials of a conjugate complex polynomial can be partitioned by conjugate pairs and self - conjugate monomials , the result follows immediately . to proceed to the ` only if ' part of the theorem , let us first consider an easier case of univariate conjugate polynomials . [ thm : uni - zero ] a univariate conjugate complex polynomial @xmath118 for all @xmath119 if and only if all its coefficients are zeros , i.e. , @xmath120 for all @xmath121 and @xmath122 . let @xmath123 with @xmath124 and @xmath125 , and the identity can be rewritten as @xmath126 for any fixed @xmath127 , the function can be viewed as a polynomial with respect to @xmath128 . therefore the coefficient of the highest degree monomial @xmath129 must be zero , i.e. , @xmath130 consequently we have for any @xmath125 , @xmath131 the first and second parts of can be respectively simplified as @xmath132 and @xmath133 by the orthogonality of the trigonometric functions , the above further leads to @xmath134 similarly , implies @xmath135 combining the above two sets of identities yields @xmath136 the degree of the function in ( in terms of @xmath128 ) is then reduced by 1 . the desired result follows obviously . let us now extend lemma [ thm : uni - zero ] to general multivariate conjugate polynomials . [ thm : multi - zero ] an @xmath33-dimensional @xmath24-th degree conjugate complex polynomial @xmath137 for all @xmath21 if and only if all its coefficients are zeros , i.e. , @xmath138 for all @xmath139 , @xmath122 , @xmath140 and @xmath141 . we shall prove the result by induction on the dimension @xmath33 . the case @xmath142 is already shown in lemma [ thm : uni - zero ] . suppose the claim holds for all positive integers no more than @xmath143 . then for the dimension @xmath33 , the conjugate polynomial @xmath29 can be rewritten according to the degrees of @xmath144 and @xmath145 as @xmath146 for any given @xmath147 , taking @xmath148 as a univariate conjugate polynomial of @xmath145 , by lemma [ thm : uni - zero ] we have @xmath149 for any given @xmath150 , as @xmath151 is a conjugate polynomial of dimension at most @xmath143 , by the induction hypothesis all the coefficients of @xmath152 are zeros . observing that all the coefficients of @xmath148 are distributed in the coefficients of @xmath152 for all @xmath150 , the result is proven for dimension @xmath33 . with lemma [ thm : multi - zero ] at hand , we can finally complete the ` only if ' part of theorem [ thm : realvalue ] . suppose a conjugate polynomial @xmath27 is real - valued for all @xmath21 . clearly we have @xmath153 for all @xmath21 , i.e. , @xmath154 by lemma [ thm : multi - zero ] it follows that @xmath155 for all @xmath139 , @xmath122 , @xmath140 and @xmath141 , proving the ` only if ' part of theorem [ thm : realvalue ] . with theorem [ thm : realvalue ] , in particular corollary [ thm : condition ] , we are in a position to characterize the tensor representations for real - valued conjugate forms . before concluding this subsection , let us present an alternative representation of real - valued symmetric conjugate forms , as a consequence of corollary [ thm : condition ] . [ thm : polyrealvalued ] a symmetric conjugate form @xmath62 is real - valued if and only if @xmath156 where @xmath157 is a complex form and @xmath158 for all @xmath159 . the ` if ' part is trivial . next we prove the ` only if ' part of the proposition . if @xmath62 is real - valued , by corollary [ thm : condition ] we have . then for any @xmath160 and @xmath161 , the sum of the conjugate pair satisfies @xmath162 summing up all such pairs ( taking half if it is a self - conjugate pair ) , the conclusion follows . similarly we have the following result for general conjugate forms . a general conjugate form @xmath88 is real - valued if and only if @xmath163 where @xmath157 is a complex polynomial and @xmath158 for all @xmath159 . as any symmetric conjugate form uniquely defines a partial - symmetric tensor ( lemma [ thm : tensors ] ) , it is interesting to see more structured tensor representations for real - valued symmetric conjugate forms . [ thm : cps ] an even order tensor @xmath71 is called conjugate partial - symmetric if + ( i ) @xmath164 for all @xmath165 and @xmath82 , and + ( ii ) @xmath166 + hold for all @xmath167 and @xmath168 . we remark that when @xmath169 , a conjugate partial - symmetric tensor is simply a hermitian matrix . for a general even degree , the square matrix flattening of a conjugate partial - symmetric tensor ; i.e. , flattening a tensor in @xmath170 to a matrix in @xmath171 by grouping the tensor s first @xmath24 modes into the rows of the matrix and its last @xmath24 modes into the columns of the matrix , is actually a hermitian matrix . the conjugate partial - symmetric tensors and the real - valued symmetric conjugate forms are connected as follows . [ thm : conju - mapping ] any @xmath33-dimensional @xmath63-th order conjugate partial - symmetric tensor @xmath71 uniquely defines ( under @xmath74 ) an @xmath33-dimensional @xmath63-th degree real - valued symmetric conjugate form , and vice versa ( under @xmath172 ) . for any conjugate partial - symmetric tensor @xmath5 , @xmath173 satisfies @xmath174 implying that @xmath79 is real - valued . on the other hand , for any real - valued symmetric conjugate form @xmath62 in , it follows from corollary [ thm : condition ] that @xmath175 holds for all the possible @xmath176 . by , its tensor representation @xmath177 with @xmath178 satisfies the 2nd condition in definition [ thm : cps ] , proving the conjugate partial - symmetricity of @xmath5 . below is a useful property for conjugate partial - symmetric tensors , in the same vein as proposition [ thm : polyrealvalued ] for the real - valued symmetric conjugate forms . [ thm : conju - decomp ] an even order tensor @xmath71 is conjugate partial - symmetric if and only if @xmath179 where @xmath180 is symmetric and @xmath158 for all @xmath159 . according to definition [ thm : cps ] , it is straightforward to verify that @xmath181 is conjugate partial - symmetric , proving the ` if ' part of the proposition . let us now prove the ` only if ' part . by proposition [ thm : conju - mapping ] , @xmath182 is a real - valued symmetric conjugate form . further by proposition [ thm : polyrealvalued ] , @xmath182 can be written as @xmath183 where @xmath157 is a complex form and @xmath158 for @xmath184 . let @xmath180 be the symmetric complex tensor associated with the complex form @xmath157 for @xmath184 ; i.e. , @xmath185 we have @xmath186 thus , the symmetric conjugate form @xmath187 satisfies @xmath188 i.e. , @xmath189 . as @xmath74 is bijective , we have @xmath190 . similar as for real - valued symmetric conjugate forms , we have the following tensor representations for real - valued general conjugate forms . [ thm : css ] an even dimensional tensor @xmath91 is called conjugate super - symmetric if + ( i ) @xmath5 is symmetric , and + ( ii ) @xmath191 holds for all @xmath192 with @xmath193 for @xmath41 . we remark that the conjugate super - symmetricity is actually _ stronger _ than the ordinary symmetricity for complex tensors since a second condition in definition [ thm : css ] is required . actually , this condition is to ensure that the general conjugate form @xmath194 is real - valued . this is because if @xmath193 holds for @xmath41 , then the monomial with coefficient @xmath195 and the monomial with coefficient @xmath196 in the above form are actually a conjugate pair by noticing that the position of a conjugate variable @xmath59 and that of a usual variable @xmath6 in the vector @xmath197 differs exactly by @xmath33 for every @xmath7 . under the mapping @xmath103 defined in section [ sec : gform ] , it is straightforward to verify the following tensor representations for real - valued general conjugate forms . [ thm : gform - rv ] any @xmath104-dimensional @xmath24-th order conjugate super - symmetric tensor @xmath91 uniquely defines ( under @xmath103 ) an @xmath33-dimensional @xmath24-th degree real - valued general conjugate form , and vice versa ( under @xmath198 ) . as mentioned earlier , lim @xcite and qi @xcite independently proposed to systematically study the eigenvalues and eigenvectors for real tensors . subsequently , the topic has attracted much attention due to the potential applications in magnetic resonance imaging , polynomial optimization theory , quantum physics , statistical data analysis , higher order markov chains , and so on . after that , this study was also extended to complex tensors @xcite without considering the conjugate variables . zhang and qi in @xcite proposed the so - called q - eigenvalues of complex tensors . [ thm : qeigen ] a scalar @xmath199 is called a q - eigenvalue of a symmetric complex tensor @xmath200 , if there exists a vector @xmath3 called q - eigenvector , such that @xmath201 throughout this paper , the notation ` @xmath202 ' stands for a position left for a vector entry . in definition [ eq : qeigen ] , as the corresponding complex tensor does not have conjugate - type symmetricity , the q - eigenvalue does not specialize to the classical eigenvalues of hermitian matrices . in particular , @xmath203 is required in the system . later on , ni et al . @xcite defined the notion of unitary symmetric eigenvalue ( us - eigenvalue ) and demonstrated a relation with the geometric measure of quantum entanglement . [ thm : useigen ] a scalar @xmath199 is called a us - eigenvalue of a symmetric complex tensor @xmath200 , if there exists a vector @xmath3 called us - eigenvector , such that @xmath204 in fact , the q - eigenvalue and the us - eigenvalue are essentially the same . @xmath205 is a pair of q - eigenvalue and q - eigenvector if and only if @xmath206 is a pair of us - eigenvalue and us - eigenvector . first , definition [ thm : useigen ] implies that a us - eigenvalue is always real . to see this , pre - multiplying @xmath207 to the first equation of gives @xmath208 and pre - multiplying @xmath209 to the second equation of yields @xmath210 therefore @xmath211 and so @xmath203 . this actually implies that the first and second equations of are the same by applying the conjugation to the second one . thus , is equivalent to @xmath212 the claimed equivalence is obvious by comparing the above system with . in terms of eigenvalues , definitions [ thm : qeigen ] and [ thm : useigen ] are the same . now with all the new notions introduced in the previous sections in particular the bijection between conjugate partial - symmetric tensors and real - valued symmetric conjugate forms , and the bijection between conjugate super - symmetric tensors and real - valued general conjugate forms we are able to present new definitions and properties of eigenvalues for complex tensors , which are naturally related to that of hermitian matrices . let us first introduce two types of eigenvalues for conjugate partial - symmetric tensors and conjugate super - symmetric tensors . [ thm : ceigen ] @xmath213 is called a c - eigenvalue of a conjugate partial - symmetric tensor @xmath5 , if there exists a vector @xmath47 called c - eigenvector , such that @xmath214 [ thm : geigen ] @xmath213 is called a g - eigenvalue of a conjugate super - symmetric tensor @xmath5 , if there exists a vector @xmath47 called g - eigenvector , such that @xmath215 in fact , the two types of eigenvalues defined above are always real , although they are defined in the complex domain . this property generalizes the well - known property of hermitian matrices . in particular , definition [ thm : ceigen ] includes eigenvalues of hermitian matrices as a special case when @xmath169 . [ thm : eigen - real ] every @xmath31-eigenvalue of a conjugate partial - symmetric tensor is always real ; so is every @xmath89-eigenvalue of a conjugate super - symmetric tensor . suppose @xmath216 is a c - eigenvalue and c - eigenvector pair of a conjugate partial - symmetric tensor @xmath5 . multiplying @xmath217 on both sides of the first equation in , we get @xmath218 as @xmath5 is conjugate partial - symmetric , the left hand side of the above equation is real - valued , and so is @xmath199 . next , suppose @xmath216 is a g - eigenvalue and g - eigenvector pair of a conjugate super - symmetric tensor @xmath5 . multiplying @xmath219 on both sides of the first equation in yields @xmath220 as @xmath5 is conjugate super - symmetric , the left hand side of the above equation is real - valued , and so is @xmath199 . as a consequence of proposition [ thm : eigen - real ] , one can similarly define the c - eigenvalue @xmath203 and its corresponding c - eigenvector @xmath21 for a conjugate partial - symmetric tensor @xmath5 equivalently as follows . [ thm : ceigen2 ] @xmath213 is a c - eigenvalue of a conjugate partial - symmetric tensor @xmath5 , if and only if there exists a vector @xmath21 , such that @xmath221 one important property of the z - eigenvalues for real symmetric tensors is that they can be fully characterized by the kkt solutions of a certain optimization problem @xcite . at a first glance , this property may not hold for c - eigenvalues and g - eigenvalues since the real - valued complex functions are not analytic . therefore , direct extension of the kkt condition of an optimization problem with such objective function may not be valid . however , this class of functions is indeed analytic if we treat the complex variables and their conjugates as a whole due to the so - called wirtinger calculus @xcite developed in the early 20th century . in the optimization context , without noticing the wirtinger calculus , brandwood @xcite first proposed the notion of complex gradient . in particular , the gradient of a real - valued complex function can be taken as @xmath222 . interested readers are referred to @xcite for more discussions on the wirtinger calculus in optimization with complex variables . with the help of wirtinger calculus , we are able to characterize c - eigenvalues and c - eigenvectors in terms of the kkt solutions . therefore many optimization techniques can be applied to find the c - eigenvalues / eigenvectors for a conjugate partial - symmetric tensor . [ prop : c - eigen ] @xmath21 is a c - eigenvector associated with a c - eigenvalue @xmath203 for a conjugate partial - symmetric tensor @xmath5 if and only if @xmath3 is a kkt point of the optimization problem @xmath223 with lagrange multiplier being @xmath224 and the corresponding objective value being @xmath199 . by the multilinearity of @xmath5 , the gradient on @xmath3 of the real - valued symmetric conjugate form associated with @xmath5 is given by @xmath225 where the equality is due to the partial - symmetry of @xmath5 . denote @xmath226 to be the lagrange multiplier associated with the constraint @xmath227 . the kkt condition gives rise to the equations @xmath228 the conclusion follows immediately by comparing the above with and . similarly , we have the following characterization . [ prop : g - eigen ] @xmath21 is a g - eigenvector associated with a g - eigenvalue @xmath203 for a conjugate super - symmetric tensor @xmath5 if and only if @xmath3 is a kkt point of the optimization problem @xmath229 with lagrange multiplier being @xmath224 and the corresponding objective value being @xmath199 . although the definitions of the c - eigenvalue , the g - eigenvalue , and the previously defined q - eigenvalue and the us - eigenvalue involve different tensor spaces , they are indeed closely related . our main result in this section essentially states that the q - eigenvalue and the us - eigenvalue are special cases of the c - eigenvalue , and the c - eigenvalue is a special case of the g - eigenvalue . [ thm : q - c - eigen ] denote @xmath230 to be a complex tensor and define @xmath231 . it holds that + ( i ) @xmath200 is symmetric if and only if @xmath5 is conjugate partial - symmetric ; + ( ii ) if @xmath200 is symmetric , then all the c - eigenvalues of @xmath5 are nonnegative ; + ( iii ) if @xmath200 is symmetric , then @xmath232 is a c - eigenvalue of @xmath5 if and only if @xmath199 is a q - eigenvalue ( or a us - eigenvalue ) of @xmath200 . \(i ) this equivalence can be easily verified by the definition of conjugate partial - symmetricity ( definition [ thm : cps ] ) . \(ii ) let @xmath21 be a c - eigenvector associated with a c - eigenvalue @xmath203 of @xmath5 . by multiplying @xmath3 on both sides of the first equation in , we obtain @xmath233 \(iii ) since the q - eigenvalue is the same as the us - eigenvalue , we only prove the former case . suppose @xmath21 is a q - eigenvector associated with a q - eigenvalue @xmath203 of @xmath200 . by we have @xmath227 and @xmath234 , and so @xmath235 by the similar derivation in the proof of ( ii ) , we get @xmath236 implying that @xmath232 is a c - eigenvalue of @xmath5 . on the other hand , suppose @xmath21 is a c - eigenvector associated with a nonnegative c - eigenvalue @xmath232 of @xmath5 . then by we have @xmath227 and @xmath237 where the first equality is due to the symmetricity of @xmath200 . this leads to @xmath238 . let @xmath239 with some fixed @xmath125 , and further define @xmath240 . we then get @xmath241 now we are able to verify that @xmath242 is a q - eigenvector associated with q - eigenvalue @xmath199 of @xmath200 . observing @xmath243 , and by , @xmath244 we finally get @xmath245 as we saw in section [ sec : preparation ] , by definition , a symmetric conjugate form is a special general conjugate form . hence in terms of their tensor representations , a conjugate partial - symmetric tensor is a special case of conjugate super - symmetric tensor , although they live in different tensor spaces . to study the relationship between the c - eigenvalue and the g - eigenvalue , let us embed a conjugate partial - symmetric tensor @xmath71 to the space of @xmath246 . the conjugate super - symmetric tensor @xmath247 corresponding to @xmath5 is then defined by @xmath248 for example when @xmath169 , a conjugate partial - symmetric tensor is simply a hermitian matrix @xmath55 . then its embedded conjugate super - symmetric tensor is @xmath249 , and clearly we have @xmath250 in general it is straightforward to verify that @xmath251 based on this , we are led to the following relationship between the c - eigenvalue and the g - eigenvalue . [ thm : c - g - eigen ] if @xmath252 is a conjugate super - symmetric tensor induced by a conjugate partial - symmetric tensor @xmath253 according to , then @xmath199 is a c - eigenvalue of @xmath5 if and only if @xmath254 is a g - eigenvalue of @xmath255 . first , by taking the gradient @xmath256 on both sides of , we have that @xmath257 next , according to definition [ thm : ceigen ] and proposition [ thm : ceigen2 ] , @xmath199 is a c - eigenvalue of @xmath5 if and only if there exists a vector @xmath21 such that @xmath258 finally , according to definition [ thm : geigen ] , @xmath254 is a g - eigenvalue of @xmath255 if and only if there exists a vector @xmath21 such that @xmath259 the conclusion follows immediately by combining the above three facts . a classical result originally due to banach @xcite states that if @xmath260 is a continuous symmetric @xmath24-linear form , then @xmath261 in the space of real tensors where @xmath262 and @xmath263 is a multilinear form defined by a real symmetric tensor @xmath264 , states that the largest singular value @xcite of @xmath263 is equal to the largest eigenvalue @xcite ( in the absolute value sense ) of @xmath263 , i.e. , @xmath265 alternatively , is essentially equivalent to the fact that the best rank - one approximation of a real symmetric tensor can be obtained at a symmetric rank - one tensor @xcite . a recent development on this topic for special classes of real symmetric tensors can be found in @xcite . in this section , we shall extend banach s theorem to symmetric conjugate forms ( the conjugate partial - symmetric tensors ) and general conjugate forms ( the conjugate super - symmetric tensors ) . let us start with conjugate super - symmetric tensors , which are a generalization of conjugate partial - symmetric tensors . a key observation leading to the equivalence ( theorem [ thm : equal - con ] ) is the following result . [ lemma : super - sym ] for a given real tensor @xmath266 , if @xmath267 for every @xmath268 and every permutation @xmath269 of @xmath270 , then @xmath5 is symmetric . our first result in this section extends to any conjugate super - symmetric tensors in the complex domain . [ thm : equal - con ] for any conjugate super - symmetric tensor @xmath271 , we have @xmath272 let @xmath273 for @xmath274 . we observe that @xmath275 is also a multilinear form with respect to @xmath276 . as a result , we are able to find a real tensor @xmath277 such that @xmath278 as @xmath255 is conjugate super - symmetric , for any @xmath279 and any permutation @xmath269 of @xmath280 , one has @xmath281 by lemma [ lemma : super - sym ] we have that the real tensor @xmath5 is symmetric . finally , noticing that @xmath282 for @xmath41 , the conclusion follows immediately by applying to @xmath5 and then using the equality . for extending banach s theorem to a conjugate partial - symmetric tensor @xmath71 , one could hope to proceed as follows . since it is a special case of the conjugate super - symmetric tensor , one can embed @xmath5 into a conjugate super - symmetric tensor @xmath247 using . then , by applying theorem [ thm : equal - con ] to @xmath255 and rewriting the real part of its associated multilinear form @xmath283 in terms of @xmath5 , we may have an equivalent expression as . however , this expression is not succinct . taking the case @xmath284 ( degree 4 ) for example , it is straightforward to verify from that @xmath285 and this would lead to @xmath286 instead , one would hope to get @xmath287 however , this does not hold in general . the main reason is that @xmath288 which is easily observed since its left hand side is invariant under the permutation of @xmath289 while its right hand side is not . in particular , only holds for @xmath169 , viz . hermitian matrices ; see the following proposition and example [ ex : quarticfail ] . [ degree2 ] for any hermitian matrix @xmath290 , it holds that @xmath291 furthermore , for any optimal solution @xmath292 of @xmath293 with @xmath294 , @xmath295 is an optimal solution of @xmath296 as well . denote @xmath297 and @xmath298 to be the optimal values of @xmath296 and @xmath293 , respectively . noticing that @xmath299 , by the optimality condition of @xmath293 we have that @xmath300 where @xmath199 and @xmath226 are the lagrangian multipliers of the constraints @xmath227 and @xmath301 , respectively . pre - multiplying the first two equations in with @xmath302 and @xmath303 respectively , and summing them up , lead to @xmath304 similarly , the summation of the third and fourth equations in leads to @xmath305 which further leads to @xmath306 moreover , the summation of the first and fourth equations in yields @xmath307 which further leads to @xmath308 let @xmath309 . clearly @xmath310 is a feasible solution of @xmath296 . by and we have @xmath311 this implies that @xmath312 . notice that @xmath293 is a relaxation of @xmath296 and hence @xmath313 . therefore we conclude that @xmath314 , and an optimal solution @xmath310 of @xmath296 is constructed from an optimal solution @xmath292 of @xmath293 . [ ex : quarticfail ] let @xmath315 with @xmath316 and other entries being zeros . clearly @xmath5 is conjugate partial - symmetric . in this case fails to hold because : + ( i ) @xmath317 for any @xmath318 with @xmath227 . + ( ii ) @xmath319 for @xmath320 and @xmath321 . thus , banach s theorem does not hold in general for conjugate partial - symmetric tensors . a natural question arises : is there any reasonable condition to ensure the identity to hold ? recall from proposition [ thm : conju - decomp ] that every conjugate partial - symmetric tensor can be written as @xmath181 where @xmath180 is symmetric and @xmath158 for all @xmath159 . if further we have all @xmath322 s being nonnegative , then is true . before presenting this result , we first need the following type of banach s theorem for symmetric complex tensors , whose proof can be constructed almost identically to that of theorem [ thm : equal - con ] . [ thm : complexbanach ] if @xmath323 is symmetric , then @xmath324 [ thm : banach2 ] if a conjugate partial - symmetric tensor @xmath71 written as @xmath325 satisfies that @xmath326 for all @xmath159 , then @xmath327 let us first introduce a sandwiched optimization model : @xmath328 denote @xmath329 , @xmath330 and @xmath331 to be the optimal values of @xmath332 , @xmath333 and @xmath334 , respectively . clearly , @xmath334 is a relaxation of @xmath333 , and @xmath333 is a relaxation of @xmath332 , implying that @xmath335 . next , let @xmath336 be an optimal solution of @xmath334 . consider the following problem : @xmath337 whose optimal solution is denoted by @xmath338 . noticing that @xmath339 is symmetric , by proposition [ thm : complexbanach ] , we have @xmath340 for the same reason , we have @xmath341 implying that @xmath342 . finally , let @xmath343 be an optimal solution of @xmath333 . since @xmath326 for all @xmath344 , we have @xmath345 remark that the positivity of @xmath322 s is exploited when invoking the triangle inequality in the first inequality above . this implies that either @xmath346 or @xmath347 attains @xmath330 , proving that @xmath348 . therefore we have @xmath349 . we remark that the condition for @xmath322 s being nonnegative in @xmath5 in theorem [ thm : banach2 ] is actually the condition for the real - valued symmetric conjugate form @xmath182 being a sum of squares ( sos ) of complex polynomials ; see the relation between propositions [ thm : polyrealvalued ] and [ thm : conju - decomp ] . in the field of polynomial optimization , checking whether a polynomial is sos can be done by the feasibility of a semidefinite program . in fact , there is an easy sufficient condition for the condition on @xmath5 in theorem [ thm : banach2 ] to hold : the square matrix flattening of @xmath5 is hermitian positive semidefinite . interested readers are referred to @xcite for details . the theoretical results developed in the previous sections are also useful in practice . in this section , we shall discuss some applications that can be formulated as real - valued complex polynomial optimization models . in particular , these problems can be cast as finding the largest c - eigenvalue of a conjugate partial - symmetric tensor or the largest g - eigenvalue of a conjugate super - symmetric tensor . one challenge of these eigenvalue optimization problems is that the variables are coupled in the complex polynomial objective function . however , the extended banach s theorem in section [ sec : banach ] , specifically theorems [ thm : equal - con ] and [ thm : banach2 ] , guarantee that we can separate the variables without losing the optimality . this enables us to focus on the multilinear ( block ) optimization model @xmath350 for a conjugate super - symmetric tensor @xmath255 , or @xmath351 for certain conjugate partial - symmetric tensor @xmath5 . one great advantage of the above models is that the optimization over one block variable is easy when other blocks are fixed . therefore , some efficient solution methods tailored for these models can be applied , such as the block coordinate decent method @xcite and the maximum block improvement method @xcite . conversely , the extended banach s theorem in section [ sec : banach ] provides an alternative way to solve the symmetric multilinear optimization model by resorting to some approaches tailored for symmetric tensor problems such as the power method @xcite and the semidefinite programming method @xcite . in particular , as the search space can be restricted to symmetric solutions , the latter equivalent model significantly reduces the number of decision variables , which is beneficial to many practical algorithms such as semidefinite programs . the ambiguity function of the waveform is often used to probe the environment in radar system . by controlling both the doppler and the range resolutions of the system , it can regulate the interference power produced by unwanted returns @xcite . to be specific , suppose @xmath352 is the normalized target doppler frequency and @xmath353 is the radar code to be optimized . there are @xmath354 interfering scatterers and the matrix @xmath355 for @xmath356 is defined as @xmath357 the ambiguity function of @xmath358 for the time - lag @xmath356 and the normalized doppler frequency @xmath359 $ ] is given by @xmath360 where @xmath361 and @xmath362 denotes the hadamard product ; interested readers are referred to @xcite for more details of the ambiguity function and radar waveform design . denote by @xmath363 the time - lag of the @xmath364-th scatterer , and let @xmath365 be the normalized doppler frequency of the @xmath364-th scatterer . the latter is usually modeled as a random variable uniformly distributed around a mean frequency @xmath366 with some tolerance @xmath367 , i.e. , @xmath365 is a uniform distribution in @xmath368 $ ] . consequently , the disturbance power at the output of the matched filter is given by @xmath369 + \sigma^2\|{\boldsymbol{s}}\|^2,\ ] ] where @xmath370 is the variance of the circular white noise , and @xmath371 is the echo mean power produced by the @xmath364-th scatterer . to simplify the notation , all the following normalized doppler frequencies are expressed in terms of the difference with respect to @xmath352 . we discretize the normalized doppler interval @xmath372 into @xmath373 bins , denoted by discrete frequencies @xmath374 for @xmath375 . let @xmath376 \neq \varnothing \right\}.\ ] ] then the above statistical expectations can be approximated by the sample means over @xmath377 , i.e. , @xmath378 \approx \frac{1}{|\delta_k|}\sum_{j \in \delta_k } g_{{\boldsymbol{s } } } ( r_k , x_{j } ) , \ ] ] plugging the above expression into , the total disturbance power at the output of the matched filter can be rewritten as @xmath379 where @xmath380 with @xmath381 being the kronecker delta and @xmath382 being an indicator function . to obtain phase - only modulated waveforms , an optimization model to minimize @xmath383 subject to constant modulus constraints was proposed in @xcite : @xmath384 . another modeling strategy is to account for the finite energy transmitted by the radar and assume that @xmath385 . however , this single constraint does not provide any kind of control on the shape of the resulting coded waveform . to circumvent this drawback , one practical approach is to enforce a similarity constraint ( see @xcite for more details ) : @xmath386 where @xmath387 is a known code which shares some nice properties such as a constant modula and a reasonable range resolution . moreover , since any feasible @xmath358 satisfies @xmath388 and @xmath389 therefore , @xmath390 is equivalent to @xmath391 . typically , the similarity constraint is not a hard constraint , it aims to restrict the searching area within some neighborhood of @xmath387 and the size of the neighborhood is controlled by @xmath392 . motivated by the aforementioned equivalence , a similar result can be achieved by penalizing the quantity @xmath393 in the objective and we arrive the following formulation @xmath394 with penalty parameter @xmath128 . notice that the objective function in is a real - valued quartic conjugate complex form . if @xmath395 is the optimal solution and so is @xmath396 , then we can choose one of them to make sure that @xmath397 . the model is obviously finding the smallest c - eigenvalue of a conjugate partial - symmetric tensor , which can also be viewed as finding the smallest g - eigenvalue of a conjugate super - symmetric tensor as mentioned in theorem [ thm : c - g - eigen ] . many modern engineering problems can be cast as multilinear least squares regression given as @xmath398 where @xmath399 is a given nonzero complex tensor . for instance , in quantum entanglement the geometric measure of a given @xmath24-partite pure state @xmath5 is defined by ; see @xcite for details . in fact , can be also categorized as a g - eigenvalue problem for a conjugate super - symmetric tensor . to see this , first it is easy to see that is equivalent to @xmath400 when all @xmath401 s with @xmath402 for @xmath403 are fixed , the optimal @xmath199 satisfies @xmath404 therefore , by multilinearity , is equivalent to @xmath405 let us now consider a relaxation of the above model @xmath406 a key observation is that this relaxation is actually tight . to see this , suppose @xmath407 is an optimal solution of . trivially we have @xmath408 as @xmath5 is nonzero and so @xmath409 for @xmath41 . by noticing @xmath410 we have that @xmath411 and so @xmath412 therefore , the feasible solution @xmath413 of is already optimal to the relaxation model , proving the equivalence between and . finally , to formulate as a g - eigenvalue optimization problem , let us denote @xmath414 and construct a symmetric complex tensor @xmath415 such that @xmath416 thus , can be rewritten as @xmath417 where @xmath418 is a conjugate super - symmetric tensor . the multilinear least square model is shown to be a special case of the g - eigenvalue optimization problem . this paper focuses on complex polynomial functions that incorporate conjugate variables . we introduced two types of conjugate complex forms and their symmetric tensor representations . necessary and sufficient conditions for these conjugate complex forms being real - valued are presented , based on which two types of symmetric complex tensors are introduced . we present new definitions of eigenvalues / eigenvectors , namely the @xmath31-eigenvalue and the @xmath89-eigenvalue , which generalize the existing concepts of eigenvalues in the literature . extensions of banach @xcite type s theorem on these complex tensors are discussed as well . to give the readers a holistic picture table [ table ] summarizes the main contents . .summary of the symmetric conjugate form and the general conjugate form [ cols="<,<,<",options="header " , ] an important aspect of polynomials is the theory of nonnegativity . most existing results only apply for polynomials in real variables , for the reason that such polynomials are real - valued . since we have the full characterization of real - valued conjugate complex polynomials introduced in this paper , the question about their nonnegativity naturally arises , in particular , the relationship between nonnegativity and sos . in the real domain , this problem was completely solved by hilbert @xcite in 1888 . however , relationship between nonnegative complex polynomials and sos has not been established explicitly in the literature as far as we know . this would be one of the future research using the notion of conjugate polynomials . moreover , the new notions of symmetric complex tensors and the eigenvalues / eigenvectors would hopefully attract future modelling opportunities , and the newly developed properties , in particular the extension of banach s result would be helpful in solution methods for complex polynomial optimization . the authors would like to thank the anonymous referees and the associated editor for their insightful comments , which helped to significantly improve this paper from its original version . we are also grateful to augusto aubry for discussions on radar waveform design with similarity constraint . a. aubry , a. de maio , a. farina , and m. wicks , _ knowledge - aided ( potentially cognitive ) transmit signal and receive filter design in signal - dependent clutter _ , ieee transactiopns on aerospace and electronic systems , 49 , 93117 , 2013 . lim , _ singular values and eigenvalues of tensors : a variational approach _ , proceedings of the ieee international workshop on computational advances in multi - sensor adaptive processing , 1 , 129132 , 2005 . l. sorber , m. van barel , and l. de lathauwer , _ optimization - based algorithms for tensor decompositions : canonical polyadic decomposition , decomposition in rank-@xmath420 terms and a new generalization _ , siam j. optim . , 23 , 695720 , 2013

the idea behind flow control in data networks is to regulate the source rates in order to prevent network overload , and provide fair allocation of resources . in recent years , extensive research has been devoted to the problem of network utility maximization , with the objective of optimizing network resource allocation through a combination of source - based flow control , routing , and scheduling . the common theme is to assign a utility , as a function of the source rate , to a flow specified by a source - destination pair , and formulate an optimization problem that maximizes the sum of utilities ( e.g. , see @xcite ) . the optimal network control policy is revealed as the algorithm that solves the utility maximization problem . source - based flow control implicitly requires all sources to react properly to congestion signals such as packet loss or delay . congestion - insensitive traffic , however , is pervasive in modern networks . udp - based applications such as video streaming or voip are increasingly popular and do not respond to congestion . moreover , greedy or malicious users can inject excessive data into the network to either boost self - performance or bring down high - profile websites . in these circumstances , the network can be temporarily overloaded and congestion - aware applications , e.g. , tcp - based traffic , may be adversely affected or even starved . in this context , source - based flow control may not be adequate . there are other scenarios in which optimizing the source rates of data flows on a per - flow basis is ineffective . the internet nowadays is vulnerable to distributed denial of service ( ddos ) attacks @xcite , in which an attacker creates a large number of flows with different source addresses to overwhelm prominent websites . similarly , in a multi - player online game , thousands of users require continuous access to the game servers . there are also occasions in which a large number of users seek access to a website that broadcasts live events @xcite ; this is the so - called _ flash crowd _ phenomenon @xcite . in these situations , source - based flow control is ineffective because each individual flow uses little bandwidth , but their aggregate traffic can lead to severe link congestion near the service provider , and may starve other users in the network . a flow control scheme that can optimize a utility assigned to a collection of flows as opposed to optimizing the sum of per - flow utilities can be used to cope with such scenarios . moreover , in order to cope with uncooperative flows , it is necessary to relocate the flow control functionality from untrusted network hosts to more secure ones , such as the web servers that provide service at the receiver end . in this paper , we develop such _ receiver - based flow control _ policies using tools from stochastic network optimization @xcite . our main contributions are three - fold . first , we formulate a utility maximization problem that can assign utilities to an aggregate of flows , of which the usual per - flow - based utility maximization is a special case . second , given an arbitrary arrival rate matrix ( possibly outside the network s stability region ) , we characterize the corresponding achievable throughput region in terms of queue overflow rates . third , using a novel decomposition of the utility functions , we design a network control policy consisting of : ( i ) a set of flow controllers at the receivers ; ( ii ) packet dropping mechanism at internal nodes ; and ( iii ) back - pressure routing at intermediate nodes . the receiver - based flow controllers adjust throughput by modifying the differential backlogs between the receivers and their neighboring nodes a small ( or negative ) differential backlog is regarded as a _ push - back _ mechanism to slow down data delivery to the receiver . to deal with data that can not be delivered due to network overload , we design a threshold - based packet dropping mechanism that discards data whenever queues grow beyond certain thresholds . surprisingly , we show that our threshold - based packet dropping scheme , without the use of flow control , is sufficient to maximize the weighted sum throughput . moreover , the combined flow control and packet dropping mechanism has the following properties : ( i ) it works with finite - size buffers . ( ii ) it is nearly utility - optimal ( throughput - optimal as a special case ) and the performance gap from the optimal utility goes to zero as buffer sizes increase . ( iii ) it does not require the knowledge of arrival rates and therefore is robust to time - varying arrival rates that can go far beyond the network s stability region . in addition , our control policy can be implemented only in parts of the network that include the receivers , treating the rest of the network as exogenous data sources ( see fig . [ fig:105 ] for an example ) , and thus might be an attractive flow control solution for web servers . , @xmath0 , and @xmath1 on the right , where @xmath1 is the only receiver . the rest of the network on the right may be controlled by another network operator or follow a different network control scheme.,width=336 ] there has been a significant amount of research in the general area of stochastic network control . utility - optimal policies that combine source - end flow control with back - pressure routing have been studied in @xcite ( and references therein ) . these policies optimize per - flow utilities and require infinite - capacity buffers . however , they are not robust in the face of uncooperative users who may not adhere to the flow control scheme . a closely related problem to that studied in this paper is that of characterizing the queue overflow rates in lossless networks in overload . in a single - commodity network , a back - pressure policy is shown to achieve the most balanced queue overflow rates @xcite , and controlling queue growth rates using the max - weight policy is discussed in @xcite . the queue growth rates in networks under max - weight and @xmath2-fairness policies are analyzed in @xcite . we finally note that the importance of controlling an aggregate of data flows has been addressed in @xcite , and rate - limiting mechanisms in front of a web server to achieve some notion of max - min fairness against ddos attacks have been proposed in @xcite . an outline of the paper is as follows . the network model is given in section [ sec:601 ] . we formulate the utility maximization problem and characterize the achievable throughput region in terms of queue overflow rates in section [ sec:602 ] . section [ sec:801 ] introduces a threshold - based packet dropping policy that maximizes the weighted sum throughput without the use of flow control . section [ sec:501 ] presents a receiver - based flow control and packet dropping policy that solves the general utility maximization problem . simulation results that demonstrate the near - optimal performance of our policies are given in sections [ sec:801 ] and [ sec:501 ] . we consider a network with nodes @xmath3 and directed links @xmath4 . assume time is slotted . in every slot , packets randomly arrive at the network for service and are categorized into a collection @xmath5 of classes . the definition of a data class is quite flexible except that we assume packets in a class @xmath6 have a shared destination @xmath7 . for example , each class can simply be a flow specified by a source - destination pair . alternatively , computing - on - demand services in the cloud such as amazon ( elastic compute cloud ; ec2 ) or google ( app engine ) can assign business users to one class and residential users to another . media streaming applications may categorize users into classes according to different levels of subscription to the service provider . while classification of users / flows in various contexts is a subject of significant importance , in this paper we assume for simplicity that the class to which a packet belongs can be ascertained from information contained in the packet ( e.g. , source / destination address , tag , priority field , etc . ) . let @xmath8 be the number of exogenous class @xmath9 packets arriving at node @xmath10 in slot @xmath11 , where @xmath12 is a finite constant ; let @xmath13 for all @xmath11 . we assume @xmath14 are independent across classes @xmath9 and nodes @xmath15 , and are i.i.d . over slots with mean @xmath16 = { \lambda_{n}^{(c)}}$ ] . in the network , packets are relayed toward the destinations via dynamic routing and link rate allocation decisions . each link @xmath17 is used to transmits data from node @xmath10 to node @xmath18 and has a fixed capacity @xmath19 ( in units of packets / slot ) . under a given control policy , let @xmath20 be the service rate allocated to class @xmath9 data over link @xmath21 in slot @xmath11 . the service rates must satisfy the link capacity constraints @xmath22 at a node @xmath15 , class @xmath9 packets that arrive but not yet transmitted are stored in a queue ; we let @xmath23 be the backlog of class @xmath9 packets at node @xmath10 at time @xmath11 . we assume initially @xmath24 for all @xmath10 and @xmath9 . destinations do not buffer packets and we have @xmath25 for all @xmath9 and @xmath11 . for now , we assume every queue @xmath23 has an infinite - capacity buffer ; we show later that our control policy needs only finite buffers . to resolve potential network congestion due to traffic overload , a queue @xmath23 , after transmitting data to neighboring nodes in a slot , discards @xmath26 packets from the remaining backlog at the end of the slot . the drop rate @xmath26 is a function of the control policy to be described later and takes values in @xmath27 $ ] for some finite @xmath28 . the queue @xmath23 evolves over slots according to @xmath29^{+ } \\ & \quad+ { a_{n}^{(c)}}(t ) + \sum_{a } \mu_{an}^{(c)}(t ) , \quad \forall c,\ \forall n\neq d_{c } , \end{split}\ ] ] where @xmath30 . this inequality is due to the fact that endogenous arrivals may be less than the allocated rate @xmath31 when neighboring nodes do not have sufficient packets to send . for convenience , we define the maximum transmission rate into and out of a node by @xmath32 throughout the paper , we use the following assumption . [ assu:601 ] we assume @xmath33 . from , the sum @xmath34 is the largest amount of data that can arrive at a node in a slot ; therefore it is an upper bound on the maximum queue overflow rate at any node . assumption [ assu:601 ] ensures that the maximum packet dropping rate @xmath28 is no less than the maximum queue overflow rate , so that we can always prevent the queues from blowing up . we assign to each class @xmath35 a utility function @xmath36 . given an ( unknown ) arrival rate matrix @xmath37 , let @xmath38 be the set of all achievable throughput vectors @xmath39 , where @xmath40 is the aggregate throughput of class @xmath9 received by the destination @xmath7 . note that @xmath38 is a function of @xmath41 . we seek to design a control policy that solves the global utility maximization problem @xmath42 where the region @xmath38 is presented later in lemma [ lem:201 ] . we assume all @xmath36 functions are concave , increasing , and continuously differentiable . for ease of exposition , we also assume the functions @xmath36 have bounded derivatives such that @xmath43 for all @xmath44 , where @xmath45 are finite constants . that have unbounded derivatives as @xmath46 can be approximated by those with bounded derivatives . for example , we may approximate the proportional fairness function @xmath47 by @xmath48 for some small @xmath49 . ] as an example , consider the tree network in fig . [ fig:102 ] that serves three classes of traffic destined for node @xmath1 . class @xmath50 data originates from two different sources @xmath51 and @xmath52 , and may represent the collection of users located in different parts of the network sending or requesting information from node @xmath1 . if class @xmath50 traffic is congestion - insensitive and overloads the network , without proper flow control class @xmath53 and @xmath54 will be starved due to the presence of class @xmath50 . a utility maximization problem here is to solve @xmath55 where @xmath56 denotes the throughput of class @xmath50 data originating from @xmath51 ; @xmath57 , @xmath58 , and @xmath59 are defined similarly . note that this utility maximization - is very different from , and generalizes , the traditional per - flow - based utility maximization . the next lemma characterizes the set @xmath38 of all achievable throughput vectors in . [ lem:201 ] under i.i.d . arrival processes with an arrival rate matrix @xmath60 , let @xmath38 be the closure of the set of all achievable throughput vectors @xmath39 . then @xmath61 if and only if there exist flow variables @xmath62 and queue overflow variables @xmath63 such that @xmath64 in other words , @xmath65 see appendix [ appendix:201 ] . in lemma [ lem:201 ] , equation is the flow conservation constraint stating that the total flow rate of a class into a node is equal to the flow rate out of the node plus the queue overflow rate . equation is the link capacity constraint . the equality in shows that the throughput of a class is equal to the sum of exogenous arrival rates less the queue overflow rates . lemma [ lem:201 ] is closely related to the network capacity region @xmath66 defined in terms of admissible arrival rates ( see definition [ defn:101 ] ) ; their relationship is shown in the next corollary . [ defn:101 ] the capacity region @xmath66 of the network is the set of all arrival rate matrices @xmath67 for which there exists nonnegative flow variables @xmath68 such that @xmath69 [ lem:501 ] an arrival rate matrix @xmath70 lies in the network capacity region @xmath66 if there exist flow variables @xmath71 such that flow conservation constraints and link capacity constraints hold with @xmath72 for all @xmath10 and @xmath9 . corollary [ lem:501 ] shows that @xmath70 is achievable if and only if there exists a control policy yielding zero queue overflow rates . in this case the throughput of class @xmath9 is @xmath73 . we remark that the solution to the utility maximization - activates the linear constraints ; thus the problem - is equivalent to @xmath74 let @xmath75 be the optimal throughput vector that solves - . if the arrival rate matrix @xmath70 is in the network capacity region @xmath66 , the optimal throughput is @xmath76 from corollary [ lem:501 ] . otherwise , we have @xmath77 , where @xmath78 is the optimal queue overflow rate . our control policy that solves - has two main features . first , we have a packet dropping mechanism discarding data from the network when queues build up . an observation here is that , in order to optimize throughput and keep the network stable , we should drive the packet dropping rate to be equal to the optimal queue overflow rate . second , we need a flow controller driving the throughput vector toward the utility - optimal point . to convert the control objective into these two control features , we define , for each class @xmath35 , a utility function @xmath79 related to @xmath36 as @xmath80 where @xmath81 are control parameters to be decided later . using , we have @xmath82.\ ] ] since @xmath83 are unknown constants , maximizing @xmath84 is the same as maximizing @xmath85.\ ] ] this equivalent objective can be optimized by jointly maximizing the new utility @xmath86 at the receivers and minimizing the weighted queue overflow rates ( i.e. , the weighted packet dropping rates ) @xmath87 at each node @xmath10 . optimizing the throughput at the receivers amounts to controlling the amount of data _ actually delivered_. this is difficult because the data available to the receivers at their upstream nodes is highly correlated with control decisions taken in the rest of the network . optimizing the packet dropping rates depends on the data available at each network node , which has similar difficulties . to get around these difficulties , we introduce auxiliary control variables @xmath88 and @xmath89 and consider the optimization problem @xmath90 \label{eq:901 } \\ \text{subject to } & \quad r_{c } = \nu_{c } , \quad \forall\ , c\in { \mathcal{c}}\label{eq:210 } , \\ & \quad { q_{n}^{(c)}}\leq { \varphi_{n}^{(c ) } } , \quad \forall\ , c\in { \mathcal{c}},\ n\neq d_{c}. \label{eq:211 } \\ & \quad \text{\eqref{eq:213}-\eqref{eq:605 } hold . } \label{eq:902}\end{aligned}\ ] ] this is an equivalent problem to - . the constraints and can be enforced by stabilizing virtual queues that will appear in our control policy . the new control variables @xmath91 and @xmath92 to be optimized can now be chosen freely unconstrained by past control actions in the network . introducing auxiliary variables and setting up virtual queues are at the heart of using lyapunov drift theory to solve network optimization problems . for ease of exposition , we first consider the special case of maximizing the weighted sum throughput in the network . for each class @xmath35 , we let @xmath93 for some @xmath94 . we present a threshold - based packet dropping policy that , together with back - pressure routing , solves this problem . surprisingly , flow control is not needed here . this is because maximizing the weighted sum throughput is equivalent to minimizing the weighted packet dropping rate . indeed , choosing @xmath95 in , we have @xmath96 for all classes @xmath9 , under which maximizing the equivalent objective is the same as minimizing @xmath97 . in the next section , we will combine the threshold - based packet dropping policy with receiver - based flow control to solve the general utility maximization problem . to optimize packet dropping rates , we set up a _ drop queue _ @xmath98 associated with each queue @xmath23 . the packets that are dropped from @xmath23 in a slot , denoted by @xmath99 , are first stored in @xmath98 for eventual deletion . from , we have @xmath100.\ ] ] note that the quantity @xmath99 is the _ actual _ packets dropped from @xmath23 , which is strictly less than the allocated drop rate @xmath101 if queue @xmath23 does not have sufficient data . packets are permanently deleted from @xmath98 at the rate of @xmath102 $ ] in slot @xmath11 . the queue @xmath98 evolves according to @xmath103^{+}+ { \widetilde{d}_{n}^{(c)}}(t).\ ] ] assume initially @xmath104 for all @xmath10 and @xmath9 , where @xmath105 is a control parameter . to be finite . our choice of @xmath106 avoids unnecessary packet dropping in the initial phase of the system . ] if queue @xmath98 is stabilized , then minimizing the service rate of @xmath98 effectively minimizes the time average of dropped packets at @xmath23 . we propose the following policy . ' '' '' overload resilient algorithm ( @xmath107 ) ' '' '' _ parameter selection : _ choose @xmath108 for all classes @xmath35 , where @xmath109 . choose a parameter @xmath105 . _ backpressure routing : _ over each link @xmath110 , let @xmath111 be the subset of classes that have access to link @xmath112 . compute the differential backlog @xmath113 for each class @xmath114 , where @xmath115 at the receiver @xmath7 . define @xmath116 we allocate the service rates @xmath117 let @xmath118 for all classes @xmath119 . _ packet dropping : _ at queue @xmath23 , allocate the packet dropping rate @xmath101 ( see ) according to @xmath120 where @xmath121 is a constant chosen to satisfy assumption [ assu:601 ] . at the drop queue @xmath98 , allocate its service rate @xmath122 according to @xmath123 _ queue update : _ update queues @xmath23 according to and update queues @xmath98 according to - in every slot . ' '' '' the packet dropping subroutine in this policy is threshold - based . the @xmath107 policy uses local queueing information and does not require the knowledge of exogenous arrival rates . it is notable that network overload is autonomously resolved by each node making local decisions of routing and packet dropping . [ lem:402 ] for each class @xmath35 , define the constants @xmath124 in the @xmath107 policy , queues @xmath23 and @xmath98 are deterministically bounded by @xmath125 in addition , we have @xmath126 for all @xmath10 , @xmath9 , and @xmath11 . see appendix [ appendix:601 ] . in lemma [ lem:402 ] , the value of @xmath127 is the finite buffer size sufficient at queue @xmath23 . the parameter @xmath128 controls when queue @xmath23 starts dropping packets . indeed , due to @xmath129 , the @xmath107 policy discards packets from @xmath23 only if @xmath130 . the quantity @xmath131 is a controllable threshold beyond which we say queue @xmath23 is overloaded and should start dropping packets . as we see next , the performance of the @xmath107 policy approaches optimality as the buffer sizes increase . [ thm:501 ] define the limiting throughput of class @xmath9 as @xmath132,\ ] ] where @xmath133 denotes the class @xmath9 packets received by node @xmath7 over link @xmath134 . the @xmath107 policy yields the limiting weighted sum throughput satisfying @xmath135 where @xmath75 is the optimal throughput vector that solves - under the linear objective function @xmath136 , @xmath105 is a control parameter , and @xmath0 is a finite constant defined as @xmath137,\ ] ] where @xmath138 denotes the cardinality of a set @xmath139 . we omit the proof of theorem [ thm:501 ] because it is similar to that of theorem [ thm:502 ] presented later in the general case of utility maximization . from , the @xmath107 policy yields near - optimal performance by choosing the parameter @xmath128 sufficiently large . correspondingly , a large @xmath128 implies a large buffer size of @xmath140 . as shown in corollary [ lem:501 ] , if the arrival rate matrix @xmath70 lies in the network capacity region @xmath66 , then the optimal throughput for class @xmath9 is @xmath76 and reduces to @xmath141 that we can choose @xmath128 arbitrarily large leads to the next corollary . [ cor:601 ] the @xmath107 policy is ( close to ) throughput optimal . we conduct simulations for the @xmath107 policy in the network shown in fig . [ fig:101 ] . the directed links @xmath142 and @xmath143 have the capacity of @xmath50 packet / slot . there are three classes of traffic to be served ; for example , class 1 data arrives at node @xmath0 and is destined for node @xmath52 . classes @xmath50 and @xmath53 compete for service over @xmath143 ; classes @xmath53 and @xmath54 compete for service over @xmath142 . each simulation below is run over @xmath144 slots . in each class , we assume a bernoulli arrival process whereby @xmath145 packets arrive to the network in a slot with probability @xmath146 , and no packets arrive otherwise . the arrival rate of each class is 2 packets / slot , which clearly overloads the network . let @xmath40 be the throughput of class @xmath9 . consider the objective of maximizing the weighted sum throughput @xmath147 ; the weights are rewards obtained by serving a packet in a class . the optimal solution is : ( i ) always serve class @xmath50 at node @xmath0 because it yields better rewards than serving class @xmath53 . ( ii ) always serve class @xmath54 at node @xmath51although class @xmath53 has better rewards than class @xmath54 , it does not make sense to serve class @xmath53 at @xmath51 only to be dropped later at @xmath0 . the optimal throughput vector is therefore @xmath148 . consider another objective of maximizing @xmath149 . here , class @xmath53 has a reward that is better than the sum of rewards of the other two classes . thus both nodes @xmath51 and @xmath0 should always serve class @xmath53 ; the optimal throughput vector is @xmath150 . table [ table:601 ] shows the near - optimal performance of the @xmath107 policy in both cases as @xmath128 increases . we show that the @xmath107 policy is robust to time - varying arrival rates . suppose class @xmath50 and @xmath54 have a fixed arrival rate of @xmath151 packets / slot . the arrival rate of class @xmath53 is @xmath53 packets / slot in the interval @xmath152 and is @xmath146 packets / slot elsewhere . we consider the objective of maximizing @xmath149 . the network is temporarily overloaded in the interval @xmath153 ; the optimal time - average throughput in @xmath153 is @xmath150 as explained in the above case . the network is underloaded in the interval @xmath154 , in which the optimal throughput is @xmath155 . we use the following parameters here : @xmath156 , @xmath157 , @xmath158 , and @xmath159 . table [ table:603 ] shows the near - optimal throughput performance of the @xmath107 policy . figure [ fig:103 ] shows the sample paths of the queue processes @xmath160 , @xmath161 , @xmath162 , and @xmath163 in the simulation . clearly the queues suddenly build up when the network enters the overload interval @xmath153 , but the backlogs are kept close to the upper bound @xmath140 without blowing up . .the throughput performance of the @xmath107 policy under time - varying arrival rates . [ cols= " < , < , < " , ] in this paper we develop a receiver - based flow control and an in - network packet dropping strategy to cope with network overload . our scheme is robust to uncooperative users who do not employ source - based flow control and malicious users that intentionally overload the network . a novel feature of our scheme is a receiver - based backpressure / push - back mechanism that regulates data flows at the granularity of traffic classes , where packets can be classified based on aggregates of data flows . this is in contrast to source - based schemes that can only differentiate between source - destination pairs . we show that when the receiver - based flow control scheme is combined with a threshold - based packet dropping policy at internal network nodes , optimal utility can be achieved . 10 [ 1]#1 url@samestyle [ 2]#2 [ 2]l@#1=l@#1#2 m. chiang , s. h. low , a. r. calderbank , and j. c. doyle , `` layering as optimization decomposition : a mathematical theory of network architectures , '' _ proc . ieee _ , vol . 95 , no . 1 , 255312 , jan . 2007 . kelly , a. k. maulloo , and d. k. h. tan , `` rate control in communication networks : shadow prices , proportional fairness and stability , '' _ journal of the oper . _ , vol . 49 , pp . 237252 , 1998 . kelly , `` charging and rate control for elastic traffic , '' _ european trans . telecommunications _ , vol . 8 , pp . 3337 , 1997 . s. h. low and d. e. lapsley , `` optimization flow control i : basic algorithm and convergence , '' _ ieee / acm trans . _ , vol . 7 , no . 6 , pp . 861874 , dec . a. l. stolyar , `` maximizing queueing network utility subject to stability : greedy primal - dual algorithm , '' _ queueing syst . _ , vol . 50 , no . 4 , pp . 401457 , 2005 . m. j. neely , e. modiano , and c .- p . li , `` fairness and optimal stochastic control for heterogeneous networks , '' _ ieee / acm trans . _ , vol . 16 , no . 2 , pp . 396409 , apr . a. eryilmaz and r. srikant , `` joint congestion control , routing , and mac for stability and fairness in wireless networks , '' _ ieee j. sel . areas commun . _ , 24 , no . 8 , pp . 15141524 , aug . , `` fair resource allocation in wireless networks using queue - length - based scheduling and congestion control , '' _ ieee / acm trans . _ , vol . 15 , no . 6 , pp . 13331344 , dec . 2007 . x. lin and n. b. shroff , `` joint rate control and scheduling in multihop wireless networks , '' in _ ieee conf . decision and control ( cdc ) 2004 , pp . 14841489 . r. k. c. chang , `` defending against flooding - based distributed denial - of - service attacks : a tutorial , '' _ ieee commun . mag . _ , vol . 40 , no . 4251 , oct . 2002 . a. srivastava , b. b. gupta , a. tyagi , a. sharma , and a. mishra , `` a recent survey on ddos attacks and defense mechanisms , '' in _ advances in parallel distributed computing _ communications in computer and information science.1em plus 0.5em minus 0.4emspringer berlin heidelberg , 2011 , vol . 570580 . j. borland , `` net video not yet ready for prime time , '' feb . [ online ] . available : http://news.cnet.com/2100-1033-221271.html r. mahajan , s. m. bellovin , s. floyd , j. ioannidis , v. paxson , and s. shenker , `` controlling high bandwidth aggregates in the network , '' _ acm computer communication review _ 32 , pp . 6273 , 2002 . l. georgiadis , m. j. neely , and l. tassiulas , `` resource allocation and cross - layer control in wireless networks , '' _ foundations and trends in networking _ , vol . 1 , no . 1 , 2006 . m. j. neely , _ stochastic network optimization with application to communication and queueing systems_.1em plus 0.5em minus 0.4em morgan & claypool , 2010 . l. georgiadis and l. tassiulas , `` optimal overload response in sensor networks , '' _ ieee trans . inf . theory _ 52 , no . 6 , pp . 26842696 , jun . 2006 . c. w. chan , m. armony , and n. bambos , `` fairness in overloaded parallel queues , '' 2011 , working paper . d. shah and d. wischik , `` fluid models of congestion collapse in overloaded switched networks , '' _ queueing syst . _ , 69 , pp . 121143 , 2011 . r. egorova , s. borst , and b. zwart , `` bandwidth - sharing in overloaded networks , '' in _ conf . information science and systems ( ciss ) _ , princeton , nj , usa , mar . 2008 , pp . 3641 . d. k. y. yau , j. c. s. lui , f. liang , and y. yam , `` defending against distributed denial - of - service attacks with max - min fair server - centric router throttles , '' _ ieee / acm trans . _ , vol . 13 , no . 1 , 2942 , feb . 2005 . c. w. tan , d .- chiu , j. c. s. lui , and d. k. y. yau , `` a distributed throttling approach for handling high bandwidth aggregates , '' _ ieee trans . parallel distrib . _ , vol . 18 , no . 7 , pp . 983995 , jul . s. chen and q. song , `` perimeter - based defense against high bandwidth ddos attacks , '' _ ieee trans . parallel distrib . _ , vol . 16 , no . 6 , pp . 526537 , jun . 2005 . m. j. neely , e. modiano , and c. e. rohrs , `` dynamic power allocation and routing for time - varying wireless networks , '' _ ieee j. sel . areas commun . _ , 23 , no . 1 , 89103 , jan . 2005 . m. j. neely , `` super - fast delay tradeoffs for utility optimal fair scheduling in wireless networks , '' _ ieee j. sel . areas _ , vol . 24 , no . 8 , pp . 14891501 , first we show - are necessary conditions . given a control policy , let @xmath164 be the amount of class @xmath9 packets transmitted over link @xmath165 in the interval @xmath166 $ ] , and @xmath23 be the class @xmath9 packets queued at node @xmath10 at time @xmath11 . from the fact that the difference between incoming and outgoing packets at a node in @xmath166 $ ] is equal to the queue backlog at time @xmath11 , we have @xmath167 which holds for all nodes @xmath15 for each class @xmath168 . taking expectation and time average of , we obtain @xmath169 } = \frac{{\mathbb{e}\left[{q_{n}^{(c)}}(t)\right]}}{t } + \frac{1}{t } { \mathbb{e}\left[\sum_{b } f_{nb}^{(c)}(t)\right]}.\ ] ] the link capacity constraints lead to @xmath170 } \leq { \mu_{\text{max}}}^{ab } , \quad\forall\ , ( a , b)\in { \mathcal{l}}.\ ] ] consider the sequences @xmath171/t$ ] and @xmath172/t$ ] indexed by @xmath11 in . for each @xmath35 and @xmath173 , the sequence @xmath174/t , t\in{\mathbb{z}}^{+}\}$ ] is bounded because the capacity of each link is bounded . it follows from that the sequence @xmath175/t , t\in{\mathbb{z}}^{+}\}$ ] is also bounded . there is a subsequence @xmath176 such that limit points @xmath68 and @xmath177 exist and satisfy , as @xmath178 , @xmath179 } & \to f_{ab}^{(c ) } , \quad \forall\,c\in{\mathcal{c}},\ \forall\ , ( a , b)\in{\mathcal{l } } , \label{eq:601 } \\ \frac{1}{t_{k}}{\mathbb{e}\left[{q_{n}^{(c)}}(t_{k})\right ] } & \to { q_{n}^{(c ) } } , \quad \forall\,c\in{\mathcal{c}},\ \forall\,n\neq d_{c}. \label{eq:602}\end{aligned}\ ] ] applying - to - results in and . define the throughput @xmath40 of class @xmath9 as @xmath180}.\ ] ] the inequality in follows and . the equality in results from summing over @xmath15 . to show the converse , it suffices to show that every interior point of @xmath38 is achievable . let @xmath181 be an interior point of @xmath38 , i.e. , there exists @xmath182 such that @xmath183 . there exist corresponding flow variables @xmath68 and @xmath177 such that @xmath184 in the flow system - , by removing the subflows that contribute to queue overflows , we obtain new reduced flow variables @xmath185 and @xmath186 such that @xmath187 , @xmath188 , and @xmath189 define @xmath190 \frac{r_{c}}{\sum_{n } \hat{\lambda}_{n}^{(c)}-\epsilon/2}.\ ] ] from , we have @xmath191 and @xmath192 . it is not difficult to check that @xmath193 , @xmath194 , and @xmath195 . combined with - , we obtain @xmath196 where the first inequality is a strict one . these inequalities show that the rate matrix @xmath197 is an interior point of the network capacity region @xmath66 in definition [ defn:101 ] , and therefore is achievable by a control policy , such as the back - pressure policy @xcite . therefore , the aggregate rate vector @xmath39 , where @xmath198 , is also achievable . we prove lemma [ lem:402 ] by induction . first we show @xmath98 is deterministically bounded . assume @xmath199 for some @xmath11 , which holds at @xmath200 because we let @xmath106 . consider the two cases : 1 . if @xmath201 , then from we obtain @xmath202 where the second inequality uses . 2 . if @xmath203 , then the @xmath107 policy chooses @xmath204 at queue @xmath98 and we have @xmath205^{+ } + { d_{\text{max}}}\\ & \leq \max[{d_{n}^{(c)}}(t ) , { d_{\text{max } } } ] \leq { d_{\text{max}}}^{(c)},\end{aligned}\ ] ] where the last inequality uses the induction assumption . from these two cases , we obtain @xmath206 . similarly , we show @xmath23 is bounded . assume @xmath207 for some @xmath11 , which holds at @xmath200 because we let @xmath208 . consider the two cases : 1 . if @xmath209 , then from we get @xmath210 where the second inequality uses assumption [ assu:601 ] . 2 . if @xmath211 , the @xmath107 policy chooses @xmath212 at queue @xmath23 and we have @xmath213 where the third inequality uses induction assumption . we conclude that @xmath214 . finally , we show @xmath129 for all slots . assume this is true at some time @xmath11 ; this holds when @xmath200 because we assume @xmath106 . consider the two cases : 1 . if @xmath215 , the @xmath107 policy chooses @xmath216 at queue @xmath98 and we have @xmath217 by induction assumption . 2 . if @xmath218 , the @xmath107 policy chooses @xmath219 and we have @xmath220 the proof is complete . [ [ appendix:501 ] ] we construct a proper lypuanov drift inequality that leads to the @xmath221 policy . let @xmath222 be the vector of all physical and virtual queues in the network . using the parameters @xmath223 and @xmath224 given in the policy , we define the lyapunov function @xmath225^{2 } + \frac{1}{2}\sum_{c , n\neq d_{c } } \bigl[d_{n}^{(c)}(t)\bigr]^{2 } \\ + \sum_{c\in{\mathcal{c } } } \big(e^{w{\big(z_{c}(t)-q\big ) } } + e^{w{\big(q - z_{c}(t)\big ) } } \big).\end{gathered}\ ] ] the last sum is a lyapunov function whose value grows exponentially when @xmath226 deviates in both directions from the constant @xmath224 . this exponential - type lyapunov function is useful for both stabilizing @xmath226 and guaranteeing there is sufficient data in @xmath226 . such exponential - type lyapunov functions are previously used in @xcite to study the optimal utility - delay tradeoff in wireless networks . we define the lyapunov drift @xmath227},\ ] ] where the expectation is with respect to all randomness in the system in slot @xmath11 . define the indicator function @xmath228 if @xmath229 and @xmath230 otherwise ; define @xmath231 . define @xmath232 . the next lemma is proved in appendix [ appendix:202 ] . [ lem:401 ] the lyapunov drift @xmath233 under any control policy satisfies @xmath234 } \notag \\ & + v\sum_{nc}\theta_{c } \,{\mathbb{e}\left[{\varphi_{n}^{(c)}}(t)\mid h(t)\right ] } \leq b + \sum_{nc } { q_{n}^{(c)}}(t ) { \lambda_{n}^{(c)}}\notag\\ & -\sum_{nc } { q_{n}^{(c)}}(t ) { \mathbb{e}\left[\sum_{b } \mu_{nb}^{(c)}(t ) + d_{n}^{(c)}(t ) - \sum_{a } \mu_{an}^{(c)}(t ) \mid h(t)\right ] } \notag \\ & - \sum_{nc } { d_{n}^{(c)}}(t ) { \mathbb{e}\left[{\varphi_{n}^{(c)}}(t)-{d_{n}^{(c)}}(t ) \mid h(t)\right ] } \notag \\ & -v \sum_{c } { \mathbb{e}\left[h_{c}\big(\nu_{c}(t)\big ) \mid h(t)\right ] } + v\sum_{nc } \theta_{c}\ , { \mathbb{e}\left[{\varphi_{n}^{(c)}}(t)\mid h(t)\right ] } \notag \\ & - w \sum_{c } { 1_{c}^{\text{r}}}(t ) e^{w(z_{c}(t)-q ) } \big({\mathbb{e}\left[\delta_{c}(t)\mid h(t)\right ] } - \frac{\epsilon}{2 } \big ) \notag \\ & + w \sum_{c } { 1_{c}^{\text{l}}}(t ) e^{w(q - z_{c}(t))}\big({\mathbb{e}\left [ \delta_{c}(t ) \mid h(t)\right ] } + \frac{\epsilon}{2 } \big),\label{eq:227}\end{aligned}\ ] ] where @xmath0 is a finite constant defined by @xmath235 + \\ & \quad 2 { \left|{\mathcal{n}}\right|}{\left|{\mathcal{c}}\right| } { d_{\text{max}}}^{2 } + { \left|{\mathcal{c}}\right|}\big [ w(2{\delta_{\text{max}}}+\epsilon ) + e^{w({\nu_{\text{max}}}+{\mu_{\text{max}}^{\text{in } } } ) } + e^{wq } \big ] . \end{split}\ ] ] by isolating decisions variables in , it is not difficult to verify that the @xmath221 policy observes the current network state @xmath236 and minimizes the right - hand side of in every slot . [ [ appendix:202 ] ] we establish the lyapunov drift inequality in . applying to the fact that @xmath237^{+ } = ( a - b - c)^{+}$ ] for nonnegative reals @xmath238 , @xmath239 , and @xmath9 , and using lemma [ lem:101 ] in appendix [ appendix:203 ] , we have @xmath240^{2 } - \big[q_{n}^{(c)}(t)\big]^{2 } \big ) \leq b_{q } - \\ & q_{n}^{(c)}(t)\ , \big ( \sum_{b } \mu_{nb}^{(c)}(t ) + d_{n}^{(c)}(t ) - a_{n}^{(c)}(t ) - \sum_{a } \mu_{an}^{(c)}(t ) \big ) , \end{split}\ ] ] where @xmath241 is a finite constant . from and , we get @xmath242^{+}+ { d_{n}^{(c)}}(t).\ ] ] similarly , we obtain for queue @xmath98 @xmath243^{2 } - \big[d_{n}^{(c)}(t)\big]^{2 } \big ) \\ \leq b_{d } - d_{n}^{(c)}(t)\ , \big ( \varphi_{n}^{(c)}(t ) - d_{n}^{(c)}(t ) \big),\end{gathered}\ ] ] where @xmath244 is a finite constant . [ lem:202 ] given fixed constants @xmath245 and @xmath246 , we define @xmath247 then @xmath248,\end{gathered}\ ] ] @xmath249.\end{gathered}\ ] ] define @xmath250 and @xmath251 ; we have @xmath252 and @xmath253 . equation yields @xmath254 since @xmath255 , we have @xmath256 because the first term is bounded by the second term if @xmath257 , and is bounded by the third term otherwise . multiplying both sides by @xmath258 yields @xmath259,\end{gathered}\ ] ] where the last term follows the taylor expansion of @xmath260 . if we have @xmath261 then plugging into leads to . indeed , by definition of @xmath223 in we get @xmath262 which uses @xmath263 . also , leads to @xmath264 define the event @xmath265 as @xmath266 i.e. , all upstream nodes of @xmath7 have sufficient data to transmit , in which case we must have @xmath267 . it follows @xmath268 where the second case follows that queue @xmath226 is always non - negative . similar as , from we obtain @xmath269 , \end{split}\ ] ] in which we use the taylor expansion of @xmath270 . bounding the last term of by and multiplying the inequality by @xmath271 , we obtain . now , define the indicator function @xmath228 if @xmath229 and @xmath230 otherwise . define @xmath231 . then for any real number @xmath238 , @xmath272 similarly , we have @xmath273 using @xmath274 in and @xmath275 in , inequalities and lead to @xmath276,\end{gathered}\ ] ] @xmath277.\end{gathered}\ ] ] finally , summing and over @xmath15 for each @xmath168 , summing and over @xmath168 , and taking conditional expectation , we have @xmath278 } \\ - \sum_{nc } { d_{n}^{(c)}}(t ) { \mathbb{e}\left[{\varphi_{n}^{(c)}}(t)-{d_{n}^{(c)}}(t ) \mid h(t)\right ] } \\ - w \sum_{c } { 1_{c}^{\text{r}}}(t ) e^{w(z_{c}(t)-q ) } \big({\mathbb{e}\left[\delta_{c}(t)\mid h(t)\right ] } - \frac{\epsilon}{2 } \big ) \\ + w \sum_{c } { 1_{c}^{\text{l}}}(t ) e^{w(q - z_{c}(t))}\big({\mathbb{e}\left[\delta_{c}(t)\mid h(t)\right ] } + \frac{\epsilon}{2 } \big),\end{gathered}\ ] ] where @xmath0 is a finite constant defined as @xmath279 . \label{eq:1208}\end{aligned}\ ] ] the constants @xmath280 and @xmath281 are defined right after and , respectively . adding to both sides of @xmath282 } + v\sum_{nc } \theta_{c}\ , { \mathbb{e}\left[{\varphi_{n}^{(c)}}(t)\mid h(t)\right]}\ ] ] yields . the boundedness of the queues @xmath23 and @xmath98 follows the proof of lemma [ lem:402 ] . for the queues @xmath226 , we again prove by induction . suppose @xmath283 , which holds at @xmath200 because we let @xmath284 . consider the two cases : 1 . if @xmath285 , then by @xmath286 2 . if @xmath287 , then from we obtain @xmath288 where the last inequality follows our choice of @xmath128 satisfying @xmath289 as mentioned in section [ sec:1201 ] . as a result , in we have @xmath290 where the inequality follows @xmath291 in . since all queues @xmath23 for @xmath15 are deterministically bounded by @xmath127 , we have @xmath292 for all nodes @xmath10 such that @xmath293 . consequently , the @xmath221 policy does not transmit any class @xmath9 packets over the links @xmath294 and the virtual queue @xmath226 has no arrivals . therefore @xmath295 . we conclude that @xmath296 . the proof is complete . let the throughput vector @xmath75 and the flow variables @xmath297 be the optimal solution to the utility maximization - . we consider the stationary policy that observes the current network state @xmath236 and chooses in every slot : the first part is a feasible allocation because the value of @xmath78 is at most @xmath302 , which is less than or equal to @xmath28 by assumption [ assu:601 ] . the second part is feasible because the link rates @xmath303 satisfy the link capacity constraints . the third part is feasible because we assume @xmath304 . under this stationary policy , by using the equality in and using , we have @xmath305 in every slot . as a result , @xmath306 } & = \mathbb{e}\big[\nu_{c}^{*}(t ) - \sum_{a } \mu_{ad_{c}}^{(c)*}(t ) \mid z_{c}(t ) \geq q\big ] \\ & = { \mathbb{e}\left[\nu_{c}^{*}(t ) - r_{c}^ { * } \mid z_{c}(t ) \geq q\right ] } = \frac{\epsilon}{2 } , \\ { \mathbb{e}\left[\delta_{c}^{*}(t)\mid z_{c}(t ) < q\right ] } & = 0.\end{aligned}\ ] ] the @xmath221 policy minimizes the right - hand side of in every slot . using the decisions in the stationary policy , we can upper bound under the @xmath221 policy by @xmath307 } + v\sum_{nc } \theta_{c}\,{\mathbb{e}\left[{\varphi_{n}^{(c)}}(t)\mid h(t)\right ] } \notag \\ & \leq b + { \left|{\mathcal{c}}\right| } \frac{w\epsilon}{2 } e^{wq}- v \sum_{c } h_{c}(\nu_{c}^{*}(t ) ) + v \sum_{nc } \theta_{c}\ , q_{n}^{(c ) * } \notag \\ & \ -\sum_{nc } { q_{n}^{(c)}}(t ) \big ( \sum_{b } f_{nb}^{(c ) * } + q_{n}^{(c ) * } - { \lambda_{n}^{(c)}}- \sum_{a } f_{an}^{(c)*}\big ) \notag \\ & = b_{1 } - v \sum_{c } h_{c}(\nu_{c}^{*}(t ) ) + v \sum_{nc}\theta_{c}\ , q_{n}^{(c ) * } , \label{eq:404}\end{aligned}\ ] ] where we define @xmath308 and @xmath0 is given in . the last equality in is because the variables @xmath309 and @xmath78 satisfy the flow conservation constraints . the utility functions @xmath36 are assumed to have bounded derivatives with @xmath43 for all @xmath44 . thus , the utility functions @xmath79 have bounded derivatives with @xmath310 for all @xmath44 . using @xmath311 , we have @xmath312 if @xmath301 then it is trivial . if @xmath313 then by mean - value theorem there exists @xmath314 such that @xmath315 using in yields @xmath307 } + v\sum_{nc } \theta_{c}\,{\mathbb{e}\left[{\varphi_{n}^{(c)}}(t)\mid h(t)\right ] } \notag \\ & \leq b_{1 } - v \sum_{c } h_{c}(r_{c}^ { * } ) + v \sum_{nc } \theta_{c}\,q_{n}^{(c ) * } + \frac{v\epsilon}{2 } \sum_{c } ( m_{c}+\theta_{c } ) \notag \\ & = b_{1 } - v\sum_{c } g_{c}(r_{c}^ { * } ) + v\sum_{nc } \theta_{c}\,{\lambda_{n}^{(c)}}+ \frac{v\epsilon}{2 } \sum_{c } ( m_{c}+\theta_{c } ) , \label{eq:905}\end{aligned}\ ] ] where the last equality uses @xmath316 and @xmath317 in . the next lemma is useful . the value of @xmath321 is the solution to the convex program - . if @xmath229 , then we have @xmath322 and @xmath323 . if @xmath324 , the problem - reduces to @xmath325 for all @xmath326 , we have by mean - value theorem @xmath327 where the first inequality results from the choice of @xmath328 that yields @xmath329 for all @xmath330 . rearranging terms in , for all @xmath326 we have @xmath331 therefore , the solution to the problem - must satisfy @xmath332 . since @xmath333 , we conclude @xmath320 . now , from lemma [ lem:403 ] we have @xmath334 if @xmath335 then this is trivial . otherwise , if @xmath336 then by mean - value theorem we obtain for some @xmath337 that @xmath338 plugging into and rearranging terms yield @xmath339 } \notag \\ & \qquad+ v\sum_{nc } \theta_{c}\,{\mathbb{e}\left[{\varphi_{n}^{(c)}}(t)\mid h(t)\right ] } \notag \\ & \leq b_{1 } + \frac{3v \epsilon}{2 } \sum_{c } ( m_{c}+\theta_{c } ) - v\sum_{c } g_{c}(r_{c}^ { * } ) + v\sum_{nc } \theta_{c}\,{\lambda_{n}^{(c)}}. \label{eq:411}\end{aligned}\ ] ] define the time average @xmath340}.\ ] ] define @xmath341 , @xmath342 , and @xmath343 similarly . taking expectation and time average over @xmath344 in , dividing by @xmath128 , rearranging terms , and applying jensen s inequality to the functions @xmath79 , we get @xmath345}}{vt } - \frac{3\epsilon}{2 } \sum_{c}(m_{c}+\theta_{c}).\end{gathered}\ ] ] adding and subtracting @xmath346 at the left - hand side of and using the definition of @xmath79 yield @xmath347}}{vt } - \frac{3\epsilon}{2 } \sum_{c } ( m_{c}+\theta_{c}).\end{gathered}\ ] ] from and we have @xmath348 taking expectation and time average yields @xmath349/t , \label{eq:412 } \\ \overline{\widetilde{\nu}_{c}(t ) } = \sum_{a } \overline{\widetilde{\mu}_{ad_{c}}^{(c)}(t ) } + \mathbb{e}\big[z_{c}(0)-z_{c}(t)\big]/t . \label{eq:413}\end{gathered}\ ] ] now , by the law of flow conservation , the sum of exogenous arrival rates must be equal to the sum of delivered throughput , time averages of dropped packets , and queue growth rates . in other words , we have for each class @xmath168 and for all slots @xmath11 @xmath350.\ ] ] combining - yields @xmath351 } + \frac{1}{t } \sum_{n } { \mathbb{e}\left[{d_{n}^{(c)}}(t)+{q_{n}^{(c)}}(t)\right]}.\end{gathered}\ ] ] finally , using the boundedness of queues @xmath23 , @xmath98 , and @xmath226 in lemma [ lem:502 ] and the continuity of @xmath36 , we obtain from and that @xmath352 where the last equality of uses the definition in . taking a limit of as @xmath353 and using and , we obtain @xmath354 where the constant @xmath355 , defined in , is @xmath356 \\ & \quad + { \left|{\mathcal{c}}\right| } \big [ w ( 2{\delta_{\text{max}}}+ \epsilon ) + e^{w({\nu_{\text{max}}}+ { \mu_{\text{max}}^{\text{in } } } ) } + \frac{w\epsilon}{2}e^{wq } + e^{wq}\big ] . \end{split}\ ] ] the proof is complete . [ lem:101 ] if a queue process @xmath357 satisfies @xmath358 where @xmath359 and @xmath360 are nonnegative bounded random variables with @xmath361 and @xmath362 , then there exists a positive constant @xmath0 such that @xmath363

adaptive filtering has attracted much research interest in both theoretical and applied issues for a long time@xcite . due to its good performance , easy implementation , and high robustness , least mean square ( lms ) algorithm @xcite has been widely used in various applications such as system identification , channel equalization , and echo cancelation . the unknown systems to be identified are sparse in most physical scenarios , including the echo paths @xcite and digital tv transmission channels @xcite . in other words , there are only a small number of non - zero entries in the long impulse response . for such systems , the traditional lms has no particular gain since it never takes advantage of the prior sparsity knowledge . in recent years , several new algorithms have been proposed based on lms to utilize the feature of sparsity . m - max normalized lms ( mmax - nlms)@xcite and sequential partial update lms ( s - lms)@xcite decrease the computational cost and steady - state mean squared error ( mse ) by means of updating filter tap - weights selectively . proportionate nlms ( pnlms ) and its improved version@xcite accelerate the convergence by setting the individual step size in proportional to the respective filter weights . sparsity in adaptive filtering framework has been a long discussed topic @xcite . inspired by the recently appeared sparse signal processing branch @xcite , especially compressive sampling ( or compressive sensing , cs)@xcite , a family of sparse system identification algorithms has been proposed based on @xmath1 norm constraint . the basic idea of such algorithms is to exploit the characteristics of unknown impulse response and to exert sparsity constraint on the cost function of gradient descent . specially , za - lms @xcite utilizes @xmath2 norm and draws the zero - point attraction to all tap - weights . @xmath0-lms @xcite employs a non - convex approximation of @xmath0 norm and exerts respective attractions to zero and non - zero coefficients . the smoothed @xmath0 algorithm , which is also based on an approximation of @xmath0 norm , is proposed in @xcite and analyzed in @xcite . besides lms variants , rls - based sparse algorithms @xcite and bayesian - based sparse algorithms @xcite have also been proposed . it is necessary to conduct a theoretical analysis for @xmath0-lms algorithm . numerical simulations demonstrate that the mentioned algorithm has rather good performance compared with several available sparse system identification algorithms @xcite , including both accelerating the convergence and decreasing the steady - state msd . @xmath0-lms performs _ zero - point attraction _ to small adaptive taps and pulls them toward the origin , which consequently increases their convergence speed and decreases their steady - state bias . because most coefficients of a sparse system are zero , the overall identification performance is enhanced . it is also found that the performance of @xmath0-lms is highly affected by the predefined parameters . improper parameter setting could not only make the algorithm less efficient , but also yield steady - state misalignment even larger than the traditional algorithm . the importance of such analysis should be further emphasized since adaptive filter framework and @xmath0-lms behave well in the solution of sparse signal recovery problem in compressive sensing@xcite . compared with some convex relaxation methods and greedy pursuits @xcite , it was experimentally demonstrated that @xmath0-lms in adaptive filtering framework shows more robustness against noise , requires fewer measurements for perfect reconstruction , and recovers signal with less sparsity . considering its importance as mentioned above , the steady - state performance and instantaneous behavior of @xmath0-lms are throughout analyzed in this work . one contribution of this work is on steady - state performance analysis . because of the nonlinearity caused by the sparsity constraint in @xmath0-lms , the theoretical analysis is rather difficult . to tackle this problem and enable mathematical tractability , adaptive tap - weights are sorted into different categories and several assumptions besides the popular independence assumption are employed . then , the stability condition on step size and steady - state misalignment are derived . after that , the parameter selection rule for optimal steady - state performance is proposed . finally , the steady - state msd gain is obtained theoretically of @xmath0-lms over the tradition algorithm , with the optimal parameter . another contribution of this work is on instantaneous behavior analysis , which indicates the convergence rate of lms type algorithms and also arouses much attention @xcite . for lms and most of its linear variants , the convergence process can be obtained in the same derivation procedure as steady - state misalignment . however , this no longer holds for @xmath0-lms due to its nonlinearity . in a different way by utilizing the obtained steady - state msd as foundation , a taylor expansion is employed to get an approximated quantitative analysis of the convergence process . also , the convergence rates are compared between @xmath0-lms and standard lms . in order to theoretically characterize the performance and guide the selection of the optimal algorithm parameters , the mean square analysis has been conducted for standard lms and a lot of its variants . to the best of our knowledge , widrow for the first time proposed the lms algorithm in @xcite and studied its performance in @xcite . later , horowitz and senne @xcite established the mathematical framework for mean square analysis via studying the weight vector covariance matrix and achieved the closed - form expression of mse , which was further simplified by feuer and weinstein @xcite . the mean square performance of two variants , leaky lms and deficient length lms , were theoretically investigated in similar methodologies in @xcite and @xcite , respectively . recently , dabeer and masry @xcite put forward a new approach for performance analysis on lms without assuming a linear regression model . moreover , convergence behavior of transform - domain lms was studied in @xcite with second - order autoregressive process . a summarized analysis was proposed in @xcite on a class of adaptive algorithms , which performs linear time - invariant operations on the instantaneous gradient vector and includes lms as the simplest case . similarly , the analysis of normalized lms has also attracted much attention @xcite . however , the methodologies mentioned above , which are effective in their respective context , could no longer be directly applied to the analysis of @xmath0-lms , considering its high non - linearity . admittedly , nonlinearity is a long topic in adaptive filtering and not unique for @xmath0-lms itself . researchers have delved into the analysis of many other lms - based nonlinear variants @xcite . nevertheless , the nonlinearity of most above references comes from non - linear operations on _ the estimated error _ , rather than _ the adaptive tap - weights _ that @xmath0-lms mainly focuses on . we have noticed that the mean square deviation analysis of za - lms has been conducted @xcite . however , this work is far different from the reference . first of all , the literature did not consider the transient performance analysis while in this work the mean square behavior of both steady - state and convergence process are conducted . moreover , considering @xmath0-lms is more sophisticated than za - lms , there are more parameters in @xmath0-lms than in za - lms , which enhances the algorithm performance but increases the difficulty of theoretical analysis . last but not least , taking its parameters to a specific limit setting , @xmath0-lms becomes essentially the same as za - lms , which can apply the theoretical results of this work directly . a preliminary version of this work has been presented in conference@xcite , including the convergence condition , derivation of steady - state msd , and an expression of the optimal parameter selection . this work provides not only a detailed derivation for steady - state results , but also the mean square convergence analysis . moreover , both the steady - state msd and the parameter selection rule are further simplified and available for analysis . finally , more simulations are performed to validate the results and more discussions are conducted . this paper is organized as follows . in section [ section : review ] , a brief review of @xmath0-lms and za - lms is presented . then in section [ section : preliminaries ] , a few assumptions are introduced which are reasonable in a wide range of situations . based on these assumptions , section [ section : analysis ] proposes the mean square analysis . numerical experiments are performed to demonstrate the theoretical derivation in section [ section : simulation ] and the conclusion is drawn in section [ section : conclusion ] . the unknown coefficients and input signal at time instant @xmath3 are denoted by @xmath4^{\rm t}$ ] and @xmath5^{\rm t}$ ] , respectively , where @xmath6 is the filter length . the observed output signal is @xmath7 where @xmath8 denotes the additive noise . the estimated error between the output of unknown system and of the adaptive filter is @xmath9 where @xmath10^{\rm t}$ ] denotes the adaptive filter tap - weights . in order to take the sparsity of the unknown coefficients into account , @xmath0-lms @xcite inserts an @xmath0 norm penalty into the cost function of standard lms . the new cost function is @xmath11 where @xmath12 is a factor to balance the estimation error and the new penalty . due to the np hardness of @xmath0 norm optimization , a continuous function is usually employed to approximate @xmath0 norm . taking the popular approximation@xcite and making use of the first order taylor expansion , the recursion of @xmath0-lms is @xmath13 where @xmath14^{\rm t}$ ] and @xmath15 the last item in ( [ w_update ] ) is called _ zero - point attraction _ @xcite , @xcite , because it reduces the distance between @xmath16 and the origin when @xmath17 is small . according to ( [ l_0gradapprx ] ) and fig . [ fig : g_l0l1](a ) , obviously such attraction is non - linear and exerts varied affects on respective tap - weights . this attraction is effective for the tap - weights in the interval @xmath18 , which is named _ attraction range_. in this region , the smaller @xmath17 is , the stronger attraction affects . za - lms ( or sparse lms ) @xcite runs similar as @xmath0-lms . the only difference is that the sparse penalty is changed to @xmath2 norm . accordingly the zero - point attraction item of the former is defined as @xmath19 which is shown in fig . [ fig : g_l0l1](b ) . the recursion of za - lms is @xmath20 where @xmath21 is the parameter to control the strength of sparsity penalty . comparing the sub figures in fig . [ fig : g_l0l1 ] , one can readily accept that @xmath22 exerts the various attraction to respective tap - weight , therefore it usually behaves better than @xmath23 . in the following analysis , one will read that za - lms is a special case of @xmath0-lms and the result of this work can be easily extended to the case of za - lms . as its improvement , reweighted za - lms ( rza - lms ) is also proposed in @xcite , which modifies the zero - point attraction term to @xmath24 where parameter @xmath25 controls the similarity between ( [ rzagradapprx ] ) and @xmath0 norm . please refer to fig . [ fig : g_l0l1](c ) for better understanding the behavior of ( [ rzagradapprx ] ) . in section v , both za - lms and rza - lms are simulated for the purpose of performance comparison . -lms , ( b ) za - lms , ( c ) rza - lms.,width=307 ] denote @xmath26 and @xmath27 as the steady - state msd and instantaneous msd after @xmath3 iterations for lms with zero - mean independent gaussian input , respectively . the steady - state msd has the explicit expression @xcite of @xmath28 where @xmath29 and @xmath30 denote the power of input signal and additive noise , respectively , and @xmath31 is a constant defined by ( [ definedeltal ] ) in appendix [ const ] . for the convergence process , the explicit expression of instantaneous msd is implied in @xcite as @xmath32 the next one turns to za - lms , @xmath33 is used to denote the steady - state msd with white gaussian input . reference @xcite reaches the conclusion that @xmath34 where @xmath35 is the solution to @xmath36 where @xmath37 denotes the number of non - zero unknown coefficients and @xmath38 is a constant defined by ( [ definedelta0 ] ) . @xmath0-lms employs steepest descent recursively and is applicable to solving sparse system identification . more generally , steepest ascent iterations are used in several algorithms in the field of sparse signal processing . for example , researchers developed smoothed @xmath0 method @xcite for sparse decomposition , whose iteration includes a steepest ascent step and a projection step . the first step is defined as @xmath39 where @xmath40 serves as step size , @xmath41^{\rm t}$ ] denotes the negative derivative to an approximated @xmath0 norm and takes the value @xmath42 after ( [ sl0_1 ] ) , a projection step is performed which maps @xmath43 to @xmath44 in the feasible set . it can be seen that ( [ sl0_1 ] ) performs steepest ascent , which is similar to zero - point attraction in @xmath0-lms . the iteration details and performance analysis of this algorithm are presented in @xcite and @xcite , respectively . another algorithm , named iterative bayesian@xcite , also enjoys steepest ascent iteration as @xmath45 where @xmath40 denotes the step size and @xmath6 is a log posterior probability function . analysis of this algorithm and its application to sparse component analysis in noisy scenario are presented in @xcite . considering the nonlinearity of zero - point attraction , some preparations are made to simplify the mean square performance analysis . because various affects are exerted in @xmath0-lms to the filter tap - weights according to their respective system coefficients , it would be helpful to classify the unknown parameters , correspondingly , the filter tap - weights , into several categories and perform different analysis on each category separately . according to the attraction range and their strength , all system coefficients are classified into three categories as @xmath46 where @xmath47 . obviously , @xmath48 and @xmath49 . in the following text , derivations are firstly carried out for the three sets separately . then a synthesis is taken to achieve the final results . the following assumptions about the system and the predefined parameters are adopted to enable the formulation . * input data @xmath50 is an _ _ zero - mean gaussian signal . * tap - weights @xmath51 , input vector @xmath52 , and additive noise @xmath8 are mutually independent . * the parameter @xmath53 is so small that @xmath54 . assumption ( i ) commonly holds while ( ii ) is the well - known independence assumption @xcite . assumption ( iii ) comes from the experimental observations , i.e. , a too large @xmath53 can cause much bias as well as large steady - state msd . therefore , in order to achieve better performance , @xmath53 should not be too large . besides the above items , several regular patterns are supposed during the convergence and the steady state . * all tap - weights , @xmath51 , follow gaussian distribution . * for @xmath55 , the tap - weight @xmath56 is assumed to have the same sign with the corresponding unknown coefficient . * the adaptive weight @xmath56 is assumed out of the attraction range for @xmath57 , while in the attraction range elsewhere . assumption ( iv ) is usually accepted for steady - state behavior analysis @xcite . assumption ( v ) and ( vi ) are considered suitable in this work due to the following two aspects . first , there are few taps violating these assumptions in a common scenario . intuitively , only the non - zero taps with rather small absolute value may violate assumption ( v ) , while assumption ( vi ) may not hold for the taps close to the boundaries of the attraction range . for other taps which make up the majority , these assumptions are usually reasonable , especially in high snr cases . second , assumptions ( v ) and ( vi ) are proper for small steady - state msd , which is emphasized in this work . the smaller steady - state msd is , the less tap - weights differ from unknown coefficients . therefore , it is more likely that they share the same sign , as well as on the same side of the attraction range . based on the discussions above , those patterns are regarded suitable in steady state . for the convergence process , due to fast convergence of lms - type algorithms , we may suppose that most taps will get close to the corresponding unknown coefficients very quickly , so these patterns are also employed in common scenarios . as we will see later , some of the above assumptions can not always hold in whatever parameter setting and may restrict the applicability of some analysis below . however , considering the difficulties of nonlinear algorithm performance analysis , these assumptions can significantly enable mathematical tractability and help obtain results shown to be precious in a large range of parameter setting . thus , we consider these assumptions reasonable to be employed in this work . based on the assumptions above , the mean and mean - square performances of @xmath0-lms are analyzed in this section . define the misalignment vector as @xmath58 , combine ( [ expout ] ) , ( [ esterror ] ) , and ( [ w_update ] ) , one has @xmath59 taking expectation and using the assumption ( ii ) , one derives @xmath60 where _ overline _ denotes expectation . * for @xmath57 , utilizing assumption ( vi ) , one has @xmath61 . * for @xmath62 , combining assumptions ( iii ) , ( v ) and ( vi ) , it can be derived that @xmath63 * for @xmath64 , noticing the fact that @xmath65 has the opposite sign with @xmath66 in interval @xmath67 and using assumptions ( iv ) and ( vi ) , it can be derived that @xmath61 . thus , the bias in steady state is obtained @xmath68 in steady state , therefore , the tap - weights are unbiased for large coefficients and zero coefficients , while they are biased for small coefficients . the misalignment depends on the predefined parameters as well as unknown coefficient @xmath69 itself . the smaller the unknown coefficient is , the larger the bias becomes . this tendency can be directly read from fig . [ fig : g_l0l1](a ) . in the attraction range , the intensity of the zero - point attraction increases as tap - weights get more closing to zero , which causes heavy bias . thus , the bias of small coefficients in steady state is the byproduct of the attraction , which accelerates the convergence rate and increases steady - state msd . the condition on mean square convergence and steady - state msd are given by the following theorem . [ theo1 ] in order to guarantee convergence , step - size @xmath40 should satisfy @xmath70 and the final mean square deviation of @xmath0-lms is @xmath71 where @xmath72 are defined in ( [ beta1])@xmath73([beta3 ] ) in appendix [ const ] , respectively . the proof of theorem [ theo1 ] goes in appendix [ proof_steady_msd ] . _ remark 1 _ : the steady - state msd of @xmath0-lms is composed of two parts : the first item in ( [ msd_expression ] ) is exactly the steady - state msd of standard lms ( [ msdlms ] ) , while the latter two items compose an additional part caused by zero - point attraction . when @xmath53 equals zero , @xmath0-lms becomes the traditional lms , and correspondingly the additional part vanishes . when the additional part is negative , @xmath0-lms has smaller steady - state msd and thus better steady - state performance over standard lms . consequently , it can be deduced that the condition on @xmath53 to ensure @xmath0-lms outperforms lms in steady - state is @xmath74 _ remark 2 _ : according to theorem [ theo1 ] , the following corollary on parameter @xmath53 is derived . [ coro : opt_kappa ] from the perspective of steady - state performance , the best choice for @xmath53 is @xmath75{\frac{\beta_1+\beta_2}{\beta_1-\beta_2}}-\sqrt[4]{\frac{\beta_1-\beta_2}{\beta_1+\beta_2}}\right ) , \label{kappabest}\ ] ] and the minimum steady - state msd is @xmath76 the proof of corollary @xmath77 is presented in appendix [ best_parameters ] . please notice that in ( [ msd_opt ] ) , the first item is about standard lms and the second one is negative when @xmath78 is less than @xmath6 . therefore , the minimum steady - state msd of @xmath0-lms is less than that of standard lms as long as the system is not totally non - sparse . _ remark 3 _ : according to the theorem , it can be accepted that the steady - state msd is not only controlled by the predefined parameters , but also dependent on the unknown system in the following two aspects . first , the sparsity of the system response , i.e. @xmath78 and @xmath6 , controls the steady - state msd . second , significantly different from standard lms , the steady - state msd is relevant to the _ small coefficients _ of the system , considering the attracting strength appears in @xmath79 and @xmath80 . here we mainly discuss the effect of system sparsity as well as the distribution of coefficients on the minimum steady - state msd . based on the above results , the following corollary can be deduced . [ coro : steady_q ] the minimum steady - state msd of ( [ msd_opt ] ) is monotonic increasing with respect to @xmath78 and attracting strength @xmath81 . the validation of corollary @xmath82 is performed in appendix [ effect_sparsity ] . the _ zero - point attractor _ is utilized in @xmath0-lms to draw tap - weights towards zero . the more sparse the unknown system is , the less steady - state msd is . similarly , small coefficients are biased in steady state and deteriorate the performance , which explains that steady - state msd is increasing with respect to @xmath81 . _ remark 4 _ : according to ( [ convcondition ] ) , one knows that @xmath0-lms has the same convergence condition on step size as standard lms@xcite and za - lms@xcite . consequently the effect of @xmath40 on steady - state performance is analyzed . it is indicated in ( [ msdlms ] ) that the standard lms enhances steady - state performance by reducing step size@xcite . @xmath0-lms has a similar trend . for the seek of simplicity and practicability , a sparse system of @xmath78 far less than @xmath6 is considered to demonstrate this property . utilizing ( [ convcondition ] ) in such scenario , the following corollary is derived . [ coro : steady_mu ] for a sparse system which satisfies @xmath83 the minimum steady - state msd in ( [ msd_opt ] ) is further approximately simplified as @xmath84 where @xmath85 and @xmath86 are defined by ( [ eta6 ] ) in appendix [ const ] , and @xmath81 , defined by ( [ attractingstrength ] ) , denotes the attracting strength to the zero - point . furthermore , the minimum steady - state msd increases with respect to the step size . the proof of corollary [ coro : steady_mu ] is conducted in appendix [ effect_mu ] . due to the stochastic gradient descent and zero - point attraction , the tap - weights suffer oscillation , even in steady state , whose intensity is directly relevant to the step size . the larger the step size , the more intense the vibration . thus , the steady - state msd is monotonic increasing with respect to @xmath40 in the above scenario . _ remark 5 _ : in the scenario where @xmath87 remains a constant while @xmath88 approaches to zero , it can be readily accepted that ( [ w_update ] ) becomes totally identical to ( [ w_update_l1 ] ) , therefore @xmath0-lms becomes za - lms in this limit setting of parameters . in appendix [ relationship_sparse_standard_lms ] it is shown that the result ( [ shi_sparse_lms_msd ] ) for steady - state performance @xcite could be regarded as a particular case of theorem [ theo1 ] . as @xmath88 approaches to zero in @xmath0-lms , the attraction range becomes infinity and all non - zero taps belong to small coefficients which are biased in steady state . thus , za - lms has larger steady - state msd than @xmath0-lms , due to bias of all taps caused by uniform attraction intensity . if @xmath53 is further chosen optimal , the optimal parameter for za - lms is given by @xmath89 ( notice that @xmath90 approaches @xmath91 as @xmath88 tending to zero , as makes @xmath92 finite ) , and the minimum steady - state msd of @xmath0-lms ( [ msd_opt ] ) converges to that of za - lms . to better compare the three algorithms , the steady - state msds of lms , za - lms , and @xmath0-lms are listed in table [ table1 ] , where that of za - lms is rewritten and @xmath93 is defined in ( [ gamma_za ] ) in appendix [ relationship_sparse_standard_lms ] . it can be accepted that the steady - state msds of both za - lms and @xmath0-lms are in the form of @xmath94 plus addition items , where @xmath94 denotes the steady - state msd of standard lms . if the additional items are negative , za - lms and @xmath0-lms exceed lms in steady - state performance . [ cols="^,^,^,^,^ " , ] in the first experiment , the steady - state performance with respect to @xmath53 is considered . referring to fig . [ fig : steady_msd_sweep_kappa_40db ] , the theoretical steady - state msd of @xmath0-lms is in good agreement with the experiment results when snr is @xmath95db . with the growth of @xmath53 from @xmath96 , the steady - state msd decreases at first , which means proper zero - point attraction is helpful for sufficiently reducing the amplitude of tap - weights in @xmath97 . on the other hand , larger @xmath53 results in more intensity of zero - point attraction item and increases the bias of small coefficients @xmath98 . overlarge @xmath53 causes too much bias , thus deteriorates the overall performance . from ( [ kappabest ] ) , @xmath99 produces the minimized steady - state msd , which is marked with a square in fig . [ fig : steady_msd_sweep_kappa_40db ] . again , simulation result tallies with analytical value well . when snr is @xmath100db , referring to fig . [ fig : steady_msd_sweep_kappa_20db ] , the theoretical result also predicts the trend of msd well . however , since the assumptions ( v ) and ( vi ) do not hold well in low snr case , the theoretical result has perceptible deviation from the simulation result . -lms ( with respect to different @xmath53 ) , where snr is @xmath95db and the solid square denotes @xmath101.,title="fig:",width=384 ] + -lms ( with respect to different @xmath53 ) , where snr is @xmath100db and the solid square denotes @xmath101.,title="fig:",width=384 ] + in the second experiment , the effect of parameter @xmath88 on steady - state performance is investigated . please refer to fig . [ fig : steady_msd_sweep_alpha_40db ] for results . rza - lms is also tested for performance comparison , with its parameter @xmath21 chosen as optimal values which are obtained by experiments . for the sake of simplicity , the parameter @xmath25 in ( [ rzagradapprx ] ) is set the same as @xmath88 . simulation results confirm the validity of the theoretical analysis . with very small @xmath88 , all tap - weights are attracted toward zero - point and the steady - state msd is nearly independent . as @xmath88 increases , there are a number of taps fall in the attraction range while the others are out of it . consequently , the total bias reduces . besides , the results for za - lms are also considered in this experiment , with the optimal parameter @xmath21 proposed in _ remark 5_. it is shown that @xmath0-lms always yields superior steady - state performance than za - lms ; moreover , in scenario where @xmath88 approaches @xmath102 , the msd of @xmath0-lms tends to that of za - lms . in the parameter range of this experiment , @xmath0-lms shows better steady - state performance than rza - lms . ) , and @xmath0-lms ( with respect to different @xmath88 ) , where @xmath25 equals @xmath88 . parameters @xmath21 and @xmath53 are chosen as optimal for rza - lms and @xmath0-lms , respectively.,title="fig:",width=384 ] + the third experiment studies the effect of non - zero coefficients number on steady - state deviation . please refer to fig . [ fig : steady_msd_sweep_q ] . it is readily accepted that @xmath0-lms with optimal @xmath53 outperforms traditional lms in steady state . the fewer the non - zero unknown coefficients are , the more effectively @xmath0-lms draws tap - weights towards zero . therefore , the effectiveness of @xmath0-lms increases with the sparsity of the unknown system . when @xmath78 exactly equals @xmath6 , its performance with optimal @xmath53 already attains that of standard lms , indicating that there is no room for performance enhancement of @xmath0-lms for a totally non - sparse system . -lms ( with respect to different total non - zeros taps @xmath78 ) , where @xmath53 is chosen as optimal.,title="fig:",width=384 ] + the fourth experiment is designed to investigate convergence process with respect to @xmath53 . also , the learning curve of the standard lms is simulated . when snr is @xmath95db , the results in fig . [ fig : converegence_msd_sweep_kappa_40db ] demonstrate that our theoretical analysis of convergence process is generally in good accordance with simulation . it can be observed that different @xmath53 results in differences in both steady - state msd and the convergence rate . due to more intense zero - attraction force , larger @xmath53 results in higher convergence rate ; but too large @xmath53 can have bad steady - state performance for too much bias of small coefficients . moreover , @xmath0-lms outperforms standard lms in convergence rate for all parameters we run , and also surpasses it in steady - state performance when @xmath53 is not too large . when snr is 20db , fig . [ fig : converegence_msd_sweep_kappa_20db ] also shows similar trend about how @xmath53 influences the convergence process ; however , since the low snr scenario breaks assumptions ( v ) and ( vi ) , the theoretical results and experimental results differ to some extent . -lms ( with respect to different @xmath53 ) , where snr is @xmath95db.,title="fig:",width=384 ] + -lms ( with respect to different @xmath53 ) , where snr is @xmath100db.,title="fig:",width=384 ] + the fifth experiment demonstrates convergence process for various step sizes , with the comparison of lms and @xmath0-lms . please refer to fig . [ fig : converegence_msd_sweep_mu ] . similar to traditional lms , smaller step size yields slower convergence rate and less steady - state msd . therefore , the choice of step size should seek a balance between convergence rate and steady - state performance . furthermore , the convergence rate of @xmath0-lms is faster than that of lms when their step sizes are identical . -lms with respect to different step sizes @xmath40 , where @xmath53 is chosen as optimal for @xmath0-lms.,title="fig:",width=384 ] + the comprehensive mean square performance analysis of @xmath0-lms algorithm is presented in this paper , including both steady - state and convergence process . the adaptive filtering taps are firstly classified into three categories based on the zero - point attraction item , and then analyzed separately . with the help of some assumptions which are reasonable in a wide range , the steady - state msd is finally deduced and the convergence of instantaneous msd is approximately predicted . moreover , a parameter selection rule is put forward to minimize the steady - state msd and theoretically it is shown that @xmath0-lms with optimal parameters is superior than traditional lms for sparse system identification . the all - round theoretical results are verified in a large range of parameter setting through numerical simulations . in order to make the main body simple and focused , the explicit expressions of some constants used in derivations are listed here . all through this work , four constants of @xmath103 are used to simplify the expressions . to evaluate the zero - point attracting strength , with respect to the sparsity of the unknown system coefficients , two kinds of strengthes are defined as @xmath104 which are utilized everywhere in this work . considering the attraction range , it can be readily accepted that these strengthes are only related to the small coefficients , other than the large ones and the zeros . in lemma [ lemma ] , @xmath105 is defined as @xmath106,\label{matrix_a}\ ] ] and @xmath107^{\rm t},\label{vector_b}\ ] ] where @xmath108 where @xmath109 is the solution to ( [ quadratic_equ ] ) . in theorem [ theo2 ] , the constants @xmath110 and @xmath111 are @xmath112 in corollary [ coro : opt_kappa ] , the constants @xmath72 are @xmath113 in appendix [ effect_sparsity ] and [ effect_mu ] , the constants @xmath114 are @xmath115 denote @xmath116 to be msd at iteration @xmath3 , and @xmath117 to be the second moment matrix of @xmath118 , respectively , @xmath119 substituting ( [ h_update ] ) into ( [ secondmoment ] ) , and expanding the term @xmath120 into three second moments using the gaussian moment factoring theorem @xcite , one knows @xmath121 using the fact that @xmath122 , one has @xmath123 consequently , the condition needed to ensure convergence is @xmath124 and ( [ convcondition ] ) is derived directly , which is the same as standard lms and similar with the conclusion in @xcite . next the steady - state msd will be derived . using ( [ r_update ] ) and considering the @xmath125th diagonal element , one knows @xmath126 to develop @xmath127 , one should first investigate two items , namely @xmath128 and @xmath129 in ( [ msd_comp_steadystate ] ) . for @xmath57 , from assumption ( vi ) one knows @xmath130 , thus @xmath131 for small coefficients , considering assumptions ( v ) and ( vi ) , formula ( [ l_0gradapprx ] ) implies @xmath132 is a locally linear function with slope @xmath133 , which results in @xmath134 thus , it can be shown @xmath135 where @xmath136 is derived in ( [ h_steadystate_detail ] ) . then turning to @xmath64 , it is readily known that @xmath137 in this case . thus , from assumptions ( iv ) and ( vi ) , the following results can be derived from the property of gaussian distribution , @xmath138 combining assumption ( iii ) , ( [ h_steadystate_detail ] ) and ( [ ehgegg_lc])@xmath73([egg_sc ] ) , one can know the equivalency between ( [ msd_comp_steadystate ] ) and following equations for @xmath125 in @xmath139 , @xmath98 , and @xmath97 , respectively , @xmath140 where @xmath109 denotes @xmath141 for simplicity . summing ( [ msd_comp_steadystate_lc ] ) and ( [ msd_comp_steadystate_sc ] ) for all @xmath142 , and noticing that @xmath143 it could be derived that @xmath144 where @xmath81 is introduced in ( [ attractingstrength ] ) . combining ( [ msd_expression_old ] ) and ( [ msd_comp_steadystate_zc ] ) , it can be reached that @xmath109 is defined by the following equation @xmath145 finally , ( [ msd_expression ] ) is achieved after solving the quadratic equation above and a series of formula transformation on ( [ msd_expression_old ] ) . thus , the proof of theorem [ theo1 ] is completed . the vector @xmath187 in ( [ vector_b ] ) could be denoted as @xmath188 , \label{b_simp}\ ] ] where @xmath110 is defined in ( [ lambda3 ] ) and @xmath189 are constants . take @xmath190-transform for ( [ d_omega_iter ] ) , it can be derived that @xmath191 = ( z\mathbf{i}-\mathbf{a})^{-1 } z \left[\!\!\ ! d_0\\ \omega_0\end{array}\!\!\!\right ] + ( z\mathbf{i}-\mathbf{a})^{-1}{\mathbf{b}(z)},\ ] ] where @xmath192 . then combine the definition of @xmath193 in theorem [ theo2 ] and the above results , it is further derived @xmath194 where @xmath195 and @xmath196 are constants . take the inverse @xmath190-transform and notice the definition of @xmath197 , it finally yields @xmath198 thus we have completed the proof of ( [ d_conv_closedform ] ) . by forcing the equivalence between ( [ d_conv_closedform ] ) and lemma [ lemma ] , the expression of @xmath111 could be solved as ( [ c3 ] ) . by defining @xmath146 , ( [ msd_expression ] ) becomes @xmath147 where @xmath148 is defined as @xmath149 next we want to find @xmath150 which minimizes @xmath148 . forcing the derivative of @xmath148 with respect to @xmath66 to be zero , it can be obtained that @xmath151 combining @xmath152 and substituting @xmath153 in ( [ msd_theta ] ) , corollary [ coro : opt_kappa ] can be finally achieved . from ( [ msdlms ] ) , ( [ msd_opt ] ) , ( [ beta1 ] ) , ( [ beta2 ] ) , and ( [ eta0 ] ) , it can be obtained that @xmath154 note neither the @xmath26 defined in ( [ msdlms ] ) nor @xmath155 defined in ( [ eta0 ] ) is dependent on @xmath78 or @xmath156 , thus the focus of the proof is the denominator in ( [ msd_opt_eta ] ) . in the following , we will analyze the two items in the denominator separately and obtain their monotonicity . the first item in the denominator is @xmath157 from ( [ beta1_coro : steady_q ] ) , it has already shown that @xmath158 is increasing with respect to @xmath78 and @xmath156 . next we consider the second item . it can be obtained beforehand that @xmath80 and @xmath159 equal @xmath160 and @xmath161 , respectively . thus , one has @xmath162 further notice that @xmath163 it can be proved that all of the three items in the square root of ( [ beta1beta2_coro : steady_q ] ) are increasing with respect to @xmath78 and @xmath81 ; thus the second item in the denominator is monotonic increasing with respect to @xmath78 and @xmath81 . till now , the monotonicity of @xmath164 with respect to @xmath78 and @xmath81 has been proved . last , in the special scenario where @xmath78 exactly equals @xmath6 , it can be obtained that @xmath164 is identical to @xmath165 ; thus @xmath165 is larger than the minimum steady - state msd of the scenario where @xmath78 is less than @xmath6 . in sum , corollary [ coro : steady_q ] is proved . for a sparse system in accordance with ( [ sparse_condition ] ) , @xmath114 defined in appendix [ const ] are approximated by @xmath166 substituting @xmath114 in @xmath72 of ( [ msd_opt_eta ] ) , with the approximated expressions above , ( [ msd_simple_sparse_smallmu ] ) is finally derived after calculation . next we show @xmath164 in ( [ msd_simple_sparse_smallmu ] ) is monotonic increasing with respect to @xmath40 . since ( [ msd_simple_sparse_smallmu ] ) is equivalent with @xmath167 it can be directly observed from ( [ eta6 ] ) that larger @xmath40 results in larger numerator as well as smaller denominator in ( [ msd_simple_appdendix ] ) , which both contribute to the fact that @xmath164 is monotonic increasing with respect to @xmath40 . thus , the proof of corollary [ coro : steady_mu ] is arrived . define function @xmath199 then the roots of @xmath200 are eigenvalues of matrix @xmath201 . from ( [ matrix_a ] ) , it can be shown @xmath202 and @xmath203 where @xmath204 denote the entries of @xmath205 . thus , we know @xmath206 and @xmath207 , which indicates that one root of quadratic function @xmath200 is within the interval @xmath208 $ ] . similarly , another root of @xmath200 is in @xmath209 . thus , it can be concluded that the eigenvalues of @xmath201 are both in @xmath210 and satisfy @xmath211 for large step size scenario of @xmath212 , ( [ lambda3 ] ) and ( [ lambda12_bound_full ] ) yield @xmath213 through comparison between ( [ d_conv_closedform ] ) and ( [ d_conv_closedform_lms ] ) , one can know for large @xmath40 , all the three transient items in msd convergence of @xmath0-lms has faster attenuation rate than lms , leading to acceleration of convergence rate . when @xmath87 remains a constant while @xmath88 approaches zero , from ( [ w_update ] ) , ( [ l_0gradapprx ] ) , and ( [ l_1gradapprx ] ) , it is obvious that the recursion of @xmath0-lms becomes that of za - lms . furthermore , one can see that @xmath168 equals @xmath169 . from the definition , it can be shown @xmath139 is an empty set when @xmath88 approaching zero . consequently , @xmath170 combine ( [ msd_expression_old ] ) , ( [ quadratic_equ ] ) , and ( [ gsc_sum_l1 ] ) , then after quite a series of calculation , the explicit expression of steady - state msd becomes @xmath171 where @xmath93 is the discriminant of quadratic equation ( [ quadratic_equ ] ) , @xmath172 through a series of calculation , it can be proved that ( [ msd_l1_expression ] ) is equivalent with ( [ shi_sparse_lms_msd ] ) obtained in @xcite . thus , the steady - state msd in za - lms could be regarded as a particular case of that in @xmath0-lms . from ( [ r_update ] ) , the update formula is @xmath173 since lms algorithm has fast convergence rate , it is reasonable to suppose most filter tap - weights will get close to the corresponding system coefficient very quickly ; thus , the classification of coefficients @xmath69 could help in the derivation of the convergence situation of @xmath174 . for @xmath57 , from assumption ( vi ) , ( [ coeff_update ] ) takes the form @xmath175 for @xmath62 the mean convergence is firstly derived and then the mean square convergence is deduced . take expectation in ( [ h_update ] ) , and combine assumptions ( iii ) , ( v ) , and ( vi ) , one knows @xmath176 since @xmath177 , one can finally get @xmath178 combining ( [ coeff_update ] ) , ( [ mean_conv_sc ] ) and employing assumption ( iii ) , it can be achieved @xmath179 next turn to @xmath64 . from assumption ( iv ) , the following formula can be attained employing the steady state result and first - order taylor expansion @xmath180 where @xmath181 , which is the solution to equation ( [ quadratic_equ ] ) . finally , with assumption ( iii ) we have @xmath182 considering @xmath183 , and combine ( [ lc_conv]),([mean_conv_sc]),([sc_conv ] ) , and ( [ zc_conv ] ) , one can obtain ( [ d_omega_iter ] ) after a series of derivation . as for the initial value , since @xmath184 , by definition we have @xmath185 and @xmath186 . thus , lemma [ lemma ] is reached . the authors wish to thank laming chen and four anonymous reviewers for their helpful comments to improve the quality of this paper . d. angelosante , j. a. bazerque , and g. b. giannakis , `` online adaptive estimation of sparse signals : where rls meets the @xmath2-norm , '' _ ieee trans . signal process . 58 , no . 7 , pp . 3436 - 3447 , jul . 2010 . g. mileounis , b. babadi , n. kalouptsidis , and v. tarokh , `` an adaptive greedy algorithm with application to nonlinear communications , '' _ ieee trans . signal process . 6 , pp . 2998 - 3007 , jun . 2010 . y. kopsinis , k. slavakis , and s. theodoridis , `` online sparse system identification and signal reconstruction using projections onto weighted @xmath2 balls , '' _ ieee trans . signal process . 905 - 930 , mar . 2011 . e. cands , j. romberg , and t. tao , `` robust uncertainty principles : exact signal reconstruction from highly incomplete frequency information , '' _ ieee trans . inf . theory _ , 489 - 509 , feb . 2006 . h. zayyani , m. babaie - zadeh , and c. jutten , `` an iterative bayesian algorithm for sparse component analysis ( sca ) in presence of noise , '' _ ieee trans . signal process . 4378 - 4390 , oct . j. jin , y. gu , and s. mei , `` a stochastic gradient approach on compressive sensing signal reconstruction based on adaptive filtering framework , '' _ ieee j. sel . topics signal process . _ , 409 - 420 , apr . 2010 . l. horowitz and k. senne , `` performance advantage of complex lms for controlling narrow - band adaptive arrays , '' _ ieee trans . acoust . , speech , signal process . assp-29 , no . 722 - 736 , jun . 1981 . s. k. zhao , z. h. man , s. y. khoo , and h. r. wu,``stability and convergence analysis of transform - domain lms adaptive filters with second - order autoregressive process , '' _ ieee trans . signal process . 119 - 130 , jan . 2009 . a. ahlen , l. lindbom , and m. sternad , `` analysis of stability and performance of adaptation algorithms with time - invariant gains , '' _ ieee trans . signal process . 103 - 116 , jan . 2004 . s. c. chan and y. zhou , `` convergence behavior of nlms algorithm for gaussian inputs : solutions using generalized abelian integral functions and step size selection , '' _ journal of signal processing systems _ , vol . 3 , pp . 255 - 265 , 2010 . s. c. chan and y. zhou , `` on the performance analysis of the least mean m - estimate and normalized least mean m - estimate algorithms with gaussian inputs and additive gaussian and additive gaussian and contaminated gaussian noises , '' _ journal of signal processing systems _ , vol . 81 - 103 , 2010 . s. koike , `` convergence analysis of a data echo canceler with a stochastic gradient adaptive fir filter using the sign algorithm , '' _ ieee trans . signal process . 2852 - 2861 , dec . 1995

at the time when giant planets form , the mass ( 99% ) of the protoplanetary disk is dominated by gas . dust is a minor constituent of the mass of the disk . still , it dominates the opacity ; consequently , it is much easier to observe . therefore , most of the observational constraints of disks have been deduced from the study of dust emission ( e.g. , see chapter by henning et al . ) . in contrast , direct observational constraints of the gas in the disk are relatively scarce . nonetheless , to obtain direct information from the gas content of the disk is crucial for answering fundamental questions in planet formation such as : how long does the protoplanetary disk last ? , how much material is available for forming giant planets ? , how do the density and the temperature of the disk vary as a function of the radius ? , what are the dynamics of the disk ? . although we have learned important insights from disks from dust observations ( see for example the reviews by henning et al . 2006 and natta et al . 2007 ) , dust presents several limitations : ( i ) dust spectral features are broad ; consequently , dust emission does not provide kinematical information ; ( ii ) dust properties are expected to change during the planet formation process ; therefore , quantities such as the gas - to - dust ratio ( needed to derive the disk mass from dust continuum emission in the ( sub)-mm ) are expected to strongly vary with respect to the conditions of the interstellar medium ( ism ) . in addition , dust signatures are strongly related to dust size . in particular , as soon a dust particle reaches a size close to a decimeter it becomes practically `` invisible '' . for example , a crucial quantity such as the disk dissipation time scale is normally deduced from the decline of the fraction of the sources presenting near - ir ( jhkl ) excess in clusters of increasing ages ( e.g. , haisch et al . in reality , the time scale that we want to constrain is not the time scale in which the small warm dust particles disappear from the surface layers in the inner disk , but the time scale in which the gas in the disk disappears . in summary , independent information of the gas is crucial to understand protoplanetary disk structure and _ combined studies of gas and dust in the disk are needed_. in particular , we are interested in deriving constraints for the inner disk , the region where the planets are expected to form ( r@xmath020 au ) . the observational study of the gas in the inner disk is a relatively new emerging topic . only with the advent of ground - based high - spectral resolution infrared spectrographs and space - born infrared spectrographs has this research become possible . the main tool used to study the gas in the disk is molecular spectroscopy . the gas in the disk is heated by collisions with dust , shocks or uv or x - rays from the central star . the heated molecules populate different rotational and vibrational levels and when they de - excite emission lines are produced . the key is that these emission lines have the imprint of the physical conditions of the gas where they originated . important constraints can be derived from emission lines : ( i ) line ratios of different transitions constrain the excitation mechanism ( shocks , uv or x - rays ) and the temperature of the gas responsible of the emission ; ( ii ) if the emitting gas is optically thin , then the measured line fluxes are proportionally related to the amount of molecules present ; therefore , line fluxes can be used to derive column densities and gas masses ; ( iii ) line shapes and spatial extended provide constraints on the size and geometry of emitting region as well as the dynamics of the emitting gas . figure 1 displays a pedagogical example of how a line profile changes as a function of the inclination and size of the emitting region in the disk . the line width and double peak separation depend on the inclination . the line profile wings depend on the innermost radius of the emission . for the interested reader , recent reviews by najita et al . ( 2000 & 2007a ) , blake ( 2003 ) and carr ( 2005 ) cover the subject of observational studies of gas in disks . 20 au ) and a cold outer region ( r@xmath120 au ) . near - ir and mid - ir diagnostics probe the inner disk , ( sub)-mm diagnostics probe the outer disk . the disk has a vertical structure : a dense mid - plane and a hotter less dense surface layer . near - ir and mid - ir gas and dust emission features originate in the optically thin surface layer of the inner disk . near - ir and mid - ir continuum originates from the optically thick interior layer . at ( sub)-mm wavelengths we observe line emission from optically thick co gas and optically thin emission of cold dust located in the outer disk.,title="fig:",scaledwidth=50.0% ] [ fig:2 ] an important element that one should always bear in mind when studying molecular emission from disks is the region where the emission observed is produced . as the disk has a radial and vertical temperature and density gradient ( hotter closer to the star and higher in the disk , denser closer to the star and deeper in the disk ) we need to have clear three concepts : ( i ) inner and outer disk , ( ii ) interior and surface layer , ( iii ) optical depth of the medium ( i.e. , optically thin or thick ) and its wavelength and molecular species dependency . the first concept is related to the distance of the central star ( i.e. , the radius ) and the radial temperature gradient . different gas tracers probe different radial regions of the disk . the hot and warm ( t@xmath2100 - 2500 k ) _ inner disk _ ( r@xmath020 au ) is traced by lines in the uv , near- and mid - ir . the cold ( t@xmath0100 k ) _ outer disk _ ( r@xmath120 au ) is probed by lines in the ( sub)-mm . the second and third concepts are related to the vertical temperature and density gradient . as a first approximation , we can understand a protoplanetary disk at each radius as composed of two vertical layers : a dense colder _ interior layer _ and less dense hotter _ surface layer _ ( chiang and goldreich 1997 ) . at near - ir and mid - ir wavelengths ( that probe r@xmath020 au ) , the dust and gas in the interior layer are optically thick and the interior layer radiates as a black body ( i.e. , continuum equal to the blackbody level ) . gas lines -and dust emission features- in the uv , near - ir and mid - ir arise from the optically thin surface layer of the inner disk . they and are observed on top of the continuum of the interior layer . emission is observed from the surface layer because the surface layer is hotter than the interior layer . the gas emission from the surface layer could be optically thin or optically thick depending on the gas density and the molecular species ( for example co emission is generally optically thick , and h@xmath3 emission is optically thin ) . optically thick gas lines from the surface layer of the inner disk are observable because the dust in the surface layer is optically thin . at ( sub)-mm wavelengths ( that probe r@xmath120 au ) , the dust become optically thin in the interior layer as well ( i.e. , continuum below the black - body level ) , and we can observe optically thick emission from - co - gas from the interior layer on top of the optically thin dust continuum . optically thin gas emission can be observed from other less abundant molecular species such as @xmath4co . figure 2 presents a cartoon of the structure of a disk and the observable properties of the medium . in addition to the surface layer , in particular cases such as transitional disks or disks in close binaries , gas emission lines ( optically thin or thick depending on the gas density ) are expected from the inner region of the disk when the dust is optically thin . given that the disk has a radial temperature gradient , different tracers are employed to probe the gas at each location in the disk . figure 3 displays a summary of disks gas diagnostics . in the following sections , i will explain one by one each of them highlighting their strengths and limitations . i will follow a discussion inspired by the historical development of the field . properly speaking h@xmath5 in emission ( 6590 ) from young stars does not trace the gas in the disk . however , since h@xmath5 emission is one of the primary indicators of accretion of gas from the disk onto the star ( e.g. , hartmann et al . 1994 ; muzerolle et al . 1998b ) , h@xmath5 is a robust indicator of a gaseous disk . in particular , when any other gas tracer fails , h@xmath5 emission tells us unambiguously that there is gas . its main advantages are : ( i ) it is easy to observe ( a small telescope with a medium resolution r@xmath23000 spectrograph can do it ) ; ( ii ) it is a robust diagnostic ; ( iii ) it can be studied in large samples of objects of diverse brightness in relatively short time . its main limitations are : ( i ) it does not provide information on the physical conditions of the gas in the disk ( h@xmath5 tells us that there is gas and allows us to deduce the accretion rate , but not much more ) ; ( ii ) if the accretion rate is below 10@xmath6m@xmath7/yr , h@xmath5 in emission is too weak to be observed ; ( iii ) in low resolution spectra of late type stars ( k and later ) , h@xmath5 in emission from accretion can be confused with h@xmath5 due to stellar activity ( typically a line width higher than 100 km / s is required for inferring accretion ) . h@xmath5 in emission allowed us to confirm that very low mass stars have disks , that relatively old objects still have gaseous disks , and that several transition disks still have gas in their inner disks ( e.g. , sicilia - aguilar et al . . note that other h lines such as the weaker br@xmath8 line in the near - ir have been used for deducing accretion rates -and the presence of a gaseous disk- in young stars ( e.g. , najita et al . 1996a , muzerolle et al . 1998a , natta et al . 2006 ) . one of the earliest indicators of gas in a disk was the detection of co overtone ( @xmath10 = 2 ) emission at 2.3 @xmath9 m in t tauri stars ( scoville et al . 1983 ; geballe and persson 1987 ; carr et al . 1989 , 1992 & 1993 ; chandler 1993 ; najita et al . more recently thi et al . ( 2005a ) reported co 2.3 @xmath9 m emission in the herbig ae / be star 51 oph ( see also berthoud et al . 2007 and tatulli et al . 2008 ) . co bandhead emission traces very hot ( t@xmath12000 k ) and dense gas in the innermost part of the disk ( r@xmath00.1 au ) in objects exhibiting high accretion . this line provided the first evidence of keplerian rotation of gas in the inner disk of t tauri stars . the co band - head emission is characterized by a broad feature at 2.3 @xmath9 m . its particular shape can be modeled by the convolution of @xmath11=2 - 0 co ro - vibrational emission lines at rest with the double peaked line profile characteristic of emission from a disk . the main advantages of this diagnostic are : ( i ) it can be studied with a low resolution near - ir spectrographs and ( ii ) it traces the innermost part of the disk . the main disadvantage is that the densities and temperatures require to be very high to produce the emission . this is the main reason for why this line has been observed towards only a hand - full of objects . finally , note that in general the co overtone emission at 2.3 @xmath9 m is optically thick emission . for the interested reader this diagnostic and its modeling is treated in detail in carr et al . ( 1993 ) . progress in the sensitivity of millimeter radio telescopes led to the detection in the early 90 s of pure rotational emission ( @xmath12 ) from cold co from t tauri stars ( e.g. , weintraub et al . 1989 ; koerner et al . 1993 ; skrutskie et al . 1993 ; beckwith & sargent 1993 ; guilloteau & dutrey 1994 ) . those observations revealed the classical double peaked profile expected from gas rotating in keplerian disk ( see figure 1 ) . the detections with single dish telescopes were followed by studies with millimeter interferometers that allowed to spatially resolve the emission in t tauri stars ( e.g. , dutrey et al . 1996 ; guilloteau & dutrey 1998 ; dutrey et al . 1998 ) and herbig ae / be stars ( e.g. , mannings & sargent 1997 ) . in the particular case of herbig ae stars , millimeter interferometry observations of co permitted to settle the controversy about the existence of circumstellar disks around these objects . since then a multitude of studies in the ( sub)-mm have been undertaken in protoplanetary disks thanks to development of sensitive millimiter and sub - mm arrays ( e.g. , plateau de bure , sma , carma , see recent review by dutrey et al . progress has not only been done in tracing cold co , but also in detecting many other molecules such as cn , hcn , hco@xmath13 , h@xmath3co , etc ( e.g. , thi et al . 2004 ) . these detections have opened the door to observational studies of chemistry in disks ( e.g. , review by bergin et al . 2007 ) . the principal advantages of co are : ( i ) co is the second most abundant molecule after the h@xmath3 . ( ii ) co ( sub)-mm emission is bright . these not trivial characteristics had allowed the survey of relatively large samples of young stars with disks ( e.g. , dent et al . 2005 , andrews & williams 2005 ) , and have permitted the spatially resolved study of the disk structure in several objects ( e.g. , pitu et al . the main limitation of co emission and other gas tracers in the ( sub)-mm is that they trace the cold gas in the outer part of the disk ( r@xmath120 au ) . therefore , they are not suitable to trace the disk region ( r@xmath020 au ) where planets are expected to form . nonetheless , most of the mass of the disk is located in the cold outer part of the disk . co emission in the ( sub)-mm has important limitations as tracer of gas mass : ( i ) cold co is expected to freeze - out onto dust grain surfaces at low temperatures ( t@xmath020 k ) and high densities ( n@xmath14 @xmath15 ) ; therefore , co is expected to be depleted ; ( ii ) co emission can be optically thick . both limitations introduce a large uncertainty on the conversion factor ( up to a 1000 ) between the co observed and the real amount of h@xmath3 present in the outer disk . typically , the ism co / h@xmath3 conversion factor of 10@xmath16 is used . however , this value is most likely incorrect for disks . the large uncertainty in the determination of mass of gas affects seriously the empirical estimation of the gas - to - dust ratio . finally , note that there are sources which are known to have a disk - by the detection of ir & ( sub)-mm excess and h@xmath5 in emission - , but , because of sensitivity limitations or depletion , we do not detect co emission . in the near future , significant progress is expected in the study of disks in the ( sub)-mm domain thanks to the atacama large millimiter / submillimiter array ( alma ) , which will start operations soon . alma will increase the sensitivity and spatial resolution by an order of magnitude with respect to those obtained with present facilities . in addition , the far - infrared space telescope herschel will allow the study of disks at frequencies which are impossible to observe from the ground . a detailed description of gas diagnostics in the ( sub)-mm and chemistry in disks deserves a review in its own . for the interested reader , the reviews by dutrey et al . ( 2007 ) and bergin et al . ( 2007 ) cover these topics . m vs corotation radius of the sources . in black squares co 4.7 @xmath9 m detections ( najita et al . 2007a ) , in filled circles the inner dust radius from near - ir interferometry ( akeson et al . 2005a , b. ) , in open circles r@xmath17 from sed modeling ( muzerolle et al . in several sources co gas is present inside the corotation radius and inward the dust sublimation radius . _ right panel : _ distributions of the gas inner radius from co 4.7 @xmath9 m observations and the distribution of the semi - mayor axis of short period extrasolar planets . the minimum radius of both distributions is similar . both figures are taken from the review from najita et al . ( 2007a).,scaledwidth=50.0% ] the advent of high - resolution ( r@xmath110000 ) spectrographs in the near - ir opened the door to the study of the co ro - vibrational emission band ( @xmath18 ) at 4.7 @xmath9 m . co emission at 4.7 @xmath9 m probes the gas at temperatures ranging from hundred to thousand of degrees . this diagnostic is important because the gas at these temperatures is located in the terrestrial planet forming region of the disk ( r @xmath05 au ) . co 4.7 @xmath9 m emission from disks has been studied in t tauri stars ( e.g. , carr et al . 2001 ; najita et al . 2003 ; brittain et al . 2005 & 2007b ; rettig et al . 2005 ; brown et al . 2005 ) , herbig ae / be stars ( e.g. , brittain et al . 2002 , 2003 & 2007a ; blake & boogert 2004 ; carmona et al . 2005 ; goto et al . 2006 ; van der plas et al . 2008b ) and in a growing number of transitional disks ( rettig et al . 2004 ; salyk et al . 2007 ; najita et al . 2008 ; pontoppidan et al . 2008 ) . the detection of the co band at 4.7 @xmath9 m allowed in the first place to constrain the temperature of the gas by means of rotational diagrams . the gas temperatures observed varied from 100 k to 1000 k and in a few cases up to 3000 k. in several sources , for example ab aur ( brittain et al . 2003 ) , the shape of the rotational diagram allowed to infer co gas at two different temperatures . the shape of the co lines was used to set constraints on the inclination of inner disks ( e.g. , blake & boogert 2004 ) . in general , within the errors , the inclinations deduced from co 4.7 @xmath9 m emission were consistent with the inclinations derived from ( sub)-mm observations ( note that to derive the disk inclination , the co emitting region needs to be assumed ) . modeling of the shape of the line profile allowed to estimate the innermost radius where the co emission is produced ( see figure 1 ) . najita et al . ( 2007a ) found that the inner radius of the co 4.7 @xmath9 m emission in t tauri stars extends inside the dust sublimation radius and , typically , inside the corotation radius too . those authors compared the distribution of the inner radii deduced from the co 4.7 @xmath9 m lines and the distribution of the orbital radii of short period extrasolar planets . they found that both distributions are very similar ( see figure 4 ) . the gas required for giant planet migration was put in evidence by the co fundamental transitions . = 2 - 1 emission at 4.7 @xmath9 m from the transition disk sr 21 . data in three slit orientations is displayed . in the bottom of the right panel , a detail of the transitions ( r and p ) employed for calculating the composite profile is shown . in all the panels the continuous line is a keplerian disk model . the line shape and spectro - astrometric signal observed are consistent with co gas in keplerian rotation . adapted from pontoppidan et al . ( 2008 ) . , scaledwidth=50.0% ] the most recent development has been to combine high spectral resolution with the enhanced spatial resolution provided by adaptive optics ( ao ) . goto et al . ( 2006 ) spatially resolved the co @xmath11=2 - 1 4.7 @xmath9 m emission from the herbig ae / be star hd 141569 and measured an inner clearing of radius 11@xmath192 au . pontoppidan et al . ( 2008 ) and van der plas et al . ( 2008 ) observed young stars with disks combining ao and very high spectral resolution ( r@xmath2100000 ) . those authors , employing the spectroastrometry technique , were able to measure the displacement of the psf center as a function of wavelength in the line , and obtained robust evidence of emission from keplerian disks ( see figure 5 ) . pontoppidan et al . ( 2008 ) detected gas emission from the inner gaps of transitional disks ( sr 21 , hd 135344b an tw hya ) . besides the fact that co 4.7 @xmath9 m emission probes the terrestrial planet forming region of the disk , one clear advantage of co 4.7 @xmath9 m emission is that it is detectable in a large number of sources with present instrumentation . this opens the perspective for developping large surveys . for example , co 4.7 @xmath9 m emission has been detected in a vast majority of herbig ae / be stars that display excess @xmath201 ( brittain et al . co 4.7 @xmath9 m emission studies have in consequence an important potential for the future . a second advantage comes from the fact that several lines are observable at the same time in a single setup even with echelle spectrographs . this is important to make reliable rotational diagrams and to perform analyses involving the averaging of line profiles ( i.e. , determine r@xmath17 or the inclination ) . the principal limitations of this diagnostic are : ( i ) it requires high - spectral resolution . thus it is limited to relatively bright targets ( for the moment stars fainter than m=10 are challenging ) ; ( ii ) observations in the m band are relatively time intensive and observations of telluric standard stars are required before or after the science observations ; ( iii ) co 4.7 @xmath9 m lines lie very close , if not inside co atmospheric features . hence , observations should be performed in an epoch such that the velocity shift due to the orbital motion of the earth is maximized ; ( iv ) one should keep in mind that co 4.7 @xmath9 m emission does not trace in general ( unless we see emission from a gap ) the giant planet forming region of the disk ( r@xmath15 - 10 au ) . from previous sections , the reader perhaps realizes that principally co has been discussed . the discussion started with co at 2.3 @xmath9 m , it was followed with co in the ( sub)-mm , and finally co emission at 4.7 @xmath9 m was treated . but what about molecular hydrogen ? h@xmath3 is by far more abundant than co ( in the ism h@xmath3/co is typically 10@xmath21 ) . although h@xmath3 is the principal constituent of the gas in the disk , it is very challenging to detect . this is due to the h@xmath3 physical nature : h@xmath3 is an homonuclear molecule that - in contrast to co - has not permanent dipole moment . its fundamental transitions are quadrupole in nature . hence , their einstein spontaneous emission coefficients are small and give rise to weak lines . in the context of circumstellar disks , one extra challenge is present . the weak h@xmath3 lines should be detected on top of the strong dust continuum emission . thus , high spectral resolution is a must to disentangle the weak lines . to complicate a little bit the life of astronomers , the energy levels of h@xmath3 are widely spaced in energy ( for example the first energy level starts at 510 k ) . that means that a dedicated observation is required for each h@xmath3 line in order to construct rotational diagrams ( note that in contrast , co lines are closely spaced and several lines are observable within one observation set up ) . moreover , since h@xmath3 emission is produced in warm t@xmath22 k gas , h@xmath3 emission can also originate from shocked gas in outflows and not in the disk ( note that historically h@xmath3 emission has been used to trace shocks ) . finally , the mid - ir region is a challenging window to perform high - sensitivity observations from the ground . nevertheless , besides being the most abundant molecule , h@xmath3 has some advantages : ( i ) h@xmath3 is optically thin up to column densities of 10@xmath23 @xmath24 ; thus , if the observed emission is thermal , we can derive direct constrains to the mass of the emitting gas ; ( ii ) h@xmath3 self - shields against photo - dissociation ; ( iii ) h@xmath3 traces the inner disk in the giant planet formation region ; ( iv ) the h@xmath3 condensation temperature is 2 k. hence , h@xmath3 does not freeze onto dust grains surfaces -as happens with co- ; ( v ) the velocity shift of h@xmath3 lines and their spatial extent allow to distinguish if the emission observed is from a disk or from an outflow . to discuss the observations of h@xmath3 emission in disks , i will start by addressing the near - ir ro - vibrational lines , later i will discuss the mid - ir pure rotational lines and finally i will discuss the h@xmath3 electronic transition lines in the uv . h@xmath3 lines in the near - ir probe the inner disk from a fraction of au up to a few au . they trace gas at temperatures of 1000 k and higher . they are sensitive up to a few lunar masses of gas . near - ir emission arises from ro - vibrational transitions between the rotational levels ( j ) of the first and second vibrational states ( @xmath11=1,2 ) and the rotational levels of the ground vibrational state ( @xmath11=0 ) . typical lines observed are the @xmath11=1 - 0 s(1 ) line ( j=3 - 1 ) at 2.12 @xmath9 m , the 1 - 0 s(0 ) line at 2.22 @xmath9 m , and the 2 - 1 s(1 ) line at 2.24 @xmath9 m . the h@xmath3 1 - 0 s(1 ) line was first observed in outflows in t tau stars ( e.g. , beckwith et al . 1978 ; van langevelde et al . 1994 ) . only recently , h@xmath3 1 - 0 s(1 ) quiescent emission ( i.e. , at the velocity of the star ) from a disk was reported in the t tauri star tw hya ( weintraub et al . later on bary et al . ( 2002 ) detected the same line in the weak line t tauri star(wtts ) doar 21 . after these initial studies , quiescent h@xmath3 1 - 0 s(1 ) emission from t tauri stars has been reported by bary et al . ( 2003 ) , itoh et al . ( 2003 ) , weintraub et al . 2005 , ramsay howat & greaves ( 2007 ) , carmona et al . ( 2007 , 2008a ) and bary et al . ( carmona et al . ( 2007 ) detected for the first time the h@xmath3 1 - 0 s(0 ) line in the disk of lkh@xmath5 264 and demonstrated empirically that h@xmath3 near - ir from disks originates in gas at temperatures around 1000 k. equivalent width vs x - ray luminosity in ctts and wtts in which 1 - 0 h@xmath3 s(1 ) emission at 2.12 micron has been detected . filled circles represent detections , empty circles non - detections , left arrow upper limits in x - ray luminosity , asterisks transition disks . the boundary ctts / wtts is set represented by the horizontal dashed line at ew 10 . the probability of detecting h@xmath3 in ctts seems to be correlated with accretion signatures ( h@xmath5 , @xmath25 excess ) . in wtts h@xmath3 emission is observed preferentially towards sources with high x - ray luminosity . figure taken from bary et al . ( 2008 ) . , scaledwidth=45.0% ] in classical t tauri stars ( ctts ) h@xmath3 near - ir lines are preferentially observed in objects that exhibit signatures of active accretion , that is , objects with large h@xmath5 equivalent widths and strong uv excess ( carmona et al . 2007 , bary et al . 2008 , see figure 6 ) . in ctts there is no apparent correlation between the x - ray luminosity , disk mass and the presence of the near - ir h@xmath3 lines ( carmona et al . the h@xmath3 1 - 0 s(1 ) line has been detected in a few wtts ( bary et al . 2002 , bary et al . 2008 , see figure 5 ) . this indicates that a fraction of such objects still have a gaseous disk . the wtts in which the h@xmath3 1 - 0 s(1 ) line has been detected are sources exhibiting large x - ray luminosity ( bary et al . 2002 , bary et al . 2008 , see figure 6 ) . near - ir lines have been observed in relatively old objects such as echaj08843.3 - 7905 in the @xmath26 chamaleontis cluster ( 6 myr , ramsay howat & greaves 2007 ) and tw hya ( bary et al . 2000 ) in the tw hya association ( 8 - 10 myr ) . h@xmath3 lines have been detected in transitional disks in chamaleon ( bary et al . 2008 ) , and in the close binary star gv tau ( doppmann et al . 2008 ) . in the transitional disks in chameleon in which h@xmath3 near - ir emission has been observed , [ ne ii ] emission at 12 @xmath9 m has been also detected by spitzer ( bary et al . this suggests a common excitation mechanism for both lines ( see section 3.7 ) . the main strength of the h@xmath3 lines in the near - ir lies in their ability to probe very small quantities of hot gas in the inner disk if the right excitation conditions are present ( accretion , uv photons , x - rays ) . this strength can be exploited to search for gas in environments where other gas diagnostics give negative results ( i.e. , wtts ) . in a similar way , non - detections allow to set stringent upper limits on the presence of hot gas in the inner disk ( e.g. , carmona et al . 2007 found no evidence of h@xmath3 emission in the debris disk 49 cet , and constrained the mass in the inner most disk to be lower than a few lunar masses ) . observationally , h@xmath3 lines in the near - ir have the advantage that they are observed in the k band . the transmission of the atmosphere in the k band is relatively good ; hence , h@xmath3 lines in the near - ir can be searched in relatively faint objects with large telescopes . there is potential for perfoming large surveys for h@xmath3 emission in the near - ir . the main disadvantage of h@xmath3 near - ir emission is that it is also produced by shocks in outflows . thus , to reduce the risk of confusion with shocks , high spectral resolution observations with precision higher than 5 km / s and observations at high angular resolution with ao are needed . for the interested reader , beck et al . ( 2008 ) and gustafsson et al . ( 2008 ) present examples of h@xmath3 integral field spectroscopy obtained with ao in t tauri stars . in the sources observed by those autors , we can observe the contribution of outflow emission to the h@xmath3 1 - 0 s(1 ) observed . finally , note that near - ir h@xmath3 only trace the hot h@xmath3 in the innermost disk ( r@xmath0few au ) . consequently , an additional limitation is that h@xmath3 lines in the near - ir can _ not _ be used to derive the total disk mass or the gas mass in the giant planet forming region of the disk ( note that most of the mass of the disk is in cold gas at r@xmath120 au ) . h@xmath3 pure - rotational emission in the mid - ir traces warm gas at temperatures generally from 150 k up to 1000 k. thus , h@xmath3 mid - ir lines have the unique potential of tracing the gas in the giant planet formation region of the disk at r @xmath1 2 au . with present sensitivities ( 5@xmath27 erg / s/@xmath28 ) a few earth masses of warm gas in sources at 140 pc can be probed . mid - ir emission arises from transitions between the rotational levels ( @xmath29j=2 ) of the ground vibrational state ( @xmath11=0 ) . typical lines searched are the @xmath11=0 - 0 s(0 ) line ( j=2 - 0 ) at 28 @xmath9 m ( visible only from space ) , the 0 - 0 s(1 ) line at 17 @xmath9 m , the 0 - 0 s(2 ) line at 12 @xmath9 m and the 0 - 0 s(3 ) line at 9.6 @xmath9 m . following the initial detection claims of h@xmath3 0 - 0 s(0 ) and 0 - 0 s(1 ) emission from observations with iso ( thi et al . 2001 ) , several ground - based ( richter et al . 2002 ; sheret et al . 2003 ; sako et al . 2005 , martin - zadi et al . 2007 & 2008b ; bitner et al . 2007 & 2008 ; carmona et al . 2008b ) and space ( hollenbach et al . 2005 ; pascucci et al . 2006 ; lahuis et al . 2007 ) searches for h@xmath3 mid - ir emission have been performed . iso detections remain controversial , as all subsequent efforts do not confirm the high h@xmath3 emission fluxes reported by thi et al . at the time of writing , positive detections of h@xmath3 mid - ir emission have been reported only in a handful of objects : the herbig ae / be stars ab aur ( bitner et al . 2007 ) and hd 97048 ( martin - zadi et al . 2007 ) , and the t tauri stars sz 102 , ec 74 , ec 82 , ced_110_irs6 , ec92 , iso - cha237 ( lahuis et al . 2007 ) , doar 21 , elias 29 , gss 30 . tau n and hl tau ( bitner et al . 2008 ) . in total , h@xmath3 pure - rotational emission has been detected only in 13 young stars with disks from @xmath2 114 objects searched ( here we exclude 14 iso h@xmath3 detections in 18 stars observed , and the searches in 16 optically thin disks of hollenbach et al . 2005 and pascucci et al . the main result of these searches are : ( i ) h@xmath3 mid - ir emission is relatively rare ( @xmath30 of the sources ) , and ( ii ) in the sources where the emission is detected , the gas is heated by an additional mechanism . the numerous non detections can be understood in a first approximation under the frame of the chiang & goldreich ( 1997 ) two - layer model of an optically thick disk . in this model , only emission from the surface molecular layer of the disk is observed ( the mid - plane is optically thick ) . since the amount of gas in the surface layer is very small , if the dust and gas are at equal temperatures and the gas - to - ratio is equal to 100 , then the expected thermal emission of h@xmath3 from the surface layer is too weak to be detected ( @xmath210@xmath31 - 10@xmath32 erg / s/@xmath28 , carmona et al . 2008 ) . in the context of the two - layer model , the few detections can be explained if the temperature of the gas is allowed to be at twice the temperature of the dust and if the gas - to - dust ratio is larger than 1000 ( carmona et al . this can occur under certain physical circumstances such as a high uv or x - ray radiation fields or if dust coagulates and sediments towards the mid - plane of the disk . the two - layer model is just an approximation to the real structure of the disk ; however , similar conclusions have been reached from more detailed modeling . sophisticated models of h@xmath3 emission from disks ( nomura et al . 2005 , 2007 ) assuming typical x - ray and uv - fluxes from pre - main sequence stars predict flux levels below present detection limits ( both in line flux and in the line contrast with the continuum ) . models from optically thick disks with enhanced levels of uv and x - ray radiation ( e.g. , gorti & hollenbach 2008 ) can account for the h@xmath3 fluxes reported . finally , note that h@xmath3 mid - ir emission has been reported up to the 0 - 0 s(9 ) line ( indication that an extra source of heating is present ) and that it has been observed in a few transitional disks ( e.g. , sz 102 ) and in the wtts doar 21 ( bitner et al . 2008 ) . in summary , h@xmath3 mid - ir lines have the potential of tracing the gas in the giant planet forming region of the disk , but their main limitation is that they are too weak to be detected in a large number of sources . in addition , if the disks observed are optically thick , then mid - ir h@xmath3 lines only trace the small amount of gas in the surface layer of the disk . therefore , h@xmath3 mid - ir lines in optically thick disks can not be used to derive constraints in the total disk mass . in the objects in which we are able to detect the emission , h@xmath3 appears be heated by an additional mechanism . in this case the lte approximation is no longer valid and the h@xmath3 masses derived can be overestimated . mid - ir emission studies from the ground are so far limited to relatively bright sources with fluxes typically larger than 1 jy . future mid - ir infrared instrumentation such as high - resolution mid - ir spectrograph exes in the airplane observatory sofia , and the mid - ir spectrograph miri in the james - webb space telescope will increase the sensitivity and will allow the study of h@xmath3 mid - ir emission in larger number of objects , in particular in ctts . h@xmath3 electronic transitions in the uv probe two kinds of gas : ( i ) in emission when the disk is seen face on , they trace highly excited gas at temperatures of a few thousand kelvin ( r@xmath33 1 au ) . ( ii ) in absorption when the disk is seen close to edge - on , they probe colder gas at few hundred k. h@xmath3 uv lines are observed in absorption against the bright continuum in hot stars or against broad emission lines ( e.g. , ovi at @xmath34 1032 - 1038 ) in cool stars . electronic transitions occur between the vibrational levels of the first @xmath35 or the second @xmath36 electronic states and the ground vibrational level of the ground electronic state ( @xmath37 ) . given the low transmission of the atmosphere at uv - wavelengths , h@xmath3 uv transitions are studied primarily with space instrumentation ( e.g. , iue , hst , fuse ) . h@xmath3 lines are excited by the far - uv ( fuv ) flux of the central star , or by shocked gas in outflows . warm h@xmath3 gas can absorb photons all over the fuv ; consequently , a flat radiation field would excite a very large number of densely packed h@xmath3 lines and a pseudocontinuum is produced . in the case of ctts , the fuv flux is dominated by emission lines , in particular by ly@xmath5 . hence , in ctts , we observe principally ly@xmath5-band ( b - x ) transitions . nevertheless , werner - band transitions ( c - x ) are observed as well . h@xmath3 electronic transitions have been searched towards t tauri stars ( valenti et al . 2000 ; errico et al . 2001 ; ardila et al . 2002 ; wilkinson et al . 2002 ; herczeg et al . 2002 , 2004 , 2005 & 2006 ; walter et al . 2003 ; bergin et al . 2004 ) , herbig ae / be stars ( roberge et al . 2001 ; lecavelier des etangs et al . 2003 ; bouret et al . 2003 ; grady et al . 2005 ; martin - zadi et al . 2004 , 2005 & 2008a ) , accreting brown dwarfs ( gizis et al . 2005 ) and in close to edge - on debris disks ( @xmath38 pic , lecavelier des etangs et al . 2001 , martin - zadi et al . 2008 ; au mic , roberge et al . 2005 , france et al . 2007 ) . in the case of t tauri stars , h@xmath3 uv lines have been detected in several ctts and in at least one wtts ( v836 tau ) . nontheless , a large fraction of the h@xmath3 detections in ctts appears to be emission from outflows . the observed lines are blueshifted ( e.g. , ardila et al . 2002 ) and/or spatially extended ( e.g. , herczeg et al . so far , only in very few cases quiescent spatially unresolved h@xmath3 emission in the uv has been reported . the most notably example is the t tauri star tw hya . in this source , the lines observed are excited by ly@xmath5 fluorescence . the emission is consistent with emission from the surface of a hot inner disk ( herczeg et al . 2002 ) in the case of herbig ae / be stars , h@xmath3 uv lines have been studied mostly in absorption at 1032 - 1038 . evidence for different origin / excitation for the lines is observed as function of the stellar mass . for the massive herbig be stars ( spectral types b2 to b8 ) , the h@xmath3 lines are consistent with a photodissociation region ( pdr ) in large circumstellar envelopes ( martin - zadi 2008a ) . for herbig ae stars , the line of sight generally appears _ not _ to pass through the disk . the h@xmath3 lines observed are consistent with warm gas in the circumstellar environment of the stars ( e.g. , an envelope in ab aur , roberge et al . 2001 ) or cold gas in the line of sight ( e.g. , hd 141569 martin - zadi et al . 2005 ) . in herbig ae stars , the h@xmath3 absorption lines are not consistent with pdr models ( martin - zadi 2008a ) . in some cases , if the inclination and flaring of the disk are adequate , we may see a thin layer of a flared circumstellar disk ( e.g. , hd 100456 & hd 163296 , lecavelier des etangs et al . 2003 ) . in the case of the debris disk @xmath38 pic , no strong h@xmath3 lines either in emission or absorption have been detected . only a marginal detection of the j=0 h@xmath3 line was reported ( lecavelier des etangs et al . this was interpreted as evidence of lack of gas in the @xmath38 pic disk . in the case of au mic h@xmath3 was not detected in absorption , but it was detected in emission . the estimated gas mass was of @xmath39 ( france et al . 2007 ) . the main advantage of the h@xmath3 lines in the uv are : ( i ) if observed in emission , they can probe small amount of very hot gas ( t@xmath22500 k ) in the innermost part of the disk ; ( ii ) if observed in absorption , they can trace small amounts of colder ( t@xmath2 50-hundreds of k ) gas . h@xmath3 lines in the uv have allowed us to unveil the presence of hot gas in the inner disk in wtts and transitional disks . their main limitations are : ( i ) in most of the ctts , the h@xmath3 emission in the uv originates in outflows ; ( ii ) in most of the herbig ae stars where h@xmath3 has been observed in absorption , the probed gas is not in the disk , but instead in circumstellar ambient material or cold gas along the line of sight ; ( iii ) absorption studies probe a narrow pencil - beam that limits its application in sources observed far from edge - on ; ( iv ) in the case of detections in emission , the interpretation of the results ( i.e. , column densities ) requires complex modeling based on the poorly constrained fuv field of the central star . finally , uncertain conversion factors are needed to extrapolate the mass of gas probed to the total amount of gas in the disk . [ oi ] emission at 6300 is a tracer of atomic gas in the surface layers of flared disks . the line traces up to a 100 au depending on the geometry of the disk . [ oi ] forbidden emission at 6300 is a diagnostic that started to be explored recently . it has been studied in a growing number of herbig ae / be stars with high - resolution optical spectroscopy ( acke & van den ancker 2005 & 2006 ; van der plas et al . 2008a ; fedele et al . 2008 ) . the majority of the studied sources display narrow ( width @xmath0 50 km / s ) single peaked profiles ; however , several objects exhibit double - peaked profiles . in such objects , the low velocities with respect to the star s rest velocity , the symmetry of the features and the peak - to - peak separation are consistent with line emission from a circumstellar disk in keplerian rotation . the strength of the [ oi ] line appears to be strongly correlated with the sed shape of the sources ( i.e. , disk geometry ) . following the sed classification scheme of meeus et al . ( 2001 ) , acke et al . 2005 found that herbig ae / be from group i ( flared disks ) exhibit a [ oi ] emission stronger than those of group ii ( self - shadowed disks ) . in addition , the feature was found to be absent in an important fraction ( 40% ) of group ii sources . acke et al . found also that the [ oi ] luminosity is correlated to the pah luminosity of the sources . acke et al . proposed that the [ oi ] line arises from the photodisociation by uv radiation of oh and h@xmath3o molecules in the surface layer of flared disks . the conclusions shown by acke et al . require fractional oh abundance @xmath40(oh ) 10@xmath41 - 10@xmath42 . based on the analysis of the [ oi ] emission line profile in hd 100546 , acke et al . ( 2006 ) found evidence of a gap present at 10 au in hd 100546 s disk . they suggested that such a gap was likely induced by a planetary - mass companion of 20 m@xmath43 located at 6.5 au . temporal changes in the [ oi ] line profile were reported as well . acke et al . concluded that such changes are related to inhomogeneities in the [ oi ] disk emitting region . by analyzing the [ oi ] line , van der plas et al . ( 2008a ) found evidence of the existence of a puffed - up inner rim followed by a shadow in hd 101412 . fedele et al . ( 2008 ) employed a combined analysis of the [ oi ] emission and mid - ir interferometry observations found evidence of different dust and gas vertical structure inside 2 au in the disks of hd 101412 and hd 135344b . fedele et al . interpreted their findings as the result of dust / gas decoupling in the inner disk . they suggested that they may exist an evolutionary sequence from flared disks to flat disks due to the combined action of gas - dust decoupling , grain growth , and dust settling . [ oi ] emission at 6300 has the advantange of being observable in the optical ; therefore , statistical studies of large numbers of objects are possible at high - spectral resolution in the future . one limitation of the diagnostic is that [ oi ] emission can produced by shocked gas in an outflow . thus high spatial and spectral resolution observations are required to avoid confusion with emission from outflows . recently , pascucci et al . ( 2007 ) and lahuis et al . ( 2007 ) reported the detection of [ ne ii ] 12.8 @xmath9 m emission from t tauri stars with spitzer . pascucci et al . ( 2007 ) detected the line in four out of six targets ( all with faint mid - ir continuum with respect to classical t - tauri stars of the same spectral type ) and lahuis et al ( 2007 ) in 15 out of 76 t tauri stars . these first detections were followed by ground - based detections in tw hya by herczeg et al . the [ ne ii ] line has been observed towards optically thick disks , as well as in the transitional disk gm aur and the circumbinary disk cs cha ( espaillat et al . 2007 ) . given that neon can not be ionized by photons with energies of less than 21.4ev , the detection of [ ne ii ] emission from disks attracted the attention of the disk community . the [ ne ii ] line could provide evidence of higher energy photons irradiating the disk , in particular extreme - uv ( euv ) photons ( gorti & hollenbach , 2008 ; alexander et al . 2008 ) or x - rays ( glassgold et al . 2007 , meijerink et al . nonetheless , [ ne ii ] emission can be produced by shocked gas in outflows ( e.g. , hollenbach & mckee 1989 ) . at the present , it is not well constrained which of these scenarios is taking place . spitzer s spatial and spectral resolution are too modest to distinguish between the different scenarios proposed . models of x - ray irradiated disks by glassgold et al . ( 2007 ) predicted that the ionization fraction of neon is proportional to the square root of the x - ray luminosity . however , models by meijerink et al . ( 2008 ) rather found that there is a linear proportionality between the [ neii ] line luminosity and the x - ray luminosity . pascucci et al ( 2007 ) reported a tentative correlation between the [ ne ii ] and the x - ray luminosities . however , only about 30% of the sources with [ ne ii ] detections from lahuis et al . ( 2007 ) are identified as x - ray sources ( either due to the lack of x - ray emission , sensitivity limited x - ray searches , or object geometry ) . in the uv model of gorti & hollenbach ( 2008 ) , euv photons from the stellar chromosphere and/or from accretion create a region similar to a hii region at the surface of the disk . the euv are absorbed in the inner 10 au of the disk , and a euv photon luminosity @xmath210@xmath44 erg / s would be able to produce detectable [ neii ] line fluxes . but , in the euv model , a stellar wind greater than @xmath45 could drastically reduce the uv - photons reaching the disk . stellar winds of this magnitude correspond to typical ctts accretion rates . thus , lower [ ne ii ] luminosities are expected for ctts stars simple due to the accretion and the related wind rates . nonetheless , several detection from lahuis et al . ( 2007 ) are from ctts . in an alternative model , alexander et al . ( 2008 ) suggested that the [ ne ii ] line is produced in a photoevaporating wind . alexander et al . predicted broad ( 30 - 40 km / s ) double peaked [ ne ii ] lines when the stars are viewed close to edge - on and narrower ( @xmath210 km / s ) , slightly blue - shifted lines when viewed face - on . high spatial and spectral resolution is required for testing this model . in an additional scenario , high - velocity shocks can entrain ionized lines ( e.g. , hollenbach & mckee 1989 ) . van den ancker et al . ( 1999 ) detected [ ne ii ] emission in the iso spectra of the t tau triplet . this system contains three stars of which at least one is a strong x - ray source , as well as regions of shocked gas in the immediate vicinity . due to the large beam ( 27 `` @xmath46 15 '' ) it was not clear from where in the system the emission comes . van den ancker et al . proposed that the emission arose in j - type ( dissociative ) shocks resulting from the interaction of the outflow of t - tau south with ambient material . t tau exhibits a very high @xmath47}}/l_{\rm{x}}$ ] ratio , hinting that processes other than x - ray irradiation are important . recent high spectral and spatial resolution observations of the [ ne ii ] line in t tau with the vlt spatially resolved the various components of the system ( van boekel et al . in prep . ) . van boekel et al . found that the vast majority of the [ ne ii ] flux appeared to be associated with an outflow from t tau s , and that only a small fraction ( @xmath25@xmath4810% ) of the [ ne ii ] emission was directly related to the x - ray bright northern component . van boekel et al . observations showed that if strong accretion / outflow activity is present , shocks were the main mechanism for producing [ ne ii ] emission in t tau . finally , note that in three young stars that exhibit the _ spitzer _ [ ne ii ] line , the h@xmath3 1 - 0 s(1 ) line at 2.12 @xmath9 m was detected as well : tw hya ( bary et al . 2003 ) , gm aur ( shukla et al . 2003 ) and cs cha ( bary et al . 2008 ) . bary et al . ( 2008 ) pointed out that simultaneous presence of the [ ne ii ] line and _ quiescent _ h@xmath3 disk emission suggests a shared excitation mechanism of the gas in the disk by the central star s high - energy photons . the same high - energy photons that stimulate the [ ne ii ] emission could also produce the _ quiescent _ h@xmath3 emission observed . however , as previously discussed , h@xmath3 near - ir emission can be also produced by shocks . in such a case , objects displaying the [ ne ii ] lines , may show the h@xmath3 near - ir emission velocity shifted or extended . in such objects , h@xmath3 and [ ne ii ] emission will be produced by shocks . spectrally and spatially resolved observations are required to test both scenarios . in summary , [ neii ] emission has the potential of unveiling the effects of x - ray or uv - ray in the surfaces of disks . but , since [ neii ] emission can be produced by shocks , it is necessary to perform observations with higher spatial and spectral resolution to determine the dominant excitation mechanism . in addition , observations of larger samples are required to establish statistically meaningful correlations between the presence / strength of [ neii ] emission and physical properties of the central star ( i.e. , spectral type , accretion rate , age , x - ray and uv - luminosity ) . how water is transported to the surface of habitable planets is one fundamental and fascinating question in planet formation theory . water and other simple organic molecules are expected to be abundant -in gas phase- in the inner regions of the disk ( r@xmath05 au ) . nevertheless , observational measurements of water in the planet forming region of disks are relatively recent . first detections of hot water vapor were reported by carr et al . ( 2004 ) and thi & bik ( 2005b ) in the near - ir spectrum of young stellar objects exhibiting the 2.3 @xmath9 m co bandhead emission ( e.g. , svs 13 , dg tau and 51 oph ) . the observed h@xmath3o displays a characteristic excitation temperature of @xmath21500 k. since this temperature is cooler than the temperature of the co overtone emission ( 2500 k ) the observed water emission should be produced in a region exterior to the co 2.3 @xmath9 m emitting region . simultaneous modeling of the co and steam emission revealed that the h@xmath3o / co ratio is lower to that expected in chemical equilibrium ( carr et al . 2004 ) . physical processes other than equilibrium chemistry should be responsible of the mesaured h@xmath3/co abundances . thi & bik ( 2005 ) reproduced the h@xmath3/co abundance observed in 51 oph with gas at 1600 k and an enhanced uv field over gas density ratio . these first detections of h@xmath3o were followed by detections in the ir of simple molecules such as c@xmath3h@xmath3 and hcn . they were observed in absorption in the near - ir spectrum of the t tauri binary gv tau n ( gibb et al . 2007 & 2008 ) and in absorption in the mid - ir spectrum of the low - mass young stellar object irs 46 ( lahuis et al . 2006 ) . the temperaures measured were of a few hundred k. the lines originate most likely in the inner disk ( r@xmath06 au ) of the sources . however , there is also the possibility that they originate in a disk wind . the high sensitivity reached with ir spectroscopy from space and ground - based high spectral resolution allowed the detection of emission of simple organic molecules from the terrestrial planet region of the disk ( r@xmath05 au ) . carr & najita ( 2008 ) reported the detection of hcn , c@xmath3h@xmath3 , co@xmath3 water vapor and oh emission in the spitzer s mid - ir spectrum of the t tauri star aa tau . modeling of the observed profiles indicates that the observed gas has temperatures from 500 to 900 k and is located within 0.5 and 3 au . one important finding is that the abundances relative to co found are much higher than those obtained from models of hot molecular cores , this suggests that substantial molecular synthesis occurs within the disk . salyk et al . ( 2008 ) reported the detection of h@xmath3o emission at 3 @xmath9 m and 10 - 20 @xmath9 m and oh emission at 3 @xmath9 m in the t tauri stars as 205a and dr tau . they measured excitation temperatures of 1000 k. based on the velocity wings of the lines , they concluded that the observed emission arises up to radius of 1 au . mandell et al . ( 2008 ) found oh emission at 3 @xmath9 m in the herbig ae stars ab aur and mwc 758 . optically thin lte models of the observed lines revealed excitation temperatures @xmath2700 k and an oh emitting region extending up to 1 au . the observed oh lines can be explained by collisional excitation models , and in the case of ab aur by a uv fluorescence model too . pascucci et al . ( in prep . ) detected with spitzer c@xmath3h@xmath3 and hcn at 7 - 14 @xmath9 m in a large number of young stellar objects with disks they found that the abundance of these simple organic molecules appears to be different between stars cooler than the sun and sun - like stars . the study of simple organic molecules in the near and mid - ir is growing . the main advantage of these diagnostics are : ( i ) they trace the gas in the terrestrial planet forming region of the disks , and ( ii ) they have the potential of constraining the chemical processes in this region . from the observational point of view , an additional advantage is that within a single observation set - up , multiple tracers are observable simultaneously . in addition , the fact that some of the lines are observable from the ground at 3 @xmath9 m open the door to the study of a large number of objects at high spectral and spatial resolution . one limitation of these gas tracers is that the interpretation of the observations generally requires complex modeling that involves : ( i ) assumptions in the structure of the inner disk ; ( ii ) poorly constrained physical quantities ( e.g. , the uv field , inclination of the inner disk ) ; and ( iii ) poorly constrained physical processes ( e.g. , excitation mechanisms , chemistry , radial and vertical mixing in the disk ) . on the other hand , these limitations can become an advantage , because , even if observations can be interpreted with different models , observations can also rule out scenarios that are not consistent with the measurements . it is interesting that future observations of simple molecules in the inner disk could reveal a variety of disk chemistry until now unsuspected . finally , note that in very short time spitzer will not offer mid - ir spectroscopy anymore . however , future mid - ir spectrograph miri planned for jwst will be an ideal tool for this research . in this contribution i have discussed several observational diagnostics that allow to study the gas in protoplanetary disks highlighting the strengths and limitation of each diagnostic . here , i summarize important general points that one should bear in mind about observational constraints of gas in disks . * disks have a radial temperature structure ; therefore , different gas diagnostics are employed for probing different regions of the disk . uv probes the hottest gas in the innermost disk ( r@xmath331 au , t@xmath12000 k ) , nir and mir probes hot and warm gas in the inner disk ( r@xmath20.1 - 20 au , t@xmath2100 - 2000 k ) , ( sub)-mm probes cold gas in the outer disk ( r@xmath49 au , t@xmath220 - 100 k ) . note that disks have typical sizes of several hundreds of au . * to observe a line from gas ( in emission or absorption ) the dust in the medium where the line is produced must be optically thin at the observed wavelength , or the gas should be at a higher temperature than the dust . * an optically thick disk has a vertical temperature structure . the surface layer is hotter than the interior mid - plane layer . the dust in the surface layer is optically thin , the dust in the interior layer ( of the inner disk ) is optically thick . in optically thick disks , hot and warm gas emission lines from the uv to the ir are produced in the surface layer . note that the amount of mass in the surface layer is much smaller than the amount of mass in the interior layer . gas lines in the ir probe only a limited amount of material . h@xmath3 lines in the ir are optically thin . co lines in the ir can be optically thick . * sub - mm lines trace only the cold outer regions of the disk at r@xmath120 au . planets usually are not expected to form at this distances . still , most of the mass of the disk is at r@xmath120 au . * co in the ( sub)-mm is not a reliable disk gas mass tracer because ( i ) co freezes onto dust grain surfaces , ( ii ) the co / h@xmath3 conversion factor is unknown in disks ( iii ) the emission is likely optically thick . * be aware that so far we do not measure directly the mass of gas in disks . we deduce the disk mass from dust continuum emission in the ( sub)-mm from cold dust in the outer disk ( r@xmath120 au ) . the deduced mass depends on the dust opacity , dust temperature and the dust - to - gas ratio assumed . the disk mass observed is from material located at r@xmath150 au , and disk evolutionary models are typically of disks of r@xmath050 au . the mass and surface density profiles of gas at r@xmath020 au are poorly constrained from observations . * modeling is required to interprete line observations . conventional assumptions are that the observed gas is at lte and is optically thin . the rotational diagrams employed for deducing the temperature of the observed gas make use those two assumptions . to deduce the inner most radius of the observed emission ( i.e. , r@xmath17 ) from line profile modeling , it is required to assume the disk inclination and the dependence of the intensity as a function of radius ( @xmath50 ) . inclinations are typically deduced from ( sub)-mm imaging and extended to the inner disk . on the other hand , if we want to deduce independently the inclination ( e.g. , to search for evidence of warps ) we need to assume ( or calculate ) r@xmath17 and @xmath50 . * most of the gas tracers ( e.g. , h@xmath3 , [ ne ii ] ) can be produced by shocked emission from outflows as well . high spectral / spatial resolution is required for determine whether the emission is observed at the velocity of the central star and whether it is spatially extended or not . * gas emission could be excited by collisions with dust grains , shocks and uv or x - rays . line ratios are generally employed to discriminate between the different excitation mechanisms . * gas could be present in a disk even when dust emission is weak or not present . gas emission lines have been observed in the inner regions of transitional disks and in several wtts . * the study of gas in disks acquired momentum with the advent of high - resolution ground ir spectrographs and high - sensitivity ir spectrographs in space . * a new chapter in gas studies was recently written with the detection of emission of simple molecules in the inner disks . at the beginning of the paper i highlighted some fundamental science questions that required direct observational constraints from the gas in the disk . to conclude , i would like to address them again from the actual state of the observational study of gas . _ ( i ) how long does the protoplanetary disk last ? _ in general terms , disk gas diagnostics are not detected when the signatures of the dust in the disk have disappeared . thus , as a first approximation the gas in the disk disappears almost simultaneously with the signatures from small dust particles . nonetheless , we should be aware of a few details : ( a ) with the exception of h@xmath5 emission , there are many sources for which we know that they have a disk , but given sensitivity issues , we are not able to detect any gas tracer ; ( b ) gas has been observed in several wtts stars ( e.g. , h@xmath3 in the uv and nir ) , and in transitional disks ( e.g. , h@xmath3 in the near - ir , co at 4.7 @xmath9 m , [ ne ii ] at 12 @xmath9 m ) ; ( c ) unbiased surveys for hot / warm gas in a large sample of wtts have yet not been performed ; ( d ) surveys of h@xmath5 at high spectral resolution need to be done in larger samples of wtts in star - forming regions of different ages . if the accretion rate and spectral resolution are low , one can easily miss objects accreting gas at low accretion rates . in summary , an estimate of disk gas lifetime independent of dust is still required . _ ( ii ) how much material is available for forming giant planets ? _ on this aspect our hopes are fainter . several surveys for h@xmath3 emission in the mir showed that the emission is observed only in a handful of objects , and that in these objects the gas is heated by an additional ( to dust collisions ) heating mechanism ( uv , x - ray excitation ) . although these results confirm our two - layer picture of the disk , in practice these results mean that we are blind to the disk s mid - plane and that we are unable to know directly how much mass is in the giant planet forming region of the disk . the good news are two fold : ( a ) in objects with data at several wavelengths , an educated guess can be found from modeling the surface density ( see next point ) ; ( b ) at some point the disk should become optically thin -by dust evolution- then we will be able to measure the gas mass . _ ( iii ) how do the density and temperature of the disk vary as a function of the radius ? _ there is a growing number of sources ( e.g. , tw hya , ab aur ) for which we start to have information from several gas tracers at wavelengths from the uv to the ( sub)-mm . in such sources , we can attempt to constrain the disk structure employing a disk model aimed to fit all the available data . therefore , at least for a handful of objects , by modeling multiwavelength spectroscopy , we have the potential to constrain the density and temperature as a function of radius . however , we should always keep in mind that we observe emission from the disk s surface layer and that the mid - plane temperature and density will be deduced from the model . hence , the final gas mass deduced will be model dependent . _ ( iv ) what are the dynamics of the disk ? _ in this aspect we had good news . to detect gas lines in the ir we commonly use the highest spectral resolution available . at the present , the resolution is sufficiently high to spectrally resolve the lines . this allows the modeling of the line shape ; therefore , constraints in the dynamics of the emitting gas can be derived . recently , with the combination of ao and high - resolution spectroscopy , it has been possible to spatially resolve gas disk emission directly or to spatially resolve the spectroastrometric signal with spectra taken at different slit orientations . so far the observed gas is consistent with gas in keplerian rotation . this kind of studies should be extended in the future to a larger number of systems . i would like to thank the organizers of the ascona planet formation meeting for inviting me to give the observations of gas in disks talk , and to r. nelson , c.p . dullemond , c. clarke and s. fromang for kindly accepting that i present my talk as in individual contribution in this volume . i am grateful for the detailed suggestions in the manuscript provided by m. audard , c. baldovin - saavedra and f. fontani . i would like to warmly thank m. van den ancker , y. pavluychenkov , m. goto , c.p . dullemond , th . henning , c. martin - zadi , g. van der plas and d. fedele for all the discussions about gas in disks during the last years . part of this work is supported by a swiss national science fundation grant to m. audard . acke , b. , & van den ancker , m. e. , _ `` resolving the disk rotation of hd 97048 and hd 100546 in the [ o i ] 6300 ; line : evidence of a giant planet orbiting hd 100546 '' _ , astronomy and astrophysics , 449 , 267 , 2006 ardila , d. r. , basri , g. , walter , f. m. , valenti , j. a. , & johns - krull , c. m. , _ `` observations of t tauri stars using hubble space telescope ghrs . i. far - ultraviolet emission lines '' _ , astrophysical journal , 566 , 1100 , 2002 bary , j. s. , weintraub , d. a. , shukla , s. j. , leisenring , j. m. , & kastner , j. h. , _ `` quiescent h@xmath3 emission from pre - main - sequence stars in chamaeleon i '' _ , astrophysical journal , 678 , 1088 , 2008 beck , t. l. , mcgregor , p. j. , takami , m. , & pyo , t .- s . , _ `` spatially resolved molecular hydrogen emission in the inner 200 au environments of classical t tauri stars '' _ , astrophysical journal , 676 , 472 , 2008 bitner , m. a. , richter , m. j. , lacy , j. h. , greathouse , t. k. , jaffe , d. t. , & blake , g. a. , _ `` texes observations of pure rotational h@xmath3 emission from ab aurigae '' _ , astrophysical journal , 661 , l69 , 2007 brittain , s. d. , rettig , t. w. , simon , t. , kulesa , c. , disanti , m. a. , & dello russo , n. , _ `` co emission from disks around ab aurigae and hd 141569 : implications for disk structure and planet formation timescales '' _ , astrophysical journal , 588 , 535 , 2003 brittain , s. , rettig , t. w. , simon , t. , balsara , d. s. , tilley , d. , gibb , e. , & hinkle , k. h. , _ `` post - outburst observations of v1647 orionis : detection of a brief warm molecular outflow '' _ , astrophysical journal , 670 , l29 , 2007b bouret , j .- c . , martin , c. , deleuil , m. , simon , t. , & catala , c. , _ `` far uv spectroscopy of the circumstellar environment of the herbig be stars hd 259431 and hd 250550 '' _ , astronomy and astrophysics , 410 , 175 , 2003 carmona , a. , van den ancker , m. e. , thi , w .- f . , goto , m. , & henning , t. , _ `` upper limits on co 4.7 @xmath9 m emission from disks around five herbig ae / be stars '' _ , astronomy and astrophysics , 436 , 977 , 2005 carmona , a. , van den ancker , m. e. , henning , t. , goto , m. , fedele , d. , & stecklum , b. , _ `` a search for near - infrared molecular hydrogen emission in the ctts lkh@xmath5 ; 264 and the debris disk 49 ceti '' _ , astronomy and astrophysics , 476 , 853 , 2007 carmona , a. , van den ancker , m. e. , henning , t. , goto , m. , fedele , d. , & stecklum , b. , _ `` erratum : a search for near - infrared molecular hydrogen emission in the ctts lkh@xmath5 264 and the debris disk 49 ceti '' _ , astronomy and astrophysics , 478 , 795 , 2008a carr , j. s. , mathieu , r. d. , & najita , j. r. , _ `` evidence for residual material in accretion disk gaps : co fundamental emission from the t tauri spectroscopic binary dq tauri '' _ , astrophysical journal , 551 , 454 , 2001 cieza , l. a. , swift , j. j. , mathews , g. s. , & williams , j. p. , _ `` the masses of transition circumstellar disks : observational support for photoevaporation models '' _ , astrophysical journal , 686 , l115 , 2008 dutrey , a. , guilloteau , s. , duvert , g. , prato , l. , simon , m. , schuster , k. , & menard , f. , _ `` dust and gas distribution around t tauri stars in taurus - auriga . i. interferometric 2.7 mm continuum and @xmath4co j=1 - 0 observations '' _ , astronomy and astrophysics , 309 , 493 , 1996 fedele , d. , et al . , _ `` the structure of the protoplanetary disk surrounding three young intermediate mass stars . spatially resolved dust and gas distribution '' _ , arxiv e - prints , arxiv:0809.3947 , 2008 grady , c. a. , et al . , _ `` coronagraphic imaging of pre - main - sequence stars with the hubble space telescope space telescope imaging spectrograph . i. the herbig ae stars '' _ , astrophysical journal , 630 , 958 , 2005 goto , m. , usuda , t. , dullemond , c. p. , henning , t. , linz , h. , stecklum , b. , & suto , h. , _ `` inner rim of a molecular disk spatially resolved in infrared co emission lines '' _ , astrophysical journal , 652 , 758 , 2006 herczeg , g. j. , linsky , j. l. , valenti , j. a. , johns - krull , c. m. , & wood , b. e. , _ `` the far - ultraviolet spectrum of tw hydrae . i. observations of @xmath52 fluorescence '' _ , astrophysical journal , 572 , 310 , 2002 herczeg , g. j. , wood , b. e. , linsky , j. l. , valenti , j. a. , & johns - krull , c. m. , _ `` the far - ultraviolet spectra of tw hydrae . models of h@xmath3 fluorescence in a disk '' _ , astrophysical journal , 607 , 369 , 2004 herczeg , g. j. , linsky , j. l. , walter , f. m. , gahm , g. f. , & johns - krull , c. m. , _ `` the origins of fluorescent h@xmath3 emission from t tauri stars '' _ , astrophysical journal supplement series , 165 , 256 , 2006 henning , th . , _ `` dust in protoplanetary disks '' _ , meteoritics & planetary science , vol . 41 , supplement , proceedings of 69th annual meeting of the meteoritical society , held august 6 - 11 , 2006 in zurich , switzerland . , p.5392 , 41 , 5392 , 2006 lahuis , f. , van dishoeck , e. f. , blake , g. a. , evans , n. j. , ii , kessler - silacci , j. e. , & pontoppidan , k. m. , _ `` c2d spitzer irs spectra of disks around t tauri stars . iii . [ ne ii ] , [ fe i ] , and h@xmath3 gas - phase lines '' _ , astrophysical journal , 665 , 492 , 2007 mandell , a. m. , mumma , m. j. , blake , g. a. , bonev , b. p. , villanueva , g. l. , & salyk , c. , _ `` discovery of oh in circumstellar disks around young intermediate - mass stars '' _ , astrophysical journal , 681 , l25 , 2008 mannings , v. , & sargent , a. i. , _ `` a high - resolution study of gas and dust around young intermediate - mass stars : evidence for circumstellar disks in herbig ae systems '' _ , astrophysical journal , 490 , 792 , 1997 martin - zadi , c. , deleuil , m. , simon , t. , bouret , j .- c . , roberge , a. , feldman , p. d. , lecavelier des etangs , a. , & vidal - madjar , a. , _ `` fuse observations of molecular hydrogen on the line of sight towards hd 141569a '' _ , astronomy and astrophysics , 440 , 921 , 2005 martin - zadi , c. , lagage , p .- o . , pantin , e. , & habart , e. , _ `` detection of warm molecular hydrogen in the circumstellar disk around the herbig ae star hd 97048 '' _ , astrophysical journal , 666 , l117 , 2007 martin - zadi , c. , van dishoeck , e. f. , augereau , j .- c . , lagage , p .- o . , & pantin , e. , _ `` searching for molecular hydrogen mid - infrared emission in the circumstellar environments of herbig be stars '' _ , astronomy and astrophysics , 489 , 601 , 2008b meeus , g. , waters , l. b. f. m. , bouwman , j. , van den ancker , m. e. , waelkens , c. , & malfait , k. , _ `` iso spectroscopy of circumstellar dust in 14 herbig ae / be systems : towards an understanding of dust processing '' _ , astronomy and astrophysics , 365 , 476 , 2001 najita , j. , carr , j. s. , glassgold , a. e. , shu , f. h. , & tokunaga , a. t. , _ `` kinematic diagnostics of disks around young stars : co overtone emission from wl 16 and 1548c27 '' _ , astrophysical journal , 462 , 919 , 1996b nomura , h. , aikawa , y. , tsujimoto , m. , nakagawa , y. , & millar , t. j. , _ `` molecular hydrogen emission from protoplanetary disks . ii . effects of x - ray irradiation and dust evolution '' _ , astrophysical journal , 661 , 334 , 2007 pontoppidan , k. m. , blake , g. a. , van dishoeck , e. f. , smette , a. , ireland , m. j. , & brown , j. , _ `` spectroastrometric imaging of molecular gas within protoplanetary disk gaps '' _ , astrophysical journal , 684 , 1323 , 2008 sako , s. , yamashita , t. , kataza , h. , miyata , t. , okamoto , y. k. , honda , m. , fujiyoshi , t. , & onaka , t. , _ `` search for 17 @xmath9 m h@xmath3 pure rotational emission from circumstellar disks '' _ , astrophysical journal , 620 , 347 , 2005 salyk , c. , pontoppidan , k. m. , blake , g. a. , lahuis , f. , van dishoeck , e. f. , & evans , n. j. , ii , _ `` h@xmath3o and oh gas in the terrestrial planet - forming zones of protoplanetary disks '' _ , astrophysical journal , 676 , l49 , 2008 sicilia - aguilar , a. , hartmann , l. w. , frsz , g. , henning , t. , dullemond , c. , & brandner , w. , _ `` high - resolution spectroscopy in tr 37 : gas accretion evolution in evolved dusty disks '' _ , astronomical journal , 132 , 2135 , 2006 scoville , n. , kleinmann , s. g. , hall , d. n. b. , & ridgway , s. t. , _ `` the circumstellar and nebular environment of the becklin - neugebauer object - 2 - 5 micron wavelength spectroscopy '' _ , astrophysical journal , 275 , 201 , 1983 shukla , s. j. , bary , j. s. , weintraub , d. a. , & kastner , j. h. , _ `` further detections of rovibrational h@xmath3 emission from the disks of t tauri stars '' _ , bulletin of the american astronomical society , 35 , 1209 , 2003 thi , w. f. , et al . , _ `` h@xmath3 and co emission from disks around t tauri and herbig ae pre - main - sequence stars and from debris disks around young stars : warm and cold circumstellar gas '' _ , astrophysical journal , 561 , 1074 , 2001 valenti , j. a. , johns - krull , c. m. , & linsky , j. l. , _ `` an iue atlas of pre - main - sequence stars . i. co - added final archive spectra from the swp camera '' _ , astrophysical journal supplement series , 129 , 399 , 2000 van den ancker , m. e. , wesselius , p. r. , tielens , a. g. g. m. , van dishoeck , e. f. , & spinoglio , l. , _ `` iso spectroscopy of shocked gas in the vicinity of t tauri '' _ , astronomy and astrophysics , 348 , 877 , 1999 van der plas , g. , van den ancker , m. e. , fedele , d. , acke , b. , dominik , c. , waters , l. b. f. m. , & bouwman , j. , _ `` the structure of protoplanetary disks surrounding three young intermediate mass stars . i. resolving the disk rotation in the [ oi ] 6300 line '' _ , astronomy and astrophysics , 485 , 487 , 2008a van der plas , g. , van den ancker , m. e. , acke , b. , carmona , a. , dominik , c. , fedele , d. , & waters , l. b. f. m. , _ `` spatially resolved 4.7 @xmath9 m co fundamental emission in two protoplanetary disks '' _ , arxiv e - prints , arxiv:0810.3417 , 2008b van langevelde , h. j. , van dishoeck , e. f. , van der werf , p. p. , & blake , g. a. , _ `` the spatial distribution of excited h2 in t tau : a molecular outflow in a young binary system '' _ , astronomy and astrophysics , 287 , l25 , 1994

radio pulsars show a wide variety of modulations of their pulse amplitudes , including bursts and nulls , that affect their detectability in surveys . phenomena seen in some pulsars include short - period nulling , in which a pulsar is not detected for several pulse periods , only to reappear with full strength ( @xcite ) ; eclipses , in which a companion star or its wind or magnetosphere absorbs or disperses the pulsar signal ( @xcite,@xcite,@xcite,@xcite ) ; long - term nulling or intermittent behavior , in which a pulsar is quiescent for days or weeks ( @xcite ) ; and rotating radio transients ( rrats , @xcite ) , pulsar - like objects from which only occasional radio bursts are detected . this paper describes analysis of a large - scale survey using the arecibo telescope that is sensitive to both periodic and aperiodic signals . rrats were first discovered in archive parkes multibeam survey data ( @xcite ) . eleven objects , with periods ranging from 0.7 to 7 s and pulse widths of @xmath0 ms , were found using a single pulse search algorithm ( @xcite ) . the longer periods of rrats compared with the general pulsar population suggest similarities with the x - ray populations of x - ray dim isolated neutron stars ( xdinss ) and magnetars . rrat j1819@xmath11458 has been detected at x - ray energies ( @xcite ) with properties that are similar to those of xdinss and high magnetic field radio pulsars . a different type of pulse modulation is observed in the case of pulsars emitting giant pulses . such pulses are tens to thousands of times brighter and an order of magnitude or more narrower than the average pulse ( see @xcite for an overview ) . giant pulses from the crab pulsar have substructure on timescales of 2 ns ( @xcite ) , and psr b1937 + 21 emits giant pulses as narrow as 16 ns ( @xcite ) . giant micro - pulses from the vela pulsar have widths @xmath2s ( @xcite ) , and the slowly rotating pulsars psr b1112 + 50 , psr b0031@xmath107 , and psr j1752 + 2359 occasionally emit bright pulses which are @xmath3 ms wide , @xmath4 times narrower than the average pulse ( @xcite ) . the detection of giant pulses is a potentially powerful method for finding extragalactic pulsars too distant for their normal emission to be detectable by periodicity searches ( @xcite ) . a variety of energetic phenomena other than pulsar emission can give rise to fast transients potentially detectable in radio pulsar surveys . within the solar system , transient radio events may be generated by energetic particles impacting the earth s atmosphere , solar flares , and decameter radio flares originating in jupiter s atmosphere . analogously to the latter , extrasolar planets with strong magnetic fields are expected to be detectable in the 10 - 1000 mhz range ( @xcite , @xcite , @xcite ) . magnetic activity on the surfaces of brown dwarfs and particle acceleration in the magnetic fields of flare stars are also known radio flare progenitors ( @xcite , @xcite , @xcite , @xcite ) . gamma - ray bursts are predicted to have detectable radio emission at @xmath5 mhz ( @xcite , @xcite ) , and radio flares have been observed from some x - ray binaries ( @xcite , @xcite ) . among the most energetic and exotic events in the universe , supernovae , merging neutron stars and coalescing black holes may produce wide - band radio bursts detectable at extragalactic distances ( @xcite ) . in this paper we describe an ongoing survey for pulsars and transient radio sources with the arecibo telescope . the survey addresses outstanding questions about the nature and emission mechanisms of intermittent radio sources . in [ section_surveyparams ] we present the palfa survey parameters , and in [ section_spmethods ] we describe the single pulse search and radio frequency interference excision algorithms which are part of the survey data processing pipeline . section [ section_palfastats ] contrasts palfa detection statistics on known pulsars and new discoveries , and [ section_biases ] examines selection effects influencing the detection and classification of transient sources . in [ section_intermittency ] , we apply an intermittency measure method for comparing the efficiency of periodicity and single pulse pulsar searches . in [ section_objects ] , we discuss the properties of individual intermittent objects discovered by palfa . in [ section_faroutmodel ] and [ section_constraints ] , we apply constraints derived from the survey sensitivity and results to the detectability of various energetic phenomena expected to emit radio bursts . finally , in [ section_conclusions ] , we present our main conclusions . the palfa survey started in 2004 , shortly after the installation of the seven - beam alfa receiver on the arecibo telescope . the survey searches for pulsars and transients in the inner and outer galactic plane regions accessible to arecibo ( see below ) . the alfa receiver is well - suited for survey observations , allowing simultaneous data collection from seven fields , each @xmath6 ( fwhm ) across . taking into account the hexagonal arrangement of the beams on the sky and the near sidelobes , the combined power - pattern is approximately @xmath7 ( @xcite ) . we observe a 100 mhz passband centered on 1440 mhz in each of the seven telescope beams . wideband arecibo pulsar processors ( wapps ; @xcite ) are used to synthesize 256-channel filterbanks spanning these bands at intervals of 64 @xmath8s . during observations , full - resolution data are recorded to disk and , in parallel , decimated down to a time resolution of 1 ms and searched for periodic signals and single pulses by a quick - look processing pipeline running in real time at the arecibo observatory ( @xcite ) . this approach allows for immediate discovery of relatively bright pulsars with periods longer than a few milliseconds . searching full - resolution data allows detection of millisecond pulsars and narrower single pulses and is done offline at participating palfa institutions as the processing is much more computationally intensive . table [ pointing_stats ] lists various alfa system and survey parameters , including the sky area corresponding to processed and inspected data reported on in the present paper and the total sky area observed to date ( see @xcite for a detailed explanation of other parameters ) . standard observation times are 268 s for inner galaxy pointings ( @xmath9 ) and 134 s for outer galaxy pointings ( @xmath10 ) . some early observations had a duration of 134 s and 67 s for inner and outer galaxy pointings , respectively . the quoted system temperature of 30 k is measured looking out of the galactic plane . the initial threshold of @xmath11 for the single pulse search is used when selecting events from dedispersed time series based on their signal - to - noise ratios only , before any filtering is applied . while there are a significant number of events due to random noise above this threshold , identification of an event as a genuine pulse takes into account not only its signal - to - noise ratio but also the fact that it is detected at a contiguous range of trial dispersion measures , which is in general not true of spurious events . thus weak pulses can be correctly identified , while they would be excluded if the threshold was set according to gaussian noise statistics . lccc center frequency ( ghz ) & 1.440 + total bandwidth ( mhz ) & 100 + channel bandwidth ( mhz ) & 0.39 + sampling time ( @xmath8s ) & 64 + nominal system temperature ( k ) & 30 + gain ( k / jy ) & + center pixel & 10.4 + ring pixels & 8.2 + beam width ( ) & + main beam ( 1 pixel ) & 3.5 + main beam plus near sidelobes ( 1 pixel ) & 18 + main beam plus near sidelobes ( 7 pixels ) & @xmath12 + inner galaxy & + observation time per pointing ( s ) & 268 , 134 + observed ( deg@xmath13 ) & 156 + processed ( deg@xmath13 ) & 99 + anticenter & + observation time per pointing ( s ) & 134 , 67 + observed ( deg@xmath13 ) & 119 + processed ( deg@xmath13 ) & 87 + detection threshold ( @xmath14 ) & + single pulse search ( initial threshold ) & 5 + fourier transform periodicity search & 7.5 here we compute the maximum distance at which sources of a given luminosity are be detectable by the palfa survey , @xmath15 . we then compare the sensitivity of the palfa survey to previous work done with the parkes multibeam system ( @xcite , @xcite ) . the rms noise in a radio transient search , where the effective observation time is equal to the pulse width , is _ n = , where @xmath16 is the system - equivalent flux density , @xmath17 is the number of polarization channels summed and @xmath18 is the bandwidth . a pulse s observed width @xmath19 may be broadened compared to its intrinsic width @xmath20 by several effects . after dedispersing the raw data and obtaining a dedispersed time series , there is residual dispersive broadening due to the finite width of a frequency channel and the error of the trial dispersion measure ( dm ) used compared to the actual pulsar dm . scatter broadening is not correctable and will have a contribution that depends on observing frequency and varies with direction on the sky . in general , w ( w_i^2 + t_dm , ch^2 + t_dm , err^2 + t_sc^2)^1/2 , where @xmath21 is the dispersive broadening across a frequency channel of width @xmath22 ( mhz ) for an observing frequency @xmath23 ( ghz ) , @xmath24 is the dispersive broadening due to the difference between the trial and actual dm of the source , and @xmath25 is the scattering broadening . broadening conserves pulse area , so that the intrinsic and observed peak flux densities are related through @xmath26 . if @xmath27 is the detection threshold , the minimum detectable _ intrinsic _ peak flux density is s_p , i , min = ( ) . for a one steradian pulsar radio beam , a source of intrinsic peak luminosity @xmath28 can be detected out to a maximum distance of d_max = ( ) ^1/2 = l_p , i^1/2 ( ) ^1/2 . for a steadily emitting pulsar with period - averaged luminosity @xmath29 and duty cycle @xmath30 , we have @xmath31 . the amount of pulse broadening depends on system parameters as well as dispersion and scattering , which vary with direction on the sky so that @xmath32 . we use the ne2001 model of galactic ionized electron density ( @xcite ) to calculate representative results for @xmath15 in the direction @xmath33 , a region of overlap between palfa and the parkes multibeam survey . [ fig : dmax_vs_lp ] shows @xmath15 vs. @xmath28 detection curves using a threshold @xmath34 for both surveys . for lower luminosities , sources are not visible to large enough distances for scattering to affect detectability and the inverse square law dominates the detection curve so that @xmath35 . for larger distances and smaller intrinsic pulse widths , scattering and ( residual ) dispersion smearing make pulses increasingly harder to detect and @xmath15 increases more slowly with @xmath28 . is about twice as large for palfa than for the parkes multibeam survey because of the larger sensitivity of the arecibo telescope . breaks in the curves correspond to a transition from luminosity - limited to scattering - limited detection for palfa and to a dispersion - limited regime for parkes . the difference is due to the smaller channel width of palfa ( 0.4 mhz ) compared to parkes ( 3 mhz ) . ] this section describes processing methods employed by the cornell pulsar search pipeline at the cornell center for advanced computing and the swinburne university of technology . the presto search code is run independently at west virginia university , university of british columbia , mcgill university and the university of texas at brownsville . presto uses a similar matched filtering algorithm to the one described below , but a different rfi excision scheme and trial dm list . we dedisperse raw data with 1272 trial dms in the range @xmath36 pc @xmath37 . in order to find individual pulses from intermittent sources we operate on dedispersed time series with two time domain algorithms : matched filtering , which has been the standard in single pulse searching so far , and a friends - of - friends algorithm . after single pulse candidates have been identified we use a stacking method in the time - frequency plane to verify that the pulses are dispersed and the sweep observed across frequencies follows the dispersion relation , as expected for non - terrestrial sources . certain types of terrestrial signals , e.g. swept - frequency radars , have pulses whose appearance in the time - frequency plane may approximate dispersion by ionized gas , but on closer inspection such pulses generally deviate significantly from the cold plasma dispersion relation . matched filtering detection of broadened pulses relies on the convolution of a pulse template with the dedispersed time series . ideally , a pulse template would consist of one or several superimposed gaussian templates , reflecting the diversity of pulsar pulse profiles , some of which exhibit multiple peaks . in addition , multipath propagation due to scattering adds an exponential tail to the pulse profile . we approximate true matched filtering by smoothing the time series by adding up to @xmath38 neighboring samples , where @xmath39 , and selecting events above a threshold after each smoothing iteration . the smoothing is done in pairs of samples at each @xmath40-stage , so that the resulting template is a boxcar of length @xmath38 samples ( @xcite ) . the sampling time used in palfa observations is 64 @xmath8s and therefore our matched filtering search is most sensitive to pulse widths of 64 @xmath8s to 8.2 ms . this search strategy is not optimal for single pulses from heavily scattered pulsars because the templates are symmetric , while the scattered pulse shape with an exponential tail is not , and significantly scattered pulses can be wider than @xmath41 ms . two issues make a complement to matched filtering necessary . the pulse templates described above have discrete widths of @xmath42 samples by algorithm design and there is decreased sensitivity to pulses with widths that are significantly different . in addition , the matched filter search output may be dominated by bright , wide pulses from radio frequency interference ( rfi ) . in that case , a single rfi burst is detected as an overwhelming number of individual events instead of a single event . the cluster algorithm is similar to the friends - of - friends search algorithm used to find galaxies in optical images ( @xcite ) , and complements matched filtering by not restricting the width of expected pulse detections . the dedispersed time series is processed sequentially . each event above a threshold which is found is designated as the first of a cluster . a cluster of events is augmented by broadening it to include any adjacent samples above a threshold . gaps of @xmath43 samples are allowed within a cluster , and palfa data is processed with @xmath44 . the brightest sample of a cluster is recorded as the event amplitude , and the total number of samples in the cluster as its width . this approach is less sensitive to weak , narrow pulses but results in significantly fewer spurious events due to rfi . a limiting factor for the largest pulse width detectable by both the matched filtering and friends - of - friends search algorithms is the fact that data are analysed in blocks much shorter than the complete time series length . the mean and standard deviation of a block are used for thresholding events found within the block . this approach minimizes the effect of baseline variations with time scales much longer than typical pulsar pulse widths of a few to a few tens of ms . the disadvantage is that only pulses with @xmath45 can be detected . in processing palfa data , we use blocks of length @xmath46s @xmath47 s. according to the ne2001 model of ionized gas in the galaxy ( @xcite ) , a scattering broadening time of that magnitude in the inner - galaxy part of the palfa survey corresponds to @xmath48 pc @xmath37 and a maximum search distance well outside of the milky way for most directions we survey . however , lines of sight intersecting hii regions can result in large dispersion measures and scattering times and are therefore selected against . any intergalactic scattering that broadens pulses beyond about 0.1 s will also cause events to be selected against . [ j0627best ] shows standard single pulse search output for the discovery observation of psr j0627 + 16 ( 8.1 ) . the main panel shows signal - to - noise ratio ( s / n ) vs. dm and time for events with @xmath49 . the panels on top illustrate event statistics for the observation : from left to right , number of events vs. s / n and dm , and dm vs. s / n . the pulsar detection is manifested as a peak in the histogram of number of events vs. dm . there is a corresponding peak in the dm vs. s / n plot , since the several detected pulses are significantly brighter than background noise in the dedispersed time series . vs. time and dm with larger circles denoting brighter bursts . panels on top from left to right show histograms of the number of events vs. s / n and dm , and a scatter plot of event dm vs. s / n . a fit to the narrow peak of event dm vs. s / n indicates a pulse width @xmath50 ms ( @xcite ) . ] [ rratquick ] shows events in dm - time space for another palfa single pulse discovery , psr j1928 + 15 ( 8.5 ) . in this case , three closely spaced bursts were found at @xmath51 s , @xmath52 pc @xmath37 . the pulsar is detected at a range of dms , with the signal to noise ratio increasing as trial dm values approach the actual pulsar dm and decreasing as trial dms further on recede from the pulsar dm . in contrast , events due to interference from terrestrial signals at @xmath53 s span the entire range of trial dms and their signal to noise ratios do not show a significant variation with dm . if an excess of candidate pulses is identified in the dedispersed time series for a particular trial dm , we use the expected dispersion sweep across the frequency band in order to test if the pulses are from non - terrestrial origin . for a pulse with @xmath54 in the time series , we extract a chunk of raw data centered on its time of arrival and look for a sweep across frequency that follows the dispersion relation ( fig . [ 4dynspecs ] a , b ) . in the case of multiple weaker pulses detected in the same beam , we extract a raw data chunk centered on the arrival time of each pulse , stack the chunks and look for a dispersion sweep in the resulting cumulative dynamic spectrum ( fig . [ 4dynspecs ] c , d ) . stacking can induce spurious dispersion sweeps in the time - frequency plane even when the data consist of only gaussian - distributed noise . single pulse events identified in dedispersed time series by definition correspond to sums over a given dispersion path that are above average . when these are used to select chunks for summing , the statistical fluctuations will build up to show a dispersed pulse in the time - frequency plane that follows the cold - plasma dispersion law perfectly . we simulated this effect by generating a time - frequency plane of gaussian noise , dedispersing it , and selecting events above threshold from the resulting time series . we extracted chunks from the fake time - frequency plane centered on each event and stacked them , thus reproducing the procedure used on real data . selecting events with signal - to - noise above a threshold @xmath55 in the dedispersed time series yields a biased set of noise - only samples in the time - frequency plane . on the other hand , when stacking dynamic spectra , points in the stacked dynamic spectrum must have a minimum signal - to - noise @xmath56 for a dispersion sweep to be discernible by eye . the rms noise in the dedispersed time series after summing @xmath57 channels is a factor of @xmath58 larger than the rms noise in the time - frequency plane . the rms noise in the stacked dynamic spectrum after stacking @xmath59 chunks is a factor of @xmath60 larger than the rms noise in the time - frequency plane . we calculate the average deviation from the mean in the time - frequency plane implied by the two thresholds @xmath55 and @xmath61 , equate them , and obtain n_c ( m_c / m_d)^2n_ch . for the palfa survey and our processing pipeline , @xmath62 and @xmath63 . therefore if @xmath64 data chunks are added there will be a spurious , dispersed pulse in the resulting stacked dynamic spectrum even if the chunks contain only noise . since the noise decorrelates over @xmath65 samples , the pulse will have a width on the order of @xmath66s . to avoid the induced spurious - event effect , the width and s / n of stacked dynamic spectra should be compared with those that would result from noise only . all of our detections that use stacking satisfy the criteria @xmath67s and @xmath68 . typically we sum chunks centered on the brightest @xmath69 pulses for a single pulse candidate , which is safely below the @xmath70 limit . in the case of psr j1909 + 06 ( fig . [ 4dynspecs]d ) only 2 chunks were added and the dispersed swath in the time - frequency plane is 1 ms wide . the rfi environment at the arecibo telescope and the 7-beam configuration of the alfa receiver present both challenges and opportunities for rfi mitigation to facilitate searching for single pulses . palfa survey data are dedispersed with trial dispersion measures of @xmath36 pc @xmath37 . rfi pulse intensity typically peaks at @xmath71 pc @xmath37 , and incidental low - intensity pulses whose s / n peaks at dm = 0 pc @xmath37 tend to be smeared below the detection threshold for dedispersed time series with trial @xmath72 pc @xmath37 . a more complex signature is observed for federal aviation administration ( san juan airport ) radar pulses , which are unfortunately common in arecibo observations . the radar rotation period is @xmath73 s , and each pulse has an envelope that is @xmath74 s wide and consists of sub - pulses with variable period on the order of @xmath75 ms . depending on telescope orientation , radar pulses may be detected in all , some , or none of the alfa beams due to reflections off the telescope structure . radar pulses are up to two orders of magnitude brighter than pulsar pulses and , without mitigation , can completely dominate single pulse search results . in addition , the modulation of the radar signal is manifested as detections with @xmath76 pc @xmath37 so that unlike other non - radar rfi , their s / n vs. dm signature can not be used for excision . we exploit the known radar characteristics as well as the pattern of pulse detection in the 7 alfa beams to excise both radar and non - radar rfi . the first part of our excision algorithm targets radar pulses . after a list of single pulse events is generated for all trial dms , we bin the events for trial @xmath77 pc @xmath37 in time ( by 0.1 s ) and record the number of events in each time bin . then we treat the histogram as a time series and perform a discrete fourier transform . a peak near the radar rotation frequency indicates a significant number of radar pulses in the data . from the fourier components we extract the phase and find the arrival time of the earliest radar pulse , @xmath78 . since the envelope width is @xmath74 s , events within 0.5 s of that location are excised , and the procedure is repeated for events near @xmath79 , where @xmath80 is an integer . the more radar pulses are present , the better the performance of this technique because the radar peak in the dft is more prominent and the pulses arrival times are determined more precisely . however , if only a couple of strong radar pulses are present within the typical 268 s palfa integration time , they may be bright enough to dominate event statistics , yet the dft method does not excise non - periodic incidental rfi . consequently , after applying the dft - based method we use an additional rfi filter that handles aperiodic cases . the second filter uses the number and proximity of beams in which an event is detected in order to determine if it is due to rfi . again events for trial @xmath77 pc @xmath37 from each beam are binned in time . after detecting peaks indicating an excess of events for the respective time bins , the algorithm cross - checks between results for all seven beams , and each event falling within a histogram peak receives a penalty grade based on how many beams histograms exhibit a peak and how close to each other they are on the sky . most pulsars detected blindly via a single pulse search appear in one beam or two adjacent beams , and very bright pulsars may be detected in three or four adjacent beams . we set the excision penalty threshold just below the value corresponding to the latter configurations and excise events accordingly . this method complements the dft cleaning scheme and effectively excises sparse radar blasts as well as non - periodic rfi detected in multiple beams . the application of the two excision algorithms makes a marked difference in the final single pulse search output for pointings contaminated with rfi ( fig . [ spcleaned ] ) . the figure shows some false positives , for example a clump of pulsar pulses in beam 4 around @xmath81 s are excised due to low - level rfi at low dms occurring in non - adjacent beams in the same 0.1 s time bins as the pulses . lowering the threshold according to which an excess of events is defined and time bins are marked for excision reduces false positives but also diminishes the effectiveness of rfi excision . while tens of pulses are detected within 134 s from the bright pulsar b2020 + 28 ( fig . [ spcleaned ] ) , for sources discovered via a single pulse search the number is an order of magnitude smaller ( see 7 ) . therefore the chance of rfi occurring simultaneously with pulses of intermittent sources and causing them to be excised is much lower . vs. dm channel and time , number of events vs. dm channel , and event s / n vs dm channel . the pulsar detection in beams 3 and 4 is evident after rfi events are excised . ] vs. dm channel and time , number of events vs. dm channel , and event s / n vs dm channel . the pulsar detection in beams 3 and 4 is evident after rfi events are excised . ] the palfa survey has made a total of 354 blind detections of 172 pulsars up to late 2008 . following quick - look processing of the data with degraded time and frequency resolutions at the time that data are acquired ( @xcite ) , data are shipped to external sites for processing through the cornell pulsar search code and presto . table [ detection_stats ] shows a breakdown of cornell detections , including pulsars originally discovered by the quick - look pipeline , by the periodicity ( fft ) and single pulse ( sp ) search algorithms for known pulsars and new discoveries , and table [ psr_params ] lists parameters for palfa single pulse discoveries that are discussed in more detail in [ section_objects ] . columns in table [ psr_params ] are ( 1 ) pulsar name ; ( 2)-(3 ) : equatorial coordinates ; ( 4 ) period ; ( 5 ) pulse width ( fwhm ) ; ( 6 ) dispersion measure ; ( 7 ) distance ; ( 8) peak flux density of the brightest detected pulse ; ( 9 ) total observation time ; ( 10 ) total number of pulses detected ; and ( 11 ) implied average pulse rate . the last column ( 12 ) points to the paper section that discusses the object . ccccc known & 48 ( 38% ) & 5 ( 4% ) & 73 ( 58% ) & 126 + new & 28 ( 61% ) & 6 ( 13% ) & 12 ( 26% ) & 46 lllcccccrrrc j0627 + 16 & 06:27:13(7 ) & 16:12(2 ) & 2.180 & 0.3 & 113 & 3.2 & 150 & 7454 & 48 & 23 & [ 0627sec ] + j0628 + 09 & 06:28:33(7 ) & 09:09(2 ) & 1.241 & 10 & 88 & 2.5 & 85 & 1072 & 42 & 141 & [ 0628sec ] + j1854 + 03 & 18:54:09(7 ) & 03:04(2 ) & 4.559 & 50 & 216 & 5.5 & 14 & 388 & 9 & 84 & [ 1854sec ] + j1909 + 06 & 19:09:24(7 ) & 06:40(2 ) & 0.741 & 1.5 & 35 & 2.2 & 82 & 536 & 10 & 67 & [ 1909sec ] + j1919 + 17 & 19:19:47(7 ) & 17:44(2 ) & 2.081 & 100 & 148 & 5.3 & 12 & 393 & 35 & 320 & [ 1919sec ] + j1928 + 15 & 19:28:20(7 ) & 15:13(2 ) & 0.403 & 5 & 242 & 7.4 & 180 & 2880 & 3 & 4 & [ 1928sec ] + j1946 + 24 & 19:46:00(7 ) & 23:58(2 ) & 4.729 & 4 & 96 & 4.3 & 101 & 268 & 4 & 54 & [ 1946sec ] + more than half of the detected known pulsars and only a quarter of the newly discovered objects were seen by both the periodicity and single pulse search algorithms . there is also a significant difference in the fraction of pulsars detected only via a single pulse search . in total , 74% of the new pulsars as opposed to 42% of the known pulsars were detected either only by fft search or only by single pulse search . the much higher percentage of new pulsars to be detected only by one algorithm means that most of the palfa discoveries are either periodic emitters too weak to show up in single pulse searches or intermittent objects whose emission is too heavily modulated be detected by a periodicity search . considering that bright pulsars are more likely to be seen by both the fft and sp search , and they are also more likely to have already been found by previous , less sensitive surveys , this indicates that palfa is probing deeper and finding pulsars that are farther away or less bright than the known objects in the same region of the galactic plane . single pulse searches have not been routinely carried out on survey data in the past , with the exception of @xcite , @xcite , who discovered psr j1918 + 08 via a single pulse search , and @xcite , who found 11 rrats . therefore , the proportion of single - pulse - only detections is higher for the palfa discoveries than the known pulsars . the five known pulsars detected only by single pulse search are known steady emitters which were seen away from the beam center and therefore with reduced sensitivity . the sample of palfa single pulse discoveries ( [ section_objects ] ) includes one object which has not been detected again despite reobservations ( j1928 + 15 ) , two sources detectable in re - observations through their time - averaged emission ( psr j0628 + 09 and psr j1909 + 06 ) , and four long - period objects , one of which has the smallest known duty cycle for any radio pulsar ( psr j0627 + 16 ) . among the 11 parkes rrats , 10 were successfully redetected in subsequent parkes observations , and one ( j1839@xmath101 ) has not been redetected despite multiple attempts ( @xcite ) . psr j0848@xmath143 and psr j1754@xmath130 can sometimes be detected through their time - averaged emission with sensitive , low - frequency observations ( @xcite ) . six of the sources have periods greater than 2 s. in this section we discuss selection effects which may account for some of the observational characteristics of intermittent sources . characteristics of this population which remain unexplained by observational biases may indicate underlying intrinsic differences between these objects and conventional radio pulsars . they may belong to a genuinely intermittent and physically different class of neutron stars . we can define a general model for observed signal intensity in terms of the time and frequency - dependent flux density @xmath82 as i(t , ) = g_t(t , ) g_iss(t , ) s_i(t , ) + n(t , ) + rfi(t , ) , where @xmath83 is the telescope gain , @xmath84 is the variation factor due to interstellar scintillation , @xmath85 is radiometer noise , and @xmath86 is the contribution from terrestrial radio frequency interference . variations in telescope gain across the beams of the multi - beam receiver or within the power pattern of a single beam impact the detectability of a pulsar as either a single pulse or periodic source . as shown by the reobservations of psr j0627 + 16 , psr j1909 + 06 , psr j0848@xmath143 and psr j1754@xmath130 , classifying a source as intermittent may be due to insufficient sensitivity to detect its periodic emission during the discovery observation . the effectiveness of single pulse vs. fft - based periodic search depends on the pulse amplitude distribution , and the observed pulse amplitude distribution in turn depends on the pulsar location with respect to the telescope beam power pattern . [ pulsepdf ] shows the probability density function ( pdf ) of event amplitude for psr b1933 + 16 , a strong pulsar detected in two beams of the same palfa pointing . one beam of the alfa receiver is pointing directly at the pulsar and the low - intensity tail of the distribution is clearly visible since all pulses are above the detection threshold . the pulsar is also detected in the first sidelobe of an adjacent beam , in which case only the brightest pulses are above the detection threshold and only the high - intensity tail of the distribution is visible . thus an off - axis detection of a steadily emitting pulsar may misrepresent it as an intermittent source . a different set of effects which affect detectability in single pulse vs. periodic search is due to intervening ionized gas . turbulence in the ionized interstellar gas and the motion of the pulsar with respect to the gas introduce phase modulations in the pulsar emission which cause the observed intensity to change over time and frequency . in general , the scintillation gain factor has contributions from both diffractive ( diss ) and refractive ( riss ) interstellar scintillation , such that @xmath87 @xcite . the detectability of nearby pulsars with low dms can be significantly affected by diss , which can have time scales comparable to typical survey pointing times ( minutes ) and frequency scales @xmath88 . in terms of the observing frequency @xmath23 and the pulsar distance @xmath89 , the diss time scale @xmath90 . the scintillation bandwidth @xmath91 and scattering timescale @xmath92 are related through @xmath93 . an empirical fit from @xcite for @xmath23 in ghz and @xmath92 in ms gives an estimate for @xmath92 based on dm : _ s = -6.46 + 0.154 dm + 1.07 ( dm)^2 - 3.86 _ f _ [ eqn_taus ] , but there is significant scatter about this relationship . when the scintillation bandwidth and time scale are small , averaging over the observation time and bandwidth can quench diss . for refractive scintillation , @xmath94 and @xmath95 . unlike for diss , the time scale of riss increases with distance and therefore it can become very long for distant pulsars with relatively high dms . both riss and diss can either help or hinder pulsar detection due to the time - variable constructive or destructive interference of wave fronts they cause . for a single - pass survey , this means that some weak pulsars may be detected because they are modulated above the detection threshold , and some pulsars that are otherwise bright enough may be missed because of scintillation modulation . @xcite discuss in detail the detection probability for scintillating pulsars by single - pass and multi - pass surveys . since palfa is a single - pass survey , it may have missed pulsars that scintillate on time scales comparable to or longer than the observation time . none of our discoveries in table [ psr_params ] appear to be affected by diss , though riss over long time scales may affect the objects with higher dms . the detectability of a pulsar as a periodic source depends on the integrated pulse flux within the observation , which is determined by the number of periods within the observation time , @xmath96 , along with the pulse amplitude distribution . for the same integrated flux per pulse , fewer periods within a fixed @xmath97 mean that long - period pulsars are less likely to be detected than short - period pulsars by any method relying on detecting integrated flux such as an fft - based search or a continuum imaging survey . for fft searches in particular , this effect is compounded by the fact that when @xmath98 is small , the harmonics of interest are in the low - frequency part of the power spectrum , which is often dominated by non - gaussian red noise and rfi . hereafter we use the term `` small-@xmath98 bias '' to refer to the combined influence of these effects on pulsar detectability . both the periodicity and single pulse searches perform with varying efficiency depending on an object s degree of intermittency . in this section we present an intermittency measure method to quantify the relative performance of the two search algorithms , apply it to results from the parkes multibeam and palfa surveys , and discuss general implications for surveys . we can compare the detectability of objects by periodicity and single pulse search and attempt to narrow down different classes of intermittent emitters by calculating the intermittency ratio r = ( s / n)_sp/(s / n)_fft for each object in our sample of palfa and parkes detections . @xcite derive r = ( ) [ eqn_r ] , where @xmath99 and @xmath100 for a gaussian pulse shape , @xmath101 is the maximum expected pulse intensity within @xmath102 periods and @xmath103 is a modified average pulse peak intensity . the intermittency ratio is a system - independent measure of the efficiency of the two types of search for objects with different degrees of intermittency . in fig . [ r_vs_np ] , expected values of @xmath104 vs. @xmath105 are shown for an exponential pulse amplitude pdf and power - law distributions with various indices . lower limits on @xmath104 are given for intermittent sources which were not detected via periodicity search . in addition , we show @xmath104 values for pulsars that were detected using both the periodicity and single - pulse search algorithms in the palfa and parkes multibeam surveys . a total of 283 palfa pulsar detections and 255 parkes multibeam detections are shown in the plot , including multiple detections of some objects . vs @xmath96 for 283 palfa ( filled hexagons ) and 255 parkes multibeam ( squares , open hexagons ) pulsar detections . squares denote the 10 parkes rrats with period measurements , including two which can sometimes be detected through their time - averaged emission . the criterion for inclusion of canonical pulsars was a simultaneous periodic and single pulse detection . palfa points include both blind survey detections and targeted test observations of known pulsars . arrows denote lower limits for palfa and parkes intermittent sources . points with bars show the average @xmath104 for a set of observations of the same object , with the bars denoting the maximum and minimum @xmath104 values . millisecond pulsar detections appear at lower right . dotted lines show @xmath104 for power - law pulse amplitude pdfs with indices from 0.5 to 2.0 . a ratio of @xmath106 was assumed for the cutoff intensities @xmath107 and @xmath108 ( @xcite ) . the dashed line shows @xmath104 for an exponential pdf . a single pulse search performs better than an fft periodicity search for @xmath109 . ] in fig . [ r_vs_np ] , long - period pulsars are predominantly found in the upper left ( small @xmath105 , large @xmath104 ) and millisecond pulsars in the lower right ( large @xmath105 , small @xmath104).psr b0525 + 21 , with @xmath110 , has @xmath111 s and @xmath112 pc @xmath37 , which make it susceptible to both the small-@xmath105 bias and diffractive scintillation and therefore it is detected more readily as a single pulse source . at the other extreme is psr b1937 + 21 , a millisecond pulsar which emits giant pulses but is nevertheless detected with a much higher s / n in the periodicity search than in the single pulse search ( @xmath113 ) because its period is small compared to the palfa observation time and the normal pulse amplitude is exceptionally steady ( @xcite ) . this indicates that the effect of @xmath105 on detectability can overshadow individual properties , and in survey mode millisecond pulsars will show up overwhelmingly as periodic sources . in that case , processing the data with a single pulse search algorithm can reveal the presence of giant pulses in an otherwise steady emitter . in - between psr b0525 + 21 and psr b1937 + 21 , from upper left to lower right is a relatively smooth distribution of pulsars whose intermittency ratios are predominantly determined by the influence of @xmath105 on the detectability of all types of periodic emitters . we look for unusual objects among the outliers from that trend . there are a number of pulsars detected both in periodicity and single pulse search in the region where @xmath114 , including psr b0656 + 14 , which occasionally emits bursts much brighter than the average pulse and would arguably be classified as a rrat if located farther away ( @xcite ) . interspersed among them in the region are most of the palfa and parkes intermittent sources . psr j1909 + 06 , which was not detected in a periodicity search but is visible as a periodic source when folded with its known period , has @xmath115 , and the long - period palfa intermittent sources psr j1854 + 03 and psr j1946 + 24 have @xmath116 and 1.8 , respectively . of the parkes intermittent sources , seven have @xmath117 . psr j1754@xmath130 ( @xmath118 ) and psr j0848@xmath143 ( @xmath119 ) were detected through their time - averaged emission in low frequency follow - up observations . in the @xmath120 region we find only five objects : psr b0525 + 21 , the crab pulsar , parkes intermittent sources psr j1317@xmath15759 and psr j1819@xmath11458 , and palfa intermittent source psr j1928 + 15 . the crab pulsar has a high intermittency ratio due to its giant pulses but is otherwise a relatively steady emitter as periodic emission is consistently detected even if the giant pulses are ignored . of the remaining three intermittent sources none are young pulsars like the crab . the pulse amplitude distributions of psr j1317@xmath15759 and psr j1819@xmath11458 are described by power laws with indices @xmath74 ( @xcite ) , smaller than measured giant pulse indices of @xmath121 ( e.g. @xcite , @xcite ) . we conclude from fig . [ r_vs_np ] is that periodicity and single pulse searches should be used in combination by pulsar surveys since there are types of sources which are best detected by either method . as more intermittent objects are found , placing them in @xmath122 space may help us identify physically distinct populations . in addition , repeated sky coverage can probe intermittency at time scales of days to months . table [ psr_params ] lists parameters of the seven pulsars discovered by palfa via a single pulse search . in this section we discuss each object s properties in detail and describe the steps taken in addition to the palfa processing pipeline in order to verify the signals celestial origin and calculate the pulsar period from times of arrival of single pulses . psr j0627 + 16 was discovered when nine single pulses were detected at a trial @xmath123 pc @xmath37 ( fig . [ j0627best ] ) . the period estimated from the pulse arrival times in the discovery observation is @xmath124 s. what is unusual about this object is its very narrow peak in s / n vs. dm space , indicating a narrow pulse ( @xcite ) . fig . [ 4dynspecs]c shows the cumulative dynamic spectrum after stacking raw data chunks aligned around the five brightest pulses . a fit to the dispersion sweep in the time - frequency plane gives a best dm of 113 pc @xmath37 and when the raw data chunk around the brightest pulse is dedispersed with this dm , the fwhm pulse width is only 0.3 ms . two follow - up observations at 0.33 and 1.4 ghz yielded 17 pulses in 50 minutes and 22 pulses in 72 minutes , respectively . after removing bright single pulses from the dedispersed time series and replacing the respective time samples with a running average , periodic emission with @xmath124 s was detected throughout the 0.33 ghz observation , confirming the period estimate from the discovery observation at 1.4 ghz and the presence of underlying normal emission . many of the rrats detected by @xcite have short duty cycles , calculated as the ratio between the average width of single pulses and the period . psr j0627 + 16 has the smallest known duty cycle , @xmath125 by this definition . however , for steady emitters the duty cycle is typically defined as the ratio of the folded pulse profile fwhm and the period . individual pulses may occur in a phase window wider than any individual pulse and corresponding to the folded profile width . at 1.4 ghz , the individual 0.3 ms pulses of j0627 + 16 occur in a 8 ms window , while at 0.33 ghz individual pulses have widths of @xmath126 ms and occur in a 15 ms window . the folded pulse profile at 0.33 ghz has fwhm of 60 ms , suggesting that bright pulses may be confined to a narrower window than pulses comprising the underlying normal emission . the detection of bright single pulses and a weak periodic signal from psr j0627 + 16 when data are folded with an accurate period estimate , along with the absence of a periodic detection in search mode weighs in favor of the proposition of @xcite that some rrats may have periodic emission with highly variable pulse amplitudes . psr j0628 + 09 was discovered by detecting three pulses at @xmath127 pc @xmath37 . the period estimated from the pulse times of arrival was 2.48 s. in follow - up observations the pulsar was also detected in periodicity searches as it emits on average several bursts per minute . the periodicity detections and the larger number of single pulses in subsequent observations allowed the actual pulsar period of 1.241 s to be determined ( @xcite ) . psr j1854 + 03 was discovered in 2008 via a single pulse search performed on full - resolution survey data . four pulses were detected at @xmath128 pc @xmath37 during the 268 s observation . the period @xmath129 s was estimated by taking the smallest difference between times of arrival ( toas ) of two consecutive pulses and verifying that intervals between all four toas are integer multiples of it . a confirmation observation with the more sensitive central alfa beam yielded within 120 s five pulses whose arrival times match the estimated period . psr j1909 + 06 was observed in 2006 but not identified as a candidate until 2007 when the full - resolution data were searched for single pulses . two pulses with a width of @xmath74 ms and signal to noise ratio of 6 and 9 were detected at @xmath130 pc @xmath37 . the stacked dynamic spectrum clearly shows a dispersion sweep ( fig . [ 4dynspecs]d ) . psr j1909 + 06 was discovered in data from the off - axis beam 2 of the multi - beam alfa receiver . since the on - axis gain of the center beam is 10.4 k / jy compared to 8.2 k / jy for the other six beams ( @xcite ) , we aimed the center beam at the discovery coordinates for a confirmation observation . for the same integration time of 268 s , 8 pulses with @xmath131 were detected in the confirmation observation . we used the pulse arrival times to determine a period and arrived at a best estimate of @xmath132 s. during the 2007 discovery observation of psr j1919 + 17 , multiple pulses at @xmath133 pc @xmath37 were detected in beam 4 of the alfa receiver , with no corresponding periodic detection . a fast folding analysis ( @xcite , @xcite ) was used to narrow down the period to @xmath134 s , which was confirmed in 2008 , when the pulsar was detected as a normal periodic emitter with the more sensitive central alfa beam . the pulsar has a double - peak profile with the two peaks @xmath5 ms apart , which most likely made it difficult to find its period from pulse arrival times alone . psr j1928 + 15 was discovered in 2005 by detection of what looked like a single bright pulse at @xmath135 pc @xmath37 in a 120 s observation ( fig . [ rratquick ] , top ) . more detailed analysis revealed that the event was in fact composed of 3 separate pulses occurring at intervals of 0.403 s , with the middle pulse being brighter by an order of magnitude than the other two ( fig . [ rratquick ] , bottom ) . in fig . [ 4dynspecs]a the dispersion of the brightest pulse by ionized interstellar gas is shown in the time - frequency plane , evidence of the non - terrestrial origin of the pulses . a fit to the pulse signal in the time - frequency plane resulted in a refined estimate of @xmath136 pc @xmath37 . despite multiple follow - up observations , the source has not been detected again . given the dm of this source , 242 pc @xmath37 , it is unlikely that the non - detection can be attributed to diffractive scintillation . since the three pulses are equally spaced , they can be interpreted as a single event seen in successive rotations of a neutron star . this signature might be accounted for by an object that is dormant or not generally beamed toward the earth and whose magnetosphere is perturbed sporadically by accretion of material from an asteroid belt @xcite . psr j1946 + 24 has @xmath137 pc @xmath37 and @xmath138 s and was discovered by detecting 4 individual pulses . the intervals between detected pulses in this case were all @xmath139 s. the brightest detected pulse of psr j1946 + 24 has @xmath140 and its dispersion sweep is visible in the time - frequency plane without any stacking ( fig . [ 4dynspecs]b ) . single pulse arrival times collected over multiple follow - up observations were used to obtain partial timing solutions for psr j0628 + 09 , psr j1909 + 06 , psr j1919 + 17 , psr j1854 + 03 and psr j1946 + 24 and verify their estimated periods . the periods of psr j0627 + 16 and psr j1909 + 06 were verified by manually folding raw data and detecting a periodic signal . finally , despite the fact that psr j1928 + 15 was not detected after the discovery observation , the @xmath141 s intervals between the three pulses detected from that source differ by only @xmath50 ms , which is approximately half the fwhm pulse width and therefore provides a period estimate to that precision . we are currently timing the six sources which were successfully redetected . obtaining phase - connected timing solutions will enable us to compare them to other neutron star populations and will yield positions which will facilitate observations at higher wavelengths . the palfa survey was designed to detect pulsars and pulsar - like objects by identifying either ( or both of ) their periodic or single - pulse emission , as discussed earlier in this paper . the same data may also be used to detect transient events from other galactic and extragalactic sources . at minimum , the palfa survey can place limits on the rate and amplitude distribution of radio transients that are shorter than about 1 second in duration . possible source classes for short - duration transient events , to name a few , include flare stars , magnetar bursts , gamma - ray bursts ( grbs ) , and evaporating black holes . grbs have diverse time scales spanning @xmath41 ms to 1000 s. mergers of double neutron - star ( dns ) binaries and neutron star - black hole ( ns - bh ) binaries are proposed sources for short bursts , and they may also emit contemporaneous radio pulses ( @xcite , @xcite ) . merger rates in the galaxy are estimated to be @xmath74 to 150 myr@xmath142 for dns binaries from pulsar surveys and a factor of 20 - 30 less for ns - bh binaries from population synthesis studies ( e.g. , @xcite ) . event rates of detectable events depend critically on the luminosity of radio bursts but are likely to be comparable to the rate of short - duraction grbs , and thus smaller than @xmath74 event day@xmath142 hemisphere@xmath142 . another phenomenon that may produce rare , bright pulses is the annihilation of primordial black holes ( @xcite ) . @xcite estimate the galactic rate for such events to be @xmath143 kpc@xmath144 yr@xmath142 . other possibilities are discussed in @xcite . motivated by these considerations , we discuss the constraints that can be made on transient events from palfa survey data obtained to date . first , we note that most but not all of the aperiodic events detected to date in palfa data ( as reported in [ section_objects ] ) and in the parkes multibeam survey by @xcite , appear to be periodic when they are reobserved sufficiently to establish a periodicity . additionally , _ all _ events detected so far in both surveys have dispersion measures that can be accounted for by ionized gas in the galaxy ( using the ne2001 model ) , suggesting that the emitting sources are closely related to the standard galactic pulsar population . by constrast , @xcite recently detected a strong ( 30 jy ) , isolated burst with duration @xmath145 ms that shows dm @xmath146 pc @xmath37 , too large to be accounted for by modeled foreground material in the galaxy or by plasma in the small magellanic cloud , which is somewhat near the direction the event was found . in addition to detecting radio sources on axis in data streams from each of the seven beams of the alfa system , it is possible to detect strong events off axis with low but non - zero sensitivity over almost the entire hemisphere centered on the zenith . consequently , it is possible to detect relatively low - rate but very bright transients off axis as well as the much weaker but higher rate single pulses that we have detected in the on - axis parts of the beam pattern . to establish that a given event is celestial in nature , we need to distinguish it from terrestrial interference . for on - axis events , this classification process is aided by the directionality of the seven beams and the expectation that most real events are likely to be seen in three or fewer contiguous beams rather than in all seven beams . bright events may violate this expectation , however , because near - in sidelobes are fairly large ( c.f . figure 1 of @xcite ) . celestial and terrestrial radio signals can also enter the seven beams of the alfa receiver indirectly far off the nominal pointing axis through reflection and scattering off telescope support structures , as with any antenna . extremely bright signals ( celestial or terrestrial ) may be detected after scattering into one or more beams , and the effective gain is approximately that of an isotropic radiator with small gain for events incident far off axis from the pointing direction . for the arecibo telescope , which has an intricate support structure with many possible scattering surfaces , such wide - angle incidence is likely to provide multiple paths for radiation to enter the feed optics . radiation arriving along multiple paths will constructively and destructively interfere across alfa s focal plane , so that the event s strength will be nonuniform in the different data streams . when radiation enters the telescope optics from wide angles , it is difficult to distinguish celestial from terrestrial signals based on the nominal pointing direction of the telescope and on whether they are detected in a single beam , a subset of beams , or all beams . conceivably , extremely bright events from a single radio source could be emitted at very rare intervals that would be detected when the telescope is pointed nominally at directions separated by tens of degrees ( but within the same hemisphere ) . such sources and events will be exceedingly difficult to disentangle from terrestrial rfi . of course the most powerful confirmation of a particular event is the detection of similar events in reobservations of the same sky position . for most of the events reported in [ section_objects ] and by @xcite , redetections have been made in this manner . they have also relied on establishing that their signatures in the frequency - time plane conform to what is expected . the simplest celestial signals are narrow pulses that show differential arrival times in accord with the cold - plasma dispersion law ( e.g. @xcite ) . celestial objects may also show the effects of multipath propagation through the interstellar medium ( ism ) either in the frequency structure that is produced in the spectra of nearby pulsars or in the asymmetric broadening of the pulses from more distant objects . for cases where spatial information associated with the nominal telescope pointing direction is insufficient to constrain the celestial nature of a particular event , we must rely especially on a detailed study of the event s structure in the time - frequency plane . indeed the time - frequency structure of the event discussed by @xcite was a key part of the argument for its classification as a celestial event . a given reflector and feed antenna have a net antenna power pattern with a main lobe and near - in sidelobes that is similar to an airy function with a main - lobe angular width @xmath147 , where @xmath148 is the effective reflector diameter . the wide - angle part of the antenna pattern has far - out sidelobes that are independent of @xmath148 and have average gains @xmath149 , similar to that of an isotropic antenna . however peaks in the far - out sidelobe pattern can have gains significantly larger than @xmath150 , corresponding to special directions where scattering from the telescope support structure is especially efficient . for the arecibo telescope , we expect sidelobes to have the same general properties , with the caveat that support structures will introduce further complexity in the far - out sidelobe pattern that varies as the telescope is used to track a source . tracking a sky position is done by rotating the azimuth arm and moving the gregorian dome along the azimuth arm , thus changing the scattering geometry . we therefore expect the sidelobe structure to change significantly with azimuth and elevation of the targeted position on the sky . for the seven - beam alfa system , the main beams by design sample different parts of the sky , as do the near - in sidelobes . at wider angles , the power pattern of each feed overlaps the others but with considerable variation in gain from feed to feed . to derive simple constraints on burst amplitudes and rates , we expand the gain for each feed into inner beam and wide - angle terms : g ( , ) = g_max p_n ( , ) + ( 1 - _ b ) , where @xmath151 and @xmath152 are polar and azimuthal angles , respectively ; @xmath153 is the power pattern for the main beam and near - in sidelobes ; @xmath154 is the main - beam efficiency , and @xmath155 is the average level of the far - out sidelobes . the boresight gain is g_max = = = , [ eqn_gmax ] where @xmath156 is the effective telescope area , and @xmath157 and @xmath158 are the solid angles of the antenna power pattern and of the main beam , respectively . the system equivalent flux density is s_sys = = , [ eqn_ssys ] , where @xmath159 is the boltzmann constant and @xmath160 is the system temperature . the minimum detectable flux density is s_min = , [ eqn_smin ] which is a function of the event duration , @xmath19 , as well as system parameters . it is useful to define @xmath161 as the minimum detectable flux density for unit gain using eq . [ eqn_ssys ] with @xmath162 . substituting into eq . [ eqn_smin ] yields , for two polarizations and @xmath163 mhz , s_min , 1 = 10 ^ 4.9 jy ( ) ( ) ^1/2 ( ) ; we note that for some directions , the system temperature can differ from the fiducial 30 k used . we now derive limits that can be placed on the rate and amplitudes of transient events using our simple model for the antenna power pattern . let @xmath164 be the event rate per unit solid angle from a population of sources , which implies a total number @xmath165 in a total observation time @xmath97 . if no events are detected , we can place an upper limit on the rate @xmath166 for minimum flux density , @xmath167 . we treat the response of the antenna power pattern as azimuthually symmetric about the bore axis , so @xmath168 , and calculate the gain as a function of polar angle from the bore axis . for each annulus of solid angle @xmath169 , the upper bound on the rate is _ s , max , for flux densities greater than s_min = . the multiplier @xmath170 accounts for whether a given alfa feed and receiver combination samples the same patch of sky or not . for the main lobes , @xmath171 while for the far - out sidelobes , @xmath172 . we evaluate fiducial values by integrating respectively over the main beams and over the remainder of the power patterns . for a main - lobe beam width of 3.5 arc min , @xmath173 deg@xmath13 and @xmath174 . using the average gain over the main beam out to the half - power point , @xmath175 , we have s_min = 11 mjy ( ) ( ) ^1/2 ( ) and _ s , max = 0.12 hr^-1 deg^-2 ( ) . integrating over wide angles , and using @xmath176 , we estimate s_min 10 ^ 5.4 jy ( ) ( ) ^1/2 ( ) , where the far - out sidelobes have an average gain of @xmath177 , and _ s , max = hr^-1 deg^-2 ( ) . where @xmath178 is the fraction of the hemisphere actually covered by the far - out sidelobes . results for palfa data are shown in fig . [ farout_palfa ] . within the half - power beam width the constraints are strong on the flux density but weak on the event rate while the opposite is true for the far - out sidelobes , represented by the part of the curve below @xmath179 h@xmath142 deg@xmath180 . the asympotic value of the curve corresponding to the maximum , boresight gain differs from the fiducial value calculated above because the latter is an average over the main beam . in our simple model the near - in sidelobes are not included explicitly so our curves overestimate @xmath167 at the corresponding polar angles . however , the limits change rapidly , so this does not change the salient features of the figure , which are the seven order - of - magnitude difference in constrained amplitudes and the eight order of magnitudes on the event rates between the main lobs and the far - out sidelobes . h and an assumed pulse width of @xmath181 ms . the three points , from left to right , correspond to the far - out sidelobes , the break between the main beam and far - out sidelobes ( at @xmath182 h@xmath142 deg@xmath180 ) , and the boresight gain . the shaded region is excluded by the palfa data . [ farout_palfa ] ] a brief summary of our results is that ( a ) a small number of events has been detected from sources that are consistent with their membership in the radio pulsar population , but with a much greater deal of modulation in some cases ; ( b ) no very strong pulses have been identified in the far - out sidelobes of the arecibo telescope and ( c ) no bursts have been detected that are extragalactic in origin . here we compare our results with those of @xcite , who surveyed high - galactic - latitude regions that included the magellanic clouds and who discovered a single reported pulse , with amplitude 30 jy . they surveyed @xmath183 deg@xmath13 with a detection threshold @xmath184 mjy , about 100 times fainter than the detected pulse amplitude . to compare any two surveys , we initially assume a population of radio sources that is homogeneously distributed in euclidean space and we ignore propagation effects that might limit the detectability of bursts from distant sources . the number of events detected @xmath185 for a survey of duration @xmath186 that samples instantaneously a solid angle @xmath187 . for multibeam systems with @xmath188 pixels , @xmath189 , where @xmath190 is the solid angle of a single beam . we assume events have the same peak luminosity and pulse width @xmath19 and are detectable to a distance @xmath15 . the ratio of number of detected events in two surveys with total observation times @xmath191 and @xmath192 , and bandwidths @xmath193 and @xmath194 is & = & ( ) ( ) ( ) ^3 + & = & ( ) ( ) ( ) ^3/2 ( ) ^3/4 . [ eqn_nbna ] as pointed out by @xcite , the same survey that detected the 30-jy pulse should have yielded many more detections if the assumptions of a volume limited sample apply . using eq . [ eqn_nbna ] , the lone 30 jy pulse implies there should have been @xmath195 additional detections above a 100-times fainter threshold . in the 90 h of follow up observations , there should have been @xmath196 additional detections . comparing the parkes and palfa surveys , a system - equivalent flux density of 4 jy , and 100 mhz bandwidth . for the parkes survey we use 13 beams of 14 arcmin diameter , 480 h , a threshold of 600 , @xmath197 jy , and 288 mhz bandwidth . ] , we estimate that the palfa survey should have identified @xmath198 pulses above threshold . the palfa survey s null result on extragalactic events is therefore in accord with lorimer et al.s non - detection of events weaker than the single strong event they reported . with the assumptions made , these results suggest that the non - detections of pulses weaker than the 30-jy pulse are inconsistent with the source population having a distribution that is homogeneous and isotropic . there are caveats on these results related to the assumptions about the source population and to the possible role of propagation effects in limiting detectability . the minimum detectable flux density for the palfa survey is 30 mjy for a 5-ms pulse , implying that a pulse of 30 jy strength emitted at a distance @xmath199 is detectable to a distance @xmath15 given by @xmath200 . if the lone event detected by lorimer et al . is from a source at the attributed distance @xmath201 gpc ( based on assigning about half of the dm to the intergalactic medium ) , the palfa survey would detect fewer pulses than estimated above ( for constant luminosity ) because the universe is not old enough and from redshifting of the spectrum if it is steep in frequency . nonetheless , a cosmological population of constant - luminosity sources would extend to @xmath202 times the attributed distance , implying from the first form of eq . [ eqn_nbna ] that @xmath203 pulses should have been detected in the palfa survey , on average , while @xmath204 weaker pulses should have been seen in the discovery survey with parkes . the difference in yield under this alternative scaling results from the fact that the palfa survey can detect sources well beyond the cosmological population while the parkes survey is comparatively shallower ( by a factor of ten in flux density , i.e. 300 compared to 30 mjy ) , but covers a factor of 30 greater solid angle . consideration of a broad luminosity function does not alter these conclusions qualitatively . however , we would generally expect the pulse amplitude distribution to extend to _ lower _ flux densities ( from lower luminosities ) given that the sole reported event was much larger than the threshold and that , therefore , many more pulses are expected compared to the constant - luminosity assumption . pulse broadening from multipath propagation in either the ism or the intergalactic medium ( igm ) can also limit the numbers of events detected in a survey because the broadening increases with source distance . the maximum detectable distance is therefore diminished , as demonstrated for the ism in figure [ fig : dmax_vs_lp ] . lorimer et al . argued that the strong frequency dependence of the width of the 30-jy pulse was consistent with multipath propagation through a turbulent , ionized igm . if so , sources from a greater distance would be less - easily detectable . in the simplest scattering geometries , scattering conserves the area of the pulse , so matched filtering yields a test statistic with s / n @xmath205 when the pulse broadening time @xmath206 is much larger than the intrinsic pulse width ( @xcite ) . the 30-jy pulse therefore could have been broadened by an additional factor of @xmath207 and the s / n still would have been above threshold . the much longer pulses ( @xmath208 ms = 50 s ) would have been difficult if not impossible to identify in the time series owing to high - pass filtering used to mitigate baseline fluctuations . the high - pass filtering was explicit in the parkes multibeam data acquisition but is part of the post - processing in the palfa analysis . such very broadened pulses could be detected by imaging surveys with relatively short integration times like the nrao vla sky survey ( nvss ) and future surveys with the australian square kilometer array pathfinder or the allen telescope array . in between this extreme case of 50-s broadening and the reported pulse width of @xmath145 ms is a substantial search volume in which more distant sources could have produced detectable events . broadening from sources at redshifts greater than one is increased by the larger electron density but is lessened by the frequency redshift , so the net scaling of the pulse broadening with redshift depends on how the scattering material is distributed along the line of sight . we have examined the current state of knowledge about various classes of intermittent radio - loud neutron stars , summarized selection effects that may affect the classification of intermittent sources , and presented our data processing methods and results from single pulse searches performed on palfa survey data . one of our single - pulse discoveries is a relatively persistent emitter that , like many other pulsars , shows a broad amplitude distribution , from which a few bright pulses were detected . two sources were detected as normal periodic emitters when observed at a lower frequency ( psr j0627 + 16 ) and with the more sensitive central beam of the alfa receiver ( psr j1919 + 17 ) . four objects were most likely not detected via periodicity search because of their long periods ( @xmath209 s ) compared to the palfa observation time of @xmath210 s , and one intermittent object ( psr j1928 + 15 ) was discovered by detecting three pulses emitted on successive rotations but it was not detected again despite multiple reobservations . most of the palfa sources that were discovered solely through the single - pulse analysis have subsequently been found to be _ sporadic _ but with measureable and consistent rates of detectable pulses . the parkes survey had similar results . however , two sources ( the parkes source psr j1839@xmath101 and the palfa source psr j1928 + 15 ) are much more irregular even though an underlying periodicity has been found in each case . we therefore distinguish between cases where the apparent sporadic behavior is due to a detection threshold that selects only the strongest pulses from a broad distribution ( e.g. @xcite ) and those where emission is truly _ intermittent_. a major issue is related to the question , if we see transients from a given sky direction only once , how can we figure out what phenomena produce them . we must archive radio survey data and characterize the wide variety of detected radio transients outside the framework of rotating neutron stars if necessary . archived data may yield new discoveries if reprocessed with improved search algorithms and a database of transient signals would be an invaluable reference as theoretical understanding of energetic radio events and prediction of their signatures improves . the palfa survey is taking steps in both directions : raw data are stored in a tape archive and data processing products are indexed in a database at the cornell center for advanced computing and currently work is under way to make them publicly available via a web portal . we thank the staff at naic and atnf for developing the alfa receiver and the associated backend systems . in particular , we thank jeff hagen , bill sisk and steve torchinsky at naic and graham carrad at the atnf . we are also grateful to the parkes multibeam survey team for providing the results incorporated in fig . [ r_vs_np ] . this work was supported by the nsf through a cooperative agreement with cornell university to operate the arecibo observatory . nsf also supported this research through grants ast-02 - 05853 and ast-05 - 07376 ( columbia university ) , ast-02 - 06035 and ast-05 - 07747 ( cornell university ) and ast-06 - 47820 ( bryn mawr college ) . , d.r.l . and p.c.c.f . are supported by a wvepscor research challenge grant . is an alfred p. sloan research fellow . f. crawford is supported by grants from research corporation and the mount cuba astronomical foundation . j.w.t.h . is a nwo veni fellow . pulsar research at ubc is supported by nserc and the canada foundation for innovation . i.h.s . acknowledges support from the atnf distinguished visitor program and the swinburne university of technology visiting distinguished researcher scheme . l.e.k . held an nserc canada graduate scholarship while most of this work was performed . v.m.k . holds a canada research chair and the lorne trottier chair , and is supported by nserc , fqrnt , cifar , and by the canada foundation for innovation . the parkes telescope is part of the australia telescope , which is funded by the commonwealth government for operation as a national facility managed by csiro . nrao is a facility of the nsf operated under cooperative agreement by associated universities , inc . research in radio astronomy at the nrl is supported by the office of naval research . mclaughlin , m. a. et al . 2006 , nature , 439 , 817 mclaughlin , m. a. , rea , n. , gaensler , b. m. , chatterjee , s. , camilo , f. , kramer , m. , lorimer , d. r. , lyne , a. g. , israel , g. l. , possenti , a. 2007 , apj , 670 , 1307

in the context of a graphical model , the concept of local constancy has been coined recently in honorio et al . , 2009 . the goal of the present paper is to penetrate this idea by providing a thorough analysis and an extension of this concept . our approach is using the idea propounded by meinshausen and bhlmann ( 2006 ) for structure learning in gaussian graphical models , and it incorporates an additional fusion penalty term that aims to enforce the structural constraint of _ local constancy_. when the nodes of a graph have spatial information attached , then local constancy can be interpreted as a certain type of spatial regularity ( see below ) . conceptually , the entire neighborhood graph given by the graphical model is split into two subgraphs : ( a ) a _ local _ graph consisting of nodes that are spatially close , and ( b ) a _ non - local _ graph , where the edges connect nodes of two different spatial clusters of nodes . our approach then assumes that we have prior structural knowledge about the local graph that we aim to incorporate into the estimation / model selection approach . it is well known that a @xmath1-dimensional graphical model is given by a @xmath1-dimensional gaussian distribution with non - singular covariance matrix @xmath2 of a random vextor @xmath3 . the conditional independence of the components of @xmath3 can be represented by occurrence of zero entries of precision matrix @xmath4 . the non - zero entries of the precision matrix define the edges of the graph corresponding to this multivariate gaussian distribution . neighborhood selection algorithms aim to find all the neighbors of a node @xmath5 in a graphical model based on an i.i.d meinshausen & bhlmann @xcite showed that this problem can be interpreted as an ensemble of @xmath0-penalized regressions , each of which can be solved using the lasso algorithm of tibshirani ( 1997 ) . our proposal is an extension of this approach . in order to incorporate this additional structural constraint we assume that the prior structural knowledge can be translated into a _ local _ neighborhood graph , which we think of being comprised of regular graph objects such as chains , cycles , lattices or cliques . the availability of additional locality information is critical when data is observed in a certain manifold with spatial geometry . the assumption of local constancy , in some sense , enforces spatial regularization on structure learning and thereby stimulates the search for probabilistic dependencies between local clusters of nodes . we would like to emphasize that the knowledge of this prior structural information is based on domain knowledge and hence known beforehand . we propose to use the meinshausen & bhlmann approach and to add a new penalty term which , in essence , generalizes the fused lasso penalty ( see tibshirani , 2005 ) by extending the prenalty term over differences of nodes in the local neighborhood graph , which in turn is given by the prior structural information . this leads to the _ neighborhood - fused lasso _ procedure , a model selection method for locally constant gaussian graphical models . we provide theoretical and numerical evidence that our approach outperforms competing model selection algorithms where locality information is ignored . the paper is organized as follows : in section [ review ] we first provide an overview of related work . sections [ nfl ] and [ locconstdefn ] then provide a precise definition of our approach descibed above , and a discussion of local constancy , respectively . in section [ optimization ] we propose our optimization algorithm , before we present both both finite sample and large sample properties of our appraoch in section [ theory ] . we prove theoretically that introducing a local penalty term reduces the finite sample type - i error probability in model selection and leads to equivalent accuracy with smaller sample size than competitors . we also discuss data dependent choice of the penalty parameters , which avoids the use of cross - validation based methods . we study the asymptotic @xmath0 properties of our estimator to find sufficient conditions on design matrix and regularization parameters to find nice asymptotic bounds in parameter estimation and prediction in terms of @xmath0 and @xmath6 metric , respectively . in particular , our theoretical analysis reveals that with our assumptions our proposed method displays all the desired properties like sign - consistency and model selection consistency . some numerical results can be found in section [ simulations ] . the proofs are delegated to the appendix . going back to dempster @xcite who introduced _ covariance selection _ to discover the conditional independence restrictions ( the graph ) from a set of i.i.d . observations , many methods have been proposed for sparse estimation of the precision matrix in a gaussian graphical model . the proposed procedures usually rely on optimization of an objective function @xcite . while modern technological developments and high computing power enable us to deal with high dimensional models , there still are computational challenges . usually , greedy forward - selection or backward - deletion search is used . in forward ( backward ) search , one starts with the empty ( full ) set and adds ( deletes ) edges iteratively until a suitable stopping criterion is fulfilled . the selection ( deletion ) of an edge requires an mle fit @xcite for @xmath7 many models , making it a suboptimal choice for high - dimensional models where @xmath1 is large . also , the mle might not exist in general for @xmath8 ( see @xcite ) . in contrast , neighborhood selection using lasso , as proposed by meinshausen and bhlmann @xcite , relies on optimizing a convex function applied consecutively to each node in the graph , thus fitting @xmath9 many models . fast lasso - type algorithms and data dependent choices for regularization parameter reduce the computational cost . unlike covariance selection , this algorithm estimates the dependency graph by sequential estimation of individual neighbors and subsequent combination by taking unions or intersections . other authors have proposed algorithms for the exact maximization of the @xmath0-penalized log - likelihood . yuan & lin @xcite proposed an @xmath0-penalty on the off - diagonal elements of the concentration matrix for its sparse estimation with the positive definiteness constraint . they showed that this problem is similar to the _ maxdet _ problem ( see vandenberghe et al . @xcite ) , and thus solvable by the interior point algorithm . a quadratic approximation to the objective function in their proposed method leads to a solution similar to meinshausen & bhlmann . banerjee et al . @xcite viewed this as a penalized maximum likelihood estimation problem with the same @xmath0-penalty on the concentration matrix . constructing the dual transforms the problem into sparse estimation of the covariance matrix instead of the concentration matrix . they proposed _ block coordinate descent algorithm _ to solve this efficiently for large values of @xmath1 . they also showed that the dual of the quadratic objective function in the block - coordinate step can be interpreted as a recursive @xmath0-penalized least square solution . friedman , hastie & tibshirani @xcite used this idea successfully to develop an algorithm known as _ graphical lasso_. using a coordinate descent approach to solve the lasso problem speeds up the algorithm to a considerable extent , making it quite fast and effective for a large class of high - dimensional problems . in all the aforementioned methods , information on the local geometry is not taken into consideration . often one encounters data that are measured on a certain manifold . e.g. , data describing some feature on the outline of a ( moving ) silhouette , or pixels or voxels in an 2-d or 3-d image , respectively . in most of these problems , spatially close variables have a structural resemblance in terms of probabilistic dependence . exploiting this local behavior might lead to faster and/or more efficient estimation . honorio , ortiz , samaras et al . @xcite introduced the notion of _ local constancy_. according to them , if one variable @xmath10 is conditionally ( in)dependent of @xmath11 , then a local neighbor of @xmath10 in that manifold is likely to be conditionally ( in)dependent of @xmath11 as well . they developed a__coordinate direction descent algorithm _ _ to solve a penalized mle problem that penalizes both the @xmath0-norm of the precision matrix and the @xmath0-norm of local differences of the precision matrix , expressed as its diagonal excluded product " with a local difference matrix . local geometry has been addressed , although implicitly , by tibshirani et al . @xcite in the context of fused lasso . chen et al . @xcite used this idea and proposed _ graph guided fused lasso _ for structure learning in multi - task regression , where the output space is continuous and the outputs are related by a graph . the input space is high dimensional and outputs that are connected by an edge in the graph are believed to share a common set of inputs . the goal then is to learn the underlying functional map from the input space to the output space in a way that respects the similar sparsity pattern among the covariates that are believed to affect the output variables that are connected . local smoothing by penalizing differences of neighboring nodes has been discussed in kovac and smith @xcite in the context of nonparametric regression based on observations on nodes of a graph . their algorithm aims to split the image into active sets and subsequently merge , split or amalgamate them in order to minimize a penalized weighted distance . like greenshtein & ritov @xcite and meinshausen & bhlmann @xcite we work in a set - up where the number of nodes in the graph @xmath12 and the covariance matrix @xmath13 depend on the sample size . consider the @xmath12-dimensional multivariate random variable @xmath14 . the conditional ( in)dependence structure of this distribution can be represented by the graph @xmath15 , where @xmath16 is the set of nodes corresponding to each coordinate variable and @xmath17 the set of edges in @xmath18 . a pair of nodes @xmath19 lies in @xmath17 if and only if @xmath10 is conditionally dependent of @xmath11 , given all other remaining variables @xmath20 . the neighborhood @xmath21 of a node @xmath22 is defined as the smallest subset of @xmath23 such that given @xmath21 , @xmath5 is conditionally independent of all the remaining nodes . in other words , neighbors of a certain node consists of the coordinates that are conditionally dependent on that particular node . as already mentioned , the conditional independence can be represented by occurrences of zero entries at respective cells of precision matrix @xmath24 , i.e. , @xmath25 iff @xmath26 . the neighborhood selection / model selection algorithms aim to find all the neighbors of a node @xmath5 , given @xmath27 i.i.d . observations of @xmath3 . meinshausen & bhlmann considered this as a penalized regression problem , where each variable is regressed on the remaining variables with an @xmath0 penalty on the estimated coefficients . by the above definition of neighborhood , we have , for all @xmath28 that @xmath29 an alternative definition of neighborhood of @xmath10 is given as the non - zero components of @xmath30 where @xmath30 is given by @xmath31 in light of the above definition , the set of neighbors of a node @xmath28 is @xmath32 given the domain knowledge we construct a regular graph @xmath33 that is representative of the underlying spatial geometry . we call it a _ local neighborhood graph_. for example , @xmath33 could be a chain graph , a two or three dimensional lattice or more generally a collection of cliques or wheels . let @xmath34 denote the edge set corresponding to @xmath33 , then define : [ localgraphdefn ] @xmath10 and @xmath11 are _ local neighbors _ with respect to @xmath33 if the edge connecting them belongs to @xmath35 note that the edge between @xmath10 and @xmath11 need not belong to @xmath17 . now if one incorporates the _ local constancy _ property to this graph , the neighborhood becomes more structured . by local constancy , if @xmath36 is conditionally independent of @xmath5 given the other nodes ( and hence , there is no edge between @xmath22 and @xmath37 ) , it is likely that a _ local neighbor _ @xmath38 of @xmath36 is also conditionally independent of @xmath10 , making both @xmath39 and @xmath40 equal to zero . thus , the zeroes of @xmath30 are expected to reflect the sparsity pattern . we now generalize the traditional fused lasso algorithm @xcite to satisfy our purpose of estimating the zeros and non - zeros of the precision matrix by exploiting the locality information given by the local neighborhood graph @xmath33 . before doing that we carefully define the _ difference matrix_. consider an @xmath41 matrix @xmath42 where @xmath43 is the total number of pairs of local neighbors . assuming that we have a labelled sequence of nodes @xmath44 , arrange the pairs of local neighbors in a sequence @xmath45 where @xmath46 denotes the set of local neighbors of @xmath47 . the inequality @xmath48 is included to ignore double counting . note that @xmath49 is nothing but a convenient ordering of edges in @xmath33 , and hence @xmath49 contains the same information as @xmath33 . it should also be mentioned that our results are not influenced by the particular labelling used . the @xmath50 row is given by @xmath51 where @xmath52 is the @xmath50 element of the local neighbor sequence and @xmath53 denote two canonical basis vectors of @xmath54 where the @xmath55 s occur at @xmath56 and @xmath57 position , respectively . this way each pair is represented by a row in the difference matrix . we denote by @xmath58 the @xmath59 sub - matrix of @xmath42 selecting all the rows with @xmath60 entry being 0 . in other words , @xmath58 is the difference matrix corresponding to all the local neighbor pairs not involving @xmath10 . the number of local neighbors of @xmath10 in @xmath23 is @xmath61 . it should be noted that throughout our discussion we shall assume that the local neighborhood structure is known to us , meaning that @xmath42 is known beforehand and it does not depend on the data . now we define the neighborhood - fused lasso estimate @xmath62 of @xmath63 @xmath64 where @xmath65 denotes the @xmath0-norm of @xmath66 . penalizing both the @xmath0-norms of @xmath67 and @xmath68 implies parsimony , thus ensuring sparsity and local constancy at the same time . this property helps us in variable selection and thereby leads to neighborhood selection . note that we are referring to both the non - local and the local neighbors . the neighborhood estimate of node @xmath22 is defined by the nodes corresponding to non - null coefficients when @xmath10 is regressed on the rest of the variables with the fused lasso penalty in , in other words @xmath69 where @xmath70 denotes the @xmath37-th component of the vector @xmath62 . it is clear that the selected neighborhood depends on the value of @xmath71 and @xmath72 chosen . large values of @xmath71 and @xmath72 will give rise to more sparse solutions . usually the regularization parameters are chosen by some cross validation criteria . but in this paper we will find a data driven approach for selection of regularization parameters that speeds up computation and ensures asymptotic consistency in model selection . meinshausen and bhlmann derived a data driven choice for @xmath71 in their paper . we will extend their method for simultaneous selection of @xmath71 and @xmath72 from the data . here we discuss further formalizations of the notion of local constancy . it should be noted again that the definition is linked with the ideas used by honorio et al . @xcite , but our definition is more general in nature . first we introduce the notion of _ diagonal excluded matrix product_. given a matrix @xmath73 , its zero operator is defined as @xmath74 where @xmath75 is the indicator function . the diagonal excluded product of two matrices @xmath73 and @xmath76 can now be defined as , @xmath77 where @xmath78 denotes the hadamard product of matrices . although the name does not clearly show how diagonals are removed from the product , usually if @xmath73 is taken to be a difference matrix defined above , this will eventually lead to the exclusion of diagonals of the matrix @xmath76 . first observe that the absolute values of the non - zero entries of @xmath79 are of the form @xmath80 where @xmath81 corresponds to a pair of local neighbors . in fact , we have @xmath82 equals the sum of all the absolute values of the differences @xmath83 for local neighborhood pairs @xmath52 . in other words , if we think of the entires @xmath84 as measuring conditional dependence of two nodes , then @xmath82 measures in how different the local neighborhood pairs are with respect to conditional dependence . a small value of @xmath82 indicates that the local neighbors behave similarly in this respect . thus , one way to define the local constancy property is to impose a bound on the norm of the difference of local neighbors . we say a model exhibit @xmath85-local constancy if @xmath86 where @xmath87 denotes the @xmath88 norm . for our purpose , the best candidate would be @xmath89 for certain desirable properties like sparsity and convexity . however , a different choice of norm would lead to a different solution which might be more appropriate for other types of problems . with @xmath90 , this definition coincides with honorio s local penalty criterion . it should be noted that the @xmath91 parameter controls the degree of local constancy so this is closely related to the tuning parameter we are going to use . this serves as a constraint on certain matrix parameters that we are trying to estimate . a small @xmath91 would ensure a high level of local constancy where a large @xmath91 would do the opposite . since the locality information is being used here as a prior information , it makes sense to interpret it from a bayesian perspective . let @xmath92 denote a probability distribution on the space of positive definite matrices . then one can say that an @xmath93-local constancy property holds for this model if @xmath94 here we need to assume that @xmath95 . the @xmath91 parameter controls the degree of local constancy and the @xmath96 parameter is indicative of our prior belief about the local constancy . one can also think of defining local constancy by replacing @xmath24 in sections [ locconstdefn : quant ] and [ locconstdefn : bayes ] by the partial correlation matrix @xmath97 . even though the definition makes sense , the natural transition from the difference matrix used in @xcite to our analogue @xmath58 will not be consistent with it . the difference terms corresponding to partial correlation matrix will be @xmath98 but , when we regress @xmath99 on the rest of the nodes , then the local difference we aim to penalize is @xmath100 . in order to be consistent with our definition of neighborhood fused lasso , we need to replace @xmath101 by @xmath102 , where @xmath103 . assuming that the individual conditional variances are bounded away from zero ( assumption a2 in section [ ass : variance ] ) , this rescaling does not affect the local constancy and our penalty is consistent with the definition . partial correlations have the advantage of all being on the same scale . this might in particular simplify the choice of @xmath104 first we briefly review the existing algorithms used for optimization of fused lasso type problems . broadly speaking , there are two fundamental classes of algorithms used in this type of problems : ( a ) solution path algorithms - which finds the entire solution for all values of the regularization parameters and ( b ) approximation algorithms - which attempt to solve a large scale optimization given a fixed set of regularization parameters using first order approximations . friedman et al . @xcite formulated the path - wise optimization method for the standard fused lasso signal approximation problem where the design matrix @xmath105 . the algorithm was two - step and the final solution is obtained by soft - thresholding the total - variation norm penalized estimate obtained in the first step . the basic challenge in applying the coordinate descent algorithm to a fused lasso problem is the non - separability of the total variation penalty unlike usual lasso where the @xmath0-penalty is completely separable . so they used a modified coordinate descent approach where the _ descent _ step was followed by an additional fusion and smoothing step . however , this method works only for the total variation penalty and does not extend to fused lasso regression problems with generalized fusion penalty like our situation . hoefling @xcite proposed a path - wise optimization algorithm for generalized fusion penalties in fused lasso signal approximation problem . this algorithm uses the fact that when varying the penalty parameter , the corresponding sets of fused coefficients do only change at finitely many values of the penalty parameter . tibshirani and taylor @xcite proposed another path algorithm for the fused lasso regression problem with a generalized fused lasso penalty by solving the dual optimization problem . however , the path algorithm they devised can be applied only when the design matrix is full rank and hence is not applicable for high dimensional problems . in order to resolve the problem when the matrix is not full rank , they proposed to add an infinitesimal perturbation @xmath106 . however , this does not solve the problem completely as a small @xmath91 leads to ill - conditioning and increasing number of rows in the generalized fused penalty matrix causes inefficient solutions because of an increasing number of dual variables . approximation algorithms were developed to find efficient solutions to general fused lasso problem with a fixed set of penalty parameters , regardless of its rank and they usually adapt themselves to high dimensional problems quite easily . usually these approximation algorithms are based on first order approximation type methods like gradient descent . liu et al . @xcite proposed the efficient fused lasso algorithm ( efla ) that solves standard fused lasso regression problem by replacing the quadratic error term in the optimization function by its first order taylor expansion at an approximate solution followed by an additional quadratic regularization term . the approximate objective function has a fused lasso signal approximation form and can be solved by applying gradient descent on its dual which is a box - constrained quadratic program . chen et al . @xcite proposed the _ smoothing proximal gradient _ method to solve regression problem with structured penalties that closely resemble our objective function . the basic idea is to approximate the fused penalty term @xmath107 by a smooth function and devise an iterative scheme for an approximate optimization . they use the smooth function @xmath108 convexity and continuous differentiability of this function follows from nesterov @xcite . one may now proceed using standard first order approximation approach like fista @xcite which works like efla . however this process is computationally intensive and not a feasible approach in our case where we need to do a nodewise regression . another algorithm for tackling a similar problem is known as _ split bregman _ algorithm @xcite . it was proposed for standard fused lasso regression and later extended to generalized fused lasso regression . the sb algorithm is derived from augmented lagrangian @xcite , which adds quadratic penalty terms to the penalized objective function and alternatingly solves the primal and dual starting from an initial estimate . this is also computationally intensive unless one has a simple structure constraint on the parameters . we propose a different algorithm to solve our optimization problem . we show that the neighborhood - fused lasso problem can be reparametrized into a standard lasso problem , thus simplifying the optimization procedure . the neighborhood - fused lasso estimate @xmath62 of @xmath109 can be written as @xmath110 we note that since @xmath111 has full column rank , @xmath112 has full row rank . thus , the following lemma can be applied here . [ optim ] let @xmath113 and @xmath114 also assume that @xmath115 is of full column rank . then @xmath116 one interesting observation here is that although we started with the assumption that @xmath117 , where @xmath118 denotes the column space of @xmath111 , we did not optimize over @xmath118 but did the same over all @xmath119 . see the proof of lemma [ optim ] for details . the heuristic idea behind this is that we find the minimizer on @xmath118 by projecting the global minimizer onto @xmath118 and the projection operator is given by @xmath120 . the objective function in combines the two @xmath121 penalties into a single @xmath121 penalty and thus could be easily solved by any of the standard lasso algorithms . the parameters @xmath71 and @xmath72 are chosen according to theorem [ regparam ] . we show by several simulations that the proposed method performs better than meinshausen - bhlmann s method or graphical lasso in situations where local constancy holds . from discussion in section 3 , it can be seen that using neighborhood - fused lasso leads to efficient model selection when applied successively to all the nodes . in this section , we show that our proposed method leads to asymptotically consistent model selection similar to the procedure by meinshausen and bhlmann . the choice of the regularization parameters is crucial in such cases . moreover , we show that our proposed choice of regularization parameters not only ensures convergence to the true " model but the convergence is faster than the meinhausen - bhlmann s method when the underlying model is indeed locally constant . this property is also supported by our simulation results shown in section [ simulations ] . in order to prove consistency of our method for gaussian graphical models , we need to work with the following assumptions . assumptions [ a1]-[a3 ] and [ a5]-[a7 ] are as in meinshausen and bhlmann s ( see @xcite , section 2.3 ) . however we need the two additional assumptions [ a4 ] and [ a8 ] to deal with the local constancy . [ a1 ] : : [ ass : dimensionality ] * dimensionality * : @xmath122 , such that @xmath123 as @xmath124 . note that @xmath125 is included thus allowing @xmath126 . [ a2 ] : : [ ass : variance ] for all @xmath28 and @xmath127 , @xmath128 . there exists @xmath129 , so that for all @xmath127 and @xmath28 , @xmath130 . this means that all the conditional variances are bounded away from 0 . [ a3 ] : : [ ass : sparsity ] * sparsity * : there exists some @xmath131 so that @xmath132 for @xmath124 . [ a4 ] : : [ ass : lnsparse ] * local neighborhood sparsity * : the number of local neighbors also can grow at polynomial rate of @xmath27 , i.e. , @xmath133 such that the maximum number of local neighbors of a node @xmath134 for @xmath135 . for convenience we let @xmath136 be such that @xmath137 [ a5 ] : : [ ass : bounded ] * @xmath0-boundedness * : there exists some @xmath138 so that for all neighboring nodes @xmath139 and all @xmath127 , @xmath140 . [ a6 ] : : [ ass : parcor ] * magnitude of partial correlation * : there exists a constant @xmath141 and some @xmath142 with @xmath143 as in [ a3 ] , so that for every @xmath144 , @xmath145 where @xmath146 denotes the partial correlation of @xmath10 and @xmath11 . [ a7 ] : : [ ass : ns ] * neighborhood stability * : define @xmath147 . there exists some @xmath148 so that for all @xmath149 with @xmath150 , @xmath151 [ a8 ] : : [ lnstable ] * local neighborhood stability * : let @xmath152 and @xmath153_k \theta^{b,\mathcal{l}_a}_k$ ] . there exists some @xmath154 so that for all @xmath155 such that @xmath156 , @xmath157 , where @xmath158 denotes denotes the @xmath159 column of @xmath58 . similar to meinshausen and bhlmann s interpretation , we can describe an intuitive condition which implies the last two assumptions . define @xmath160 according to the characterization of @xmath161 derived from , @xmath162 . one can think of a two - dimensional perturbation approach in which one tweaks the parameters @xmath163 and @xmath164 in a way that the perturbed neighborhood @xmath165 is identical to original neighborhood @xmath166 . the following proposition shows that the two assumptions of neighborhood stability are fulfilled under this situation . the terms @xmath167 and @xmath168 measure the sub - gradient of the lasso penalty and the neighborhood - fused lasso penalty respectively . having a small @xmath0 bound on them enforces the stability of the estimated coefficients , and hence stability of neighbors . see section [ proofs ] for proof of the proposition . [ perturb ] if there exists some @xmath169 such that @xmath170 . then , @xmath171 and @xmath172 . moreover , @xmath173 . we start the presentation of the theoretical results with a lemma ( proven in section [ proofs ] ) characterizing the minimizer of our objective function in terms of its subdifferential . [ baselemma ] given @xmath174 , let @xmath175 be a @xmath12 dimensional vector with elements @xmath176 . define @xmath177_{b } \qquad \text{and } \qquad \mathcal{l}^a ( \theta)= \{b:{\cal d}^a_b\ne 0\}.\]]a vector @xmath178 is a solution to the fused lasso problem described above iff @xmath179 this lemma builds the foundation of several of the following results and will be used frequently to prove them . sign consistency is one of the major properties that a model selection method should exhibit . before we study the model selection consistency of our estimators , we show , in the following lemma that the neighborhood - fused lasso estimator is sign - consistent . [ signconsist ] let @xmath62 be defined for all @xmath22 . under the assumptions [ a1]-[a7 ] , it holds for some @xmath180 that for all @xmath22 , @xmath181 observe that this lemma , in turn , preserves the asymptotic equality of signs of local neighbors . if @xmath11 and @xmath182 are local neighbors , then with high probability , when regressing @xmath5 on the remaining variables , the coefficients of @xmath36 and @xmath38 will have the same sign . this is a direct consequence of local constancy of the estimated regression coefficients . our results show that , just like in meinshausen and bhlmann , a rate slower than @xmath183 is necessary for the regularization parameters for consistent model selection in the high dimensional case where the dimension may increase as a polynomial in the sample size . specifically if @xmath71 decays at @xmath184 and @xmath72 decays as @xmath185 where @xmath186 are as in assumptions [ a1]-[a8 ] , the estimated neighborhood is almost surely contained in the true neighborhood . hence the type - i error probability goes to 0 . this is formally stated in the following theorem ( see section [ proofs ] for a proof ) . [ type1 ] let assumptions [ a1]-[a8 ] be fulfilled . let the penalty parameters satisfy @xmath187 and @xmath188 with some @xmath189 and @xmath190 and @xmath191 . there exists some @xmath192 such that for all @xmath28 , @xmath193 the assumptions of neighborhood stability and local neighborhood stability are not redundant . the following proposition shows that one can not relax the assumptions a7 and a8 . [ type1problem ] if there exists some @xmath139 with @xmath150 and @xmath194 and @xmath195 , then for @xmath196 as in theorem 4.6 , @xmath197 on the other hand , with the same sets of assumptions , one can show that the type ii probability , i.e. , probability of falsely identifying an edge as a potential connection exponentially goes to 0 . this has been formally stated and proved in the following theorem ( see section [ proofs ] for a proof ) . [ type2 ] with all the assumptions of theorem 4.6 and @xmath198 as before , @xmath199 it follows from the discussion so far that consistent estimation of nodes is possible using our approach . however , one of the original goals of our method is to estimate the entire underlying graphical model , which can be accomplished by combining the estimated neighborhoods in some way . two different methods of combination have been proposed @xmath200 in our simulations , we combined them following the union method . it has been observed that the differences vanish asymptotically when a regular lasso is applied . although we did not theoretically study this for our case but our experiments indicate that they do not vary much asymptotically . theoretic exploration in asymptotic domain does not cast much light on the choices of regularization parameter in real life , finite sample problem . it is hard to ensure consistency or absolute containment like theorem [ type1 ] or [ type2 ] . however , following the idea proposed by meinshausen and bhlmann , one can consider the _ connectivity component _ of a node , which is defined as the set of nodes which are connected to it through a chain of edges . generalizing the results from meinshausen & bhlmann , the following theorem shows ( see proof in section [ proofs ] ) that the estimated connectivity component derived from neighborhood - fused lasso estimate will belong to the true connectivity component with probability @xmath201 , for any chosen level of @xmath202 . [ regparam ] with all the assumptions [ a1]-[a8 ] , and the following choices of the penalty parameters , @xmath203 we have @xmath204 for all @xmath27 . @xmath205 and @xmath206 are the true and estimated connectivity components of @xmath22 , @xmath207 are certain constants and @xmath208 . the choice of @xmath209 and @xmath210 depends on the rate of growth of the local neighborhood with increasing sample size . in our simulations , we found that for models with constant dimension and increasing sample size , @xmath211 and @xmath212 works fine . + as shown in the simulations , we get a higher convergence speed as compared to meinshausen - bhlmann s method by applying a neighborhood - fused lasso penalty for nodewise regression when the underlying model exhibit local constancy . this can be proved using the following lemma ( proof in section [ proofs ] ) . [ fastconv ] the upper bound of type i error probability in neighborhood fused lasso is smaller than that of the meinshausen & bhlmann procedure . delves deep into the precise constants of in the proof of theorem [ type1 ] and shows that one could achieve smaller type 1 error probability with a finite sample size if one uses neighborhood fused lasso instead of usual lasso . ] lemma [ fastconv ] does not specify how much the maximal false positive probability is reduced . this can be understood after going through its proof . roughly speaking , the reduction in the upper bound is given by @xmath213 \cdot e^{-\frac{1}{\sigma^2_*}\left(\frac{d^2_1}{4}(1-\delta_1)^2 n^{\epsilon}\right ) } , \end{aligned}\ ] ] where @xmath214 , and @xmath215 , @xmath216 are as in the proof of theorem [ type1 ] . essentially , it is shown in the proof that the neighborhood fused lasso reduces a dominant portion of the type 1 error ( given by the second factor ) by a fraction ( given by the first factor ) shown above . looking at the proof of theorem [ type1 ] , it is seen that the term @xmath217 is the principal contributor to the probability of false positives . the corresponding term using usual lasso is @xmath218 lemma [ fastconv ] makes use of this fact and it also shows that substantial reduction takes place when the local neighborhood grows . so , this method invariably performs better than the meinshausen - bhlmann s method with respect to a minimax criterion ( in a sense that is minimizes the maximum probability of false positives ) , and the improvement is more when the local neighborhood grows faster . it also implies that it can perform as bad as usual lasso regression as the worst case scenario . in our attempt to gaussian graphical model learning , we adopt the meinshausen - bhlmann approach , i.e. , do a componentwise penalized regression . however , one should keep in mind that in the process of doing so , the design matrix ( which is random here ) changes at every iteration . in previous section we discussed the asymptotic model selection consistency of our estimator . in this section , we shall go beyond model selection and explore conditions under which our neighborhood - fused lasso estimate exhibit nice asymptotic @xmath0 properties . we shall carry out our theoretical analysis under the assumption that the linear regression model holds exactly , with some underlying true " @xmath219 . most of the theoretical results in this section have been derived in the light of discussions in @xcite . we start with a quick recapitulation of the notation we are going to use here . if @xmath220 , one assumes a node - wise regression model @xmath221 where @xmath10 denotes the @xmath22-th component , @xmath222 denotes all the components except @xmath22 and @xmath223 for some @xmath224 . also assume that @xmath225 and @xmath226 is the precision matrix . if @xmath227 denotes the edge set in the conditional independence graph then @xmath228 . the design matrix @xmath222 is obtained by deleting the @xmath22-th column of the data matrix . in the high dimensional set up , where one generally has lesser number of samples that model dimension @xmath229 , we assume an inherent sparsity in the true @xmath109 . this sparsity is also ensured if we assume that the underlying conditional independence graph is sparse . let us assume @xmath230 and @xmath231 . since @xmath232 is not known , one needs a regularization penalty . meinshausen and bhlmann chose the @xmath0 penalty , i.e. , the traditional lasso and got the estimate @xmath233.\end{aligned}\ ] ] in our situation , we have added another penalty term @xmath234 along with the lasso penalty term . hence , our estimator is given by @xmath235.\end{aligned}\ ] ] note that this new definition of our estimator is exactly same as what was defined in [ nfldefn ] . as mentioned in the earlier section , we shall develop our theory based on the assumption of a linear truth . we try to provide an upper bound on the prediction error . the following lemma forms the basis of our derivation . * the basic inequality*[basicineq ] @xmath236 it should be noted that the number of non zero elements in the column of the matrix @xmath58 denotes the number of local neighbors that particular node has ( except node @xmath22 ) . let us assume that the number of local neighbors is @xmath237 . fixing @xmath27 , let us also assume that the number of local neighbors is bounded by @xmath76 . then , it follows from lemma [ basicineq ] using @xmath238 that @xmath239 the basic objective of using a penalty parameter is to overrule the empirical process term @xmath240 . it can be easily seen that @xmath241 our goal is to choose a @xmath71 and a @xmath72 such that the probability that the empirical process term in the right hand side exceeds @xmath242 is small , so that with high probability the right hand side could be dominated by @xmath243 . to that effect , we define @xmath244 our objective is to show that for some particular choice of @xmath245 and @xmath246 , @xmath247 has high probability . now one should keep in mind that both @xmath248 and @xmath249 are random here . one can make the valid assumption of their individual gaussian law and independence . we formalize these notions in the following proposition . under the assumption of linear truth for the gaussian graphical model , i.e. , @xmath250 , we have @xmath251 the proof of this proposition is straightforward and hence skipped . the following results show that for suitable choice of @xmath245 and @xmath246 , @xmath252 is high . [ highprob ] under the assumption that @xmath253 , @xmath254 , \intertext{in particular , with $ \lambda_0 = \frac{2(t + \log p)}{n } > 0 $ and $ \mu_0 = \frac{2}{b}\sqrt{\frac{2}{n}(t + \log p)}$ , then } p(\lambda_a ) & \geq 1 - 2 e^{-t}.\end{aligned}\ ] ] [ highprobcor ] assume that @xmath253 and that @xmath255 . let the regularization parameters be @xmath256 then the following is true @xmath257 the following lemma provides an upper bound on the estimation error . [ l2ineq ] assume that @xmath253 and that @xmath255 . let the regularization parameters be @xmath258 also let @xmath259 be the smallest non - zero singular value of @xmath222 . then we have @xmath260 \geq 1 - e^{-t^2}.\end{aligned}\ ] ] we now try to derive some oracle inequalities which will provide @xmath0 bound for our neighborhood - fused lasso estimate . following bhlmann s notation , let us write , for an index set @xmath261 , @xmath262 we need some more notation . split the matrix @xmath58 as follows : @xmath263,\end{aligned}\ ] ] where @xmath264 consists of all rows such that both the non zero terms belong to @xmath265 . @xmath266 consists of all rows and columns such that exactly one of the non - zero term in that row belongs to @xmath265 . @xmath267 consists of all rows and columns such that exactly one of the non - zero term in that row belongs to @xmath268 . @xmath269 consists of all rows such that both the non - zero terms belong to @xmath265 . [ oracle1 ] on @xmath247 , if we choose @xmath270 and @xmath271 , we have @xmath272 it can be easily verified that if @xmath273 for some @xmath274 , we have @xmath275 this helps us to simplify the consequences lemma [ oracle1 ] , which implies that @xmath276 combining with the aforementioned condition , we get @xmath277 a standard way to relate the @xmath0 penalty @xmath278 to an @xmath6 penalty is to use the cauchy - schwarz inequality @xmath279 where @xmath280 is some constant . however , @xmath281 being random , this condition can not hold unanimously for all @xmath282 . bhlmann provided a _ compatibility condition _ so that this is true . in our situation , we found a similar condition like them to carry out further analysis . it should be noted that our compatibility condition is weaker than the compatibility condition bhlmann provided . however , we need to work under the assumption that @xmath283 . the definition of this _ compatibility condition _ is provided below . [ compat ] ( compatibility condition ) we say that a collection of nodes @xmath284 satsifies the @xmath285-compatibility condition if for all @xmath67 satisfying @xmath286 we have @xmath287 following bickel et al . @xcite , we shall refer to @xmath288 as the _ restricted eigenvalue_. here we provide a detailed explanation of this phenomenon . observe that the compatibility condition could be re - written as @xmath289 @xmath288 can be assumed to be the minimum eigenvalue over the restricted set @xmath290 . with this compatibility condition imposed , we derive the following oracle inequality [ oracle2 ] assume that the compatibility condition ( from definition [ compat ] ) holds for some @xmath291 . then on @xmath247 , for @xmath270 , @xmath292 and @xmath273 , we have @xmath293 combining theorem [ oracle2 ] and lemma [ highprob ] , we get , [ finalthm ] let @xmath294 and assume that @xmath295 . then with probability @xmath296 , @xmath293 theorem [ finalthm ] provides a unified way to deal with both the squared estimation error and the absolute error of the estimated parameters in neighborhood fused lasso regression . it can be seen that for increasing @xmath27 , with probability increasing to 1 , the absolute difference of our nfl estimator from underlying truth is bounded by @xmath297 for large values of @xmath27 , this is approximately equal to @xmath298 which is a positive and decreasing function of @xmath299 if @xmath300 . this is automatically satisfied for all @xmath291 . so , it can be easily seen , using the assumption @xmath301 , that the above quantity is bounded by @xmath302 in this simulation we repeat the exact scenario presented in honorio s first simulation setting . the gaussian graphical model consists of 9 variables as shown in figure [ nflvscdd ] . it deals with both local and non - local interactions . our method is compared to meinshausen - bhlmann s method and graphical lasso . we run our simulation for 4 different sample sizes @xmath303 & @xmath304 . for each sample size , we run the simulation 50 times and estimate the neighborhood for each iteration . we construct a weighted graph with edge weight corresponding to the frequency of its occurrence in all of those 50 iterations . + in this simulation study we take a 50 dimensional normal random vector with zero mean . the diagonals of the precision matrix are all @xmath55 and all the nonzero off - diagonal entries are @xmath305 . the conditional ( in)dependence graph of this vector consists of both spatial ( local ) and non - spatial neighbors where the local neighborhood structure is linear ( one dimensional lattice ) . there are two groups of distant neighbors . we generate @xmath27 i.i.d . samples from the corresponding normal distribution . we run the simulation for @xmath306 . we try to reconstruct the conditional ( in)dependence graph from the data using graphical lasso ( we use an oracle version of graphical lasso where the choice of penalty parameter is contingent on the actual number of edges in the true model ) , meinshausen - bhlmann s coordinate - wise lasso and our coordinate - wise generalized fused lasso approach . for each @xmath27 , we run the simulation 50 times and calculate the number of correctly identified edges and that of falsely identified edges for each iteration . the mean and standard deviations are shown in the table [ bigmodeltable1 ] and table [ bigmodeltable2 ] . figure [ bigmodelcomp ] shows the relative performance of the competing methods for different sample sizes . sample size increases from top to bottom . in the figure we include comparison for two additional sample values . they are @xmath307 and @xmath308 respectively . + .comparison of false positives and true positives [ cols="^,^,^,^,^,^,^,^",options="header " , ] it is quite evident from the two simulation that our method converges to the true model faster than meinshausen - bhlmann s method . it is also shown theoretically in lemma [ fastconv ] . in this paper , we have successfully extended the meinshausen - bhlmann approach of model selection in gaussian graphical models where the assumption of local constancy holds . we also provided a rigorous and generalized definition of local constancy and used it to propose neighborhood - fused lasso ( nfl ) , an algorithm to solve the problem by penalizing the differences of local neighbors given by a pre - determined graph @xmath33 . we have used a generalized version of fused lasso in order to accomplish our objective . our algorithm reduces the fused lasso problem into a regular lasso problem for a given set of regularizing parameters and thereby fast solutions are readily available . we were able to provide data dependent choices for the tuning parameters in order to establish asymptotic consistency in model selection . we substantiated our theoretical discussion on asymptotic model selection consistency with simulations that furnished desired results . we also proved , both theoretically and experimentally , that incorporating local constancy ensures faster convergence to the underlying truth , meaning that similar accuracy is achieved with smaller sample size . this phenomenon is somewhat similar to what buhl @xcite and uhler @xcite observed while investigating the existence of mle s in gaussian models with certain geometric structures . they were able to prove the existence of mle of a higher dimensional model with relatively smaller sample size . our findings are somewhat reminiscent of these results . we also discussed about the compatibility issue while doing regression with local constancy and were able to derive sufficient conditions under which @xmath0 boundedness of estimated parameters is ensured . on top of that we also provided an upper bound for the quadratic prediction error . + + there are many extensions possible for future research . firstly , instead of using the pre - defined local graph @xmath33 , one can think of alternative ways to grow or shrink the local neighborhood adaptively . secondly , as discussed in section [ locconstdefn ] , one can think of devising an algorithm that penalizes the differences of partial correlations instead of the elements of the precision matrix . thirdly , ( because of the compatibility issue ) regular lasso or fused lasso might fail to perform up to the mark if the model is ill - conditioned or the compatibility assumptions do not hold . one can think of using adaptive lasso or modify it accordingly in order to incorporate the local constancy property . some preliminary simulations in this direction look promising . thirdly , a bayesian approach could be pursued where one can start with some prior distribution on the inverse covariance matrix respecting the local constancy property and use the bayesian definition of local constancy in this paper ( see [ locconstdefn : bayes ] ) to come up with a novel alternative method . let @xmath309 . we will show that @xmath310 . we know from definition of @xmath311 that @xmath312 @xmath313 the proof can be found in [ suppa ] . the subdifferential of @xmath314 is given by @xmath315 where @xmath316\quad if \quad \theta_{b}=0\}$ ] and @xmath317\quad if \quad \alpha_{b}=0 \quad \mathrm{where } \quad \alpha = d'^{a}\mathrm{sgn}(d^{a}\theta)\}$ ] . observe that @xmath318 where @xmath319 denotes the @xmath320 column of the matrix @xmath321 . this is same as the total number of local neighbors of @xmath37 except @xmath22 . the lemma follows . this proof can be found in [ suppa ] . the event @xmath322 is equivalent to the event that there exists some node @xmath323 in the set of non - neighbors of node @xmath22 such that the estimated coefficient @xmath70 is not zero . thus , @xmath324 let @xmath325 be the event that @xmath326 conditional on @xmath325 , it follows from meinshausen - bhlmann s discussion that @xmath327 is also a solution to the fused lasso problem with @xmath328 . as @xmath329 for all @xmath330 , it follows that @xmath331 for all @xmath330 . by [ baselemma ] , @xmath332 it suffices to show that there exists a constant @xmath192 so that for all @xmath330 , @xmath333 writing @xmath334 for any @xmath323 where @xmath335 for some @xmath336 and @xmath337 is independent of @xmath338 . hence , @xmath339\end{aligned}\ ] ] by lemma [ signconsist ] , there exists a @xmath192 so that with probability @xmath340 , @xmath341 @xmath342 in this case , it holds by lemma [ baselemma ] that @xmath343_m\theta^{b,\mathrm{ne}_a}_m \mu\right| \\ & \leq \delta_1\lambda + \delta_2 b\mu\end{aligned}\ ] ] the absolute value of the coefficient @xmath344 is hence bounded by @xmath345 with probability @xmath346 . conditional on @xmath347 , the random variable @xmath348 is normally distributed with mean @xmath349 and variance @xmath350 . by definition of @xmath351 , @xmath352 since @xmath353 , @xmath354 is stochastically smaller than or equal to @xmath355 . it remains to show that for some @xmath192 and some @xmath356 , @xmath357 if @xmath10 and @xmath337 are independent , @xmath358 . using gaussianity and bounded variance of both @xmath10 and @xmath337 , we obtain existence of some @xmath359 such that @xmath360 . hence using bernstein inequality and boundedness of @xmath71 and @xmath72 , it holds for some @xmath192 that for all @xmath361 ( see lemma [ fastconv ] for detailed proof ) , @xmath362 which completes the proof . we follow the proof of theorem [ type1 ] and claim that with the above assumptions , for all @xmath363 with @xmath150 , @xmath364 for @xmath135 following similar arguments afterwards , it can be concluded that with some @xmath365 and @xmath366 as @xmath135 @xmath367 it holds for the third term that for any @xmath368 , @xmath369 as @xmath370 which combined with the previous result proves the proposition . observe that @xmath371 let @xmath325 be the event @xmath372 on @xmath325 , following similar arguments as before , we can conclude that @xmath373 . therefore @xmath374 it follows from the proof of theorem 4.6 that there exists some @xmath375 so that @xmath376 . using bonferronis inequality , it hence remains to show that there exists some @xmath375 so that for all @xmath377 , @xmath378 which follows from lemma [ signconsist ] . @xmath379 implies the existence of an edge in the estimated neighborhood that connects two nodes in two different connectivity components of the true underlying graph . hence , @xmath380 going by the same arguments used in proving theorem 4.6 , we have @xmath381 thus , it is sufficient to show that @xmath382 now observe that since @xmath11 and @xmath383 are in different connectivity component , they are , in fact , independent . therefore , conditional on @xmath384 , @xmath385 , making it stochastically smaller than @xmath386 . hence it holds for all @xmath28 and @xmath387 that @xmath388 where @xmath208 . using the @xmath71 and @xmath72 proposed , the rhs becomes @xmath389 . this is a direct application of bernstein inequality . since @xmath390 and @xmath391 as @xmath135 , for large enough @xmath27 , we have , by a version of bernstein s inequality , that @xmath392\\ & \hspace*{3cm}= \exp\left[-\frac{1}{\sigma^2 _ * } \left(\frac{d_1}{2}(1 - \delta_1)+ \frac{d_2}{2}(1 - \delta_2 ) n^{\beta_0}\right)^2 n^\epsilon\right].\\ \intertext{this proves part ( a ) . write the last expression as } & e^{-\frac{1}{\sigma^2_*}\left(\frac{d^2_1}{4}(1-\delta_1)^2 n^{\epsilon}\right ) } \cdot e^{-\frac{1}{\sigma^2_*}\left(\frac{d_1 d_2}{2}(1 - \delta_1)(1 - \delta_2)n^{\beta_0 } + \frac{d^2_2}{4}(1 - \delta_2)^2 n^{2\beta_0}\right)n^{\epsilon } } \intertext{proves part ( b ) , since the first factor is the upper bound for lasso . } \end{aligned}\ ] ] since @xmath281 is the fused lasso minimizer , we get @xmath393 the proof can be found in [ suppa ] . the proof exploits the fact that @xmath281 is the minimizer of the penalized least square by equalling the sub - differential of the objective function to 0 . we start by replacing the assumed linear truth , i.e. , @xmath394 . therefore , we get by using the notation @xmath395 and @xmath396 that @xmath397_{\theta_a = \hat{\theta}^{\lambda,\mu}_a } = 0\\ & \rightarrow\quad \frac{2}{n}\left((x^a)^2\ , ( \hat{\theta}^{\lambda,\mu}_a - \theta_0 ) \right ) - \frac{2}{n}x^{a'}\epsilon_a + \lambda\mathrm{sgn}\left(\hat{\theta}^{\lambda,\mu}_a\right ) + \mu { \mathbb d}^a = 0\\ & \rightarrow\quad ( x^a)^2 \hat{\theta}^{\lambda,\mu}_a = ( x^a)^2\theta_0 + \frac{n}{2}\left(\frac{2}{n}x^{a'}\epsilon_a\right ) - \frac{n\lambda}{2}\mathrm{sgn}\left(\hat{\theta}^{\lambda,\mu}_a\right ) - \frac{n\mu}{2 } { \mathbb d}^a \\ & \rightarrow\quad \hat{\theta}^{\lambda,\mu}_a = \left((x^a)^2\right)^+ ( x^a)^2 \theta^0_a + \frac{n}{2}\left((x^a)^2\right)^+ \left(\frac{2}{n}x^{a'}\epsilon_a - \lambda\mathrm{sgn}\left(\hat{\theta}^{\lambda,\mu}_a\right ) - \mu { \mathbb d}^a\right).\end{aligned}\ ] ] with the given choices for @xmath71 and @xmath72 , define @xmath398 , then we have , with probability @xmath399 that @xmath400 where the inequality is meant to be interpreted componentwise and @xmath401 is a vector of size @xmath1 consisting of 1 s . hence , we get @xmath402 using the facts that @xmath403 , where @xmath404 is the maximum singular value of @xmath73 and that @xmath405 is idempotent , and hence its singular values are either 0 or 1 , we can continue the above sequence of inequalities and obtain @xmath406 plugging in the result obtained from previous corollary , we get @xmath407 which implies that @xmath408 to start with , we derive a series of inequalities and apply them on the basic inequality . @xmath409 we write @xmath410 and @xmath411 so that @xmath412 . similarly we write @xmath413 . hence , we get @xmath414 \left ( \begin{matrix } \hat{\theta}^{\lambda,\mu}_{a , s_0,1}\\ \hat{\theta}^{\lambda,\mu}_{a , s^c_0,1 } \end{matrix } \right)\right|\right|_1 = \left|\left| \left ( \begin{matrix } d^a_{s_0,s_0 } \hat{\theta}^{\lambda,\mu}_{a , s_0,1}\\ d^a_{s_0,0 } \hat{\theta}^{\lambda,\mu}_{a , s_0,1 } + d^a_{0,s^c_0 } \hat{\theta}^{\lambda,\mu}_{a , s^c_0,1 } \\ d^a_{s^c_0,s^c_0 } \hat{\theta}^{\lambda,\mu}_{a , s^c_0,1 } \end{matrix } \right ) \right|\right|_1\end{aligned}\ ] ] @xmath415 similarly , we get @xmath416 \left ( \begin{matrix } \theta^0_{a , s_0,1}\\ 0 \end{matrix } \right)\right|\right|_1 \quad = \quad \left|\left| \left ( \begin{matrix } d^a_{s_0,s_0 } \theta^0_{a , s_0,1}\\ d^a_{s_0,0 } \theta^0_{a , s_0,1}\\ 0 \end{matrix } \right)\right|\right|_1\\ & = \quad \big\| d^a_{s_0,s_0 } \theta^0_{a , s_0,1}\big\|_1 + \big\| d^a_{s_0,0 } \theta^0_{a , s_0,1 } \big\|_1\end{aligned}\ ] ] plugging into the basic inequality , we get on @xmath247 , with @xmath417 and @xmath292 , @xmath418 from this , we get @xmath419 the proof is basically a continuation of what we have already shown . we have @xmath420 the authors are thankful to dr . debashis paul and dr . ethan anderes who helped shape some ideas presented in this paper with insightful comments and discussions . banerjee , o. , ghaoui , l. e. , daspremont , a. ( 2008 ) . model selection through sparse maximum likelihood estimation for multivariate gaussian or binary data " , _ journal of machine learning research _ , vol . 485 - 516 . chen , x. , lin , q. , kim , s. , carbonell , j.g . , and xing , e.p . smoothing proximal gradient method for general structured sparse regression " . _ annals of applied statistics _ , vol . 719 - 752 . tibshirani , r. , saunders , m. , rosset , s. , zhu , ji and knight , k. ( 2005 ) . sparsity and smoothness via the fused lasso " , _ journal of royal statistical society _ , series b ( methodological ) , vol . 67 , part 1 , pp .

iron - based high-@xmath0 superconductors@xcite have a complex fermi surface that is determined by several bands , an effect resulting from the hybridization of the @xmath1 orbitals of iron.@xcite band structure calculations have shown that the bands that define the two hole pockets around the @xmath2 point have mostly @xmath3 and @xmath4 character , while the two electron pockets around the m point have @xmath5 , @xmath4 , and a smaller amount of @xmath6 contributions.@xcite for this reason it is important to understand superconductivity in multiorbital systems in general terms . among the first to address this complex problem several years ago were suhl _ et al._@xcite using a model consisting of two orbitals , one @xmath7 and one @xmath8 , that did not hybridize with each other . thus , in this case each band was determined by one single orbital . they showed that bcs pairing@xcite could occur in each band and , since in the most general case the electron - phonon interaction would have different strengths for electrons in the different bands , it was proposed that two different superconducting gaps could arise . almost 50 years were needed to observe experimental evidence of this phenomenon . in 2001 , superconductivity with @xmath9 k was observed in mgb@xmath10.@xcite despite the high-@xmath11 , it became clear that the bcs mechanism@xcite was at play and for the first time two different superconducting gaps were observed.@xcite as shown in ref . , the fermi surface ( fs ) is determined by two bands : the @xmath12 band formed by the @xmath13 orbitals of b , and the @xmath14 band constituted by a linear combination of the @xmath15 and @xmath16 b - orbitals . although three orbitals determine the fs , it is interesting to notice that only two different bcs gaps are observed . this occurs because two of the three orbitals hybridize with each other and determine one single band , which couples strongly to the lattice phonons . this opens a large superconducting gap on the @xmath14 fs . the other orbital , @xmath13 , does not hybridize and forms the @xmath12 band that couples weakly to the lattice phonons determining a second , smaller , superconducting gap at the fs of the @xmath12 band . thus , the number of different gaps that can arise in a multiorbital system is related to the degree of hybridization among the orbitals . also note that in this early effort _ interband hopping of pairs of electrons belonging to the same band was included _ but the possibility of interband pairing , i.e. , _ pairs formed by electrons belonging to two different bands _ , was not considered . in this paper , the subject of superconductivity in multiorbital systems is revisited , in particular to shed light on the possible symmetry of the pairing operator of the pnictides superconductors . the motivation is that the pairing operators that have been discussed the most thus far@xcite assume that only intraband pairing should occur , namely the two electrons of the cooper pair belong to the same band.@xcite however , numerical simulations@xcite performed on a two - orbital model@xcite for the pnictides favor an interorbital pairing operator that , when transformed to the band representation , results not only in intraband pairing but it includes interband pairing as well.@xcite for the pnictides , interband hopping of pairs formed between electrons in _ the same _ band is often denoted as `` interband superconductivity''.@xcite the situation discussed in the present paper is _ different _ and involves cooper pairs where the two electrons come from two different bands , which we will call `` interband pairing '' . using symmetry arguments and mean - field approximations , the plausibility and physical meaning of such an interband pairing in multiorbital systems will be discussed . interband pairing has previously been addressed in the context of quantum chromodynamics ( qcd ) and cold atoms,@xcite heavy fermions,@xcite cuprates@xcite , and bcs superconductivity.@xcite in the case of heavy fermions , it was argued that interband pairing could occur if two fermi surfaces arising from different bands are very close to one other,@xcite while in qcd and cold atoms it was presented as a possibility for the case of sufficiently strong attractive pairing interactions , or for weaker attractions among particles with very different masses.@xcite as it will be discussed for a simple model in sec . [ sec : interband ] , three different regimes , shown schematically in fig . [ regimes ] , can result from a purely interband pairing as a function of the strength of the pairing potential @xmath17 : ( 1 ) a normal regime where the ground state is not superconducting ( namely in purely interband pairing an infinitesimal attraction does not lead to superconductivity ) , ( 2 ) an exotic superconducting `` breached '' regime where gaps open at the normal fermi surfaces while new fermi surfaces defining regions containing unpaired electrons are created , and ( 3 ) a superconducting regime resembling bcs states , at large attractive coupling.@xcite of an interband pairing attraction . the label `` normal '' denotes a non - superconducting state . the case `` breached '' is an exotic regime with superconductivity and gaps , coexisting with fermi surfaces ( or several nodes ) and electrons that do not pair . `` bcs '' is the large attraction region , where the ground state resembles that of a bcs superconductor and all electrons participate in the pairing.,title="fig:",width=302 ] 0.3 cm the paper is organized as follows . in sec . [ sec : pairings ] , the general form of pairing operators in multiorbital systems will be presented , remarking how the symmetry is determined by the spatial and the orbital characteristics of the operator . the interorbital pairing operator with @xmath18 symmetry obtained numerically in a two - orbital model for the pnictides is discussed , emphasizing that this operator presents a mixture of intra and interband pairing in the band representation . in sec . [ sec : interband ] , a simple toy model with pure interband pairing attraction is introduced . this simplified model is discussed in order to illustrate the effects of interband pairing on observables , such as the occupation number and the spectral functions @xmath19 . the stability of the interband paired state is also discussed . the occupation number , spectral functions , and the stability of the @xmath18 pairing state are the subject of sec . [ sec : interorbital ] which is directly related to the physics of pnictides , while sec . [ sec : conclusions ] is devoted to the conclusions . in single - orbital models , the symmetry of a spin - singlet pairing state is completely determined by the properties of its spatial form factor . more specifically , the pairing operator will have the form : @xmath20 where @xmath21 destroys an electron with momentum @xmath22 and spin projection @xmath14 and @xmath23 is the form factor that transforms according to one of the irreducible representations of the crystal s symmetry group . thus , @xmath23 determines the symmetry of the operator . these form factors depend on the lattice geometry and generally they may be very complex . however , in materials with short pair coherence lengths , such as the high-@xmath0 cuprates , the assumption that the two particles that form the pair can be very close to one other is usually made . the cu - oxide planes in the cuprates have the symmetry properties of the @xmath24 group and the case @xmath25 , which transforms according to the irreducible representation @xmath26 , provides the well - known @xmath8-wave symmetry pairing . in multiorbital systems , on the other hand , a spin singlet pairing operator will have both spatial and orbital degrees of freedom and it will be given by @xmath27 where @xmath28 destroys an electron with momentum @xmath22 , in orbital @xmath29 , and with spin projection @xmath14 , @xmath23 is the spatial form factor as indicated above , and @xmath30 is a matrix in the space spanned by the orbitals involved . the dimension of @xmath31 is equal to the number of orbitals that are considered to be of relevance . in this case , notice that the symmetry of the pairing operator would in general be determined by the product of the symmetry properties of @xmath32 and the symmetry of the orbital contribution @xmath30 . only if @xmath30 is the identity matrix do the orbital contribution become trivial , because the identity matrix transforms according to @xmath33 . thus , this is the only case where @xmath23 fully determines the symmetry of the pairing operator , as in the single - orbital example . the minimum model for the pnictides considers the two orbitals @xmath5 and @xmath4 , which are strongly hybridized.@xcite all the possible pairing operators , up to nearest - neighbor distance , that are allowed by the lattice and orbital symmetries have been already calculated.@xcite numerical simulations performed on the two - orbital model suggest that the favored pairing operator at intermediate couplings , where the state is both magnetic and metallic , @xcite has symmetry @xmath18 and is given by eq . ( [ 2 ] ) with @xmath34 , which transforms according to @xmath33 , and @xmath35 which transforms according to @xmath18 @xcite ( where @xmath36 are pauli matrices ) . thus , the non - trivial symmetry under rotations arises from the orbital portion of the operator . this pairing operator has been studied at the mean - field level in ref . . in the orbital representation , the bogoliubov - de gennes hamiltonian matrix is given by : @xmath37 with @xmath38 and @xmath39 where @xmath40 , with @xmath41 being the strength of the pairing interaction and @xmath42 the mean - field parameter obtained by minimizing the energy . since the two orbitals are hybridized via @xmath43 , in the band representation the hamiltonian matrix becomes : @xmath44 where @xmath45 and @xmath46 are given by @xmath47 @xmath48 and @xmath49 and @xmath50 are the elements of the change of basis matrix @xmath51 given by @xmath52 with @xmath53 . remember that @xmath45 and @xmath46 , are functions of the momentum @xmath22 , and @xmath54 . thus , it is clear that in the band representation , in addition to the intraband pairing given by @xmath45 , there is also interband pairing given by @xmath46 . among pairing operators compatible with the symmetry of the model , the ones that do not lead to interband pairing not only can not mix orbitals , but have to contain @xmath55 , i.e. , the identity matrix.@xcite one example for such an operator would be the @xmath56 pairing.@xcite the discussion above describes general properties of hybridized multi - orbital systems . if the orbitals are hybridized , but not related to one another by symmetry , there is no reason to expect that the coupling between the electrons in each orbital , and the interaction that produces the pairing , will have the same strength for all the orbitals and lead to a unit matrix in the orbital sector . then , it is expected that interband pairing will arise in general and , thus , it is important to understand its consequences , providing the main motivation for the present manuscript . to address qualitatively the issue of interband pairing , postponing the matters of stability to the next subsection ( sec . [ sec : stable ] ) , let us consider the following two - bands simplified model with interband pairing : @xmath57 where @xmath58 label two bands that are not hybridized , @xmath14 is the spin projection , and for simplicity @xmath59 which gives parabolic bands that are degenerate at @xmath60 with energy @xmath61 , and with a chemical potential @xmath62 . this can be considered as a crude representation of the two hole - pocket bands around the @xmath2 point in the pnictides , but more importantly presents a simple toy model where the effects of interband pairing can be studied . as before , the parameter @xmath63 is the product of an attractive potential @xmath41 between electrons in the two different bands and a mean - field parameter @xmath42 determined by minimizing the total energy . the band dispersion without the interaction is presented in the inset of fig . [ bands ] . ) , for the indicated values of @xmath64 ( defined in the text ) as a function of the momentum @xmath65 . the case shown is for @xmath66 , @xmath67 , and @xmath68 . inset : non - interacting band dispersion for the same parameters . , title="fig:",width=302 ] 0.3 cm the bogoliubov - de gennes matrix expressed in the basis expanded by @xmath69 has the form : @xmath70 this matrix can be diagonalized becoming @xmath71 in the basis expanded by @xmath72 and the change - of - basis matrix is given by @xmath73 with @xmath74 . the four energy eigenvalues are @xmath75 , \label{13}\ ] ] where the first positive ( negative ) sign corresponds to the eigenvalues of the upper ( lower ) block , labeled @xmath76 and @xmath77 ( @xmath78 and @xmath79 ) in eq . ( [ 12a ] ) and in fig . [ bands ] . the second sign differentiates between the two solutions in each block . figure [ bands ] shows the eigenvalues for @xmath80 and @xmath81 . when @xmath80 , then @xmath82 and @xmath83 , the two bands define two circular fermi surfaces with radius @xmath84 and @xmath85 where they cross the chemical potential ( @xmath62 ) . this is illustrated schematically in fig . [ diag ] . . @xmath84 and @xmath85 indicate the fermi momentum of the two bands when @xmath80 , while @xmath86 and @xmath87 indicate the position of the fermi momenta for the case of a finite but small pairing potential . the shaded rings indicate the regions with width @xmath88 and @xmath89 in momentum space where electrons can pair . the white region in between the rings contains unpaired electrons in band 1.,title="fig:",width=302 ] 0.3 cm notice that number operators can be defined in the two bases that are being considered here , i.e. @xmath90 and @xmath91 , which of course become equivalent when @xmath64=0 . thus , in the basis @xmath90 the number operator is , @xmath92 while in basis @xmath91 the number operator is ( @xmath93=a , b ) @xmath94 the total electronic occupation of the system is given by @xmath95 then , for @xmath80 we find that @xmath96 for @xmath97 , @xmath98 for @xmath99 since in this region @xmath100 while @xmath101 , and finally @xmath102 for @xmath103 since both orbitals are totally filled with electrons . these results are represented by the dashed lines in fig . [ nkab12 ] . for ( a ) the whole system ; ( b ) band b , ( c ) band a ; ( d ) orbital 1 ; and ( e ) orbital 2 for the indicated values of the pairing potential @xmath64 , and @xmath66 , @xmath67 , and @xmath68.,title="fig:",width=302 ] 0.3 cm in the nontrivial case of @xmath64 different from zero , the bands @xmath104 and @xmath105 result from the hybridization of the @xmath106 and @xmath107 bands due to @xmath64 . it is interesting to observe that an internal gap opens at the crossing of bands @xmath76 with @xmath77 @xmath108 the chemical potential and between @xmath78 and @xmath79 @xmath109 the chemical potential , as indicated with circles in fig . [ bands ] where results for @xmath81 are displayed . these are very important differences with respect to conventional bcs calculations were all the action is restricted to the original fermi surfaces : for multiorbital models , gaps can open in other portions of the band structure as well . while the @xmath104 bands are separated by a gap , the bands @xmath105 still cross the chemical potential determining two fermi surfaces at @xmath110 and at @xmath111 , even in this pairing state ( again , this is different from the one - orbital pairing standard bcs ideas ) . the new fermi surfaces are shown schematically in fig . [ diag ] : the interior fs has expanded and the exterior one has contracted . in fact , it will be shown below that the effect of @xmath64 is to try to equalize the two fermi surfaces , as this pairing attraction grows in magnitude . calculating @xmath112 we observe that @xmath113 for @xmath114 and @xmath115 , where @xmath116 is the element of the change - of - basis matrix in eq . ( [ 12b ] ) . this agrees with the bcs expression for @xmath112 but it jumps discontinuously to @xmath98 for @xmath117 . such a result is shown by the solid lines in fig . [ nkab12](a ) . the jumps indicate the existence of the two fermi surfaces , which are here present even in the paired state . thus , some electrons in the region in between the two fermi surfaces may behave like normal unpaired electrons . a better understanding of the electronic behavior can be achieved by studying the electronic density in the two bases @xmath91 and @xmath90 . we find that @xmath118 , as shown in fig . [ nkab12](c ) , which is the standard bcs behavior in agreement with the fact that bands @xmath104 are separated by a gap . thus , all the electrons in this band participate in the pairing and they do not have a fs . on the other hand , it can be shown that @xmath119 for @xmath114 and @xmath115 , but for @xmath117 there is a discontinuous change of behavior to @xmath120 [ fig . [ nkab12](b ) ] . thus , the fermi surfaces are determined by electrons in the band @xmath78 . however , note that in this region @xmath121 is an increasing function of @xmath122 while @xmath123 is a decreasing function of @xmath122 which satisfies @xmath124 for all @xmath122 . this behavior can be better understood by calculating the ( photoemission ) spectral functions @xmath19 , which allow us to obtain @xmath125 . for the non - interacting case , @xmath80 . ( b ) same as ( a ) but for @xmath81 . note that there are still fermi surfaces in this intermediate coupling case , namely gaps open far from the chemical potential but not at the fermi level . ( c ) same as ( a ) but for @xmath126 . the presence of a gap for all the momenta shown is now clear.,title="fig:",scaledwidth=40.0% ] for the non - interacting case , @xmath80 . ( b ) same as ( a ) but for @xmath81 . note that there are still fermi surfaces in this intermediate coupling case , namely gaps open far from the chemical potential but not at the fermi level . ( c ) same as ( a ) but for @xmath126 . the presence of a gap for all the momenta shown is now clear.,title="fig:",scaledwidth=40.0% ] for the non - interacting case , @xmath80 . ( b ) same as ( a ) but for @xmath81 . note that there are still fermi surfaces in this intermediate coupling case , namely gaps open far from the chemical potential but not at the fermi level . ( c ) same as ( a ) but for @xmath126 . the presence of a gap for all the momenta shown is now clear.,title="fig:",scaledwidth=40.0% ] 0.3 cm in the non - interacting case , shown in fig . [ akw0](a ) , the spectral function shows two peaks , corresponding to the two bands @xmath106 and @xmath107 , for each value of @xmath122 . this is the expectation for free electrons in a non - interacting multiorbital system . the two non - interacting fs s are located where each band passes across the chemical potential . when the pairing interaction becomes finite , the spectral function develops four peaks for each value of @xmath122 as it can be observed in fig . [ akw0](b ) for @xmath81 . this is also the expected result since the bcs interaction generates a `` shadow '' or bogoliubov band for each band present in the non - interacting system . for example , close to @xmath127 the bands @xmath76 and @xmath78 have almost all the spectral weight , given by @xmath128 which is very close to 1 in this region , and follow a dispersion similar to the non - interacting bands @xmath106 and @xmath107 , while the bands @xmath79 and @xmath77 appear with very small spectral weight given by @xmath129 . the latter are the bogoliubov or `` shadow '' bands . these shadow bands appear above the chemical potential for large values of @xmath122 , as expected . but what happens in the intermediate region @xmath117 ? we see that the band @xmath130 crosses with @xmath131 at @xmath86 determining a fs . most of the spectral weight below the chemical potential belongs to the band @xmath131 which contains the unpaired electrons but the band @xmath132 has appreciable shadow spectral weight indicating that there are also some paired electronic population as indicated in figs . [ nkab12](b ) and ( c ) . as @xmath133 increases , spectral weight is transferred continuously from @xmath131 to @xmath132 , behavior associated with the internal gap opened by the pairing interaction , so that when @xmath133 approaches @xmath87 most of the spectral weight below the chemical potential is in @xmath132 , paired electrons , and @xmath131 , unpaired electrons , has shadow spectral weight as seen in figs . [ nkab12](b ) and ( c ) . at @xmath134 the second fs is determined by the crossing of @xmath135 at the chemical potential indicated by the sudden jump in @xmath136 . thus , the unpaired electrons in this region coexist with paired electrons and the spectral functions do not resemble the non - interacting ones . it is also illuminating to analyze what happens with @xmath112 in the basis @xmath90 . while @xmath137 in the regions where there are no unpaired electrons , we find discontinuities associated with fermi surfaces in the two distributions that are given by @xmath138 and @xmath139 for @xmath117 [ figs . [ nkab12](d , e ) ] . this indicates that the pairing interaction @xmath64 has been able to promote some electrons from above to below @xmath85 in orbital 2 . also , electrons have been transferred from their original location in orbital 1 , in the neighborhood of @xmath85 and above @xmath84 , to both orbitals 1 and 2 around @xmath84 . these are the electrons in 1 and 2 that have become paired ( the pairing is indicated by the shadowed circular regions in fig . [ diag ] ) . but the interaction was not strong enough to provide pairing partners to all the extra electrons originally in orbital 1 and , thus , they have been left unpaired in between the two paired regions , as indicated in fig . [ diag ] . thus , the interband pairing attraction creates pairs of electrons belonging to different orbitals within an interval @xmath140 around each of the two original fermi surfaces . the width of the pairing region increases with @xmath64 . the pairing partners are obtained by promoting electrons with momentum @xmath141 and @xmath142 in both orbitals and by moving electrons from the more populated to the less populated orbital . this creates the conditions to pair electrons near both fermi surfaces . the electrons in band 1 that could not find promoted partners remain unpaired . whether this state is stable or not depends , of course , on the balance between kinetic and pairing energies , which will be discussed in sec . [ sec : stable ] . as the interaction @xmath64 increases further , the number of unpaired electrons is reduced . this means that @xmath86 and @xmath87 become closer to each other making the size of the intermediate region with unpaired electrons in fig . [ diag ] smaller . eventually , the two momenta become the same @xmath143 for @xmath144 , and for @xmath145 the region with unpaired electrons vanishes and a full gap opens in the system whose physics now resembles bcs , except for the fact that the pairs are constituted by electrons from different orbitals . the electronic population of the system in such a case , e.g. at @xmath126 , is presented in fig . [ nk1 ] and the corresponding spectral functions are shown in fig . [ akw0](c ) . it can be shown that now @xmath146 for all values of @xmath133 and all the particles around the two noninteracting fermi surfaces now participate in the pairing . the spectral functions show four peaks , i.e. bogoliubov bands for all values of @xmath133 , as it can be observed in fig . [ akw0](c ) . , and for @xmath66 , @xmath67 , and @xmath68 . the case @xmath126 illustrates the `` strong '' pairing attraction regime where gaps open in the original fermi surface.,title="fig:",width=302 ] 0.3 cm thus , notice that while band 1 contained more electrons than band 2 for @xmath80 [ see figs . [ nk1](b ) and ( c ) ] , the interband pairing mechanism transfers electrons from one band to the other so that both bands have the same number of electrons in the superconducting state . consequently , the smaller fs expands and the larger one shrinks . then , when the pairing becomes strong enough , the two fermi surfaces become equalized and no unpaired electrons remain . whether this situation can be achieved will depend on the strength of the interaction and the energy balance , as discussed in the next subsection . if the two non - interacting fermi surfaces are very close to each other in momentum space , even relatively weak pairing interaction could effectively be strong enough to make the interband pairing resemble bcs pairing as in the case of large @xmath64 in the present example . as discussed in the introduction , the possibility of interband pairing has been previously discussed in the context of qcd and cold atomic matter . similar effects on the fs as found in the present study were described , although the physics was different because in the qcd context each band contained different kinds of particles and the pairing was thus not able to promote particles from the majority to the minority band.@xcite the issue of stability was explored in the qcd framework , and it was found that a purely interband paired state could be stabilized for pairing attractions @xmath108 a certain cut - off value , which could become very small for a large difference between the masses of the two paired species.@xcite in the case of our model , however , we have found ( see below ) that the purely interband - paired state only becomes stable when the attraction is sufficiently strong that no unpaired particles are left , i.e. when the two shaded regions overlap and the unpaired region in fig . [ diag ] vanishes . this means that , although the pairs would be formed by electrons in different orbitals , the physics would be analogous to bcs . a gap will be opened in the full fermi surface of the simple model studied here . in order to study the issue of stability , let us assume that the interaction term responsible for the interband attraction is given by @xmath147 where @xmath148 and @xmath149 is the number of sites . performing the standard mean - field approximation : @xmath150 and @xmath151 and making the substitution @xmath152 ( and an analogous substitution for the product of annihilation operators ) , the mean - field results are obtained . as usual , the fluctuations around the average given by @xmath153 are assumed to be small . defining @xmath154 we obtain the following mean - field hamiltonian : @xmath155 equation ( [ eq:12 ] ) can be recovered by defining @xmath156 , and disregarding the constant last term of eq . ( [ mf ] ) . we can calculate the total energy @xmath157 for eq . ( [ mf ] ) as a function of @xmath40 . if for a given @xmath158 the energy has a minimum , this indicates that the interband - paired state is stable . note that having a term linear in @xmath42 in eq.([mf ] ) is not sufficient to conclude the appearance of a superconducting state at small @xmath42 , since the sign of the coefficient of the linear term can change sign with @xmath64 . a similar situation occurs for the magnetic state of undoped pnictides : a finite hubbard @xmath51 must be reached to stabilize the `` striped '' state @xcite . returning to superconductivity , there are two regions of interest : ( i ) @xmath159 , which corresponds to the case in which two fermi surfaces are present in the paired state ; ( ii ) @xmath145 , which corresponds to the fully gapped case . in fig . [ fig : stab](a ) , @xmath160 vs. @xmath64 is shown for different values of @xmath41 . it can be observed that a second minimum develops for @xmath161 and it becomes stable for @xmath162 . the minimum always occurs for @xmath145 which means that it corresponds to the case in which there are no unpaired electrons in the system . the results shown in the figure are robust in the sense that changes in the values of @xmath163 , @xmath164 or the chemical potential were not found to stabilize the state with unpaired electrons . thus , in this respect an attraction that is only interband can only lead to a stable superconducting state in the strong attraction region . ) per site vs. @xmath40 for different values of @xmath41 . the case @xmath144 is indicated with dashed lines : it separates the regions corresponding to nodal and nodeless states . ( b ) mean - field energy ( eq . [ mf ] ) per site with intraband pairing with strength @xmath165 vs. @xmath40 , for different values of @xmath41.[fig : stab],title="fig:",scaledwidth=40.0% ] ) per site vs. @xmath40 for different values of @xmath41 . the case @xmath144 is indicated with dashed lines : it separates the regions corresponding to nodal and nodeless states . ( b ) mean - field energy ( eq . [ mf ] ) per site with intraband pairing with strength @xmath165 vs. @xmath40 , for different values of @xmath41.[fig : stab],title="fig:",scaledwidth=40.0% ] the results of the previous paragraphs may seem negative with respect to the relevance of the `` intermediate '' state with simultaneous coexistence of pairing and fermi surfaces . however , as pointed out in sec . [ sec : pairings ] , most of the pairing operators allowed by the lattice and orbital symmetry in the pnictides are characterized by a _ mixture _ of both intra and interband pairing . thus , it is important to consider such a situation in our simple model as well . in the case of the @xmath18 pairing operator , eqs . ( [ 6 ] ) , ( [ 7 ] ) , and ( [ 8 ] ) indicate that the pairing is purely interband only for @xmath166 or @xmath12 and @xmath167 or @xmath12 , because @xmath168 along these lines . thus , let us now consider our simple model in a brillouin zone ( bz ) defined by @xmath169 and with the addition of _ intraband _ pairing with intensity @xmath165 that is vanishing for @xmath166 or @xmath12 and @xmath167 or @xmath12 , i.e. , the pairing is purely interband only along those directions . the energy bands now behave as in fig . [ bands ] only along @xmath170 and @xmath171 , while the dispersion along any other direction is shown in fig . [ bandasmix ] for @xmath80 and @xmath81 . ) ( along the main diagonal ) for the indicated values of the pairing potential @xmath64 , with the addition of an intraband pairing with strength @xmath165 , as described in the text.,title="fig:",width=302 ] 0.3 cm performing the mean - field approximation similarly as explained above , we have found that the superconducting state now becomes stable on _ both _ sides of the original critical value @xmath144 . figure [ fig : stab](b ) shows that the pairing state is stabilized for @xmath172 but the value of @xmath42 where the minimum is located is such that @xmath173 and , thus , two nodes will be present along the @xmath174 and @xmath175 axes . increasing the value of @xmath41 , the minimum eventually occurs for @xmath145 . for these larger values of @xmath41 , there would consequently be no nodes . then , in this section it has been shown using a simple model that the interorbital paired state can become stable if the attraction @xmath41 is sufficiently strong . in this case , it is the nodeless case that is stable even for purely interband attraction . in addition , the very interesting novel phase with coexisting nodes and unpaired electrons in the majority band also requires intraband pairing to be stable , with strength similar to that of the interband , at least in parts of the bz . in the previous section , a simple model was presented , both with exclusively interband pairing and with both inter and intraband pairing , and it was found that the intraband pairing stabilizes the state with a mixture of superconductivity and metallicity . as mentioned in the introduction , it is expected that in the most general cases the pairing operators allowed by the symmetry of the lattice and of the orbitals will , in the band representation , have both intra and interorbital pairing . thus , now we will present and discuss , at the mean field level , the occupation number and the spectral functions for the pairing operator obtained from the numerical study of the two - orbital model for the superconducting state of the pnictides introduced in sec . [ sec : pairings ] . for the two - orbital model eq . ( [ eq : bcs_ours ] ) with pairing interaction @xmath81 ( continuous lines ) and @xmath80 ( dashed lines ) . parameters : @xmath176 , @xmath177 , @xmath178 , @xmath179 . ( b ) same as ( a ) but for the orbital @xmath5 . ( c ) same as ( a ) but for the orbital @xmath4 . ( d ) same as ( a ) but for band 1 . ( e ) same as ( a ) but for band 2.[fig : nkxy05],scaledwidth=40.0% ] for the two - orbital model eq . ( [ eq : bcs_ours ] ) without pairing interaction for the orbital @xmath5 ( orange / light ) and the orbital @xmath4 ( red / dark ) . the ` floor ' indicates the fs in red . ( b ) same as ( a ) but for band 1 ( orange / light ) and band 2 ( red / dark)[nkxysurv0],title="fig:",scaledwidth=40.0% ] for the two - orbital model eq . ( [ eq : bcs_ours ] ) without pairing interaction for the orbital @xmath5 ( orange / light ) and the orbital @xmath4 ( red / dark ) . the ` floor ' indicates the fs in red . ( b ) same as ( a ) but for band 1 ( orange / light ) and band 2 ( red / dark)[nkxysurv0],title="fig:",scaledwidth=40.0% ] let us start with the non - interacting case in which @xmath80 . the spectral functions along high - symmetry directions in momentum space are presented in fig . [ fig : akb2g_00 ] and , as expected , they reproduce the non - interacting band dispersion . we also show the total occupation number @xmath112 along the same directions [ dashed lines in fig . [ fig : nkxy05](a ) ] , as well as the occupation number for each orbital @xmath180 [ dashed lines in panel ( b ) ] and @xmath181 [ dashed lines in panel ( c ) ] and the orbital occupation in the quadrant of the first bz defined by @xmath182 in fig . [ nkxysurv0](a ) . it is clear that the electron - like and hole - like fermi surfaces are determined by an admixture of the two orbitals . on the other hand , in the band representation , band 1 determines the electron pockets while band 2 forms the hole pockets as it can be seen from the behavior of @xmath183 and @xmath184 ( indicated by the dashed lines in panels ( d ) and ( e ) of fig . [ fig : nkxy05 ] ) and by the light ( orange ) and dark ( red ) surfaces in fig . [ nkxysurv0](b ) where the fs is also indicated . it is clear that the electronic occupation of band 1 is smaller than the electronic population of band 2 , so that unpaired electrons would be expected to belong predominantly to band 1 , as in the simple model of the previous sec . [ sec : interband ] . it is interesting to notice that in the orbital representation , on the other hand , the electrons are equally distributed among the @xmath185 and @xmath186 @xmath1 orbitals . let us discuss what occurs when the pairing interaction becomes nonzero . to simplify the discussion , define the following points in momentum space : @xmath187 , @xmath188 , @xmath189 , and @xmath190 . as remarked in ref . , the @xmath18 pairing operator always has nodes along the @xmath191 direction because the spatial form factor @xmath192 vanishes along that line . but , as soon as @xmath64 is finite , a gap opens along the @xmath193 direction [ notice that along this direction the pairing is purely intraband since in eq . ( [ 6 ] ) , @xmath194 and thus @xmath195 in eq . ( [ 8 ] ) ] . along @xmath196 , @xmath197 , @xmath198 , and @xmath199 nodes associated to the different number of electrons in band 1 and band 2 remain [ notice that along these directions the pairing is purely interband since @xmath168 because eq . ( [ eq : bcs_ours ] ) and eq . ( [ 6 ] ) become identical to one other ] . when the pairing interaction @xmath64 becomes strong enough to make @xmath146 , as described in the simplified model presented in the sec . [ sec : interband_strong ] , these nodes vanish . along any other direction in the bz a mixture of intra and interorbital pairing will be present.@xcite let us first consider a relatively small pairing @xmath81 . in fig . [ fig : nkxy05](a ) , @xmath112 is presented along high symmetry directions . along @xmath200 there is no pairing and , thus , @xmath112 is unchanged from the non - interacting case shown in the figure with dashed lines . along @xmath198 , where only interband pairing occurs , no effects are observed at the electron pocket fs but a rounding in @xmath112 indicating pairing is observed at the hole pocket fs . however , @xmath112 shows discontinuities at two points indicating the existence of nodes . along the diagonal direction @xmath201 , where all the pairing is intraband , it is found that @xmath112 exhibits standard bcs behavior at both hole fermi surfaces indicating the opening of gaps . finally , along @xmath196 it can be observed a rounding of @xmath112 at the hole fermi surfaces , indicating pairing , and a sharp jump at the electron fs . we can further analyze the pairing in the orbital representation . the occupation number for the orbitals @xmath185 and @xmath186 is shown in figs . [ fig : nkxy05](b - c ) . along @xmath200 , where the pairing is zero , we observe how the fs for the electron pocket at @xmath202 @xmath203 is totally determined by electrons in the orbital @xmath185 @xmath204 and how the population of each orbital varies smoothly between the two fermi surfaces , always satisfying @xmath205 . from @xmath206 to @xmath207 , @xmath208 for @xmath209 indicating that the electrons at the hole pocket fs are paired in a wide region around it , but the pairing region is very narrow around the electron pocket because the pairing is reduced by the small value of @xmath23 . along the diagonal , i.e. from @xmath207 to @xmath2 , standard intraband pairing occurs at both hole fermi surfaces and , thus , @xmath208 while from @xmath2 to @xmath206 @xmath208 for @xmath210 indicating pairing around the hole fs and almost no pairing occurs at the electron pocket fs . the behavior @xmath211 for all @xmath133 for which pairing occurs and @xmath212 unchanged from the non - interacting value for all other @xmath133 , was observed for all values of @xmath64 . for this reason , figures for @xmath212 in the orbital representation will not be shown for the additional values of @xmath64 discussed below . the population of the different bands is presented in fig.[fig : nkxy05 ] ( d - e ) . the fermi surfaces along @xmath200 are clearly determined only by band 1 , which is the band that forms the electron pockets . it can also be observed that there is almost negligible pairing at the electron fs along @xmath198 , but there is clear interband pairing at the hole fs . this is an indication that , due to the spatial variation of the pairing interaction , the attraction is much stronger at the hole pockets than at the electrons pockets . also notice that at the hole pocket fs the pairing is interband and thus @xmath146 , but this does not happen along the diagonal direction @xmath201 where intraband pairing occurs , and there are more paired electrons belonging to band 2 than to band 1 . along the direction @xmath213 again we observed a stronger pairing effect at the hole fs than at the electron one . in this section , we discuss the form of the spectral functions @xmath214 , which can be measured in angle resolved photoemission spectroscopy ( arpes ) experiments , for weak to strong interorbital pairing . the spectral function for @xmath81 is depicted in fig . [ fig : akb2g_05 ] . as discussed above , the interorbital @xmath18 operator leads to _ intra_-band coupling along the @xmath193 line , where one consequently clearly sees the hole pockets to be gapped . along @xmath196 , the pairing is purely _ inter_-band , and one finds a fermi surface on both the hole and electron pockets , indicating that @xmath81 corresponds to the `` breached '' phase in fig . [ regimes ] . however , the spectral weight determining the hole - pocket node is weak and might thus be missed in the analysis of experiments . the node on the electron pocket fs , on the other hand , is robust and it should be observed , if present . the same occurs along @xmath198 ; the signal for the node at the electron pocket fs should be robust while the one at the hole pocket fs will be weak . as the pairing interaction increases , the nodes resulting from the interband pairing should get closer to each other , as discussed in sec . [ sec : interband ] . figures [ fig : nk36](a - c ) show @xmath112 for @xmath215 along high - symmetry directions in the band representation . one can see that there are still unpaired electrons , and @xmath215 consequently falls into the `` breached '' region schematically represented in fig . [ regimes ] . for the two - orbital model eq . ( [ eq : bcs_ours ] ) with pairing interaction @xmath215 ( continuous lines ) and @xmath80 ( dashed lines ) . parameters : @xmath176 , @xmath177 , @xmath178 , @xmath179 . ( b ) same as ( a ) but for band 1 . ( c ) same as ( b ) but for band 2 . ( d ) total occupation number @xmath112 for the two - orbital model eq . ( [ eq : bcs_ours ] ) with pairing interaction @xmath216 ( continuous lines ) and @xmath80 ( dashed lines ) . ( e ) same as ( d ) but for band 1 . ( f ) same as ( e ) but for band 2.[fig : nk36],title="fig:",scaledwidth=40.0% ] for the two - orbital model eq . ( [ eq : bcs_ours ] ) with pairing interaction @xmath215 ( continuous lines ) and @xmath80 ( dashed lines ) . parameters : @xmath176 , @xmath177 , @xmath178 , @xmath179 . ( b ) same as ( a ) but for band 1 . ( c ) same as ( b ) but for band 2 . ( d ) total occupation number @xmath112 for the two - orbital model eq . ( [ eq : bcs_ours ] ) with pairing interaction @xmath216 ( continuous lines ) and @xmath80 ( dashed lines ) . ( e ) same as ( d ) but for band 1 . ( f ) same as ( e ) but for band 2.[fig : nk36],title="fig:",scaledwidth=40.0% ] it is interesting to notice that while @xmath217 , on the other hand @xmath218 indicating that the reconstruction around the hole pockets is much larger than around the electron pockets . this can also be observed in fig . [ fig : nk36](b ) , where we observe unpaired electrons along @xmath198 , but not along @xmath196 . the effect occurs in part due to the smaller value of @xmath23 at the electron pockets but also because , due to the band dispersions , the price in kinetic energy for interband pairing is much larger at the electron pockets than at the hole pockets . the spectral density in fig . [ fig : akb2g_3 ] further illustrates that the pairing interaction is more effective along @xmath196 than along @xmath198 . only shadow spectral weight crosses the chemical potential along @xmath196 , while strong spectral weight crosses the chemical potential twice along @xmath198 and leads to two nodes . finally , let us consider a much stronger value of the pairing , such as @xmath216 , in the bcs region of fig . [ regimes ] , for which the only nodes observed are those along @xmath191 , due to the vanishing of the pairing operator . figures [ fig : nk36](d - f ) , show that @xmath112 is discontinuous only along @xmath191 , while it is smooth along all the other directions indicating pairing . note that @xmath146 along @xmath196 and @xmath198 where interband pairing occurs , while along @xmath193 , @xmath212 is smooth but different for each band because the pairing is intraband . the behavior of the spectral functions displayed in fig . [ fig : akb2g_6 ] shows that spectral weight only crosses the chemical potential along the @xmath191 direction . in this situation , in the folded bz , nodes should occur only at the points where the two electron pockets cross with each other , as indicated in ref . . in the rest of the bz an anisotropic gap will be observed . finally , let us discuss the important issue of the stability of the @xmath18 pairing state . it will be assumed , following the notation in appendix a of ref . , that eq . ( [ eq : bcs_ours ] ) has arisen from an interorbital attractive potential of the form @xmath219 and that @xmath220 where @xmath221 for different values of the attraction @xmath222 as a function of @xmath223.[fig : ebg],scaledwidth=40.0% ] @xmath64 in eq . ( [ 5 ] ) is then given by @xmath223 and the mean - field energy @xmath157 can be calculated for a given @xmath222 as a function of @xmath64 . the results are presented in fig . [ fig : ebg ] . in this figure , a minimum for @xmath158 can be seen for @xmath224 . this indicates that if the pairing attraction overcomes a finite threshold level , the mean field results presented here with @xmath225 will be stable . thus , _ if the attraction exists , both multinodal ( breached ) and states with nodes only along x - y ( strong coupling ) are possible . _ summarizing , in this manuscript the possibility of intra and interband pairing in multiorbital systems has been discussed . while interband pairing has previously been studied in the context of qcd , cold atoms,@xcite heavy fermions,@xcite cuprates@xcite , and bcs superconductivity,@xcite most of the pairing operators proposed for the pnictides are based on the premise that the pairing has to be purely intraband , i.e. , both electrons in the cooper pairs belonging to the same band , compatible with the assumption that in these superconductors all the action must occur at the fermi surfaces . however , symmetry considerations show that in models for the pnictides interorbital pairing is allowed with the two members of the cooper pair belonging to different bands.@xcite this is not surprising since the bands that determine the electron and hole fermi surfaces consist of hybridized orbitals . these interorbital pairing operators give rise to not only intra but also interband pairing when the band representation is used . in addition , numerical calculations in a minimal two orbital model for the pnictides favor one of these non - trivial pairing operators.@xcite as a consequence , a clear discussion of the role of interband pairing is necessary . the explicit calculations shown here of the electronic occupation in the orbital and the band representations , as well as the calculation of the spectral functions showing the distribution of bogoliubov bands , may offer guidance in the interpretation of arpes experiments . in particular , most arpes measurements determine the fs in the normal state and study the opening of the superconducting gap by monitoring @xmath226 as they lower the temperature.@xcite notice that this approach would miss the nodes associated with the `` breached '' phase since in the superconducting state a gap would be observed in @xmath226 while the node would be detected in @xmath227 . thus , experimentalists should investigate the possibility of nodes at points in momentum space that do not belong to the normal state fs and they must keep in mind that some of the nodes may be determined by shadow bands with very small spectral weight . the most recent experimental results with the polarization dependence of the arpes spectra for bafe@xmath228co@xmath229as@xmath10 provided the allowed contribution of each of the five @xmath1 orbitals to the electron and hole fermi surfaces.@xcite while discrepancies with proposed four and five orbital models are remarked in that publication , it is interesting to observe that the minimal two - orbital model addressed here does not contradict the arpes findings if we disregard the additional @xmath230 hole pocket fs that they present , which is reasonable because it has @xmath231 character , an orbital not included in the minimal model considered here . in fact , along their @xmath193 direction , which corresponds to our @xmath196 , our first hole fs is purely @xmath5 as it is their @xmath232 hole fs ; our second hole fs , which arises upon folding our extended fs along @xmath191 is purely @xmath4 , as it is their @xmath233 hole fs , and our electron fs is purely @xmath4 as it is their @xmath234 electron fs . our second electron fs ( obtained upon folding ) has a purely @xmath5 character , while their electron fs @xmath235 appears to be mostly @xmath4 and @xmath6 , but with some amounts of @xmath5 as well . along the diagonal direction @xmath196 , which corresponds to our @xmath193 , the two hole fermi surfaces are a symmetric admixture of @xmath5 and @xmath4 , exactly as in the two - orbital model.@xcite thus , these similarities between the experimental results and the band composition of the simple two - orbital model offer encouragement towards exploring whether the main physics of the pnictides can be captured with such a minimum number of degrees of freedom . we have also mentioned in the text that the symmetry of the lattice and of the orbitals introduce constraints on the possible pairing operators . in general , a purely intraband pairing , such as the proposed @xmath236 state,@xcite would occur only if the coupling of the electrons with the source of the attraction is _ identical _ for all orbitals . considering the different spatial orientations of the orbitals , it is not obvious that this should be the case , since , as discussed in the introduction , phonons couple differently to electrons in the @xmath13 and @xmath15 boron orbitals in the case of mgb@xmath10 . thus , if indeed the coupling results to be the same for all @xmath1 orbitals we could use this fact to elucidate the pairing mechanism ; but , if this is not the case , we would expect some degree of interband pairing , at least in some regions of the brillouin zone and , thus , it is important to study the experimental and theoretical consequences of such a possibility . it is interesting to observe that in the undoped limit there are also three regimes : normal , coexisting magnetism and metallicity , and magnetism without a fermi surface . for details see r. yu , k. t. trinh , a. moreo , m. daghofer , j. a. riera , s. haas , and e. dagotto , phys . b * 79 * , 104510 ( 2009 ) . h. ding , p. richard , k. nakayama , k. sugawara , t. arakane , y. sekiba , a. takayama , s. souma , t. sato , t. takahashi , z. wang , x. dai , z. fang , g. f. chen , j. l. luo , and n. l. wang , epl * 83 * , 47001 ( 2008 ) . l. wray , d. qian , d. hsieh , y. xia , l. li , j.g . checkelsky , a. pasupathy , k.k . gomes , c. v. parker , a.v . fedorov , g.f . chen , j.l . luo , a. yazdani , n.p . ong , n.l . wang , m.z . hasan , phys . b * 78 * , 184508 ( 2008 ) . v. b. zabolotnyy , d. s. inosov , d. v. evtushinsky , a. koitzsch , a. a. kordyuk , j. t. park , d. haug , v. hinkov , a. v. boris , d. l. sun , g. l. sun , c. t. lin , b. keimer , m. knupfer , b. buechner , a. varykhalov , r. follath , and s. v. borisenko , nature * 457 * , 569 ( 2009 ) . y. zhang , b. zhou , f. chen , j. wei , m. xu , l. x. yang , c. fang , w. f. tsai , g. h. cao , z. a. xu , m. arita , c. hong , k. shimada , h. namatame , m. taniguchi , j. p. hu , and d. l. feng , arxiv:0904.4022 , 2009 .

_ chandra _ and _ xmm - newton _ observations now routinely allow us to measure both the inverse - compton emission from the lobes of radio galaxies and the thermal bremsstrahlung from the intra - group or intra - cluster media in which they are embedded ( e.g. , * ? ? ? * ; * ? ? ? * ) . in almost all cases studied to date , the internal pressures in the radio lobes are comparable to those in the hot , x - ray emitting phase of the external medium , rather than greatly exceeding it as had sometimes been thought to be the case ( cf . * ? ? ? we know of only a few objects the best example being the inner lobes of centaurus a @xcite in which the internal lobe pressures must be much larger than the inferred external pressure , although the limited numbers of such sources found to date is almost certainly affected by selection effects . in the objects where the pressures of the radio - emitting plasma and the x - ray - emitting external medium are comparable , the dynamics of the large - scale radio structures , the lobes and plumes , will be determined , at least in part , by the properties of the hot gas . clear evidence that this is so pre - dates the launch of _ chandra _ and _ xmm - newton _ ( e.g. , * ? ? ? * ) but much more detailed studies have been possible in recent years , particularly in nearby , low - power radio galaxies ( e.g. * ? ? * ; * ? ? ? at the same time , the lobes displace and compress the material that they expand into , as is evident from observations of ` cavities ' in the x - ray - emitting material ( e.g. , * ? ? ? * ) . this means that , when we observe a radio galaxy and its environment , and see a relationship between radio - emitting structures and those in the x - ray , it is often not clear whether what we are seeing is a pre - existing distribution of hot gas that has shaped the radio source that lies in it , or whether , on the other hand , we are seeing the _ effects _ of the radio source on what would otherwise have been a relatively structure - free external medium . in a relaxed system that does not host a powerful radio source we do not expect to find long - lasting complex structure in the hot x - ray - emitting plasma of a group or cluster atmosphere , since any such structure will be washed out on a sound - crossing time . however , complex structures can be created by dynamical processes related to the group or cluster galaxies . cluster or group mergers are known to produce large - scale asymmetries in the icm / igm which may affect the dynamics of radio sources hosted by one of the group / cluster galaxies ( e.g. , * ? ? ? * ) . on smaller scales , galaxy - galaxy interactions and mergers would be expected to give rise to structures in the group or cluster atmosphere , both because they produce a local gravitational potential that is not radially symmetrical and because frictional and tidal processes will give rise to the stripping and heating of gas from the galaxy ( e.g. , * ? ? ? * ) : the latter process is of course important if one or both of the merging systems are massive elliptical galaxies with large pre - existing galaxy - scale hot gas haloes . this makes it particularly interesting to observe radio galaxies in which a major merger is taking place . in this paper we present new _ observations of two nearby radio galaxies , 3c285 and 3c442a , which meet this selection criterion . we show that structures in the x - ray - emitting gas , likely related to the mergers , are almost certainly affecting the dynamics of the lobes of the radio galaxies , and discuss the implications of this result for the radio galaxy population as a whole . more detailed discussion of the dynamics of 3c442a in relation to the ongoing galaxy merger can be found in @xcite . 3c285 is a nearby ( @xmath0 , @xcite ) low - power frii radio galaxy , with a 178-mhz luminosity km s@xmath1 mpc@xmath1 , @xmath2 and @xmath3 . ] of @xmath4 w hz@xmath1 sr@xmath1 , just above the canonical fanaroff - riley @xcite break at @xmath5 w hz@xmath1 sr@xmath1 . at the distance of 3c285 1@xmath6 corresponds to a projected distance of 1.50 kpc : the luminosity distance is 360 mpc . although 3c285 shows jets and hotspots ( best seen in the images of @xcite ) the jets are relatively bright , and the hotspots relatively weak , compared to more luminous frii sources , suggesting that it is in some sense an intermediate object . 3c285 s host galaxy is the brightest member of a small optical group @xcite and is strongly elongated towards a fainter spiral companion 40@xmath6 ( 60 kpc ) to the nnw , suggesting a tidal interaction : an extended filament of optical emission - line gas connects the two galaxies @xcite . several other galaxies with magnitudes that make them likely to be group members lie within 200 kpc . in _ hst _ imaging @xcite 3c285 s host galaxy can be seen to be strongly disturbed , with several distinct linear dust features , suggesting a recent merger with a large , gas - rich galaxy ( cf . * ? ? ? 3c285 is best known for the association of its radio jet with a blue , star - forming region , making it one of the few objects to show direct evidence for jet - induced star formation @xcite . @xcite claim a detection of an optical counterpart to the e hotspot , which we discuss further below ( [ opt - hotspot ] ) . 3c285 had not been detected in the x - ray until the observations discussed in the present paper : the _ einstein _ hri upper limit quoted by @xcite shows that the system is not very x - ray luminous ( their limit corresponds to @xmath7 erg s@xmath1 between 0.5 and 3.0 kev within @xmath8 in the cosmology we use ) . because of its distance from 3c442a it is not detected in any of the x - ray datasets discussed in this paper . ] is associated with the interacting galaxy pair ngc 7236/7 at @xmath9 @xcite . its 178-mhz luminosity of @xmath10 w hz@xmath1 sr@xmath1 puts it below the fri / frii break . at the distance of 3c442a 1@xmath6 is 0.54 kpc : the luminosity distance is 118 mpc . in the radio @xcite , 3c442a shows a double - lobed structure with no sign of any jet , although a compact flat - spectrum radio core coincident with ngc 7237 suggests current agn activity in that object . @xcite argue that the compact features in and close to the lobes are unrelated to the source , which , if true , would make 3c442a a member of the class of ` fat doubles ' or ` relaxed doubles ' ( e.g. , * ? ? ? * ) , a number of which are found in the 3crr sample close to the fri / frii luminosity break . 3c442a is remarkable for the detection by @xcite of two filaments of steep - spectrum radio emission that cross the region containing the host galaxy , ngc 7237 . optically this galaxy , an elliptical ( assumed to be the host on the basis that it hosts a compact radio source ) appears to be interacting with an s0 neighbour of similar luminosity to the ne , ngc 7236 : the two are separated by 35@xmath6 ( 19 kpc ) and are embedded in a common , distorted stellar envelope which also includes another , smaller elliptical 38@xmath6 ( 21 kpc ) to the sw of ngc 7237 , denoted ngc 7237c hereafter @xcite . the measured radial velocities for ngc 7236 and ngc 7237 are very similar , suggesting an interaction close to the plane of the sky , which allowed @xcite to make detailed models of the system . _ hst _ imaging @xcite shows some weak evidence for dust emission in ngc 7237 , but generally the inner parts of the system appear like an undisturbed elliptical , in contrast to 3c285 s host . ngc 7236/7 lies at the center of a group of galaxies @xcite . in the x - ray , @xcite used _ rosat _ data to show that the soft x - ray emission had relatively complex structure , with a small - scale extension around ngc 7237 as well as large - scale ( 3@xmath11 , or 100 kpc ) extension in a direction perpendicular to the radio lobe axis . [ 442-intro ] 3c285 and 3c442a were both observed using the acis instrument on _ chandra_. 3c285 was observed for a total livetime of 39.3 ks in a single observation with the acis - s in the hrc guaranteed time program ( pi : stephen murray ) , as part of a larger project to complete _ chandra _ observations of the 3crr sample at @xmath12 . 3c442a was observed in four separate observations , with three distinct roll angles , with the acis - i as a guest - observer observation ( pi : ralph kraft ) for a total livetime of 93.6 ks . table [ obs ] lists the observing times and obsids for the observations . we reprocessed and filtered the data in the standard way , using the standard asca grade set . both observations were made in vfaint mode and so we used vf cleaning to reduce the background level . the 0.5-pixel event position randomization in the standard pipeline processing was also removed . we then used light curves excluding the nucleus and bright point sources to search for intervals of high background . for 3c285 this flagged a small fraction of the data , as shown in table [ obs ] . llllllll 3c285&acis - s&23678&2006 mar 18&6911&147.6&39625&39280 + 3c442a&acis - i&01236&2005 jul 27&5635&141.5&27006&27006 + & & & 2005 jul 28&6353&141.5&13985&13985 + & & & 2005 oct 07&6359&262.3&19884&19884 + & & & 2005 dec 12&6392&312.2&32694&32694 + [ obs ] both radio galaxies have been well observed with the very large array ( vla ) in the radio at a number of frequencies @xcite and images showing the overall source structure at 1.4 ghz are available from the online 3crr atlas of j.p . leahy et al .. we required high - resolution radio images at multiple frequencies to allow us to compare the radio data with the x - ray data at the highest resolution , and we also needed multi - frequency data sampling the largest angular scales of the sources to allow us to carry out internal pressure and inverse - compton calculations for the radio lobes . accordingly , we retrieved various vla datasets from the online archive ( see table [ vlaobs ] for a complete list ) and reduced them in the standard manner within aips . where the vla had observed the sources at a given frequency at more than one configuration , we reduced the data from each configuration separately , cross - calibrated for phase and then concatenated the datasets with appropriate reweighting using the aips task dbcon . mapping was carried out using the aips tasks imagr ( for the lowest resolutions ) and vtess ( for maps of combined datasets at high resolutions ) . in addition , for flux measurements of the lobes of 3c285 at low frequencies , we used giant meter - wave radio telescope ( gmrt ) data at 240 and 610 mhz taken for another purpose ( pi : martin hardcastle ) . maps from these observations will be presented elsewhere . llllll 3c285&av127&1986 may 07&1.51&a&2.6 + & av127&1986 aug 29&4.85&b&1.9 + & av127&1986 aug 29&1.51&b&1.1 + & av127&1986 dec 02&4.85&c&0.9 + & av127&1986 dec 02&1.51&c&0.5 + & as549&1995 apr 11&4.91&d&0.7 + 3c442a&ac131&1985 may 13&1.44&b&3.0 + & ac131&1985 may 13&4.85&b&2.9 + & ac131&1985 jul 19&1.44&c&2.8 + & ac131&1985 jul 19&4.85&c&2.8 + & ac131&1985 dec 06&1.44&d&1.5 + [ vlaobs ] both sources are close enough that images from the digital sky survey ( dss2 ) can be used to indicate the positions of the galaxies in the host group . in addition , both have been observed with the _ hubble space telescope _ ( _ hst _ ) , as discussed above , and with _ spitzer _ as part of a survey of the low - redshift 3crr objects ( pi : mark birkinshaw ) . we use dss2 and archival _ hst _ data , the latter processed using the standard pipeline , where optical information is important to our analysis . the _ spitzer _ observations of 3c442a are discussed by @xcite and we only refer to them in passing in the present paper . in this section we discuss the main observational results from the _ chandra _ observations . we first briefly describe the small - scale emission seen from the active nuclei of the systems , the detections of or limits on any hotspot or jet - related components , and the other galaxies in the group . we then discuss in detail the extended x - ray emission seen in the _ chandra _ data and its relationship to the radio source . all the x - ray analysis we describe was carried out using ciao 3.3 and caldb 3.2.2 for spectral extraction , and xspec 11.3 for spectral fitting . the _ specextract _ script was used for spectral extraction for extended sources , and the _ psextract _ script for point sources . galactic column densities of @xmath13 @xmath14 and @xmath15 @xmath14 are assumed for 3c285 and 3c442a respectively ( interpolated from the data of @xcite ) and are included in all spectral fits . we fitted to data in the energy range 0.47.0 kev unless otherwise stated : spectra were binned , typically to 20 net counts per bin after background subtraction , so that @xmath16 statistics could be applied . errors quoted on parameters derived from xspec fitting are the 90 per cent confidence range for one interesting parameter , unless otherwise stated , but errors quoted on measured numbers such as flux densities or numbers of counts are the @xmath17 errors . where counts are quoted the @xmath17 errors are derived from the gaussian ( @xmath18 ) approximation to the true confidence range , as calculated by the funtools code used to determine the background - subtracted values this is accurate to better than 10% for @xmath19 counts ( cf . @xcite ) , and the gaussian assumption is necessary in any case if we are to combine errors in the standard way with the uncertainty on the expected background counts . the x - ray nucleus of 3c285 has already been discussed by @xcite , using the same _ chandra _ dataset . they found that it was well modelled as a double power law , with one component being heavily absorbed ( @xmath20 @xmath14 ) and one having only galactic absorption . the nucleus was relatively weak , and so in their fits the poorly constrained power - law indices were fixed to 2.0 and 1.7 for the unabsorbed and absorbed components respectively . a combination of absorbed and unabsorbed emission is consistent with what is seen for other narrow - line frii radio galaxies . 3c442a is a low - excitation radio galaxy @xcite i.e. , a nuclear spectrum of ngc 7237 shows no strong high - excitation emission - line features and so we might expect not to see a heavily absorbed nuclear component @xcite . we extracted spectra for the x - ray source corresponding to the currently active radio core using a small ( 6 standard @xmath21 _ chandra _ pixels in radius ) circular extraction region and a concentric background annulus ( between 6 and 9 pixels ) . the nuclear spectra can be fitted acceptably ( @xmath22 for 12 degrees of freedom ) with a single power law model with galactic absorption and variable normalization ( see below ) , although the best - fitting single power - law photon index fitted to all four spectra simultaneously is rather flat ( @xmath23 ) . there is some evidence in radial profiles of the core region that even this small - scale extraction region contains extended emission , and so it is possible that the spectra are contaminated by some thermal emission , but adding a thermal component does not significantly change the best - fitting power - law spectrum . adding some intrinsic absorption ( the same column density for all four epochs ) gives a more typical power - law index ( @xmath24 ) and a column density of @xmath25 @xmath14 , with a somewhat improved fit ( @xmath26 for 11 degrees of freedom ) . this column density is high , but not unprecedentedly high , compared to other low - power sources , and may indicate some obscuration by the small - scale dust lane that is seen to cross the nucleus in the _ hst _ observation . we estimate a maximum @xmath27 for the off - nuclear extinction seen in the _ hst _ data taken with the f547 m filter , which would be consistent with the lower end of the confidence range for the x - ray absorbing column for a galactic gas to dust ratio . the core was clearly variable over the time period of our observations , and acceptable fits were not obtained unless we allowed the normalizations of the power laws fitted to each dataset to vary . the normalizations for the four epochs listed in table [ obs ] correspond to 1-kev flux densities of @xmath28 , @xmath29 , @xmath30 and @xmath31 njy respectively ( assuming galactic absorption only : flux densities would be approximately a factor 2 higher for the best - fitting absorbing column ) . variability of the photon index as well as the normalization might explain the relatively poor fit and flat photon index given by the combined dataset , but the data are not good enough to constrain this possibility . we were initially concerned that the variability might be an artefact of the data , as the core in the third observation ( obsid 6359 ) lies close to a node boundary , with many events in the core region in the level 1 events file being flagged as being on or near a bad pixel . however , we are satisfied that the ancillary response files ( arfs ) make an adequate correction for this reduced effective exposure time , and we have verified that regions taken from the thermal emission close to the core ( [ tail - fit ] ) do not show any variability in count rate or fitted parameters . we therefore infer that the core is genuinely variable , presumably because of ongoing agn activity . the 5-ghz flux density of the core of 3c442a is 3.1 mjy in the vla data we use , compared to the value of 1.9 mjy from earlier radio observations tabulated by @xcite , so that it appears that the radio core may be similarly variable on long timescales ( months in the case of the x - ray , years in the case of the radio ) . the approximate positions , given the uncertainties imposed by variability and non - simultaneous observations , of the source on the radio / x - ray core flux density and luminosity correlations for low-@xmath32 sources @xcite are consistent with expectations . the bolometric unabsorbed x - ray luminosity of the core region ranges between 2 and @xmath33 erg s@xmath1 , which puts it well above the expectation from x - ray binaries from the whole galaxy @xcite given the @xmath34-band optical luminosity of @xmath35 @xcite . the jets and hotspots of some low - power frii radio galaxies ( e.g. , 3c403 , @xcite ) show prominent synchrotron x - ray counterparts to compact radio features . 3c285 has a well - defined though faint jet and counterjet in the radio and a bright compact hotspot ( in the e lobe ) . however , none of these is detectable as an x - ray source in the chandra observation ( fig.[285-highres - overlay ] ) . the only convincingly detected compact feature in the lobes in the _ chandra _ data ( with @xmath36 counts in the 0.5 - 5.0 kev energy band ) is spatially coincident with the blue object denoted 09.6 by @xcite , lying within @xmath37 of the peak of the resolved optical emission in the _ hst _ images , with a centroiding uncertainty of @xmath38 . 09.6 was identified by @xcite as a star - forming galaxy that is at the redshift of 3c285 s host , whose star formation may be related to the nearby radio jet . roughly , the unabsorbed bolometric x - ray luminosity of this object in the _ chandra _ band is of the order of @xmath39 erg s@xmath1 ( assuming a mekal model with @xmath40 kev ) , which is a plausible luminosity for a dwarf starburst system with @xmath41 ( the absolute magnitude quoted by @xcite : our different cosmology makes no significant difference ) : for example , it is comparable to , though somewhat higher than , the luminosities derived for a sample of nearby , generally somewhat optically fainter dwarf starbursts by @xcite . the x - ray detection is thus consistent with the starburst interpretation for 09.6 . the non - detections of x - ray emission from the jets and hotspots in 3c285 correspond to upper limits on the x - ray flux density . roughly , a @xmath42 upper limit on the total background - subtracted counts in the 0.5 - 5.0 kev energy range from a region corresponding to the brightest compact hotspot ( in the e lobe ) is @xmath43 counts : this corresponds to a flux density limit of @xmath44 njy at 1 kev for a source with photon index 2.0 . however , the radio flux densities for the hotspot components are faint compared to the sources in which x - ray emission has been seen . the total 5-ghz flux density of the e hotspot is 6 mjy , which implies that @xmath45 ( where @xmath46 is the two - point radio - to - x - ray energy spectral index : @xmath47 is defined here and throughout the paper in the sense that flux density @xmath48 ) , but this is not a very strong constraint , as there are some hotspots with detected synchrotron x - ray emission that show @xmath49 @xcite , while we would expect any synchrotron self - compton emission at equipartition to be at the level of about 10 pjy . the e hotspot of 3c285 is potentially interesting as it is the one claimed by @xcite to show an optical counterpart . however , examination of the f702w _ hst _ observations ( originally discussed by @xcite ) shows that the peak of the optical emission from the hotspot region is offset by @xmath50 from the peak of the radio emission , using the default astrometry , which gives a reasonably good alignment ( offset @xmath51 ) between the radio core and the center of the galaxy ( fig . [ 285-hst ] ) . more importantly the optical object is morphologically quite different from the radio structure , and seems most likely to be a faint irregular galaxy . we thus conclude that there is no evidence for any emission from the e hotspot of 3c285 in any band other than the radio . the w hotspot is fainter in the radio , so again we can only put weak limits on the radio to x - ray spectral index , and it has no detectable optical counterparts . overall , the hotspots of 3c285 seem to present a rather different picture to those seen in other nearby frii radio galaxies of similar power such as 3c403 @xcite , 3c321 @xcite and 3c33 @xcite . the hotspots of these objects , which have radio luminosities spanning a range around that of 3c285 , often have optical / ir and x - ray counterparts ; their x - ray emission , largely inferred to be of synchrotron origin , has @xmath52 . the lack of an optical counterpart in particular may suggest that 3c285 s hotspots are not able to accelerate particles to even moderately high energies . however , the biased nature of existing observations of nearby radio galaxies , many of which have been selected for observation because of their prominent radio or optical hotspots , makes it hard to say which is more typical . [ opt - hotspot ] as discussed in [ 442-intro ] , it is not clear whether any of the bright compact radio sources in the 3c442a lobe are related to the radio galaxy itself . the x - ray data shed some light on this question , but do not completely resolve it . [ 442-highres - overlay ] shows our highest - resolution radio map overlaid on lightly smoothed _ chandra _ data in the 0.5 - 5.0 kev band , with point sources labeled using the notation of @xcite . it can be seen that the two point sources , c and e , that lie outside or just on the edge of the lobes are relatively strong x - ray sources , having respectively @xmath53 and @xmath54 net counts in total in our observations . as @xcite report , both these sources have optical counterparts , and it seems most likely that they are background agn of some kind . c has an inverted radio spectrum between 1.4 and 4.9 ghz ( @xmath55 ) while e is steep - spectrum ( @xmath56 ) . the most puzzling of the bright point sources , a , is not detected in the x - ray and has no optical counterpart on dss2 images ( neither the _ hst _ nor _ spitzer _ fields of view are large enough to give us more sensitive infrared or optical constraints ) . it is resolved at the highest radio resolution we have available ( @xmath57 ) into two components separated by @xmath58 , the northern one unresolved and containing most of the flux , the southern one slightly resolved . its radio spectrum is steep ( @xmath59 ) , consistent with the results of @xcite at lower frequencies , which might imply that it is part of the source : but the lack of any connection to the filamentary structure in the lobes remains a strong argument against this , and it might equally well be a high - redshift double or core - jet source . if the source were a hotspot , the @xmath42 upper limit of 6 net counts in our observations , corresponding to 0.1 njy at 1 kev for a source with photon index 2.0 , would imply @xmath60 , but again this is not incompatible with the behaviour of some hotspots . no x - ray counterpart is detected for the weak point source d discussed by @xcite , but it seems likely to be part of the complex filamentary structure in that part of the lobe , rather than a true point source . we do not detect the source b , referred to by @xcite , in our radio images . we used the nasa extragalactic database , ned , to search for optical objects identified as galaxies around the position of 3c285 and , where possible , to obtain their redshifts as determined by the sloan digital sky survey . in a @xmath61 search radius ( chosen to cover the acis - s3 chip : equivalent to 450 kpc at the distance of 3c285 ) there are 13 objects classified as galaxies ( including 3c285 itself ) of which 8 have measured radial velocities : 6 of these are possibly consistent with membership of the 3c285 group , while the other two are foreground objects . the positions and redshifts of these nearby galaxies are listed in table [ 285-sdss ] . a few of these galaxies are detected in x - rays in the _ chandra _ data . we tabulate counts measured from the events file in @xmath62 ( 15 kpc ) extraction regions with background subtraction from a larger local background region . upper limits are @xmath42 values based on poisson statistics and derived from the local background value . lllll 3c285 & @xmath63 & @xmath64 & 23804 & @xmath65 + sdss j132116.31 + 423551.3 & @xmath66 & @xmath67 & & @xmath68 + sdss j132114.49 + 423620.3 & @xmath69 & @xmath70 & & @xmath71 + sdss j132114.01 + 423557.5 & @xmath72 & @xmath73 & 23738 & @xmath74 + sdss j132114.75 + 423631.1 & @xmath75 & @xmath76 & 24623 & @xmath77 + sdss j132117.98 + 423657.4 & @xmath78 & @xmath79 & 24130 & @xmath80 + sdss j132130.21 + 423544.3 & @xmath81 & @xmath82 & 22643 & @xmath74 + maps - ngp o-219 - 0156129 & @xmath83 & @xmath84 & & @xmath85 + maps - ngp o-219 - 0183438 & @xmath86 & @xmath87 & & @xmath74 + sdss j132131.68 + 423208.2 & @xmath88 & @xmath89 & 10678 & @xmath74 + maps - ngp o-219 - 0156725 & @xmath90 & @xmath91 & & + sdss j132051.31 + 423550.5&@xmath92 & @xmath93 & 15351 & @xmath94 + sdss j132058.74 + 423147.6 & @xmath95 & @xmath96 & 22115 & + [ 285-sdss ] although the two bright galaxies adjacent to 3c285 have no spectroscopic information available in sdss , we can estimate the velocity dispersion of the group from the information we have . if we include all sdss velocities with @xmath97 km s@xmath1 , then the mean heliocentric radial velocity is @xmath98 km @xmath99 , and the velocity dispersion @xmath100 is @xmath101 km s@xmath1 , which would be high for a group . however , as table [ 285-sdss ] shows , the two galaxies furthest away from 3c285 have radial velocities that differ markedly from its velocity . if we consider only the four galaxies closest to 3c285 on the sky then the mean velocity is 24070 km s@xmath1 , about 270 km s@xmath1 higher than that of 3c285 itself , but the velocity dispersion is a more reasonable @xmath102 km s@xmath1 ( where the errors are determined using the method of @xcite ) . we can not draw many conclusions from such a small number of radial velocities ( a relatively short optical observation would give us much better constraints on the nature of the optical group ) but it seems likely that only the inner galaxies in this region should be considered to be bound group members . using the relation of @xcite the lower velocity dispersion would be expected to correspond to a temperature for the x - ray gas of @xmath103 kev , where the errors here are only statistical and take no account of the scatter in the @xmath100 @xmath104 relation . [ 285-vd ] as for 3c285 , we used ned to search for galaxies in the 3c442a group . as 3c442a is substantially closer to us , the data are better : in a @xmath105 ( 450 kpc ) radius aperture there are 40 galaxies , all of which have measured radial velocities from the sdss or elsewhere . 31 of these have radial velocities within 1000 km s@xmath1 of the value for ngc7236/7 , and so are potential group members . positions and radial velocities of these objects are tabulated in table [ 442a - sdss ] . lllll ngc 7237 & @xmath106 & @xmath107 & 7868 & @xmath108 + ngc 7236 & @xmath109 & @xmath110 & 7879 & @xmath111 + ngc 7237c & @xmath112 & @xmath113 & 7153 & @xmath114 + sdss j221439.24 + 134920.0 & @xmath115 & @xmath116 & 7997&@xmath117 + sdss j221437.58 + 135053.0 & @xmath118 & @xmath119 & 8375&@xmath120 + 2masx j22144047 + 1352416 & @xmath121 & @xmath122 & 8030&@xmath123 + sdss j221437.45 + 135228.9&@xmath124 & @xmath125 & 8280&@xmath120 + sdss j221432.20 + 134825.4 & @xmath126 & @xmath127 & 7679&@xmath123 + sdss j221437.45 + 135229.0&@xmath128 & @xmath129 & 7915&@xmath120 + sdss j221451.49 + 135543.8 & @xmath130 & @xmath131 & 35633 & @xmath132 + cgcg 428 - 060 & @xmath133 & @xmath134 & 7270 & @xmath135 + lsbc f673 - 02 & @xmath136 & @xmath137 & 7293 & @xmath138 + sdss j221445.84 + 134449.8&@xmath139 & @xmath140 & 8161 & @xmath120 + sdss j221509.33 + 134813.3 & @xmath141 & @xmath142 & 7952&@xmath143 + sdss j221424.81 + 134633.4 & @xmath144 & @xmath145 & 7263 & @xmath146 + sdss j221445.10 + 134304.9&@xmath147 & @xmath148 & 8207 & @xmath149 + cgcg 428 - 059 & @xmath150 & @xmath151 & 7245 & @xmath152 + sdss j221514.93 + 135419.6 & @xmath153 & @xmath154 & 17810 & @xmath155 + sdss j221413.65 + 135044.0 & @xmath156 & @xmath157 & 120542&@xmath117 + sdss j221444.51 + 135836.6 & @xmath158 & @xmath159 & 7920&@xmath123 + lcsb s2633p & @xmath160 & @xmath161 & 7222 & @xmath123 + sdss j221414.03 + 135232.4&@xmath162 & @xmath163 & 7276&@xmath117 + sdss j221435.72 + 135822.6&@xmath164 & @xmath165 & 8297&@xmath143 + cgcg 428 - 054 & @xmath166 & @xmath167 & 7795 & @xmath168 + sdss j221508.74 + 135815.1 & @xmath169 & @xmath170 & 7799 & @xmath171 + sdss j221450.80 + 134051.3 & @xmath172 & @xmath173 & 7374 & @xmath174 + sdss j221422.18 + 134248.3 & @xmath175 & @xmath176 & 18124&@xmath177 + sdss j221527.93 + 135114.3&@xmath178 & @xmath179 & 6634 & @xmath180 + sdss j221530.83 + 134912.5 & @xmath181 & @xmath182 & 115271&@xmath117 + sdss j221404.42 + 134601.3 & @xmath183 & @xmath184 & 7645 & @xmath185 + sdss j221414.10 + 135901.2 & @xmath186 & @xmath187 & 29751&@xmath188 + sdss j221514.34 + 134045.3 & @xmath189 & @xmath190 & 8356 & + sdss j221407.71 + 134230.7 & @xmath191 & @xmath192 & 7361& + sdss j221402.08 + 134424.9 & @xmath193 & @xmath194 & 120609& + sdss j221515.54 + 140059.1 & @xmath195 & @xmath196 & 7956& + sdss j221540.36 + 135255.5&@xmath197 & @xmath198 & 7434& + sdss j221509.85 + 140230.4&@xmath199 & @xmath200 & 8314&@xmath201 + sdss j221356.95 + 135601.3 & @xmath202 & @xmath203 & 31005&@xmath204 + ugc 11953 & @xmath205 & @xmath206 & 7585& + sdss j221542.30 + 134855.7&@xmath207 & @xmath208 & 7923&@xmath204 + [ 442a - sdss ] several of the other galaxies in the 3c442a group , and one or two of the background galaxies , are detected in the _ chandra _ image . as with 3c285 , we measure counts in a @xmath62 aperture with a larger local background region , and these are tabulated in table [ 442a - sdss ] . because the response is non - uniform over the field covered by the observations , as a result of the multiple _ chandra _ pointings , we have used the exposure maps to correct all the counts tabulated to be equivalent to those measured at ngc 7237 with the full exposure of the combined dataset , by scaling the measured counts and error up by the ratio of the exposure map value at the position of the source to the exposure map value at the position of ngc 7237 . where there is no detection , an upper limit is tabulated , which is the @xmath42 value determined using poisson statistics based on the local background count density and then scaled by the exposure in the same way : upper limits are thus higher for regions with shorter exposure . in all cases background - subtracted , scaled counts or limits are rounded to the nearest integer . the only galaxy other than ngc 7237 bright enough for spectroscopy is ngc 7236 . the nucleus of this source is a compact x - ray source with @xmath209 net counts in the 0.55.0 kev band ( using a @xmath210 source circle and adjacent background ) . because these counts are spread over the four observations , we had to bin the spectra to 10 net counts per bin ( rather than our usual 20 ) to give us enough bins for spectral fitting . we then fitted the joint spectrum derived from the four datasets . the spectrum is equally well fitted with a very steep power law or with a thermal ( mekal ) model with @xmath211 kev , both with galactic absorption , the latter giving @xmath212 for 6 d.o.f . with abundance fixed to 0.5 solar . the unabsorbed bolometric x - ray luminosity in the thermal model was @xmath213 erg s@xmath1 . there was no evidence for any variation of the normalization of the spectrum between the datasets . thus there is no evidence for any agn - related emission in the center of ngc 7236 . using the subset of 31 galaxies with @xmath214 km s@xmath1 we determine the mean heliocentric radial velocity of the group to be 7768 km s@xmath1 , within 100 km s@xmath1 of the velocity of ngc 7236/7 itself . the velocity dispersion is @xmath215 km s@xmath1 , so we would expect a temperature for the intragroup gas of @xmath216 kev using the @xmath100@xmath104 relation of @xcite . [ 442-vd ] both sdss quasars in the field ( sdss j221458.45 + 135344.7 , with @xmath217 , and sdss j221453.84 + 140022.2 , @xmath218 ) are clearly detected in x - rays , with @xmath219 and @xmath220 exposure - corrected 0.5 - 5.0 kev counts respectively . clear asymmetrical extended x - ray emission is visible on arcminute scales around 3c285 ( fig . [ 285-smooth ] ) . near the nucleus , the elongation is in a similar direction to the elongation of the optical axis of the host galaxy : on larger scales , it extends in the direction of the other nearby galaxies in the group , and essentially perpendicular to the radio source axis . the central region of emission , roughly corresponding to the red - colored region in ( fig.[285-smooth ] , extends for just over @xmath221 ( 90 kpc in projection ) but the x - ray emission to the nw and se of the main extended region is connected to it by other x - ray emission at @xmath222 per cent confidence , so it is possible that we should view the whole region from se to nw of the nucleus as forming one physical bar or disk of x - ray - emitting material with varying density , with a size scale of up to @xmath223 ( 220 kpc ) . for convenience , and without presupposing any particular physical interpretation , we follow @xcite and refer to the structure as the ` ridge ' in what follows . we extracted a spectrum for the ridge from a @xmath224 rectangle centred on the host galaxy , with the long axis in position angle @xmath225 ( measured n through e ) . this region contains @xmath226 net 0.55.0 kev counts , using off - source background , and we found that it was acceptably fitted ( @xmath227 for 10 d.o.f . ) with a single - temperature mekal model with fixed abundance of 0.35 solar ( we adopt this value as the data are not adequate to constrain abundance ) and @xmath228 kev . the bolometric unabsorbed x - ray luminosity of the ridge on this model is @xmath229 erg s@xmath1 . the residuals for this model show a soft excess which can be modeled as a @xmath230-kev thermal plasma , also with 0.35 abundance ( the temperature is not well constrained , but @xmath231 kev at 90 per cent confidence ) without affecting the temperature of the hot component , @xmath232 kev : this gives an improvement in the fit to @xmath233 for 8 d.o.f .. a somewhat less good fit is obtained ( @xmath234 for 9 d.o.f . ) if the second component in the two - temperature model is fixed to the temperature ( 0.64 kev ) derived from fits to the group - scale gas ( see below ) : in this case the hotter component has @xmath235 kev . the hotter component is always the dominant one in these fits , and the addition of a second thermal component reduces the emission measure by at most 10% . some weak excess emission is also spatially coincident with the lobes of 3c285 ( this is just visible in parts of the lobe in fig.[285-smooth ] ) . we extract spectra in two circular regions matched to the e and w lobes , excluding point sources and the region of the ridge described above , and taking a background annulus concentric with the lobes . we find @xmath236 counts in the e lobe and @xmath237 counts in the w lobe : thus only the e lobe is significantly detected . fitting a power - law model to the e lobe s spectrum we find a photon index of @xmath238 ( @xmath17 error : @xmath239 for 2 d.o.f . ) which is consistent with the expectations from the inverse - compton process , although of course the errors are large and the spectrum would be equally consistent with a hot thermal model . assuming the power - law model and best - fitting photon index , the 1-kev flux density of the e lobe is @xmath240 njy , and , scaling by count rate , @xmath241 njy for the w lobe . the predicted 1-kev flux density of emission from inverse - compton scattering of the cosmic microwave background radiation ( cmbr ) for equipartition between magnetic fields and radiating electrons ( using the code of @xcite ) is 0.6 njy for the e lobe and 0.9 njy for the w lobe : thus , if all the emission from the e lobe is inverse - compton in origin , the lobe must be somewhat out of equipartition , though not by a particularly unusual amount in the context of frii sources in general @xcite . on larger scales , there is some excess emission on @xmath242 ( 150-kpc ) scales even when the ` ridge ' , the lobes and all the point sources ( including all those visible on fig . [ 285-highres - overlay ] ) are masked out . in a @xmath242 source circle with an adjacent background annulus ( the largest circular region that can be examined without going off the s3 chip ) there are @xmath243 net 0.55.0 kev counts . fitting a spectrum to this region , we found it to be adequately modeled ( @xmath244 for 5 d.o.f . ) with a mekal model , again with fixed abundance of 0.35 solar , with @xmath245 kev . this temperature is consistent with the rough prediction of the group gas temperature made in [ 285-vd ] . we used radial profiling ( again masking out the lobes and elongated central region , but not masking the central few arcsec to allow fitting to the central point source ) , correcting for the small changes in _ chandra _ s response over this radial distance with an exposure map , to determine rough structural parameters of this extended component , fitting @xmath246 models convolved with the _ chandra _ psf in the manner described by @xcite ( using the psf models of @xcite ) to the small- and large - scale structure . the nucleus is faint enough that distortions of the nuclear psf as a result of pileup can be neglected . although the values of @xmath246 are not well constrained ( values @xmath247 are preferred ) we find a good fit ( @xmath248 for 7 d.o.f . ) with a combination of a point source ( for the central active nucleus ) , a small - scale @xmath246 model with @xmath249 , @xmath250 , representing the environment close to the host galaxy , and a large - scale @xmath246-model with @xmath251 , @xmath252 ( @xmath253 kpc ) , representing the group . the radial profile and best - fitting models are shown in fig . [ 285-profile ] . the number of counts in the entire @xmath246 model ( i.e. assuming spherical symmetry , neglecting the fact that some material will have been removed from the lobes , and correcting for the contribution of the @xmath246 model to the background region ) is @xmath254 0.55.0 kev counts : this number is insensitive in our fits to the exact values of @xmath246 and @xmath255 used . converting to a luminosity using the temperature fitted above , this would correspond to an undisturbed group bolometric x - ray luminosity of @xmath256 erg s@xmath1 ( note that this only includes statistical errors , and takes no account of systematics in our assumptions ) which would place the system close to the temperature - luminosity relationship for groups in general ( e.g. , * ? ? ? * ) , given the relatively large uncertainty in temperature . we therefore feel justified in assuming that this faint extended component is the group environment in which 3c285 s host currently resides . we comment later in the paper ( [ impact ] ) on the effect the current radio source is likely to have on this environment . the temperature determined above for the group - scale gas is clearly cooler than that of the ridge ( within the joint 90% confidence limits ) for our choices of a fixed abundance of 0.35 solar , which might suggest that the ridge and group - scale gas have a physically different origin . however , if the two components are fitted jointly , each with an independent normalization and temperature , lower abundances are preferred ( @xmath257 at 90% confidence ) , and for abundances fixed at a low value the best - fitting temperature of the ridge emission comes down to @xmath258 kev , consistent with that of the group gas within the errors : accordingly , acceptable , though poorer , fits are possible to the joint dataset with a single gas temperature around @xmath259 kev with these low abundances . although we adopt the temperatures determined from our original fits in subsequent analysis , we can not draw any strong conclusions from the apparent difference in temperature between the two components . 3c442a shows striking x - ray structure related to the host interacting galaxy pair ngc 7236/7237 ( fig . [ 442-centre ] ) . both galaxies appear to have ` tails ' of extended , asymmetrical x - ray - emitting material , presumably radiating thermally ( see below ) , with size scales around @xmath260 ( 20 kpc ) . the tail of ngc 7237 lies in position angle @xmath261 ( n through e ) and that of ngc 7236 in position angle @xmath262 . the tail of ngc 7237 was seen with the _ rosat _ hri by @xcite . as there is no evidence for corresponding emission in the optical image ( fig . [ 442-centre ] ) , where the distorted isophotes of the two galaxies are on smaller scales @xcite , the tails must be formed not by tidal processes but by some hydrodynamical process such as ram pressure stripping ( e.g. , * ? ? ? * ) or bondi - hoyle wake formation @xcite . the x - ray tails are in the direction expected from the compression of the outer optical isophotes , which suggests that ngc 7237 is currently moving towards the ne and ngc 7236 towards the sw @xcite . equally striking is the relationship between these tails and the filamentary structure in the radio , as seen in the right - hand panel of fig . [ 442-centre ] . the brightest filament in the center of 3c442a passes directly through the gap between the two tails ( crossing the nucleus of ngc 7236 ) , while the other bright filament in the system runs along the s boundary of the ngc 7237 tail . given this , together with the qualitative similarity between the shape of the ngc 7237 tail ( e - w at the w end , curving to ne - sw at the n end ) and those of the two filaments , it seems almost certain that the filamentary structure arises from interaction between the radio plasma and the tails . we return to this point below ( [ tails ] ) . we extracted x - ray spectra for the two tails using rectangular regions with adjacent background ( excluding emission from the nuclei of both galaxies ) , and fitted them with mekal models with free abundance . this gives a good fit to the ngc 7237 tail ( @xmath263 for 56 d.o.f . , @xmath264 kev , abundance @xmath265 solar ) and an acceptable though rather poor fit to the ngc 7236 tail ( @xmath266 for 44 d.o.f . , @xmath267 kev , abundance @xmath268 ) . the bolometric x - ray luminosity of the tail regions with these fits is @xmath269 erg s@xmath1 and @xmath270 erg s@xmath1 respectively . on larger scales , the tails merge into an elongated region that is very similar in appearance to the ` ridge ' of 3c285 , confirming the result of @xcite . to illustrate this we masked out all point sources ( including the central nuclear point sources ) and produced an exposure - corrected image from the combined dataset , extending to the edges of the combined field of view of the observations . smoothing this with a large gaussian shows the large - scale structure in the source ( fig . [ 442-large ] ) . the ` ridge ' in 3c442a extends perpendicular to the radio source axis , as in 3c285 , for 4.5 arcmin ( 150 kpc ) , again ( as in 3c285 ) terminating near a bright galaxy that is probably a group member . excess x - ray emission extends to the n and s of the sw lobe and there is a clear deficit of emission coincident with the lobe itself ( and hints of a deficit associated with the ne lobe as well ) . we extracted a spectrum for the ridge using a @xmath271 rectangle centered on ngc 7237 , excluding only the nuclear point sources of the two galaxies . this spectrum clearly contains much of the emission of the two x - ray tails discussed above . although the spectrum is roughly consistent with the type of model that might have been expected from the results of fitting to the tails , the data are not very well fitted with a single - temperature model ( @xmath272 for 146 degrees of freedom : @xmath273 kev , abundance @xmath274 solar ) and the fit is improved if a two - temperature model is used ( @xmath275 for 145 d.o.f . , @xmath276 kev , @xmath277 kev : here we fix abundance to 0.3 solar , as determined in the fits above , since it is degenerate with relative normalization of the mekal models ) . this may suggest that the ` ridge ' is a mixture of cool gas from the tails and hot gas from another source . the bolometric x - ray luminosity of the ridge in these models is @xmath278 erg s@xmath1 . azimuthally symmetrical emission on scales of @xmath279 is also clearly visible in the image of fig . [ 442-large ] . as with 3c285 , this seems likely to be thermal emission from a typical group - scale atmosphere . we extracted spectra from a @xmath280 radius source circle around the source , with background taken from an adjacent annulus , using the three of the four observations ( obsids 5635 , 6353 and 6392 ) with enough of a distance between the source and the edge of the chips to allow this . point sources and the rectangular region around the ` ridge ' were excluded , but the lobes were not , as there is no direct evidence that their x - ray emission is significant . counts from the background dominate the spectrum ( there were @xmath281 net counts but @xmath282 total counts in the extraction region ) and the spectra were heavily binned before fitting to obtain adequate signal to noise . fitting with a mekal model we obtain a reasonable fit ( @xmath283 for 39 d.o.f . ) but with a poorly constrained temperature , @xmath284 kev . abundance is not constrained in these fits and we set it to 0.3 solar , matching our results for the ridge . the temperature we determine here is appropriate for a rich group or poor cluster ( it is consistent , for example , with the virgo cluster at @xmath285 kev ) . we extracted a radial profile for the large - scale emission , again using the three obsids for which annuli out to @xmath61 could be fitted on the acis - i chips and correcting for vignetting using exposure maps . all structure on scales less than @xmath286 , as well as larger - scale emission from the ` ridge ' and point sources , were excluded from the profile . we fitted @xmath246 models to the profile , and although the restricted angular range of the profile meant that it was not possible to constrain @xmath246 ( @xmath246 values @xmath247 are preferred ) we obtained a good fit ( @xmath287 for 27 d.o.f . ) with @xmath251 , @xmath288 ( @xmath289 kpc ; error for fixed @xmath246 value only ) . fig . [ 442-profile ] shows the profile and best - fitting @xmath246 model . our fits allow us to estimate the 0.5 - 5.0 kev count rate in the three observations used , extrapolating the @xmath246 model , as @xmath290 s@xmath1 , which , assuming the temperature fits above , implies a bolometric x - ray luminosity of @xmath291 erg s@xmath1 ( again , errors quoted here come from the uncertainty in the radial profile alone and do not include systematic effects ) . given the large uncertainty on derived temperature , this is consistent with the temperature - luminosity relation for the high-@xmath292 end of the group population ( or low-@xmath292 end of the cluster population ) particularly if the true temperature is closer to the lower end of its 90 per cent confidence range . a lower temperature would be more probable given the velocity dispersion for the group calculated in [ 442-vd ] , but we note that groups around radio galaxies of similar luminosity to 3c442a often have temperatures that exceed those predicted from the temperature - luminosity relation @xcite . in this section we first consider ( [ dynamics ] ) the implications of these observations for the dynamics of the radio sources . we then go on to consider the origin of the asymmetrical ` ridge ' structures in the two systems , and show ( [ origin ] ) that it is unlikely that the radio galaxy can have given rise to these features , implying that the environment shapes the radio source , rather than vice versa . this allows us to construct simple pictures of the history of both systems in terms of the interactions of gas and radio sources . the impact of the radio sources on their group environments may be substantial , and we comment on this in ( [ impact ] ) . the process of ridge formation and interaction may be important to other sources with central gaps in the radio lobes , and we discuss these in [ gaps ] . finally , in [ tails ] , we discuss the gas being stripped from the merging galaxies hosting 3c442a and its interaction with the filamentary radio structure at the center of the source . to discuss the effects of the x - ray emitting material on the radio source it is first necessary to estimate the pressure in the radio lobes . we assume initially that only the radiating electrons / positrons and magnetic field contribute to the internal pressure : it follows ( since the sound speed in a relativistic plasma is high ) that it is appropriate to quote a single value for the internal pressure . minimum pressures can then be derived in the standard way for the lobes of both sources given their sizes and flux densities . we use an electron spectrum with @xmath293 and @xmath294 , where @xmath295 is the lorentz factor of the electrons ( @xmath296 ) : the results are insensitive to the precise choices of the low- and high - energy cutoffs . the electron spectrum has a broken powerlaw form , such that @xmath297 for @xmath298 and @xmath299 for @xmath300 . @xmath301 , the low - energy electron energy power - law index ( ` injection index ' ) is taken to be 2.0 and the value of @xmath302 is determined from a fit to the observed radio lobe spectrum . the filling factor is set to unity ( as standard for minimum - energy calculations ) and the magnetic field is assumed to be tangled on small scales . for 3c285 , the possible detection of inverse - compton emission via scattering of cmbr photons from the n lobe gives us a _ measurement _ of the pressure , again assuming that only radiating particles and field are present , since it gives a direct constraint on the number density of electrons , which in turn fixes the magnetic field strength given the observed synchrotron emissivity . for 3c442a , the non - detection of inverse - compton emission gives us limits on the extent to which the lobe pressures can be dominated by electrons . we determine pressures as a function of radius in the groups for the detected large - scale thermal components using the method of @xcite , assuming spherical symmetry . for the ` ridge ' components the appropriate assumptions are less obvious . we derive approximate pressures on the assumption of uniform particle density , converting from the xspec mekal normalization to density and thus pressure in the standard way . we use the two measured dimensions of the spectral extraction region for the ridge to estimate its size in the plane of the sky . the extension along the line of sight is uncertain , and we consider two cases , treating the ridge either as a cylinder whose long axis is in the plane of the sky , or as a disk viewed edge - on . because the density depends only on the square root of the volume , these two geometries do not give rise to markedly different densities and pressures . the minimum pressures in the ne and sw lobes of 3c285 are @xmath303 pa and @xmath304 pa respectively ( the difference arises because of the fainter flux density and larger volume of the sw lobe ) . if we assume that all the x - ray flux from the ne lobe is inverse - compton in origin , its pressure is @xmath305 pa , where the error comes from the uncertainty on the measured x - ray flux density . this is an upper limit on the pressure if the lobes are electron - dominated . for the group - scale emission , the extrapolated pressure at the base of the lobes , @xmath306 ( 15 kpc ) in projection from the nucleus , using the best - fitting @xmath246 model , is @xmath307 pa ( the dominant error here is the uncertainty on the best - fitting x - ray temperature : the uncertainties on density from the @xmath246 model fits are far smaller ) . at the far end of the lobes , @xmath308 ( 140 kpc ) from the core , the thermal pressure is @xmath309 pa . formal errors due to the arbitrarily assumed abundance of 0.35 solar can not be estimated from the data , but pressure would increase by a factor 1.5 if the abundance were 0.1 solar and decrease by a factor 1.5 if it were 1.0 solar , giving a range which is similar in magnitude to that derived from the uncertainties in temperature . even with the large uncertainties on both internal and external pressure , it is clear that 3c285 resembles other low - power frii sources in poor environments ( e.g. , * ? ? ? * ) as well as more powerful objects ( e.g. , * ? ? ? * ) in having lobes that are close to being in pressure balance with the external thermal environment . it seems likely that at their far ends the lobes are continuing to expand at subsonic or mildly supersonic speeds ( even the maximum possible external thermal pressure only just balances the internal minimum pressure of the ne lobe here ) , while at the ends closest to the nucleus they can no longer expand transversely . the pressure in the central ` ridge ' , assuming our best - fitting temperature and a uniform distribution of x - ray emitting gas , is @xmath310 pa ( cylindrical geometry ) or @xmath311 pa ( disk geometry ) , where the errors are again dominated by the errors on our spectrally determined temperature ( we use the best - fitting single - temperature model for simplicity ) and again extreme variations in abundance could cause pressure variations by a factor 1.5 in either direction . this ridge pressure is in some sense a mean pressure in the region if there is a density gradient , as the surface brightness suggests , then the pressure will be higher closer to the center and lower at the edges . thus , the fact that the ` ridge ' pressure is comparable to or higher than the lobe pressure means that if our radio pressure calculations are correct ( i.e. the lobes are not dominated in pressure terms by some non - radiating component such as protons ) then we would predict that the lobes must avoid the region of the ridge . this is consistent with the observation ( fig . [ 285-smooth ] ) that the lobe plasma largely avoids the ridge ; in fact , the extensions of the lobes at the base to the s ( ` wings ' ) are on scales that suggest that the lobe plasma may be flowing around the ridge . the anticoincidence between the lobe material and the ridge is not perfect , but this is likely to be due to projection , since the jet - counterjet asymmetry in 3c285 strongly suggests that the radio source is not precisely in the plane of the sky . for 3c442a we measure minimum pressures ( using the assumptions discussed above ) from the outer regions of the lobe only , ignoring the central region where an assumption of uniform plasma filling factor would clearly not be appropriate , so that the pressures measured here are for the parts of the source that lie outside the ridge region on fig . [ 442-large ] . the minimum pressures we obtain are @xmath312 and @xmath313 pa for the ne and sw lobes respectively . these are lower than the 3c285 pressures , since 3c442a has both lower radio luminosity and a slightly larger physical size . the external thermal pressures are much larger than the minimum pressures at all points in the radio source . at @xmath314 ( 25 kpc ) , the inner edge of the region measured for the ne lobe , the extrapolated pressure ( our @xmath246-model fit does not extend to this radius ) is @xmath315 pa ( where the large ( 90% confidence ) errors we quote come entirely from the poorly constrained temperature : the uncertainties from @xmath246-model fitting are negligible by comparison , though there is a possibility of systematic errors due to the extrapolation , and we assume that large deviations from the abundance fitted to the ridge are not likely ) . at the edges of the lobes at @xmath280 @xmath316 ( 150180 kpc ) the pressures are still @xmath317 pa , at least an order of magnitude above the radio minimum pressures . thus , although 3c442a s radio structure is not that of a classical fri radio source , it obeys a tendency that has been known for many years ( e.g. , @xcite ; @xcite ) and rediscovered recently in the context of cluster - center sources ( e.g. , * ? ? ? * ) for low - power radio sources to have minimum pressures that lie substantially ( often 12 orders of magnitude ) below the external thermal pressure . radio lobes can not evolve to a state in which they are substantially underpressured with respect to the external medium : both theory ( e.g. , * ? ? ? * ) and observation @xcite imply that they are likely to be strongly overpressured initially , and , while the internal pressure should drop as the linear size of the source increases , any drop below the external pressure will cause the external thermal material to expand into the lobe region , compressing the lobe plasma until pressure balance is restored . observations that suggest that the minimum pressure is below the external pressure therefore imply some additional contribution to the internal pressure , via a large departure from equipartition , a low filling factor , and/or a dominant contribution from non - radiating particles such as protons . a departure from equipartition in the direction of electron dominance is almost certainly ruled out , since at pressure balance it would produce a 1-kev flux density via inverse - compton scattering of the cmbr of about 90 njy per lobe , giving a count rate in _ chandra _ observations that would exceed that from the thermal emission from the entire group . in fact , if anything there is a deficit of x - ray emission from the lobes , particularly the w lobe , as mentioned above . as magnetic domination by large factors is not observed in lobes of sources where inverse - compton emission can be studied @xcite , we favor a model in which the lobe pressure is dominated by non - radiating particles . in the case of more typical fri sources it has been suggested that protons from the external thermal material could be entrained and efficiently heated to high energies by the jets ( temperatures much greater than that of the external thermal material are required to avoid the detection of this component through its thermal bremsstrahlung ) . as no jets of the sort seen in normal fris are present in 3c442a , this explanation would only be viable if the lobes were at one time driven by fri - type jets that are now no longer present , or no longer visible . alternatively , a population of relativistic protons could have been injected into the lobes along with the radiating electrons ( though this is inconsistent with the picture in frii lobes where an energetically dominant proton population is disfavored : @xcite ) or the lobes could have engulfed large regions of thermal material , though again some unknown efficient heating mechanism would be required . the pressure in the ` ridge ' in 3c442a , again assuming uniform density , is @xmath318 pa ( cylindrical geometry ) or @xmath319 pa ( disk geometry ) , so , unsurprisingly , the ridge pressure is equal to or exceeds the extrapolated central pressure of the extended emission , in spite of the lower temperature of the ridge material ( here again we are using the best - fitting single - temperature model for the ridge ) . the pressures derived by @xcite , using a @xmath246-model fit to the ridge region , are similar . thus even if the lobe is in pressure balance with the thermal material further out ( as it must be , to avoid collapse on a sound - crossing time ) the lobe pressure will be lower than the ridge pressure , and the lobe material will be excluded from the ridge region unless it can achieve higher pressures . [ 442-large ] shows that material in the lobes , particularly the w lobe , does seem to be avoiding the ridge and to be being pushed out in wings to the n and s , as well as possibly along the line of sight ( note the strong radio brightness gradient to the w of the ridge ) . there is still radio - emitting material in the ridge region , but , as discussed above , the brightest components are in the form of filaments that are likely to be interacting strongly with the small - scale material from the galaxies , and these filaments will have pressures substantially higher than the lobe material as a whole ( see below , [ tails ] , for more discussion of this point ) . [ ridge - pressure ] is the ` ridge ' structure intrinsic to the system , or could it be a result of the interaction of the radio galaxy with an initially spherically symmetrical gas distribution ? there are two strong indications that the former is the case : firstly , the ridges align well with the local distribution of galaxies and starlight , which can not be affected by the radio galaxy , in both 3c285 and 3c442a ; secondly , in 3c442a , we observe material leaving the two central bright galaxies via the tails and merging with the ridge . quantitatively , we can compare the mass in the ridges ( or rather in the somewhat smaller spectral extraction regions used to derive temperatures and pressures ) with the masses of thermal material affected by the lobes . using the approximations of uniform density from the previous section , we find that the ` ridge ' gas masses are in the range @xmath320 to @xmath321 for 3c285 and @xmath322 to @xmath323 for 3c442a . using the @xmath246-model fits to the larger - scale thermal emission , we find that for 3c285 the ridge mass is similar to , but somewhat greater than , the total mass of gas that would be present in the lobe regions if the lobes were not there : thus , since only a fraction of the displaced gas would be expected to end up in the ridges , it is essentially impossible for the radio lobes to be responsible for producing the ridge in 3c285 ( though of course here we have to assume that the distribution of the gas in the regions away from the lobes of the radio source now is similar to its undisturbed distribution ) . the constraints are less strong for 3c442a , with larger lobes , a denser thermal environment , and a smaller ridge mass : roughly @xmath324 of the displaced gas would be sufficient to produce the observed ridge . however , we know that in 3c442a at least some of the gas in the ridge did _ not _ come via this route , since some of it must have passed through the tails . _ some _ contribution from compressed , displaced gas would help to explain the overdensities seen to the n and s of the w lobe ( [ tail - fit ] and fig . [ 442-large ] ) ( although note that these are not included in our mass calculations ) . we conclude that the ridge structures in these two sources are largely or wholly independent of the presence of the radio lobes . our outline models of the evolution of the two radio sources are somewhat different , because we know that the formation of the ridge in 3c442a is still going on given the observations of tails of gas from the two central galaxies ( see [ tails ] , below ) while there is no evidence that there is ongoing gas stripping in 3c285 . 3c285 thus presents the simpler situation , and we consider it first . in this system we consider it entirely possible that the ridge was present in something like its current configuration before the radio activity began . it is possible , although the data do not require it , that some of the gas that now forms the ridge was stripped from the galaxies that merged to form the current host galaxy of 3c285 . whether this is true or not , we have seen that the ridge is in rough pressure equilibrium with the external group - scale gas , at least in the center of the group . the timescale for the relaxation of the ridge back towards a spherically symmetrical situation , assuming that the gravitational potential is spherically symmetric , is thus likely to be comparable to the sound - crossing time for the _ long _ axis of the ridge , i.e. a few @xmath325 years . if the gravitational potential is non - static , the relevant timescale is the gravitational relaxation time , whch could be longer still . since the expansion of the 3c285 lobes is likely to have been supersonic initially ( and may still just be supersonic at present ) the growth of the lobes to similar size scales can take place in a much shorter time , a few @xmath326 years , and so it seems consistent to suppose that the ridge has not evolved greatly while the lobes have expanded . the expanding lobes would initially be confined by ram pressure and so would expand preferentially into the lower - density group - scale medium . later , as the lobe internal pressure dropped , the ridge material would push any lobe material out of the center of the group , leading to the situation we now see , where the currently observed ridge is in rough pressure balance with lobes that avoid the central part of the system . the more complex situation in 3c442a is discussed in detail by @xcite . in outline , our suggested picture is that here a large , old radio source could well have been present _ before _ the start of the ongoing interaction that produced the presently visible ridge material . by stripping and disrupting the hot gas in the center of the group , the start of the interaction / merger between ngc 7236/7 may have terminated the energy supply to the lobes . in any case , the resulting ridge has now largely driven the lobe material out of the center of the group , although the interaction with radio filaments ( see [ tails ] below ) shows that this process is not yet complete . buoyancy will remove the remaining filaments on a relatively short timescale ( unless they are magnetically anchored to one or other of the galaxies ) , while the ridge may continue to expand , driving the radio lobes ahead of it , as the merger of the two galaxies proceeds to completion . crudely , the work done on the thermal material by the expanding radio galaxy ( @xmath327 ) may be estimated as @xmath328 , or @xmath329 for pressure balance . this estimate probably gives constraints that are of the right order of magnitude . for 3c285 the numbers are between @xmath322 and a few @xmath330 j , depending on whether the source is in pressure balance or mildly overpressured at the present time , while for 3c442a the energy input assuming rough pressure balance at the radio lobes is @xmath331 j. in both cases the energy stored in the lobes is 3 times this value ( we assume a relativistic equation of state ) . these numbers may be compared to the total energy ( @xmath332 ) of the thermal material in the groups ( neglecting the ridges ) integrated out to the radius of the edge of the lobes ( 150 kpc for 3c285 , 160 kpc for 3c442a ) . for 3c285 this thermal energy is @xmath333 j : thus the energy input by the lobes is certainly significant , and the total energy input by the source over its lifetime will be sufficient to unbind or move to large radii a large fraction of the currently existing group - scale gas . for 3c442a , with a denser medium and larger temperature , the thermal energy of the gas is @xmath334 j , so that the impact of the radio source is less relative to the existing thermal energy than in 3c285 : however , again , the total energy input , once the internal energy of the lobes is thermalized , should be very important . spectral aging ( via inverse - compton losses ) puts a limit on the lifetime of 3c442a of @xmath335 years , while in 3c285 we expect a lifetime of a few @xmath326 years from the dynamical arguments of the previous section . thus in both cases the rate of energy input is more than enough ( by 1 - 2 orders of magnitude ) to offset the radiative cooling of the group - scale gas given the bolometric luminosities determined in [ group - properties ] , although in the case of 3c442a some of that work probably comes not from the active nucleus but ( via the expansion of the ridge ) from the dissipation of the energy of the galaxy - galaxy merger . a significant minority of radio galaxies , particularly low - luminosity friis , have radio lobes that are apparently sharply truncated at an inner edge , as first pointed out by @xcite . the truncation means that in radio images there are well - defined sharp - edged gaps between pairs of radio lobes that typically contain no lobe - related radio emission . @xcite have compiled a sample of such objects , and suggest that the central gaps between the lobes are a result of hydrodynamical interactions with a _ cold _ thermal medium . while this may be true in some cases , our results show that asymmetrical structures in the hot phase of the igm can have the same effect . a few other radio galaxies with truncated lobes have x - ray observations , and while some ( e.g. , 3c33 , @xcite ; 3c227 , hardcastle et al.in prep . ) show no obvious signs of asymmetrical x - ray structure on scales comparable to the lobes , at least one other frii radio galaxy , 3c401 , has both elongated x - ray emission filling between the gap between the lobes @xcite , and , crucially , some evidence for elongation of the stellar population in the same e - w direction in _ hubble space telescope _ images @xcite , strongly suggesting that the ` ridge ' structure can not simply be a result of compression of pre - existing symmetrical gas in the cluster environment of the source . given the evidence that mergers are important in the environments of frii sources ( e.g. , * ? ? ? * ) it is possible that central gaps are related to interaction of the lobes with pre - existing or developing x - ray structures in many other cases . ` tail ' or ` fan'-like x - ray structures have been seen in other strongly interacting galaxy pairs . the twin tails in ngc 7236/7 ( fig.[442-centre ] ) are very similar in structure and physical size to those seen by @xcite in ngc 4782/3 , which consists of two interacting galaxies of similar mass ( one hosting the radio source 3c278 ) with a projected separation of around 13 kpc . the tails in the 3c278 system are overpressured with respect to the surrounding diffuse medium , implying that they originate in ram pressure stripping of the halos of the interacting ellipticals , and @xcite carried out a detailed analysis of the x - ray surface brightness around the leading edge of one of the galaxy profiles to estimate the approximate speed of motion of the galaxy relative to the igm . because both tails in 3c442a have some excess x - ray emission in front of the galaxies , we can not repeat the @xcite analysis for 3c442a , but we can make a less rigorous estimate by determining the pressure inside the tails . the tail of ngc 7237 has only a weak surface brightness gradient over the inner 40@xmath6 ( 20 kpc ) of its length , so we model the x - ray emission as a uniform cylinder of length 20 kpc and radius 7 kpc . our spectral fit to the tail ( [ tail - fit ] ) then gives a density of @xmath336 m@xmath337 and a pressure of @xmath338 pa ( where the dominant uncertainty comes from the small uncertainty on temperature ) . to explain why the two dominant galaxies of the group are being ram - pressure stripped _ now _ , rather than having being completely stripped at a much earlier stage of the group s evolution , we must assume that it is their current close interaction that is responsible for the stripping : in this case it is appropriate to compare their pressure and density to the external medium in the ridge , which ( as discussed above ) we infer to be made up primarily of previously stripped material , rather than the group - center hot gas : in any case , the tails are clearly embedded in the ridge . this comparison gives a density contrast of a factor 3.34.5 ( depending on the geometry assumed for the ridge , see [ ridge - pressure ] ) and a pressure contrast of a factor @xmath339 , assuming cylindrical geometry since that gives the closest pressure balance with the larger - scale medium . the tails in 3c442a are thus overpressured and the leading edge must be ram - pressure confined , consistent with the picture of @xcite following @xcite in adopting the relation of @xcite between pressure contrast and mach number @xmath340 , we find that this pressure contrast implies @xmath341 , or a speed with respect to the ridge material of @xmath342 km s@xmath1 . the models of merger dynamics by @xcite imply velocities @xmath343 km s@xmath1 , in good agreement with this result . away from the leading edge , we expect the tails to expand transversely into the medium of the ridge at the sound speed , giving rise to the short , stubby structures that are observed and possibly to the apparent flaring of the ngc 7236 tail ( fig.[442-centre ] ) if the filaments seen in the radio image are directly interacting with the tails , we expect that the pressure in the radio - emitting plasma will be comparable to the pressures in the tail . the radio filaments are marginally resolved in our highest - resolution radio images , with radii of 12 arcsec . we can use observed radio surface brightness per unit length to estimate that they would need to have a radius of @xmath344 ( 0.09 kpc ) for the _ minimum _ pressures to be comparable to the thermal pressure in the tails : this is probably inconsistent with the radio data for the southern filament , and almost certainly inconsistent with the data for the northern filament . however , if the pressure in the filaments exceeds the minimum pressure by the factor @xmath345 required in the lobes , then the radius must be @xmath346 , 0.8 kpc , for pressure balance with the brighter tail , which is similar to what we observe in the high - resolution radio images . thus it is plausible that the filaments are produced by direct interaction with the tails , as suggested by their positions ( fig.[442-centre ] ) . assuming that there are not strong variations in the ratio of magnetic field to electron energy density , the filaments must be overpressured with respect to the lobes in order to be visible at all , and so must be dynamically evolving structures . a model in which they are associated with the galaxies , which we know to be moving with respect to the radio plasma , is therefore _ a priori _ plausible . since the filaments lie to the sides of the x - ray tails ( fig . [ 442-centre ] ) they can not simply be a result of compression of the remaining central radio - emitting plasma by the moving galaxies . it is plausible that they arise in the shear layers between the stripped material in the tails and the external gas . we have presented _ chandra _ observations of the nearby radio galaxies 3c285 and 3c442a . the main results of the paper can be summarized as follows . * the large - scale environments of both sources are groups , consistent with earlier optical results . the groups have similar x - ray luminosities , physical sizes and @xmath246-model parameters , although the best - fitting temperature for 3c442a s group is hotter . both sources have elongated , 150-kpc scale excess x - ray emission which aligns with galaxies in the group , and which we refer to as the ridge . this excess x - ray emission is coincident with the regions in which there is little or no radio emission from the lobes . * using temperatures and pressures derived from the x - ray observations , we have shown that the pressure in the ridges in both sources , within the errors , is equal to or exceeds any plausible internal pressure in the lobes . thus the ridges are capable of driving , or having driven , the existing radio lobes out of the central regions of the source . * we have shown that there is direct quantitative evidence in the case of 3c285 ( [ origin ] ) , and strong qualitative evidence in the case of 3c442a , that the ridges are not the result of an interaction between pre - existing gas and the radio source , but must have arisen independently . in 3c442a , we see hot gas in the process of being stripped from the central interacting galaxies and merging into the larger - scale ridge , suggesting that the ridge has been formed by the stripping of the haloes of the two merging galaxies . in 3c285 , the optical observations strongly suggest that the host galaxy has recently merged with a gas - rich system , and there is a tidal interaction with a more distant galaxy , although there is no evidence that the ridge is currently being fed with hot gas ; it is thus possible ( though by no means required by the data ) that the ridge in 3c285 also has an origin in galaxy - galaxy interactions . we argue that interaction with similar elongated structures may be responsible for some of the observed central gaps in the lobes of other radio sources . a prediction of this picture is that there will be ridge structures in other groups that do _ not _ host large , powerful radio galaxies : so far few of these are known , but the group hosting ngc 5171 @xcite may be a good example . * our detailed view of the tails of hot gas associated with 3c442a allows us to estimate the mach number and speed of ngc 7237 ( the host of 3c442a ) with respect to the diffuse igm . we obtain results consistent with detailed modelling of the merger dynamics @xcite . the properties of the radio filaments at the center of 3c442a are consistent with their being formed by interaction with the tails of the two galaxies in the system . * the x - ray nuclei of the two sources have spectra that are consistent with what is expected for objects of their emission - line type . no convincing evidence for x - ray counterparts to compact radio features , such as jets or hotspots , is found , and we have shown that the previously claimed optical counterpart to the e hotspot of 3c285 is unlikely to be associated with the hotspot . x - ray emission is detected from the region of jet - induced star formation in 3c285 reported by @xcite . we gratefully acknowledge financial support for this work from the royal society ( research fellowship for mjh ) and nasa ( grants go5 - 6100x and nas8 - 03060 ) . the national radio astronomy observatory is a facility of the national science foundation operated under cooperative agreement by associated universities , inc . this research has made use of the nasa / ipac extragalactic database ( ned ) which is operated by the jet propulsion laboratory , california institute of technology , under contract with the national aeronautics and space administration . we thank an anonymous referee for helpful comments which enabled us to improve the presentation of the paper . acreman , d.m . , stevens , i.r . , ponman , t.j . , & sakelliou , i. , 2003 , , 341 , 1333 allen , m.g . , et al . , 2002 , , 139 , 411 baum , s.a . , heckman , t. , bridle , a. , van breugel , w. , & miley , g. , 1988 , , 68 , 643 birkinshaw , m. , laing , r.a . , & peacock , j.a . , 1981 , , 197 , 253 birkinshaw , m. , & worrall , d.m . , 1993 , , 412 , 568 brzan , l. , rafferty , d.a . , mcnamara , b.r . , wise , m.w . , & nulsen , p.e.j . , 2004 , , 607 , 800 bhringer , h. , voges , w. , fabian , a.c . , edge , a.c . , & neumann , d.m . , 1993 , , 264 , l25 borne , k.d . , 1988 , , 330 , 61 borne , k.d . , & hoessel , j.g . , 1988 , , 330 , 51 chiaberge , m. , et al . , 2005 , , 629 , 100 comins , n.f . , & owen , f.n . , 1991 , , 382 , 108 croston , j.h . , hardcastle , m.j . , birkinshaw , m. , & worrall , d.m . , 2003 , , 346 , 1041 croston , j.h . , birkinshaw , m. , hardcastle , m.j . , & worrall , d.m . , 2004 , , 353 , 879 croston , j.h . , hardcastle , m.j . , & birkinshaw , m. , 2005a , , 357 , 279 croston , j.h . , hardcastle , m.j . , harris , d.e . , belsole , e. , birkinshaw , m. , & worrall , d.m . , 2005b , , 626 , 733 danese , l. , de zotti , g. , & di tullio g. , 1980 , , 82 , 322 dunn , r.j.h . , fabian , a.c . , & taylor , g.b . , 2005 , , 364 , 1343 evans , d.a . , worrall , d.m . , hardcastle , m.j . , kraft , r.p . , & birkinshaw , m. , 2006 , , 642 , 96 fabbiano , g. , miller , l. , trinchieri , g. , longair , m. , & elvis , m. , 1984 , , 277 , 115 fanaroff , b.l . , & riley , j.m . , 1974 , , 167 , 31p feretti , l. , spazzoli , o. , gioia , i.m . , giovannini , g. , & gregorini , l. , 1990 , , 233 , 325 feretti , l. , perola , g.c . , & fanti , r. , 1992 , , 265 , 9 gehrels , n. , 1986 , , 303 , 336 gopal - krishna , wiita , p. , & 2000 , , 529 , 189 greenstein , j.l . , 1962 , , 135 , 679 hardcastle , m.j . , birkinshaw , m. , & worrall , d.m . , 1998 , , 294 , 615 hardcastle , m.j . , birkinshaw , m. , cameron , r. , harris , d.e . , looney , l.w . , & worrall , d.m . , 2002 , , 581 , 948 hardcastle , m.j . , evans , d.a . , & croston , j.h . , 2006 , , 370 , 1893 hardcastle , m.j . , harris , d.e . , worrall , d.m . , & birkinshaw , m. , 2004 , , 612 , 729 hardcastle , m.j . , sakelliou , i. , & worrall , d.m . , 2005 , , 359 , 1007 hardcastle , m.j . , worrall , d.m . , & birkinshaw , m. , 1998 , , 296 , 1098 hardcastle , m.j . , & worrall , d.m . , 1999 , , 309 , 969 hardcastle , m.j . , & worrall , d.m . , 2000 , , 319 , 562 heckman , t.m . , smith , e.p . , baum , s.a . , van breugel , w.j.m . , miley , g.k . , illingworth , g.d . , bothun , g.d . , & balick , b. , 1986 , , 311 , 526 helsdon , s.f . , & ponman , t.j . , 2000 , , 315 , 356 jenkins , c.j . , pooley , g.g . , & riley , j.m . , 1977 , , 84 , 61 jenkins , c.j . , & scheuer , p.a.g . , 1976 , , 174 , 327 kaiser , c.r . , & alexander , p. , 1997 , , 286 , 215 killeen , n.e.b . , bicknell , g.v . , & ekers , r.d . , 1988 , , 325 , 180 kraft , r.p . , vzquez , s. , forman , w.r . , jones , c. , murray , s.s . , hardcastle , m.j . , worrall , d.m . , & churazov , e. , 2003 , , 592 , 129 kraft , r.p . , hardcastle , m.j . , worrall , d.m . , & murray , s.s . , 2005 , , 622 , 149 kraft , r.p . , jones , c. , nulsen , p.e.j . , & hardcastle , m.j . , 2006 , , 640 , 762 kraft , r.p . , birkinshaw , m. , hardcastle , m.j . , evans , d.a . , croston , j.h . , worrall , d.m . , & murray , s.s . , 2007 , in press ( astro - ph/0701458 ) laing , r.a . , 1993 , in burgarella d. , livio m. , odea c.p . , eds , astrophysical jets , space telescope science institute symposium series 6 , cambridge university press , cambridge , p. 95 laing , r.a . , riley , j.m . , & longair , m.s . , 1983 , , 204 , 151 machacek , m. , kraft , r.p . , jones , c. , forman , w.r . , & hardcastle , m.j . , 2006 , submitted , astro - ph/0610904 mackay , c. , 1969 , , 145 , 31 martel , a.r . , et al . , 1999 , , 122 , 81 morganti , r. , fanti , r. , gioia , i.m . , harris , d.e . , parma , p. , & de ruiter h. , 1988 , , 189 , 11 osullivan , e. , forbes , d.a . , & ponman , t.j . , 2001 , , 328 , 461 osmond , j.p.f . , & ponman , t.j . , 2004 , , 350 , 1511 osmond , j.p.f . , ponman , t.j . , & finoguenov , a. , 2004 , , 355 , 11 ott , j. , walter , f. , & brinks , e. , 2005 , , 358 , 1453 reynolds , c.s . , brenneman , l.w . , & stocke , j.t . , 2005 , , 357 , 381 roche , n. , & eales , s.a . , 2000 , , 317 , 120 sakelliou , i. , 2000 , , 318 , 1164 sandage , a.r . , 1967 , , 150 , l145 sandage , a.r . , 1972 , , 178 , 25 saslaw , w.c . , tyson , j.a . , & crane , p. , 1978 , , 222 , 435 stark , gammie , c.f . , wilson , r.w . , bally , j. , linke , r.a . , heiles , c. , & hurwitz , m. , 1992 , , 79 , 77 taylor , g.b . , perley , r.a . , inoue , m. , kato , t. , tabara , h. , & aizu , k. , 1990 , , 360 , 41 van breugel w.j.m . , & dey , a. , 1993 , , 414 , 563 vron , p. , 1977 , , 30 , 131 vikhlinin , a. , markevitch , m. , & murray , s.s . , 2001 , , 551 , 160 worrall , d.m . , & birkinshaw , m. , 2000 , , 530 , 719 worrall , d.m . , birkinshaw , m. , & cameron , r.a . , 1995 , , 449 , 93 worrall , d.m . , birkinshaw , m. , & hardcastle , m.j . , 2001 , , 326 , l7 worrall , d.m . , kraft , r.p . , birkinshaw , m. , & hardcastle , m.j . , 2007 , l , 658 , in press ( astro - ph/0702411 )

we consider the problem of european and american option pricing and refer to @xcite and the references therein for an introduction into computational methods for option pricing . while european options can be modeled by a parabolic partial differential equation , american options result in additional inequality constraints . different models can be used to price european and american options . the simplest ones , e.g. , the black - scholes model @xcite , assume that the volatility is constant . however , in most of the cases , the real market violates this assumption due to its stochastic nature . thus , alternative models , which try to capture this phenomenon are frequently used , e.g. , the heston stochastic volatility model @xcite . another difficulty which arises with solving partial differential equations ( pdes ) for option pricing , in particular for pricing american options , is that for most of the models no closed form solution exists . thus one has to develop appropriate numerical methods . the common methods to solve pricing equations with the heston model are finite differences , cf . @xcite and finite elements , cf . we refer to @xcite for a possible numerical treatment of basket options with the black - scholes model by primal - dual finite elements and to @xcite for an abstract framework on the theory of constrained variational problems . we are interested in providing fast numerical algorithms to accurately solve the variational equality and inequality systems associated with european and american call and put options for a large variety of different parameter values such as , e.g. , interest rate , dividend and correlation . reduced basis ( rb ) methods are an appropriate means for standard parametrized parabolic partial differential equations , cf . @xcite and the references therein . these techniques are based on low - dimensional approximation spaces , that are constructed by greedy procedures . convergence behavior is known in some cases @xcite . the computational advantage of rb - methods over standard discretization methods is obtained by its possible offline / online decomposition : first , a typically expensive offline - phase involving the computation of the reduced spaces is performed . this phase only needs to be precomputed once . then , the online phase allows an extremely fast computation of the rb solutions for many new parameters as only low dimensional systems need to be solved . recently , we adopted the rb methodology to constrained stationary elliptic problems @xcite , which we extend here to the instationary case . we refer to the recent contribution @xcite for a tailored rb approach in option pricing with diffusion and jump - diffusion models , which later was generalized to basket options @xcite and , in fact , was shown to be a variant of a proper orthogonal decomposition method ( pod ) @xcite . the application of the reduced basis method can be also extended to the calibration of option pricing models , e.g. , @xcite . in contrast to our setting , no inequality constraints are taken into account there . further work relevant for rb - methods and variational inequalities comprises @xcite , which addresses a time - space formulation of the problem and corresponding analysis . also , recently , ongoing work has been presented in @xcite , which alternatively treats the inequality constraints both by a primal - dual technique and a penalization approach . one main challenge in our problem setting is the construction of a suitable low dimensional approximation of the dual cone required for the treatment of the constraints . in this work , we present an algorithm to overcome this difficulty which is based on the greedy procedure and tries to capture as much `` volume '' as possible in the construction of the dual cone . this is obtained by iteratively selecting snapshots maximizing the angle to the current space . as second main ingredient for treatment of additional inequality constraints , we provide analytical results , in particular a posteriori error control . let us briefly relate the current presentation to our previous works . in contrast to @xcite , which dealt with rb methods for stationary variational inequalities , we treat here instationary problems . in this way , we apply sharper a posteriori error estimator strategies , which can straightforwardly be applied to the stationary case and gives improvements over @xcite . in @xcite we presented an rb procedure for american option pricing with a black - scholes model and gave first simple examples . the current presentation considerably extends this by including the heston model and both european and american options and providing an analysis for the corresponding rb scheme . the content of this paper is structured as follows : in section [ sec : var_form ] , we present both a strong and a variational formulation of pricing of european and american options with the black - scholes and heston model . in section [ sec : rb ] , a reduced basis method is introduced together with the construction of the primal and dual reduced basis spaces . section [ sec : error_analysis ] contains the a posteriori error analysis induced by the method and derived from equality and inequality residuals . the implementational aspects of the reduced basis method together with the construction algorithms for reduced basis spaces are presented in section [ sec : implementation ] . numerical results , given in section [ sec : results ] , illustrate the performance of the method for pricing options in one dimensional black - scholes and two dimensional heston models . in this section , we give a brief introduction to the theory of option pricing . an option is a contract which permits its owner the right to buy or sell an underlying asset ( a stock , or a parcel of shares ) at a prespecified fixed strike price @xmath0 before or at a certain time @xmath1 , called maturity . there are two basic types of an option : a _ call _ option which gives a holder a right to buy and a _ put _ option which allows an owner to sell an underlying asset . also one distinguishes between european options , where exercise is only permitted at maturity @xmath2 and american options which can be exercised at any time before an expiration time @xmath2 . we will denote the price of the underlying asset by @xmath3 , where @xmath4 is the time to maturity @xmath2 . in standard option pricing models , e.g. , the black - scholes model @xcite , a price of the underlying asset @xmath5 follows a stochastic process , governed by the following stochastic differential equation @xmath6 with a wiener process @xmath7 , a drift @xmath8 and a volatility @xmath9 . one of the main limitations of these models is the assumption that the volatility of the return on the underlying asset is constant @xmath10 , while in financial markets , the volatility is not a constant , but a stochastic variable . the heston stochastic volatility model @xcite takes into account the randomness of the volatility and is based on the following stock price and variance dynamics @xmath11 where @xmath12 follows a square root process ( known as a cox - ingersoll - ross ( cir ) process ) with the mean variance @xmath13 , rate of mean reversion @xmath14 and so called volatility of volatility @xmath15 . the wiener processes @xmath16 and @xmath17 are correlated with the correlation parameter @xmath18 $ ] . with the use of ito s formula @xcite , each model can be formulated in terms of a partial differential equation ( pde ) . for purposes of brevity , we omit the derivation of the equations . the reader is referred to @xcite for further details . we define the spatial differential operators corresponding to the black - scholes and heston model as follows @xmath19 where @xmath20 is the interest rate and @xmath21 is the dividend payment . then the value of an option @xmath22 in the black - scholes model and @xmath23 in the heston model paying @xmath24 at maturity time @xmath2 must satisfy the following partial differential equation @xmath25 where the value @xmath24 is called a payoff function and @xmath26 , @xmath27 for the black - scholes and the heston model , respectively . [ remark : bs ] assuming a constant volatility in the heston model and no dividend payment in the black - scholes model , @xmath28 , we have @xmath29 and @xmath30 , and the heston equation reduces to the black - scholes equation with the constant volatility @xmath31 . for further derivation of the weak formulation of the problem ( [ eq : heston1 ] ) , we introduce the following notation . let @xmath32 , @xmath33 , be a bounded open domain with lipschitz continuous boundary @xmath34 . we consider @xmath35 for the black - scholes model and @xmath36 for the heston model . note that this is not a limitation for the rb - approach that we present . higher dimensional problems can readily be treated , as soon as suitable solvers for the discretized pde are available . we introduce the following functional space @xmath37 where @xmath38 denotes a dirichlet portion of the boundary @xmath39 . further @xmath40 and @xmath41 denote the inner product and norm of @xmath42 , similar for other spaces . in the error analysis , we make use of a norm bound , i.e. we denote by @xmath43 a constant that satisfies , for any @xmath44 : @xmath45 in the experiments , we use @xmath46 or weighted @xmath46 norms for @xmath42 , but for the sake of generality of our analysis , we introduce this generic constant that includes other frameworks , e.g. , @xmath47 norm for which @xmath43 is the friedrichs - poincar constant . we define a backward time variable @xmath48 which we will use throughout the paper . this transforms the pde ( [ eq : heston1 ] ) into a standard forward evolution problem . we denote by @xmath49 , by @xmath50 " the hadamard product and introduce a parameter @xmath51 , @xmath52 , which parametrizes ( [ eq : heston1 ] ) . we set @xmath53 for the black - scholes model and @xmath54 for the heston model . we start our consideration with the simplest case of pricing the european type of options . since puts and calls of the european option can be easily interchanged via a put - call parity relation @xcite , it suffices for us to consider , e.g. , only call options . in addition , there exists a semi - closed analytical solution for these options in the black - scholes and heston models . however , while the black - scholes formula @xcite provides almost the exact value of the european option , the heston semi - analytical formula @xcite requires some numerical techniques to approximate the integral . thus , it is more interesting to consider the case of the heston model . then the value of a european call option @xmath55 , @xmath56 ( the relation to @xmath57 being established soon ) satisfies the following linear equation @xmath58 , \label{eq : eo1}\end{aligned}\ ] ] subject to initial and boundary conditions @xmath59 the description of the boundary conditions can be interpreted as follows : for a large stock price @xmath60 , we use a neumann boundary condition which establishes a linear growth of an option price . when the volatility @xmath61 is 0 , we can not impose any boundary conditions and assume that the equation ( [ eq : eo1 ] ) is satisfied on the line @xmath62 . when the stock price is worthless @xmath63 , it is natural to assume that the value of the call is also worthless . the option price is increasing with the volatility but it remains bounded by the stock price , hence when the volatility @xmath61 is large , we assume the value of the option tends to reach the value of the stock price @xmath60 . note , this is not the only way to prescribe boundary conditions for the problem , other types of boundary conditions can be found , e.g. , in @xcite . the variance process ( [ eq : varprocess ] ) is strictly positive if the condition on the parameters @xmath64 is satisfied , which is often referred as the feller condition , see , e.g. , @xcite . this condition plays a crucial role in the calibration process of the heston model , and it is uncommon that the parameters violate it . thus we restrict ourselves to the choice of model parameters , such that the feller condition is fulfilled . since the operator @xmath65 in ( [ eq : heston ] ) is a degenerate parabolic differential operator , the standard way to eliminate the variable coefficient @xmath60 is to perform the log - transformation of @xmath60 by introducing a new variable @xmath66 . then we are looking for the solution @xmath67 , with the initial condition @xmath68 which satisfies the transformed heston equation @xmath69 for all @xmath70 . the operator @xmath71 corresponds to the operator @xmath65 in ( [ eq : heston ] ) with respect to a change of variables and is defined as follows @xmath72 with @xmath73 and first oder spatial differential operator @xmath74 . to perform a numerical simulation , we localize the problem ( [ eq : heston_tr ] ) to a bounded computational domain @xmath75 with the variance @xmath76 . then the boundary conditions transform into @xmath77 with @xmath78 being an outward normal vector at the boundary , @xmath31 and @xmath79 and a cumulative distribution function @xmath80 defined as @xmath81 using the test function @xmath82 yields the following weak formulation of ( [ eq : heston_tr ] ) @xmath83 where @xmath84 various methods can be applied to solve this problem numerically , e.g. , @xcite . in what follows we use a @xmath85-scheme for the discretization in time and conforming piecewise linear finite elements for the discretization in spatial direction . we divide @xmath86 $ ] into @xmath87 subintervals of equal length @xmath88 and define @xmath89 , @xmath90 , @xmath91 . then , for given @xmath92 we are looking for @xmath93 that satisfy for all @xmath82 @xmath94 in order to ensure dirichlet boundary conditions we set @xmath95 , where @xmath96 solves ( [ eq : heston_tr ] ) with homogeneous dirichlet boundary conditions and @xmath97 is a dirichlet lift function , which extends non - homogeneous boundary conditions to the interior of the domain . then ( [ eq : eodisc1 ] ) becomes @xmath98 where @xmath99 we extend our considerations to american options based on the black - scholes and heston model , for which , in general , a closed form solution does not exist . since american calls are equal to european calls on non - dividend paying stocks @xcite , we will focus only on american put options in both models . the value @xmath100 of an american put paying @xmath101 at @xmath102 is a solution of the following problem @xmath103 where the differential operator @xmath104 is defined in ( [ eq : bs ] ) and ( [ eq : heston ] ) for the black - scholes and heston model , respectively . we specify the boundary conditions for the black - scholes model @xmath105 and for the heston model @xmath106 following the same arguments as in section [ sec : eo ] , we consider the price of the american put in the log - transformed variable @xmath107 with a log - transformed payoff @xmath108 which solves the following problem @xmath109 the operator @xmath110 , @xmath27 for the black - scholes or heston model , where @xmath111 the boundary conditions for the heston model are set to @xmath112 on @xmath113 and @xmath114 on @xmath115 . for the black - scholes model we define @xmath116 and on this boundary we prescribe @xmath112 . our aim is to reformulate the system ( [ eg : ao1_trans])([eq : ao2_trans ] ) in a variational saddle point form @xcite . define @xmath117 to be the dual space of @xmath42 and @xmath118 to be a dual cone . for all @xmath119 , @xmath44 introduce a duality pairing @xmath120 , @xmath121 and @xmath122 . the bilinear form of the black - scholes equation reads @xmath123 with the use of these notations we are ready to recast the problem ( [ eg : ao1_trans])([eq : ao2_trans ] ) in a saddle point form : find @xmath124 such that @xmath125 with @xmath126 in ( [ eq : bilbs ] ) or @xmath127 in ( [ eq : bilheston ] ) for the black - scholes or heston model , respectively . to treat the problem numerically , we use a @xmath85-scheme for the discretization in time and conforming piecewise linear finite elements for the discretization in spatial direction . we divide @xmath86 $ ] into @xmath87 subintervals of equal length @xmath88 and define @xmath89 , @xmath90 , @xmath91 . in order to ensure dirichlet boundary conditions , we set @xmath128 , where @xmath129 solves with homogeneous dirichlet boundary conditions and @xmath130 is a dirichlet lift function , which extends non - homogeneous boundary conditions to the interior of the domain . for @xmath131 we introduce the operators @xmath132 and the discrete problem in a saddle point form reads : for @xmath133 and given initial data @xmath134 find @xmath135 for @xmath131 and @xmath136 satisfying @xmath137 the bilinear form @xmath138 , @xmath27 is specified for each model in or . here and in the following , we frequently omit the argument @xmath139 whenever the parameter value is clear from the context . the problem ( [ eq : aodisc1])([eq : aodisc2 ] ) can be considered as a model independent formulation for pricing american put options . moreover , the european call option is also enclosed in this formulation by the exchange of @xmath140 with @xmath141 , initial conditions and omitting @xmath142 and @xmath143 terms . therefore , in the further discussions of the implementation aspects and analysis , we will focus only on the general american put option case , and we will not distinguish the models which are used to price the option . however in section [ sec : results ] , we present and compare numerical results for european and american options in both models . by a generalized lax - milgram argument , a problem of type ( [ eq : aodisc1])([eq : aodisc2 ] ) is well - posed if the bilinear form @xmath144 is continuous and coercive , @xmath145 , @xmath146 are linear and continuous and @xmath142 is inf - sup stable . in particular , for the heston model , if @xmath147 the matrix @xmath148 in ( [ eq : heston_coef ] ) is positive definite on @xmath149 and under the suitable relation on the coefficients , we obtain the coercivity and continuity of @xmath150 . this issue for the european call option was studied in great detail in @xcite . the well - posedness of the problem in the black - scholes settings can be found , e.g. , in ( * ? ? ? * ch . 6 ) . the coercivity , continuity , and inf - sup constants are defined as follows @xmath151 also , for our choice of a dual space and cone we assume @xmath152 in this section , we provide the rb - scheme for the variational inequality problem and present the main analytical results . standard finite element approaches do not exploit the structure of the solution manifold under parameter variation and for a given parameter value , a high dimensional system has to be solved . in what follows , we introduce a specific galerkin approximation of the solution , based on the reduced basis method . the first step of the reduced basis method mainly consists in computing parametric solutions in low dimensional subspaces of @xmath42 and @xmath153 , defined in section [ sec : var_form ] , that are generated with particular solutions , the so - called _ snapshots _ , of our problem . let us explain the corresponding formulation in more detail . for @xmath154 , consider a finite subset @xmath155 with @xmath156 . the reduced spaces @xmath157 and @xmath158 are defined by @xmath159 and @xmath160 where @xmath161 and @xmath162 are suitably constructed from the large set of snapshot solutions @xmath163 , @xmath164 , @xmath165 and @xmath166 , @xmath164 , @xmath167 and the reduced dimensions @xmath168 are preferably small . both families @xmath169 and @xmath170 are supposed to be composed of linearly independent functions , hence are so - called _ reduced bases_. numerical algorithms to build these two sets will be presented in section [ sec : rbcomp ] . we define the reduced cone as @xmath171 which satisfies @xmath172 due to the assumption of @xmath173 . in this setting , the reduced problem reads : given @xmath174 , find @xmath175 and @xmath176 for @xmath177 that satisfy @xmath178 for all @xmath179 and the initial value @xmath180 is chosen as an orthogonal projection of @xmath181 on @xmath157 , i.e. , @xmath182 for all @xmath183 . by the construction procedure for @xmath184 in section [ sec : implementation ] , we will assure inf - sup stability of @xmath142 on @xmath185 and guarantee that @xmath186 , where @xmath187 which implies the well - posedness of our reduced problem ( [ eq : virb1])([eq : virb2 ] ) with identical arguments as for the detailed saddle point problem . hence , we ensure existence and uniqueness of the reduced solution . for the details of the proof , we refer to @xcite . a further useful property is a basic consistency argument , the reproduction of solutions : if for some @xmath139 holds @xmath188 and @xmath189 for @xmath131 and @xmath190 then @xmath191 we prove this property by induction . for @xmath192 , @xmath193 and for @xmath194 we have @xmath195 , for all @xmath183 . set @xmath196 , then @xmath197 which is true only for @xmath198 . for the induction step , we assume that @xmath199 , @xmath200 . then , choosing @xmath201 , @xmath202 , we directly obtain @xmath203 which implies that @xmath204 solves the reduced problem . due to the uniqueness of the solution , we obtain @xmath205 and @xmath206 . in this section , we present an a posteriori analysis of our rb - scheme . in particular , only the more challenging constraint case of american options is considered , as for the european option model , the well established standard rb - error analysis for linear parabolic problems @xcite can be applied , which we omit in this work . again , to simplify the presentation , we omit the parameter vector @xmath139 in the notation of @xmath207 . in the same way , we only consider the case @xmath208 , that is , an implicit euler time discretization . however the analysis presented hereafter holds for any @xmath209 $ ] , up to technical supplementary computations . we start by introducing relevant residuals and preliminary results . in order to evaluate the approximation errors induced by our method , we define , for @xmath210 , the equality and inequality residuals by : @xmath211 we also introduce the primal and dual errors @xmath212 using the linearity of @xmath213 , one finds that : @xmath214 as a consequence of inf - sup stability , one can bound the dual error by the primal error , as stated in the next lemma . [ lemma : pderror ] for @xmath210 , the dual error at time step @xmath215 can be bounded by the primal error as @xmath216 the inf - sup stability of @xmath142 guarantees the existence of a @xmath217 such that @xmath218 using , we find that : @xmath219 and the result then follows from . let us then introduce , for @xmath210 , the riesz - representer @xmath220 of our inequality residual : @xmath221 contrary to @xmath222 , the quantity @xmath223 is not a straightforward error estimator component because of the inequality constraint . we obviously would correctly penalize if @xmath224 for some @xmath225 as desired , but we would also penalize @xmath226 which is not necessary . hence , we need to cope with the inherent nonlinearity induced by the inequalities . for this purpose , we consider a family of projectors on the cone @xmath227 which are assumed to satisfy , for @xmath210 , @xmath228 having and the characterization of the dual cone @xmath229 we find that @xmath230 projectors satisfying improve the one proposed in @xcite , namely one term in the error bound can be cancelled . we also refer to @xcite where such techniques are applied for finite element based error estimators in contact mechanics and for obstacle problems . we are now in a position to define an a posteriori error estimator associated with our method . [ apostest ] define the a posteriori quantities @xmath231 one has : @xmath232 first , we note that we have for all @xmath233 : @xmath234 and we recall that for all @xmath235 : @xmath236 from , we then have : @xmath237 let us focus on the term @xmath238 . thanks to , , , definition of @xmath239 and the fact that @xmath240 from , this term can be simplified as follows : @xmath241 this estimate combined with and the coercivity of @xmath242 , gives rise to : @xmath243 the first term of the inequality can be expressed as @xmath244 using lemma [ lemma : pderror ] and young s inequality , we then bound @xmath245 in : @xmath246 combining this estimate with , we can simplify as follows : @xmath247 using young s inequality gives : @xmath248 we get the result by combining the latter with , summing the resulting inequalities from @xmath192 up to @xmath249 and multiplying the result by @xmath250 . we point out that @xmath251 can be regarded as measure for the violation of the constraint . thus it plays a similar role as typical penalty terms in primal methods . in this section , we mainly discuss the implementational aspects of solving a reduced basis problem ( [ eq : virb1])([eq : virb2 ] ) associated with our detailed formulation ( [ eq : aodisc1])([eq : aodisc2 ] ) . let us give a few remarks about the solvability of the detailed problem . the differential operator in ( [ eq : aodisc1 ] ) is of convection - diffusion type with constant ( in the black - scholes model ) or variable ( in the heston model ) coefficients . thus for the specific range of the parameters , the convective term may dominate the diffusive one , and applying the standard galerkin method might lead to instabilities and inaccurate results . to overcome this difficulty and enhance the quality of the discrete solution , one can use a different discretization scheme , e.g. , a streamline upwind petrov - galerkin method @xcite . however , for our numerical simulations , we restrict the parameter range such that the diffusive term is large enough compared to the convective one and a standard galerkin finite element method can be applied . for the remainder of the paper , @xmath42 is now a standard conforming piecewise linear finite element space used for the discretization of the variational inequality . more precisely , consider a triangulation @xmath252 of @xmath57 , consisting of @xmath253 simplices @xmath254 , @xmath255 , such that @xmath256 . we then use a standard conforming nodal first order finite element space . we define the discrete space @xmath257 of dimension @xmath258 . for the sake of simplicity , we omit the index @xmath259 and consider @xmath42 to be a discrete finite element space . we associate the basis functions @xmath260 with its lagrange node @xmath261 , i.e. , @xmath262 . for the discretization of the lagrange multipliers in @xmath118 , we use dual basis functions @xcite , that is , we consider a dual finite element basis @xmath263 of @xmath117 , so that @xmath264 , @xmath265 . the cone @xmath229 is defined by : @xmath266 as these spaces are assumed to be sufficiently accurate a priori , that is , the finite element discretization error is neglegible compared to the reduction error , we do not discriminate notationally between the true and the finite element spaces . however a more extensive analysis has to take into account also the quality of the detailed solution . we shall now present a method to solve the problem ( [ eq : virb1])([eq : virb2 ] ) . the approach we follow is mainly based on the primal - dual - active - set - strategy @xcite , which is equivalent to semi - smooth newton method , thus a superlinear convergence of the algorithm can be achieved . we expand the solution @xmath267 of ( [ eq : virb1])([eq : virb2 ] ) as @xmath268 and @xmath269 with the coefficient vectors @xmath270 , @xmath271 . we also introduce the following set of notations : for @xmath272 $ ] , denote by @xmath273 , @xmath274 and @xmath275 the matrices of coefficients @xmath276 , @xmath277 and @xmath278 with @xmath279 and @xmath280 , respectively . denote also by @xmath281 and @xmath282 the vectors of components @xmath283 and @xmath284 , respectively . one can then rewrite equations ( [ eq : virb1])([eq : virb2 ] ) in the matricial form : @xmath285 where `` @xmath50 '' denotes the hadamard product , @xmath286 , @xmath287 are the coefficient vectors of @xmath288 and @xmath289 in @xmath157 , respectively , @xmath290 is the coefficient vector of @xmath291 in @xmath158 and the initial data is obtained by solving @xmath292 . to solve efficiently this system , we apply a semi - smooth newton method which can easily be carried out in terms of a primal - dual active set strategy . thus , a superlinear convergence of the algorithm can be achieved . for the sake of completeness , we recall here the main features of this method . let @xmath293 and @xmath294 be generic coefficient vectors of length @xmath295 and @xmath296 , respectively , then we introduce the function @xmath297 , where @xmath298 and @xmath299 applies componentwise . with the use of these notations , the problem ( [ eq : virb1])([eq : virb2 ] ) can be rewritten in the algebraic form @xmath300 and the initial data is obtained by solving @xmath292 . to employ the the primal - dual - active - set - strategy , we introduce the active and inactive sets @xmath301 where we have denoted by @xmath302 the @xmath303-th component of a vector . the equivalence between equations ( [ eq : virb3])([eq : virb6 ] ) and equations ( [ eq : virb7])([eq : virb8 ] ) follows from the fact that if @xmath304 , @xmath305 so that @xmath306 and otherwise , if @xmath307 , @xmath308 so that @xmath309 . a newton iteration can then be applied , which gives rise to the next algorithm . given a tolerance @xmath310 and initial conditions @xmath311 , the trajectory @xmath312 , @xmath313 is computed recursively as follows . suppose that at time step @xmath314 , @xmath315 is known . 1 . set @xmath316 and @xmath317 . 2 . while @xmath318 , do 1 . define @xmath319 as the solution of : @xmath320 2 . set @xmath321 . 3 . define @xmath322 . note that @xmath323 and @xmath324 do not appear in the iteration ( except in the definitions of the sets @xmath325 and @xmath326 ) as the functions involved in ( [ eq : virb7])([eq : virb8 ] ) are either linear or piecewise linear . using the previous time - step solution as a start iterate the active set is already a very good guess and thus we can expect a good convergence of the newton iteration . for the a posteriori error estimators , we require suitable projectors on the cone . in the case of our finite element basis , we simply choose them by their discrete vectorial representation in the dual finite element space as @xmath327_+\right),\ ] ] where for @xmath328 , @xmath329 is defined by @xmath330 if @xmath331 , and @xmath332 else . given a coefficient vector @xmath333 , we denote by @xmath334_+$ ] its positive part , that is @xmath335_+\right)_p=0 $ ] , if @xmath336 and @xmath335_+\right)_p=\left(x\right)_p$ ] else . with this definition , the property ( [ prop : pi ] ) is trivially fulfilled . in this section , we present a method to build the reduced primal basis @xmath337 and dual basis @xmath338 . the approach we follow consists in building iteratively and simultaneously the reduced primal and dual basis in a greedy fashion , as presented in algorithm 2 . we consider a finite training set @xmath339 small enough such that it can be scanned quickly , but sufficiently large to represent the parameter space well . after a ( quite arbitrary ) choice of the initial reduced spaces , we proceed in a greedy loop : for a given primal and dual space at stage @xmath340 , we identify the parameter vector @xmath341 in the training set that currently leads to the worst reduced basis approximation . for this parameter , new primal and dual basis vectors are generated , and this loop is repeated @xmath342 times . to be more precise , the selection of new primal vectors is based on the idea of the pod - greedy procedure , which meanwhile is standard for rb - methods for time - dependent parametric problems @xcite and has provable convergence rates @xcite . the main step consists in computing the complete simulation of the worst - resolved trajectory , extracting the new information by an orthogonal projection on the current primal space and a compression by a proper orthogonal decomposition ( pod ) to a single vector , the so called dominant pod - mode . in the algorithm @xmath343 denotes the orthogonal projection on @xmath344 with respect to @xmath345 and the dominant pod - mode is given as @xmath346 the new dual basis vectors are selected using an angle - greedy argument , which aims at maximizing the volume of the resulting cone @xcite . this means , we include the snapshot showing the largest deviation from the current reduced dual space , that is , the vector @xmath347 that maximizes @xmath348 where @xmath349 denotes the angle between a vector @xmath119 and a linear space @xmath350 . we point out that our system has a saddle point structure . thus ignoring the dual basis in the construction of the primal basis may lead to a reduced system for which the stability can not be guaranteed . to guarantee the inf - sup stability of our approach , we follow the `` inclusion of supremizers '' idea introduced in @xcite for the stokes problem . this enrichment consists in including @xmath351 into the primal space , where @xmath352 is the solution of @xmath353 , for @xmath44 . then @xmath354 is the element that supremizes the expression @xmath355 , hence it is called a `` supremizer '' . this extension ensures that the reduced inf - sup condition ( [ eqn : reduced - inf - sup ] ) is satisfied , which can be proven following the lines of @xcite , and hence the reduced problem is well - posed . we conclude by defining the final reduced space @xmath356 of dimension @xmath357 . the selection of the `` worst '' parameter in the greedy loop requires a measure @xmath358 , which is an error estimator that can be chosen as one of these three quantities : @xmath359 hence @xmath360 is the squared true @xmath361-error , @xmath362 is the true error measured in a space - time energy , and @xmath363 is the a posteriori error bound from the previous section , cf . theorem [ apostest ] . with these notations , the resulting pod - angle - greedy algorithm is fully specified . [ alg : global - greedy ] given @xmath364 , @xmath339 1 . choose arbitrarily @xmath365 and @xmath366 2 . set @xmath367 , @xmath368 , @xmath369 , 3 . set @xmath370 , @xmath371 , 4 . for @xmath372 , define @xmath373 , 2 . find @xmath374 3 . set @xmath375 , + @xmath376 , @xmath377 , 4 . define @xmath378 , + set @xmath379 , @xmath380 , 5 . define @xmath381 , @xmath382 , @xmath383 , 6 . define @xmath384 , @xmath385 , @xmath386 . let us give some additional comments on this procedure . first note , that the extension steps in the algorithm are only performed if the sets remain linearly independent . another way of stating this is to say that if a supremizer is already contained in the current space , it will not be inserted again in order to maintain the linear independence of the reduced bases . this results in @xmath387 and @xmath388 . note also that the orthonormalization steps can either be done by a simple gram - schmidt orthonormalization , or a singular value decomposition ( svd ) . in particular the former is very attractive , as @xmath389 and @xmath390 are already orthonormal , hence in step 4.(d ) it is sufficient to orthonormalize the single vector @xmath351 . further , we emphasize that the supremizer enrichment in practice also may be skipped : this results in smaller and hence faster reduced models , by accepting the loss of a theoretical stability guarantee , see , e.g. , @xcite . using the error estimator @xmath391 or @xmath392 corresponds to work with quantities associated with the true error and is consequently more expensive , as the full trajectories for all parameters in the training set must be precomputed . this is in contrast to the case of working with @xmath393 , which only involves the reduced solver used in the online phase , and only detailed trajectories for the selected parameters must be computed . hence , when using @xmath393 , the training set @xmath394 can be potentially chosen much larger and hence much more representative for the complete parameter domain . we finally give a short comment on the european option case : in that case , we use the @xmath392 measure for selecting the worst parameter . due to the absence of the lagrange multiplier , we do not obtain a saddle point structure of the problem . hence , we omit the supremizer enrichments and the angle - greedy step , and the algorithm then reduces to the classical ( strong ) pod - greedy algorithm . we use the @xmath85-scheme presented in section [ sec : var_form ] for the time - discretization with @xmath395 for european and @xmath208 for american options . the time domain @xmath396=[0,1]$ ] is discretized with a uniform mesh of step size @xmath397 , @xmath398 . the space domain is set to @xmath399 and @xmath400 for the black - scholes and heston model , respectively . the black - scholes model is treated in the original @xmath60 variable , while the heston model is considered with respect to the log - transformation @xmath401 . as a consequence , we use the @xmath46 norm for the heston model and the weighted @xmath46 norm @xmath402 for the black - scholes model . for the black - scholes model we set @xmath403 nodes and the heston model @xmath404 nodes . to build the basis , we consider a subset @xmath394 of @xmath405 , which is specified for the black - scholes model as in and for the heston model as in for european options and as in for american options . @xmath406\times[0.0014,0.0016]\times[0.4750,0.5250 ] \label{eq : par_dombs}&\subset\mathbb{r}^3,\\ \mathcal{p}&\equiv[0.1,0.4]\times[0.21,0.9]\times[0.08,0.15]\times[1.2 , 3]\times[0.01,0.2]&\subset\mathbb{r}^5,\label{eq : par_domh}\\ \mathcal{p}&\equiv[0.6,0.9]\times[0.21,0.9]\times[0.16,0.25]\times[3 , 5]\times[0.01,0.2]&\subset\mathbb{r}^5\label{eq : par_domha}.\end{aligned}\ ] ] for the black - scholes setting we use @xmath407 composed of @xmath408 values chosen equidistantly distributed in @xmath409 , while for the heston model the set @xmath407 is varying and specified in every particular case . we define our model parameters @xmath133 as @xmath410 and @xmath411 for the black - scholes and heston models , respectively . the strike @xmath412 is not considered as a model parameter , since it scales the value of the option . we set @xmath413 and @xmath414 for the black - scholes and heston model , respectively . in order to design the reduced primal and dual bases , we use the pod - angle - greedy algorithm with different error measures @xmath358 . we first consider the numerical results and a performance of the reduced basis method for the simpler linear case of the european call option with the heston model . to illustrate the motivation to apply the reduced basis approach , we demonstrate the variability of the solution in parameter and time . in figure [ fig : snapshots_heston ] ( top left ) the detailed solution for a fixed parameter value at the final time @xmath415 is presented . the evolution of the solution for different time @xmath416 is shown in figure [ fig : snapshots_heston ] ( top right ) and the variation with respect to parameters in figure [ fig : snapshots_heston ] ( bottom ) . at @xmath417 ( left ) and a time evolution of the option at @xmath418 ( right ) . bottom : snapshots of the solution for different parameter values extracted at a fixed volatility @xmath418.,title="fig : " ] at @xmath417 ( left ) and a time evolution of the option at @xmath418 ( right ) . bottom : snapshots of the solution for different parameter values extracted at a fixed volatility @xmath418.,title="fig : " ] + at @xmath417 ( left ) and a time evolution of the option at @xmath418 ( right ) . bottom : snapshots of the solution for different parameter values extracted at a fixed volatility @xmath418.,title="fig : " ] at @xmath417 ( left ) and a time evolution of the option at @xmath418 ( right ) . bottom : snapshots of the solution for different parameter values extracted at a fixed volatility @xmath418.,title="fig : " ] after computing the set of snapshots which consists of the detailed solutions for different @xmath419 , we employ the pod - angle - greedy algorithm with @xmath392 as a selection criterion to construct the reduced basis space @xmath157 . as mentioned earlier , this procedure corresponds to the standard strong pod - greedy algorithm . we test the reduced basis approach for different dimension of the parameter domain @xmath420 , @xmath421 , namely we consider @xmath422 , @xmath423 and @xmath424 . for each choice of @xmath139 , the remaining parameter values are assumed to be fixed and taking the value from the default parameter vector @xmath425 . in our first test , we consider @xmath426 and @xmath427 equidistantly distributed points . the first six orthonormal reduced basis vectors produced by the algorithm are presented in figure [ fig : rb_2d_basis ] . to quantify the efficiency of the reduced basis method , we investigate the error decay when increasing the dimension @xmath295 . for each reduced model , we compute the maximal error @xmath430 over a random test set @xmath431 . in figure [ fig : rb_2d_error ] , we observe that the error plotted versus the dimension of the reduced model decays exponentially . for the european option with the heston model and @xmath429 . the test grid is random @xmath431 . right : plot of the selected parameters @xmath432 and their frequency of the selection in the construction of the reduced basis in algorithm [ alg : global - greedy ] . the train set composed of @xmath427 equidistantly distributed points.,title="fig : " ] for the european option with the heston model and @xmath429 . the test grid is random @xmath431 . right : plot of the selected parameters @xmath432 and their frequency of the selection in the construction of the reduced basis in algorithm [ alg : global - greedy ] . the train set composed of @xmath427 equidistantly distributed points.,title="fig : " ] the analogous results for other choices of @xmath433 and @xmath434 are presented in figure [ fig : rb_3d_error ] and figure [ fig : rb_5d_error ] . for both cases the exponentially decaying behavior of the error is shown , however it can be observed that the convergence is slower for larger dimension of the parameter , which is explained by the increasingly complex parameter dependence of the model . in figure [ fig : rb_5d_error ] , we also present the evolution of the train error @xmath435 used for the reduced basis construction in the step @xmath436 of algorithm [ alg : global - greedy ] . for the european call option with the heston model with @xmath437 . the test grid is random @xmath438 . right : plot of the selected parameters @xmath432 and their frequency of the selection in the construction of the reduced basis in algorithm [ alg : global - greedy ] . the train set composed of @xmath439 equidistantly distributed points.,title="fig : " ] for the european call option with the heston model with @xmath437 . the test grid is random @xmath438 . right : plot of the selected parameters @xmath432 and their frequency of the selection in the construction of the reduced basis in algorithm [ alg : global - greedy ] . the train set composed of @xmath439 equidistantly distributed points.,title="fig : " ] ( blue stars ) and a test error @xmath430(green crosses ) for the european call option with the heston model with @xmath434 . the test grid is random @xmath440 . the train set composed of @xmath441 equidistantly distributed points . ] in this section , we present the numerical results corresponding to the performance of the reduced basis approach for pricing american options with the black - scholes and heston model and the corresponding a posteriori error estimates . for the american option case with the heston model , we consider the settings of the parameter domain defined in . for the experiments presented in this section , we consider @xmath442 and , if not stated , the remaining parameter values are set to the corresponding entries of @xmath443 . in figure [ fig : ao_snapshots ] , we show the snapshots of the primal solutions and their lagrange multipliers at different parameter values , which motivates us to apply the reduced basis method to this model . to study the efficiency of the reduced basis approach , we consider the pod - angle - greedy algorithm with the true error indicator @xmath444 . we build the hierarchical reduced basis with @xmath445 and @xmath446 using the training set @xmath447 uniformly distributed points . the evolution of the error produced by the algorithm is depicted in figure [ fig : ao_basis2 ] . each reduced basis model is tested on the random test set @xmath448 and the corresponding error @xmath449 is also depicted in figure [ fig : ao_basis2 ] . now , we restrict ourselves to the black - scholes model . we first demonstrate the parameter and time dependence of our model . in figure [ fig : bs_snapshots ] , the examples of the primal and dual solutions at different parameter values @xmath451 and time steps are presented . at a final time step @xmath415 . the evolution of the solution for different times ( right ) . , title="fig : " ] at a final time step @xmath415 . the evolution of the solution for different times ( right ) . , title="fig : " ] at a final time step @xmath415 . the evolution of the solution for different times ( right ) . , title="fig : " ] as the next set of experiments , we consider the performance of the reduced basis approach using algorithm [ alg : global - greedy ] with the true energy error @xmath452 . we construct the bases using the proposed algorithm with @xmath453 , @xmath454 and test the reduced basis approach first for a random value of the parameter @xmath455 . some steps of the simulation are represented in figure [ fig : simu ] . the reduced basis and fine detailed simulation curves are hardly distinguishable , that reveals a good reduced basis approximation . ) at time steps @xmath456 , @xmath457 and @xmath458 for the american put option with the black - scholes model , with @xmath459 . the payoff function is represented as the black line.,title="fig : " ] ) at time steps @xmath456 , @xmath457 and @xmath458 for the american put option with the black - scholes model , with @xmath459 . the payoff function is represented as the black line.,title="fig : " ] ) at time steps @xmath456 , @xmath457 and @xmath458 for the american put option with the black - scholes model , with @xmath459 . the payoff function is represented as the black line.,title="fig : " ] we then test our algorithm on a larger set of parameters . we consider @xmath460 , a random set of @xmath461 parameters and estimate the mean value of the error @xmath391 over @xmath462 . more precisely , we evaluate @xmath463 . the results are plotted in the diagram presented in figure [ fig : errorcurve ] . we observe an error decay of several orders in magnitude when simultaneously increasing @xmath295 and @xmath296 . we now focus on the efficiency of the design procedure and an inclusion of an a posteriori error estimate using as an example the american put option with the black - scholes and heston model . for this , we consider the pod - angle - greedy algorithm and use alternatively @xmath452 and @xmath464 as selection criterion with @xmath465 , i.e. , @xmath453 , @xmath454 for both models . the other parameter settings remain the same as introduced in section [ sec : bs_results ] and section [ sec : h_results ] . the visualization of the reduced basis vectors of @xmath466 , @xmath467 for the black - scholes model with @xmath464 is represented in figure [ fig : basis_bs ] . we now comment on an employment of the a posteriori error bounds developed in section [ sec : error_analysis ] . we consider the values of the quantity @xmath468 along the iterations of algorithm [ alg : global - greedy ] , choosing either @xmath452 or @xmath464 . the corresponding results for the black - scholes model are presented in figure [ fig : convgreedy ] and for the heston model in figure [ fig : convgreedyh ] . we observe that for both models using @xmath452 as an error measure in algorithm [ alg : global - greedy ] results in less monotone error convergence of @xmath393 with respect to @xmath469 and vice versa . overall we obtain similar accuracy when using the ( cheap ) @xmath393 measure in contrast to the expensive true error @xmath391 , which illustrates the relevance of our a posteriori analysis . for the american put option with the black - scholes model during the iterations of algorithm [ alg : global - greedy ] with error estimator @xmath452 ( left ) and with @xmath464 ( right ) . blue stars : values of @xmath391 , green crosses : values of @xmath393.,title="fig : " ] for the american put option with the black - scholes model during the iterations of algorithm [ alg : global - greedy ] with error estimator @xmath452 ( left ) and with @xmath464 ( right ) . blue stars : values of @xmath391 , green crosses : values of @xmath393.,title="fig : " ] for the american put option with the heston model during the iterations of algorithm [ alg : global - greedy ] with error estimator @xmath452 ( left ) and with @xmath464 ( right ) . blue stars : values of @xmath391 , green crosses : values of @xmath393.,title="fig : " ] for the american put option with the heston model during the iterations of algorithm [ alg : global - greedy ] with error estimator @xmath452 ( left ) and with @xmath464 ( right ) . blue stars : values of @xmath391 , green crosses : values of @xmath393.,title="fig : " ] as a measure of the quality of the proposed error estimate , we define the associated effectivities @xmath470 table [ table : bs ] and table [ table : heston ] provide the maximum effectivities @xmath471 associated with the error bounds for the black - scholes and heston model using different error measures in the pod - angle - greedy algorithm . note , that the choice of @xmath358 in the algorithm does not have a significant impact on the values of the maximum effectivities . we also observe that for the heston model the effectivity values are higher than for the black - scholes one , which can be justified by a more complex nature of the model . overall the effectivities of two orders of magnitude are very well acceptable for instationary rb problems . .maximum effectivities for the american put with the black - scholes model at different time steps @xmath472 and for different error measures @xmath358 and @xmath465 in algorithm [ alg : global - greedy ] . [ cols="^,^,^,^,^,^",options="header " , ] , _ contact problems in elasticity : a study of variational inequalities and finite element methods _ , vol . 8 of siam studies in applied mathematics , society for industrial and applied mathematics ( siam ) , philadelphia , pa , 1988 . , _ an introduction to variational inequalities and their applications _ , vol . 31 of classics in applied mathematics , society for industrial and applied mathematics ( siam ) , philadelphia , pa , 2000 . reprint of the 1980 original . , _ online - efficient rb methods for contact and other problems in nonlinear solid mechanics_. slides of the presentation on workshop : numerical methods for high - dimensional problems , ecole des ponts paristech , 2014 . , _ a posteriori error bounds for reduced - basis approximation of parametrized noncoercive and nonlinear elliptic partial differential equations _ , in in proceedings of 16th aiaa computational fluid dynamics conference , 2003 . paper 2003 - 3847 .

roughly one - third of _ kepler _ s planet candidates announced by @xcite are associated with targets that have more than one candidate planet . false positives ( fps ) plague ground - based transit searches , but the exquisite quality of _ kepler _ photometry , combined with the ability to measure small deviations in center of light during transits @xcite , have been used to cleanse the sample prior to presentation in @xcite . accounting for candidates on each one s individual merit , @xcite estimated the fidelity of _ kepler _ s planet candidates ( fraction of the candidates expected to be actual planets ) to be above @xmath0 . yet the fidelity of multiple planet candidates is likely to be higher than that for singles @xcite . we show herein that the vast majority of _ kepler _ s multiple planet candidates are true multiple planet systems . the majority of _ kepler _ planet discoveries announced to date have been confirmed dynamically , using radial velocity variations of the target star or departures of observed transit times from a linear ephemeris ( transit timing variations , ttvs ) . however , three of the first sixteen _ kepler _ planets do not have dynamical evidence supporting the discovery , but rather these planets have been `` validated '' by showing that the probability that the observed signal is produced by a planet around the host star is at least 100 times as large as that of a false positive @xcite . for these three planets , the validation process was based on the results of the blender code , which performs detailed comparisons between the observed _ kepler _ lightcurve and the lightcurves predicted for an ensemble of theoretically conceivable fps to reject fp models that are not consistent with the observations . blender then sums the _ a priori _ probability for the allowed false positives and compares that to the _ a priori _ probability for the planet model . blender validates a planet when the resulting odds ratio very strongly favors the planet model . although all three of these planets orbit stars that also have dynamically - confirmed planets , planetary multiplicity was not used quantitatively to aid in their validation . @xcite do a more cursory analysis , but one that is far easier , enabling them to perform it for all of the candidate planets ( `` candidates '' , henceforth ) that are listed in the @xcite paper . the study by @xcite assumes a flat 20% planet occurrence rate as a prior for their bayesian analysis ( i.e. , the average number of planets per star is assumed to be 0.2 ) , and they do not consider planets orbiting background stars to be fps . @xcite use the depth and period of the transit , together with the magnitude , stellar properties and galactic latitude of the target star ( henceforth `` target '' ) to provide an upper bound on the fp prior ; for high snr transits , this upper bound on the fp prior can be much higher than the one computed using detailed lightcurve matching by blender to rule out many fp scenarios @xcite . indeed , @xcite do not examine _ kepler _ lightcurves , but rather assume for their calculations that various false - positive vetting procedures using _ kepler _ photometry have been done , such as elimination of candidates with v - shaped lightcurves or where the centroid of light received on the _ kepler _ focal plane is in a different position during a transit than it is outside of transit , and that the planet candidates show no evidence for being fps . however , these cuts were not applied to all of the candidates listed in @xcite . some candidates were not vetted as extensively as assumed by @xcite , and other candidates did not pass one of the cleanness tests but still remain viable , if somewhat more suspect , planetary candidates ( e.g. , grazing planetary transits , as well as grazing stellar eclipses , produce v - shaped lightcurves ) , and thus remain on @xcite s candidate list . we develop herein a quantitative method to estimate the true planet fraction of the overall sample of _ kepler _ candidate multiple planets that is based upon the observation that there are far more multiples than would be expected for a random distribution of candidates among _ kepler _ targets @xcite . this `` overabundance '' of multiples is illustrated by the fact that fewer than 1% of targets have a candidate on @xcite s candidate list , whereas more than 15% of targets with at least one candidate have multiple candidates , and more than 25% of targets with at least two candidates have a third candidate . such a result is expected if planets often appear within systems of planets , especially if such systems tend to be flat in the distribution of inclinations . high multiplicity , low inclination , planetary systems are consistent with _ kepler _ observations @xcite . we describe an approach to validation of specific multiple planet system candidates that is intermediate between those of @xcite and blender , and we apply these techniques to validate the very flat and dynamically - packed koi-707 system . in this approach , applied herein exclusively to koi-707 , we examine the target and candidates individually , supplementing the _ kepler _ input catalog ( kic , * ? ? ? * ) with spectroscopic observations of the target and high - resolution imaging , including keck guider camera exposures and speckle imaging @xcite and/or adaptive optics @xcite , to search for other stars in the target aperture . we check the lightcurve for a shape consistent with a transit . we verify that the centroid motion is nil or consistent with a transit around the target star and not a transit/ eclipse of any other stars observed nearby . but we do not perform detailed matching of observed lightcurves to fp scenarios . we then use multiplicity to effectively increase the planet priors and allow for validation in a bayesian sense . note that for candidate giant ( @xmath1 1 jupiter radius ) planets , one would need to consider the possibility of the transit signal being produced by a late m star partially eclipsing the target star ( perhaps with dilution ) ; for smaller candidates , the possibility of transits of fainter stars by larger planets must be taken into account . the probability of a candidate in a multiple system being a false positive is contrasted to that of a single candidate in section 2 . other factors involved in validating transiting planet candidates are discussed in section 3 . we apply our method to the koi-707 candidate multi - planet system in section 4 . we conclude in section 5 with a summary of our results and a discussion of their implications . _ kepler _ has found far more multiple candidate systems than would be the case if candidates were randomly distributed among target stars @xcite . false positives are expected to be nearly randomly distributed among _ kepler _ targets , whereas true transiting planets could be clustered if planets whose sizes and periods are adequate for transits to be detected often come in multiples , as is the case for planets detected by radial velocity variations @xcite , and/or if planetary systems tend to be flat , so geometry leads to higher transit probabilities for other planets if one planet is seen to transit @xcite . we quantify the excess of multiples and its implications for the overall fidelity of the sample of multiple candidates in this section . we use the following notation and input parameters : the number of planet candidates is @xmath2 , the number of targets with two or more candidates , i.e. , the number of candidate multiple planet systems , is @xmath3 , and the number of targets from which the sample is drawn is @xmath4 = 160,171 . these numerical values are taken from @xcite , who removed from the 1235 candidates listed in @xcite those objects that were estimated to have radii greater than twice that of jupiter , those for which only one transit had been observed , and those identified as fps by @xcite . we denote the number of actual planets among the candidates as @xmath5 , the number of false positives as @xmath6 , and the fidelity of the sample ( fraction of candidates that are planets ) as @xmath7 , so @xmath8 and @xmath9 . for our calculations in equations ( 19 ) , we make the following two assumptions : \i ) fps are randomly distributed among the targets ; \ii ) there is no correlation between the probability of a target to host one or more detectable planets and to display fps . in contrast , we do not assume that planets are randomly distributed among the targets . the previous verifications of the two - planet kepler-10 system @xcite , the three - planet kepler-9 @xcite and kepler-18 systems @xcite , and especially the six - planet kepler-11 system @xcite , provide strong evidence of the non - random distribution of transiting planets among _ kepler _ target stars . the above two assumptions , together with observed values of @xmath10 , @xmath11 and @xmath4 and an assumed lower bound on @xmath7 , allow us to estimate lower bounds on the fidelity of candidates in various classes of candidate multi - planet systems using simple algebra . for example , assumption ( 1 ) implies that the expected number of targets with @xmath12 false positives , @xmath13 , is given by a poisson distribution of @xmath4 members whose mean is given by the average number of false positives per target , @xmath14 . this expectation value is given by the following formula : @xmath15 where the approximation in equation ( 1 ) is valid for @xmath16 . we next compute the expected number of _ kepler _ targets with multiple candidates at least one of which is a fp . in some cases , we make approximations that yield conservative estimates of the expected numbers of true planets ( i.e. , overestimate the expected number of fps ) . our results , presented in equations ( 27 ) , are given first as general formulae , then as a function of the fraction of candidates that are true planets , @xmath7 , for the observed values of 160,171 targets , 1199 candidates , and 170 multiple candidate systems stated above , and finally these value if @xmath7 = 0.5 or 0.9 because we consider it to be a reasonable estimate of the actual planet fidelity rate , and @xmath17 to show that even for an overall fp rate far higher than studies suggest , the expected number of fps among the multis is remarkably low . ] . these final numbers are to be compared to the observation that there are 115 candidate 2-planet systems , with a total of 230 planet candidates , and 55 candidate systems of 3 or more transiting planets , with a total of 178 planet candidates . equation ( 1 ) yields estimates of the number of targets with two or three false positives : @xmath18 @xmath19 the number of targets expected to have four or more false positives is very small and can be neglected for our purposes . equation ( 1 ) also yields the expected number of targets with a single fp under the same assumptions , 597.3 or 119.8 . the assumed lack of correlation between the propensity of a target to have false positives and true planets implies that the probability that a given target hosts both a planet and one or more false positive is equal to the product of these individual probabilities . the probability of a target having a detected transiting planet is given by @xmath20 , and the probability of it showing a false positive is given by equation ( 1 ) . thus , the estimated numbers of targets with at least one planet as well as one or more false positives are given by : @xmath21 @xmath22 the number of targets expected to have both a planet and three or more false positives is very small and can be neglected for our purposes . somewhat lower estimates of the number of multi- candidate systems with one planet and at least one fp than given by equations ( 4 ) and ( 5 ) can be derived by noting that the observed multiplicity implies fewer ( than @xmath10 ) targets with a nonzero number of planet candidates . to derive an estimate of the expected number of targets with multiple true planets as well as at least one fp , replace the term in equations ( 4 ) and ( 5 ) that represents the probability of a given target having a true planet , @xmath20 , with a term representing the probability of a given target having a multi - planet system , @xmath23 . ( assuming all multis are multi - planet systems to estimate of this term is conservative in the sense that it may lead to an overestimate in the number of false positives in multi - candidate systems . ) @xmath24 @xmath25 the number of targets expected to have both multiple planets and three or more false positives is extremely small and can be neglected for our purposes . adding up the above estimates of fps of different types yields expected totals of 4.5 or 0.9 fps ( for assumed fp rates of 50% and 10% , respectively ) in doubles ( compared to a total of 230 candidates ) and 0.65 or 0.13 fps in systems with three or more candidates ( compared to a total of 178 candidates ) . these numbers suggest that validation of doubles at the 99% level by this method alone is marginal . triples and higher multiplicity candidates are more strongly validatable , enough so that the one dynamically - unstable candidate triple system noted by @xcite , koi-284 , which has a pair of candidates with a period ratio @xmath26 , should be viewed as somewhat surprising ( but see section 2.4 ) . it is of interest to compare the ( low ) false positive rates for multis estimated in the previous paragraph with the presumed fp rates for singles upon which their calculations are based . for 50% fps in singles , our estimates above correspond to 2% fps in doubles ( compare the first numbers at the end of equations 2 and 4 with the 230 candidates in _ kepler _ candidate 2-planet systems ) and 0.4% fps for targets with three or more candidates ( compare the first numbers at the end equations 3 and 57 with the 178 candidates in _ kepler _ candidate 3 or more planet systems ) ; for 10% fps in singles , we estimate 0.4% fps in doubles ( compare the final numbers at the end of equations 2 and 4 with the 230 candidates in _ kepler _ candidate 2-planet systems ) and 0.1% for targets with three or more candidates ( compare the final numbers at the end equations 3 and 57 with the 178 candidates in _ kepler _ candidate 3 or more planet systems ) . these values are consistent with a `` multiplicity boost '' in the bayesian prior for planets of @xmath1 25 for a system with one additional candidate and of @xmath1 100 for a system with more than one additional candidate . the probability that a target hosts a planet candidate , 1/150 , can be compared to the probability of a target that hosts one candidate hosts another , 1/6 , and that an additional candidate has been found for higher multiplicity systems , 1/3 . these latter values , together with the assumption that fps and true planets are uncorrelated , lead to an estimate of the multiplicity boost for candidate doubles of 30 , slightly larger than the value of 25 computed above , and to a somewhat reduced estimate of @xmath1 50 for candidates in systems of higher multiplicity . the multiplicity boost introduced in the previous paragraph can be used to quantify the increased probability that a set of transit - like signatures represent a real transiting planet if one or more additional set of transit - like signatures are detected for the same _ kepler _ target ; i.e. , to compute an estimate the probability of a particular candidate in a multiple candidate system being a true planet from an estimate of that candidate being a planet that does not account for multiplicity . we denote the probability of planethood of the candidate in question computed without factoring in multiplicity by @xmath27 and the probability of it being a planet accounting for multiplicity by @xmath28 if it is in a two candidate system and by @xmath29 if it is a member of a three or more candidate system . in our formulation , multiplicity effectively increases the bayesian estimate of the planet prior but leaves the fp prior unchanged . thus the probability of the planet interpretation accounting for multiplicity is related to that neglecting multiplicity according to the following approximate formulae : @xmath30 @xmath31 where in each case we use the smaller value computed in the previous paragraph . rough estimates for the planet probabilities of individual _ kepler _ candidate multis can be obtained by applying expressions ( 8) and ( 9 ) to the estimates of planet probability by @xcite ; the @xcite planet probabilities are given by @xmath27 in the above formulae since their numbers do not account for multiplicity . note that this multiplicity boost incorporates self - consistently both the coplanarity boost discussed in @xcite and the systems boost ( that planets tend to come in systems ) without need to disentangle these effects . the above calculations of the multiplicity boost do not assume that any of the candidates of the target under consideration have been confirmed or validated as planets . does validating one or more candidate as a true planet(s ) substantially increase the probability that other candidates of the same target are also planets ? for realistic estimates of fp rates ( e.g. , 10% ) , the equations above show that most fps in multis are likely to be combinations of 1 fp with one or more true planets , so validating one candidate does not substantially increase confidence in the others from a statistical viewpoint . nonetheless , validating one candidate does eliminate some of the non - statistical issues raised in the following three subsections . in particular , if that validated candidate is shown to not only be a planet but also to orbit the _ kepler _ target star , then the probability that the other candidates are physically - associated with the target increases . to examine the validity of the three assumptions stated above , we divide fps into two classes , ( i ) chance - alignment blends such as background eclipsing binaries ( bgebs ) and ( ii ) physical triple stars ( pts ) , where grazing eclipses of the target star by a stellar companion are considered together with pts for this purpose . we note that it is a matter of choice how to count cases where the transit signal is produced by a planet transiting a star other than the _ kepler _ target . larger planets transiting stars that are not physically associated with the star providing most of the target s flux can be considered to fall within the chance - alignment blends / bgeb class of fps ( as they must be for _ kepler _ s statistical census of planets , because background stars are not in the denominator of the ratio taken to calculate the fraction of stars with planets , so their planets must not be included in the numerator ) , or alternatively not be regarded as fps when assessing individual candidates and the reliability of the sample . planets transiting physical companions to the star providing most of the flux can either be viewed as members of the pt class of fps or not classified as fps ; how these planets are categorized for _ kepler _ s statistical census of planets is also a matter of choice . assumptions ( i ) and ( ii ) lead to the following caveats : ( i ) : the expected number of bgebs ( chance - alignment blends ) that can cause fps varies with the magnitude and galactic latitude of the target star , whereas those of pts do not ( except to the extent that the distribution of stellar types among _ kepler _ targets depends somewhat on magnitude ) . the variations in expected bgebs act to increase the expected total number of systems with 2 or more fps compared to the value given by the above formulae . when looking at individual candidate systems , the fp probability of those around faint targets and at low galactic latitude is larger , whereas the planetary interpretation becomes more likely for brighter targets and those at higher galactic latitude . however , the distributions in galactic latitude of the planet candidates and of the multiple candidate systems both track that of all _ kepler _ targets , whereas the distribution of expected bgeb fps is quite different ( figure 1 ) ; this suggests that the fraction of candidates ( both singles and multis ) that are produced by bgebs is small . binomial variation ) of each of the planet candidate curves at quartile points in their distributions . the basic shape of the three solid curves is governed by the orientation of the _ kepler _ field on the sky , where the amount of area covered increases from a point at galactic latitude @xmath32 to a region spanning @xmath33 in longitude at @xmath34 in galactic latitude and then decreases to a point at @xmath35 latitude ; the second most important factor is a weighting towards the galactic plane because of the higher spatial density of target stars in that portion of the field . the dashed curve is more heavily weighted towards the galactic plane because , to first approximation , it is weighted quadratically in the spatial density of stars ( more precisely , the product of the spatial densities of target stars and background stars ) , whereas the other curves are only weighted linearly . the spatial density of background stars drops by a factor of 17 from the lowest latitude _ kepler _ observes to the highest latitude . the similarity of curves for targets , planets and multis , and the difference of these from the curve estimating bgeb fps , suggests that the fraction of candidates that are produced by bgebs ( as well as that resulting from planets transiting background stars ) is small . ] the statistical analysis that we presented above does not discriminate between multiple candidates resulting from multiple planets orbiting a background star and a single candidate resulting from a single planet orbiting a background star . however , the correlations and lack thereof between the curves shown in figure 1 also imply that the fraction of kois caused by single or multiple planets transiting a background star is small . in some cases , it is especially unlikely that a multiple candidate system is produced by several planets orbiting a background star , either because its largest member would be greater than planetary in size , or because the requirement of dynamical stability would imply implausibly small densities for the larger planets orbiting the fainter star . additionally , as the fraction of giant planets among the multis is smaller than it is for the singles @xcite , the probability of a multi being a system of larger planets orbiting a fainter star is less than that of a single candidate orbiting a star fainter than the target . a second class of multiple fps can be produced from a single physical system , either a triple star with two sets of eclipses or a background star with multiple transiting planets . triple star systems are unstable for small period ratios and produce large eclipse timing variations for moderate period ratios , so if large timing variations are not detected , we do not need to consider triple star system fps ( nor grazing eclipsing binary stars with a transiting planet ) unless period ratios are large . quadruple star systems consisting of two pairs of eclipsing binaries ( e.g. , * ? ? ? * ) could produce a small excess probability of a pair of fps . such systems , as well as line of sight pairs of physically - unrelated ebs , also might be more likely to mimic planetary - depth transits because of mutual dilution of the fraction of light in the _ kepler _ aperture . we note also that fps are inherently more likely among both small ( earth - size and smaller ) and large ( jupiter - size and larger ) candidates than among intermediate ( neptune - size ) candidates , independent of multiplicity . most small candidates produce low - amplitude transits that reduce our ability to remove false positives from the candidate list on the basis of the shape of the lightcurve or from differences in the centroid of the location of the dip relative to that of the target . small candidates , including multis , are more likely to be planets orbiting stars fainter than the target simply because the neptune - size planets that could produce such signals are far more common according to _ kepler _ observations than are giant planets @xcite , and this difference in abundance is even larger for the multis than for the singles @xcite . at the other extreme of the planet size distribution , giant candidates appear to be more contaminated by stellar false positives @xcite of the type which plague ground - based surveys than do neptune - size candidates . ( ii ) : planets and fps of both types are more easily detectable for quieter targets ; planets and pts are more detectable for brighter targets , but bgebs that are bright enough to mimic observable transits of the target star are more likely for fainter targets . a grazing eclipsing binary with a circumstellar or circumbinary transiting planet analogous to kepler-16 @xcite would violate our assumption of a lack of correlation between planets and fps , but such a configuration requires a large period ratio for stability and an even larger period ratio for the planet to have modest ttvs . the assumption of no correlation between planets and fps would be violated if the orientations of orbital planes of pts with one star having a planet are not random . alignments of planetary orbital planes for binary stars for which both stars possess planets also violate this assumption , although members of this class are true planets that are all physically associated ( albeit less directly ) with the target star and thus would for many purposes not be classified as fps . the numbers quoted in equations ( 29 ) do not incorporate the caveats noted in this subsection . correlations between the propensity of a target to have one or more fps and to host one or more identified true planets ( ii ) are likely to be at most a factor of a few . false positives of the bgeb class are likely to be concentrated among faint target stars that are located at low galactic latitude @xcite ; this concentration increases the expected number of targets with two ( or more ) fps , but also points to a substantial portion of _ kepler _ s target list whose members have _ a priori _ fp likelihood that is lower than or comparable to those given by the numbers above . the distribution of candidate period ratios exhibits spikes near strong mean motion resonances , which implies that candidates with such period ratios are more likely to be true planets than those candidates not in or very near such resonances @xcite . this clustering can be used to increase confidence in candidates with periods near resonances ; such a `` near - resonance boost '' can be quantified using the measured values of the normalized distance from resonance computed from the distribution of planet period ratios in @xcite . candidate resonant planets can , however , arise from period aliases in analysis of light curves ; such aliases are not of the form of the astrophysical fps considered in this paper , and we note that special care to search for them is an important aspect of validating candidate resonant planets ( see the discussion of koi-730 and the note added in proof in @xcite ) . resonant period ratios increase confidence in various candidates , and as mentioned at the beginning of section 2.1 , one of the candidate systems , koi-284 , is highly suspect because two of the candidates have periods that differ from one another by @xmath36 , which would rapidly lead to a dynamical instability on a very short time scale ( but see the discussion of this system in section 2.4 ) . do non - resonant periods reduce the confidence in individual multis and do stable systems produce an increase in our confidence ? the answers to both parts of this question are yes , but in most cases these effects are minor . while resonances are overpopulated , a substantial majority of multis are not in nor near such resonances @xcite . and while drawing planetary periods randomly from the observed ensemble of transiting planet periods gives many more nominally unstable systems than are observed , the fraction of unstable configurations from such a random draw is still small . nonetheless , in high - multiplicity densely - packed systems , wherein most configurations with the same number of planets randomly spread over the same range in period would likely be unstable , the boost in confidence from stability can be significant . note that special period ratio boosts and stability boosts are in addition to the standard multiplicity boost , i.e. , they can be applied together as two multiplicative factors in the planet prior for appropriate systems . a very rough estimate of the expected number of multiple candidate systems that orbit different members of a binary / multiple star system ( rather than a single system of planets orbiting the same star ) can be made by assuming that half of _ kepler _ s targets are binaries , that each star within a binary is as likely to host both single and multiple transiting planets as are single stars , and that the probability of one star within a binary hosting transiting planets is uncorrelated with that of its binary companion . these assumptions imply @xmath10 planets spread around @xmath37 stars , and thus @xmath38 planets per star and @xmath39 cases in which multis are composed of planets transiting differing members of a stellar binary . positive correlations between the propensity of stars in a binary to have transiting planets ( resulting from the abundance of planets and/or alignment of orbital planes ) would increase this number , but the fact that a significant majority of binaries are composed of stars of substantially differing surface brightness combined with the paucity of giant planets transiting low luminosity stars should reduce the number of multis comprised of two planetary systems because of the difficulty of detecting planets around secondary stars . as mentioned above , koi-284 , with three planetary candidates having periods of 6.18 , 6.42 , 18.0 days and nominal sizes of @xmath1 2 r@xmath40 ( earth radii ) , is the only one of 170 multi - candidate system identified in @xcite that would be clearly dynamically unstable if all of the planets orbited the same star @xcite 0.0012 days . the revised period places the planets much farther apart , and the system is stable with the updated parameters . ] . analysis of a mcdonald observatory spectrum of this bright ( _ kepler _ magnitude @xmath41 ) target yields a temperature of @xmath1 6250 k and log@xmath42 , which suggest a f7v star that is twice as luminous as the sun . speckle images of koi-284 obtained on 24 june 2010 revealed two stars differing in brightness by less than one magnitude and separated from one another by slightly less than 1@xmath43 @xcite . spectroscopic observations of each individual star obtained at keck in 2011 show the two stars to have nearly identical spectra and have a difference in radial velocity of 0.94 @xmath44 0.1 km / s . the nearly identical velocities are consistent with their being gravitationally bound . their separation of @xmath45 at a distance of roughly 500 pc implies a projected separation of 450 au ; the relative orbital velocities for two 1 @xmath46 stars on a circular orbit with semimajor axis @xmath47 = 450 au is 2 km / s . these observations suggest that the three candidates may well all be transiting planets in the same stellar system , with one of the stars hosting one of the six day period planets and the other two planets orbiting the other star in the binary . thus , this system would not call into question the statistical arguments presented in section 2.1 . a detailed analysis of both _ kepler _ and ground - based data to test this hypothesis and elucidate particulars is currently underway @xcite . the numbers derived in section 2 suggest that the sample of multi - planet candidates identified by @xcite is likely to include only a small fraction of false positives , of order 1% . but because of the uncertainties associated with the caveats discussed in section 2 , as well as any factors that we may have omitted , we do not consider it appropriate to view all of _ kepler _ s multi - planet candidates to be validated planets at the 99% level . nonetheless , it is possible to identify candidates with various favorable characteristics as detailed below , and to examine the available data for systems individually for signs of potential inconsistencies with the planet hypothesis . see @xcite for a more detailed discussion of transit characteristics . the shape and duration of a candidate transit can be used to distinguish events produced by planets from false positives produced by known astrophysical sources . the blender code @xcite compares the observed lightcurve to synthetic lightcurves of both a planet transiting the target star and a vast array of possible astrophysical false positives involving binary stars and larger planets transiting fainter stars . if the planet hypothesis does not provide an acceptable fit to the data , the candidate is not viable ; candidates for which the planet fit is marginal are called into question , and such candidates are not considered further herein . astrophysical false positives are considered credible if they provide an acceptable fit ( 3@xmath48 ) to the data . then the _ a priori _ likelihoods of the planet hypothesis and that of a false positive are compared , and a probability that the apparent transit is caused by a blend is computed . the process can be very involved , especially for candidates with orbital periods long enough that astrophysical false positives could plausibly be produced by objects with eccentric ( not tidally - damped ) orbits @xcite . true planetary transits of the star that is the dominant source of light in the aperture are wavelength - independent ( neglecting the small effects of stellar limb - darkening and contamination by flux from other stars in the aperture ) , whereas false positives are not ( except in the unlikely case that the effective temperatures of the contributing stars are similar ) . by providing infrared time series spanning times of transit , the warm spitzer mission can assist in the validation of many transiting planet systems . several candidates in multiple systems have been observed , and they all exhibit achromatic transit depths , consistent with planetary signals @xcite . the durations of transit - like signatures can be used to assess the probability of the planetary interpretation , especially when more than one planet candidate is present for a given target . transit and eclipse durations depend on several factors , including the masses and radii of the two objects , the impact parameter , @xmath49 , which is defined as the minimum projected separation between the center of the smaller body ( planet in the case of a planetary transit ) and that of the larger body ( star ) measured in units of the radius of the larger body , and the period and eccentricity of the orbit . central transits ( those for which the center of the transiting body passes over the center of the transited body ) by a planet on a circular orbit have durations that vary as the cube root of the orbital period @xcite and inversely to the cube root of the stellar density @xcite ; these relationships hold for fps as long as the bright star is substantially larger and more massive than the other body . when the system s mass is dominated by the primary but the radius of the secondary can not be ignored , then the duration , @xmath50 , ( 1@xmath51 contact 4@xmath52 contact , which corresponds to the values reported in @xcite and in table 3 of this paper ) , of a central transit for a circular orbit satisfies the following proportionality relationship : @xmath53 if the transit duration is measured over the time interval when the center of the planet covers the disk of the star , the first factor on the right - hand side of expression ( 10 ) is no longer relevant and the proportionality is simply to @xmath54 . note that the orbital period , @xmath55 , is very accurately measured by _ kepler _ , as is the ratio of the planet s radius , @xmath56 , to that of the star , @xmath57 , provided dilution by nearby stars is small . planetary transits with an impact parameter smaller than half the star s radius have a duration that exceeds @xmath58 that of a central transit . such transits account for almost half of all transiting planets , and over half of the transits observed to a specified snr threshold . the fraction of transits with durations greater than half that of a central transit is @xmath59 ; again specifying a snr threshold increases the observed fraction . orbital eccentricities also influence the durations of transits and fps ; in most cases , durations are reduced because speeds are highest near periastron @xcite , where geometric considerations imply that transit probabilities are larger . there are two primary purposes of this work : demonstrating the very high credibility of the ensemble of _ kepler _ multi - planet candidates and providing a framework to validate particular systems of candidates as planets . we now turn to a specific system that is both intrinsically interesting because it is the closest analog to kepler-11 observed to date @xcite and has characteristics very favorable for validation by multiplicity . koi-707 ( kic 9458613 ) is a @xmath60 star for which four planet candidates , ranging in period from 13 to 41 days , were announced by @xcite ; a fifth candidate planet , with shorter orbital period and smaller size , has subsequently been identified . the lightcurve of this target is shown in figure 2 . we performed an analysis of pixel - level _ kepler _ spacecraft data that shows that the transit signals come from a location on the sky that is coincident with that of the target star within small uncertainty , and we obtained high - resolution images of the target that did not show any nearby stars capable of producing the transit signals ; these studies are presented in section 4.1 . our analyses of the transit lightcurves and spectra taken of the target star are presented in section 4.2 . in section 4.3 , we review the characteristics of this system that allow for its validation by multiplicity . we discuss the properties of the kepler-33 planetary system in section 4.4 . a seeing - limited image taken at lick shows a star @xmath61 to the nw that is 4.5 magnitudes fainter than koi-707 , but no stars closer to the target . on 12 june 2011 , koi-707 was imaged in the 880 nm filter ( @xmath62 = 50 nm ) of our speckle camera at the wiyn 3.5-m telescope located on kitt peak . this star showed no companions within a 2.8 @xmath63 2.8 arcsec region down to 3.1 magnitudes fainter than the target star ; stars within @xmath64 of the target star would not have been detected . details of this speckle imaging data collection and reduction precesses are fully described in @xcite . we check to see whether the transit signal is due to a source other than koi-707 using the difference image technique described in @xcite . this method fits the measured _ kepler _ pixel response function ( prf ) to a difference image formed from the average in - transit and average out - of - transit pixel images in order to calculate the position of the difference image centroid . this difference image centroid position is compared with the position of the prf - fit centroid of the average out - of - transit image , as well as to the predicted deviations in centroid locations due to scene crowding . the results for all 5 planet candidates are presented in table 1 and shown in figure 3 . uncertainty in the offset distance . the robust average of each modeled quarterly offsets is shown as a thick gray cross , with the gray dashed circle showing the average model 3@xmath48 offset distance uncertainty . when the circle showing the observed 3@xmath48 uncertainty in the difference image has significant intersection with the dashed circle showing the modeled 3@xmath48 uncertainty in the position of the target star , the data are consistent with the observed transit signal being on koi-707 . , title="fig : " ] uncertainty in the offset distance . the robust average of each modeled quarterly offsets is shown as a thick gray cross , with the gray dashed circle showing the average model 3@xmath48 offset distance uncertainty . when the circle showing the observed 3@xmath48 uncertainty in the difference image has significant intersection with the dashed circle showing the modeled 3@xmath48 uncertainty in the position of the target star , the data are consistent with the observed transit signal being on koi-707 . , title="fig : " ] + uncertainty in the offset distance . the robust average of each modeled quarterly offsets is shown as a thick gray cross , with the gray dashed circle showing the average model 3@xmath48 offset distance uncertainty . when the circle showing the observed 3@xmath48 uncertainty in the difference image has significant intersection with the dashed circle showing the modeled 3@xmath48 uncertainty in the position of the target star , the data are consistent with the observed transit signal being on koi-707 . , title="fig : " ] uncertainty in the offset distance . the robust average of each modeled quarterly offsets is shown as a thick gray cross , with the gray dashed circle showing the average model 3@xmath48 offset distance uncertainty . when the circle showing the observed 3@xmath48 uncertainty in the difference image has significant intersection with the dashed circle showing the modeled 3@xmath48 uncertainty in the position of the target star , the data are consistent with the observed transit signal being on koi-707 . , title="fig : " ] + uncertainty in the offset distance . the robust average of each modeled quarterly offsets is shown as a thick gray cross , with the gray dashed circle showing the average model 3@xmath48 offset distance uncertainty . when the circle showing the observed 3@xmath48 uncertainty in the difference image has significant intersection with the dashed circle showing the modeled 3@xmath48 uncertainty in the position of the target star , the data are consistent with the observed transit signal being on koi-707 . , title="fig : " ] the uncertainty in the quarterly centroid locations displayed in figure 3 is based upon propagating pixel - level uncertainty and does not include possible prf fit bias . sources of prf fit bias include scene crowding , because the fit is of a single prf that assumes a single star , and prf error . the centroid offset is the difference between the positions of the centroids of the difference image and out - of - transit images , so common biases such as prf error should approximately cancel . there is , however , a quarter - dependent residual bias . statistical analysis across many uncrowded _ kepler _ targets indicates that the residual bias due to prf error is zero - mean , so we average the quarterly centroid positions . specifically , we compute a robust @xmath65 minimizing best - fit position to the quarterly position data . the uncertainties of the average are estimated by taking the uncertainties of the fit and performing a bootstrap analysis . this average for each planet candidate is shown on figure 3 . bias due to crowding ( stars close enough to the target that they contribute light observed in the aperture ) , however , does not cancel because , to the extent that variations in other field stars are not correlated with transits , field stars do not contribute to the difference image . this introduces a systematic bias component that remains in the average over quarters . to investigate the possibility that the observed offsets are due to crowding bias , we model the local scene using stars from the kic and the measured prf , induced the transit on koi-707 with the ephemeris of each planet candidate , and performed the same prf fit analysis on the model difference and out - of - transit images as that performed on the pixel data . if the resulting modeled average offset matches the observed average offset , then the observed offset can be explained by crowding bias . the locations of the transit signatures ( difference image centroids ) of each of koi-707 s five planet candidates lie within @xmath66 of the predicted position , and given the uncertainties listed in table 1 , none of the signatures are significantly offset and all are fully consistent with transits on the target star . furthermore , the lack of near neighbor stars seen in the speckle image rules out some of the possible false positive scenarios that could produce a centroid shift of the magnitude suggested by the marginally significant offset observed . we obtained a spectrum of the target star koi-707/kepler-33 using the hires spectrometer at keck i 10-m telescope . classification of this spectrum using sme ( spectroscopy made easy , @xcite ) yields the stellar parameter values shown in table 2 ( the final two columns in this table show results from constrained fits described at the end of this subsection ) . the mean stellar density based on spectroscopic information was estimated by matching the determined distributions of @xmath67 and [ fe / h ] to the yonsei - yale evolution tracks . this allowed the determination of stellar mass and radius and mean stellar density posterior distribution based on spectroscopy . the spectrum of koi-707 is consistent with two sets of stellar evolution tracks ( figure 4 ) , leading to double - peaked probability distributions and asymmetric uncertainties in several stellar parameters . we plot the probability distributions of the fundamental parameters of stellar mass and age in figure 5 and those of the stellar density , which is a key input to transit fitting , in figure 6 , and we include in table 2 various measures of uncertainty beyond the simple standard deviation for the values of various derived stellar parameters that are either critical to the validation and properties of koi-707 and/or distributed highly asymmetrically . note that both solutions are somewhat above the main sequence , implying an evolved star . ( red ) , 1.2 m@xmath68 ( green ) , 1.3 m@xmath68 ( blue ) , 1.4 m@xmath68 ( cyan ) and 1.5 m@xmath68 ( magenta ) stars with [ fe / h ] = 0.15 . ages , in gyr , are marked along the tracks . note the uneven rates of evolution along these tracks . the gray dotted tracks are spaced by 0.01 m@xmath68 in stellar mass and 0.02 gyr in time to illustrate the distribution of field stars on the @xmath69 plane ; the apparent gaps / discontinuities in these tracks are caused by rapid changes at various stages of stellar evolution . the three boxes mark the 1 , 2 and 3 @xmath48 confidence regions of the spectroscopic sme analysis of the spectrum of koi-707 taken at keck . there are degeneracies in determining a unique stellar mass and radius for specified values of @xmath70 and @xmath71 . for example , a 4.4 gyr , 1.3 m@xmath68 model and a 5.5 gyr , 1.2 m@xmath68 model have nearly identical @xmath70 and @xmath71 . the low - mass / old peaks in the probability distribution functions shown in figure 5 result from the pile - up of stellar evolution tracks as core h is exhausted in the transition regime from radiative cores of 1.1 m@xmath68 and smaller stars to convective cores in 1.2 m@xmath68 and larger stars . the high - mass / young peaks result from the overlap between tracks for the early shell h burning phase of stars with convective cores with tracks for the late core h burning phase of stars that are @xmath72 more massive . ] as an independent check of the values of the sme parameters , we also derived values by matching the spectrum to synthetic spectra @xcite , and in addition we employ a new fitting scheme ( stellar parameter classification , spc ) that is currently under development and being readied for publication , allowing us to extract precise stellar parameters from the spectra . the spc analysis yielded the following parameters : @xmath73 k , @xmath74 , [ m / h]=@xmath75 km / s , rv = @xmath76 km / s . the sme and spc parameters are fully consistent , and the nominal value of the key parameter @xmath71 from spc is slightly ( @xmath77 ) larger than the sme value . in contrast , the derived stellar radius from sme is @xmath78 , substantially larger ( more than @xmath79 formal error , see table 2 ) than the radius estimate of 1.29 r@xmath68 based upon the kic that was used by used by @xcite . we measured the radial velocity of koi-707 relative to the barycenter of our solar system . the method involves measuring the doppler shift of the stellar spectrum obtained with the keck i telescope and hires spectrometer on bjd = 2455782.9441 relative to the solar spectrum by a @xmath65 fit of the two spectra . to set the zero point of the wavelength scale accurately , we measure the displacement of telluric lines from the earth s atmosphere , serving as the iodine cell " . the resulting radial velocities are accurate to 0.1 km / s , based on measurements of 2000 stars @xcite and comparison with iau standard stars . for koi-707 , the radial velocity is @xmath80 = 14.09 km / s , indicating motion of the star away from the solar system at a speed typical of the velocity dispersion of stars in the galactic disk . orbital periods of the planets quoted in table 3 are as perceived from the barycenter of our solar system . as the kepler-33 system is receding from us at @xmath81 , where @xmath82 represents the speed of light , the actual orbital periods in the rest frame of the kepler-33 system are 0.999953 times as large as the tabulated values . we used q1-q8 long - cadence _ kepler _ aperture photometry ( i.e. , only pixel - level corrections such as bias and dark smear are applied to the data and there is no co - detrending ) . at this level of data reduction , there are still instrumental artifacts on timescales similar to the planetary orbits . to detrend the data ( remove long - period variations ) , a second - order polynomial was fitted to segments of sequential _ kepler _ photometry data outside of transits ( i.e. , in - transit observations were given zero weight ) , and then this polynomial was subtracted from the data . a segment is defined as a length of data that is uninterrupted for 5 or more consecutive long cadences ( @xmath12.5 hours ) . each segment is then combined to produce a final time series that is then normalized by the median . we adopt a simple model to fit the data . we fit for the mean stellar density ( @xmath83 ) , and each planet s orbital period ( @xmath55 ) , transit epoch ( @xmath84 ) , scaled planetary radius ( @xmath85 ) and impact parameter ( @xmath49 ) . the planets are assumed to be on circular orbits ; planet - planet interactions are implicitly accounted for by allowing the center of transit time to vary in order to best - fit transit shape ( minimization of scatter on residuals ) . the transit was described with the analytic formulae of @xcite , and we adopted a non - linear limb darkening law with coefficients fixed based on the @xcite tabulation for the _ kepler _ band with stellar parameters from sme analysis . a best fit model was computed using a levenberg - marquardt algorithm to minimize the @xmath65 statistic and initial uncertainties were estimated via the construction of a co - variance matrix . the best fit model and co - variances were used to seed a hybrid - mcmc ( markov chain monte - carlo ) algorithm to determine posterior distributions of our model parameters . the model is considered hybrid , as we randomly use a gibbs - sampler and a buffer of previous chain parameters to produce proposals to jump to a new location in the parameter space . the addition of the buffer allows for a calculation of vectorized jumps that allow for efficient sampling of highly correlated parameter space . a direct comparison of the mean stellar density determined by the circular transit - model ( 0.236 @xmath44 0.080 ) and the mean stellar density determined from spectroscopy ( 0.301 @xmath44 0.067 ) differ by less than 1@xmath48 . this shows that we can produce a self - consistent circular model in which all five planets orbit the same star and produce stellar parameters that agree with spectroscopy . close - ups of the lightcurves of each of the planets near planetary transit are shown in figure 7 . transit durations of all five kois vary in proportion to the @xmath54 to within 5% , and to within 3% when planet sizes are factored in as specified in expression ( 10 ) . the remarkable similarity of these transit durations when appropriately scaled to orbital periods is apparent in figure 8 , which shows the same data as figure 7 but with the time coordinate scaled by @xmath54 . such similarity would be the consequence of an ensemble of planets transiting the same star on low eccentricity orbits with similar impact parameters . the impact parameters given in table 3 are indeed quite similar for all of the planets . but the inclinations implied to produce these parameters require a very fortuitous arrangement in which tilt to the line of sight is just enough larger in the inner planets than the outer ones . in contrast , if the system is viewed nearly edge on , i.e. , @xmath86 , then both a wider range of impact parameters would be allowed ( because transit durations are less sensitive to the precise value of @xmath49 when @xmath87 ) and similar values of @xmath49 imply a very flat system rather than a fortuitous ensemble of tilts . of the planet in question . vertical lines mark the beginning and end times for each transit . for planets on circular orbits with the same impact parameter , durations normalized in this manner should be the same for all planets . the remarkable similarity in normalized durations of these planets is very strong evidence that they all orbit the same star . ] we are thus motivated to perform a second set of fits constrained by the assumptions of circular orbits for all planets and @xmath88 for the planet with the longest normalized ( according to expression 10 ) transit duration , koi-707.01 . two such fits were performed , one for the low mass stellar solution ( where we imposed the constraint @xmath89 ) and the other for the high mass ( @xmath90 ) solution . we show in the far right columns of tables 2 and 3 both the high and low stellar mass `` flat '' fits ; no uncertainties are quoted for these fits because the @xmath88 assumption constrains results so tightly that computed uncertainties are unrealistically low . the fitting procedure that we used to compute the planetary parameters listed in table 3 did not account for the small ( 1 2% ) contribution of the second star in the aperture to the flux measured by _ this dilution implies that the planet radii quoted are overestimated by 0.5 1% , a difference far smaller than the uncertainties in their actual values . the five planetary candidates of koi-707 are extremely likely to represent a true 5 planet system orbiting the target star . there are several lines of evidence that support this conclusion . first , the analysis presented in section 2 implies that the overwhelming majority of _ kepler _ multiple - planet candidates are true planets and bound within the same physical system . second , the planetary system is already closely - packed dynamically for the sizes of planets around an undiluted target star ( see section 4.4 ) . were the planetary system to orbit another star that is not seen because it is significantly fainter than the target , the larger planets required to match the observations would necessitate implausibly low densities for the planetary system to be stable . third , the spatial location analysis presented in section 4.1 , combined with the galactic latitude distribution of targets and planet candidates shown in figure 1 and the relatively high galactic latitude of koi-707 ( 15@xmath91 ) , together imply that it is unlikely that the signals are due to a source not physically associated with the target star . fourth , the long durations of the transits relative to the planets orbital periods imply that they are produced by light being blocked from a star whose density is significantly less than those of cool dwarfs , and/or they are produced by apocentric transits of objects with substantial orbital eccentricity . substantial orbital eccentricities would make the system unstable if the planets orbited the same star . it is highly implausible that two stars in an unresolved stellar binary would be so similarly coeval to have nearly equal densities characteristic of a somewhat evolved star . fifth , the remarkable similarity of the transit durations when appropriately scaled to orbital periods ( expression ( 10 ) and figure 8) can be understood as the consequence of an ensemble of planets transiting the same star with similar impact parameters , but would otherwise be an unlikely coincidence . the above characteristics of koi-707.0105 , combined with the overall high validity of _ kepler _ multis , imply that these candidates have a well above 99% probability of composing a true multi - planet system , hence we consider these candidates to be validated planets and name the system kepler-33 . numerical integrations to test stability of the system were performed using masses of the planets from the nominal mass - radius relationship based upon planets within our solar system presented in @xcite . this relationship is given by : @xmath92 where r@xmath40 and m@xmath40 are the radius and mass of the earth , respectively . in previous work , @xcite integrated the koi-707 system using parameters based on the data presented by @xcite and assuming planar initially circular orbits ; the model system survived intact for the entire @xmath93 years simulated . all of our integrations also assume that the system is planar . for all planets having zero initial eccentricity , all three models of the system listed in tables 2 and 3 survived intact for the entire @xmath94 years simulated . of the three kepler-33 planetary system models that we integrated , the one based directly on the spectroscopic stellar parameters had the largest planetary sizes and thus largest estimated planetary masses , with the fractional differences in estimated planetary masses from one model to another being larger than those for stellar mass . thus , the spectroscopic parameters yielded the largest interplanetary perturbations , and presumably the largest ttv signatures . we examined distributions of ttv signals over 2 year periods during the first 2 myr of this simulation . these theoretical ttvs for the inner planet were smaller than 30 seconds , and those for kepler-33c less than 90 seconds , both values below current measurement capabilities . the outer three planets exhibited far larger ttvs , 9 18 minutes for kepler-33d , 23 38 minutes for kepler-33e and 23 35 minutes for kepler-33f , roughly consistent with variations detected . while this model used relatively high mass estimates , the integration began with circular orbits , which produce smaller ttvs , so the results do not yield a preference between the three models of planetary sizes listed in table 3 . a detailed analysis of planetary ttvs is beyond the scope of this paper , but we note that it may well soon be possible to derive estimates of the masses of some or all of the outer three planets using observed ttvs . the properties of the 5 planet kepler-33 system are analogous to those of the 6 planet kepler-11 system @xcite . both systems include 4 planets with orbital periods between 13 and 47 days that are comparable in radius to , or somewhat smaller in size than , neptune , as well as a smaller planet closer to the star . both systems are thus very closely - packed in a dynamical sense . the kepler-11 system is slightly more dynamically close - packed than is the kepler-33 system , both in the sense that the ratios of the orbital periods of neighboring planets are a bit larger for kepler-33 s planets @xcite and as evidenced from smaller ttvs detected in the kepler-33 system . in some aspects , the kepler-33 system can be viewed as a less extreme cousin of kepler-11 , with 5 known transiting planets rather than 6 and spacing that not quite as close , yet dynamically much tighter than most systems @xcite . however , the new system s inner planet , kepler-33b , is smaller and closer to its star than is its counterpart , kepler-11b . the high density of kepler-11b relative to its brethren , coupled with its smaller size , indicates a composition richer in heavy elements @xcite . while the mass and thus the density of kepler-33b is not known , it is unlikely that a planet this small ( @xmath95 ) and close ( 0.07 au ) to its bright star ( 3 l@xmath68 ) could have retained a substantial h / he atmosphere . this lends support to the suggestion by @xcite that planets such as kepler-11b may have once had atmospheres of light gases that subsequently escaped as a result of stellar irradiation . the slightly shorter normalized duration of the small inner planet kepler-33b suggests that this planet has somewhat larger inclination and/or eccentricity than do the outer four planets . we note that mercury , the innermost and smallest planet in our solar system , also has the largest eccentricity and inclination of the 8 planets known to orbit our sun . stability of the kepler-33 planetary system excludes large eccentricities of the four outer planets , although interplanetary perturbations must cause eccentricities to be non - zero . the remarkable similarity in the normalized transit durations of the five planets orbiting kepler-33 would be very unlikely if the eccentricities and inclinations of planetary orbits are significant ( figure 9 ) . moreover , the consequent similarity in computed impact parameters despite the planets differing orbital distances requires that either the system is viewed near edge - on , with impact parameters @xmath96 , or a very fortuitous coincidence producing an anticorrelation between inclinations and orbital periods . we view such a coincidence as unlikely , and therefore prefer the planetary parameters closer to those listed in final two columns of tables 2 and 3 ( the `` flat '' fits ) than to those listed in the second column of tables 2 and 3 . case . the blue triangles , yellow diamonds and red squares show results from simulations in which the eccentricity assigned to each planet was drawn form a rayleigh distribution with mean value @xmath97 , @xmath98 and @xmath99 , respectively . the addition of even a small range of eccentricities greatly reduces probability of highly correlated transit durations . ] the calculations presented in section 2 show that the vast majority of _ kepler _ s multi - planet candidates are indeed planets . the essence of our argument is that fewer than 10@xmath100 of _ kepler _ targets have a planet candidate listed in @xcite . if 10% or fewer of these candidates are fps , then fewer than 10@xmath101 of _ kepler _ targets have a false positive , and if fps are not correlated with other planet candidates , then fewer than 10@xmath102 of _ kepler _ s @xmath1 160,000 targets , i.e. , a total of at most 2 targets , are expected to have a false positive as well as another transit - like signature that makes them a candidate multi - planet system . this number can be compared to 170 candidate multi - planet systems with a total of 408 planet candidates announced by @xcite . accurate estimates of the false positive rate in this population require detailed analysis of both _ kepler _ and ground - based data to account for the caveats expressed in section 2.2 . nonetheless , we expect that @xmath103 of the 408 planet candidates in multis that were announced by @xcite are indeed exoplanets , although a small fraction of these planets orbit stars other than their nominal _ kepler _ target star . note that the smallest planet candidates are most prone to false positives , because low snr limits the restrictiveness in parameter space of tests of lightcurve shape and location of the centroid of the transit signal on the sky plane , and the abundance of astrophysical signals that are capable of producing lower amplitude events is greater . also note that stellar parameters , and thus planetary sizes , are prone to large uncertainties for candidates that have no high - snr spectroscopic observations from the ground . from a bayesian statistical perspective , multiplicity effectively allows us to increase the planet prior , and thus enables the validation of candidates as being true planets at a specified high probability level without going to the extreme lengths often required to provide a very small upper bound on the fp prior . we use these arguments along with a detailed analysis of _ kepler _ data to validate the kepler-11 analog koi-707 system , which we name kepler-33 . was competitively selected as the tenth discovery mission . funding for this mission is provided by nasa s science mission directorate . the authors thank the many people who gave so generously of their time to make the _ kepler _ mission a success , chief among them bill borucki , who has devoted decades to developing and implementing _ kepler_. useful comments were provided by natalie batalha , ruth murray - clay , dimitar sasselov , jason steffen , and especially by tim brown . kevin zahnle and mark marley provided constructive comments on the manuscript . d. c. f. acknowledges nasa support through hubble fellowship grant # hf-51272.01-a , awarded by stsci , operated by aura under contract nas 5 - 26555 . ccccc .01 & @xmath104 & @xmath105 & @xmath106 & @xmath107 + .02 & @xmath108 & @xmath109 & @xmath110 & @xmath111 + .03 & @xmath112 & @xmath113 & @xmath114 & @xmath115 + .04 & @xmath116 & @xmath117 & @xmath118 & @xmath119 + .05 & @xmath120 & @xmath121 & @xmath122 & @xmath123 + [ tab : centroids ] ccccccccccc spectroscopy + @xmath124 ( m@xmath68 ) & 1.291 & 0.082 & 0.063 & -0.121 & 0.212 & -0.200 & 1.271 & 1.164 + @xmath57 ( r@xmath68 ) & 1.82 & 0.16 & 0.14 & -0.18 & 0.49 & -0.42 & 1.663 & 1.615 + @xmath125 & 0.0250 & 0.0019 & & & & & 0.0248 & 0.0241 + @xmath126 ( k ) & 5904 & 47 & & & & & 5899 & 5880 + log(@xmath127/l@xmath68 ) & 0.556 & 0.080 & & & & & 0.476 & 0.446 + age ( gyr ) & 4.27 & 0.87 & 0.74 & -1.03 & 2.68 & -1.65&4.31 & 5.92 + log@xmath128 ( @xmath129/s ) & 4.027 & 0.056 & & & & & 4.0994 & 4.087 + @xmath80 ( km / s)&14.09&0.1 & & & & & & & + @xmath83 ( g/@xmath130 ) & 0.300 & 0.066 & 0.049 & -0.079 & 0.230 & -0.147 & 0.3883 & 0.3885 + transit model + @xmath83 ( g/@xmath130 ) & 0.288 & 0.081 & 0.091 & -0.066 & 0.112 & -0.215 & 0.03897&0.3897 + ccccccccccc kepler-33b = koi-707.05 + @xmath84 ( bjd-2454900 ) & 64.8981 & 0.0075 & & & & & 64.8973 & 64.8973 + @xmath55 ( days ) & 5.66793 & 0.00012 & & & & & 5.667939 & 5.667939 + @xmath49 & 0.50 & 0.16 & 0.16 & -0.20 & 0.34 & -0.41 & 0.298 & 0.298 + @xmath85 & 0.00877 & 0.00046 & & & & & 0.00853&0.00853 + @xmath56 ( r@xmath40 ) & 1.74 & 0.18 & & & & & 1.549 & 1.504 + @xmath131 ( degrees ) & 86.39 & 1.17 & 0.96 & -1.62 & 2.90 & -3.16 & 88.03 & 88.03 + @xmath132 & 7.87 & 0.87 & & & & & 8.714 & 8.714 + @xmath47 ( au ) & 0.0677 & 0.0014 & & & & & 0.06740 & 0.06544 + tr@xmath133 ( ppm ) & 86.8 & 6.8 & & & & & 87.7 & 87.7 + tr@xmath134 ( hr ) & 4.88 & 0.16 & 0.16 & -0.15 & 0.45 & -0.54 & 4.80 & 4.80 + + kepler-33c = koi-707.04 + @xmath84 ( bjd-2454900 ) & 76.6764 & 0.0042 & & & & & 76.6777 & 76.6777 + @xmath55 ( days ) & 13.17562 & 0.00014 & & & & & 13.17563 & 13.17563 + @xmath49 & 0.44 & 0.17 & 0.17 & -0.21 & 0.38 & -0.38 & 0.144 & 0.144 + @xmath85 & 0.01602 & 0.00057 & & & & & 0.01560 & 0.01560 + @xmath56 ( r@xmath40 ) & 3.20 & 0.30 & & & & & 2.833&2.751 + @xmath131 ( degrees ) & 88.19 & 0.72 & 0.58 & -1.06 & 1.55 & -1.92 & 89.46&89.46 + @xmath132 & 13.8 & 1.5 & & & & & 15.292 & 15.292 + @xmath47 ( au ) & 0.1189 & 0.0025 & & & & & 0.11287 & 0.11484 + tr@xmath133 ( ppm ) & 297.7 & 9.1 & & & & & 299.8 & 299.8 + tr@xmath134 ( hr ) & 6.700 & 0.104 & 0.096 & -0.107 & 0.325 & -0.295&6.616&6.616 + + kepler-33d = koi-707.01 + @xmath84 ( bjd-2454900 ) & 122.6342 & 0.0018 & & & & & 122.6341 & 122.6341 + @xmath55 ( days ) & 21.77596 & 0.00011 & & & & & 21.775984 & 21.7759 + @xmath49 & 0.44 & 0.17 & 0.14 & -0.23 & 0.38 & -0.40 & 0 ( assumed ) & 0 ( assumed ) + @xmath85 & 0.02667 & 0.00087 & & & & & 0.02589 & 0.02589 + @xmath56 ( r@xmath40 ) & 5.35 & 0.49 & & & & & 4.700&4.564 + @xmath131 ( degrees ) & 88.71 & 0.51 & 0.61 & -0.48 & 1.08 & -1.37 & 90 ( assumed ) & 90 ( assumed ) + @xmath132 & 19.3 & 2.1 & & & & & 21.38&21.38 + @xmath47 ( au ) & 0.1662 & 0.0035 & & & & & 0.16533 & 0.16053 + tr@xmath133 ( ppm ) & 831.4 & 10.5 & & & & & 831 & 831 + tr@xmath134 ( hr ) & 8.011 & 0.097 & 0.065 & -0.103 & 0.345 & -0.192 & 7.987 & 7.987 + + kepler-33e = koi-707.03 + @xmath84 ( bjd-2454900 ) & 68.8715 & 0.0048 & & & & & 68.8720 & 68.8720 + @xmath55 ( days ) & 31.78440 & 0.00039 & & & & & 31.78438 & 31.78438 + @xmath49 & 0.47 & 0.16 & 0.14 & -0.21 & 0.38 & -0.37 & 0.204 & 0.204 + @xmath85 & 0.02011 & 0.00072 & & & & & 0.01955 & 0.01955 + @xmath56 ( r@xmath40 ) & 4.02 & 0.38 & & & & & 3.550 & 3.446 + @xmath131 ( degrees ) & 88.94 & 0.37 & 0.45 & -0.35 & 0.84 & -1.07 & 89.576 & 89.576 + @xmath132 & 24.9 & 2.8 & & & & & 27.51 & 27.51 + @xmath47 ( au ) & 0.2138 & 0.0045 & & & & & 0.2127 & 0.20656 + tr@xmath133 ( ppm ) & 467 & 12 & & & & & 469 & 469 + tr@xmath134 ( hr ) & 8.90 & 0.13 & 0.12 & -0.13 & 0.40 & -0.38 & 8.819 & 8.819 + + kepler-33f = koi-707.02 + @xmath84 ( bjd-2454900 ) & 105.5763 & 0.0040 & & & & & 105.5757 & 105.5757 + @xmath55 ( days ) & 41.02902 & 0.00042 & & & & & 41.02903 & 41.02903 + @xmath49 & 0.43 & 0.17 & 0.15 & -0.23 & 0.39 & -0.42 & 0.130 & 0.130 + @xmath85 & 0.02227 & 0.00076 & & & & & 0.02171 & 0.02171 + @xmath56 ( r@xmath40 ) & 4.46 & 0.41 & & & & & 3.941 & 3.827 + @xmath131 ( degrees ) & 89.17 & 0.34 & 0.33 & -0.42 & 0.68 & -0.98 & 89.772 & 89.772 + @xmath132 & 29.5 & 3.3 & & & & & 32.61 & 32.61 + @xmath47 ( au ) & 0.2535 & 0.0054 & & & & & 0.2522 & 0.2449 + tr@xmath133 ( ppm ) & 579 & 12 & & & & & 581 & 581 + tr@xmath134 ( hr ) & 9.87 & 0.13 & 0.11 & -0.15 & 0.43 & -0.36 & 9.732 & 9.732 +

mobile devices such as smartphones and wearable devices are increasingly gaining popularity as platforms for collecting and sharing sensor data . these devices , which are often equipped with sensors such as accelerometer , orientation sensor , magnetometer , camera , microphone , gps and so on , are being used by mobile sensing systems to make sophisticated inferences about users . these inferences have in turn enabled an entire ecosystem of context - aware apps such as behavior - based authentication @xcite , fitness monitoring based on activity tracking @xcite , speech translation @xcite and traffic / environmental monitoring @xcite . while shared data has enabled context - aware apps to provide utility to the users , the same data can also be used by an adversary to make inferences that are possibly sensitive to the user such as speaker identity @xcite , keystroke detection @xcite , location tracking @xcite , device placement @xcite , onscreen taps recognition @xcite , onset of stress @xcite and detection of emotional state @xcite . therefore , there exist fundamentally conflicting requirements between protecting privacy of the sensitive information contained in mobile sensor data and preserving utility of the same data for authorized inferences . we consider the following case studies that further illustrate this privacy - utility tradeoff : 1 . case study 1 : a user shares accelerometer data with an authentication app that performs keyless validation of user identity ( _ useful inference _ ) @xcite . but the same data can also be used to infer the activity mode of the user if she is walking , or standing still , or moving up or down the stairs ( _ sensitive inferences _ ) @xcite which in turn may lead to more serious inferences regarding living habits and health conditions . case study 2 : a user shares location and accelerometer data with a real - time life - logging app that helps with travel planning ( _ useful inferences _ ) @xcite . however , the same data can also be maliciously used to extract sequences of the user s entered text ( e.g. , passwords or pins ) ( _ sensitive inference _ ) @xcite . case study 3 : a user shares speech data with a voice - based search app that translates speech to text for web searches ( _ useful inference _ ) @xcite . but the same data can also be used to recognize the user s identity ( _ sensitive inference _ ) @xcite , compromising her privacy . recently , several sensor privacy protection mechanisms ( sppms ) have been proposed @xcite . however , most existing sppms only provide binary access control over the sensor data that requires users to choose between sharing or blocking a sensor limiting their applicability and adoption @xcite . a limited set of sppms that do focus on obfuscating sensor data , however , lack provable privacy guarantees @xcite . in this paper , we propose a new framework called deeprotect that mediates privacy - preserving access to mobile sensor data . deeprotect enables a new permission model where instead of blocking or allowing sensors , users specify their utility and privacy preferences in terms of inferences that can be derived from mobile sensor data . the overall information flow of deeprotect is summarized in figure [ fig : motivation ] . in deeprotect , users can specify as input a set of useful / authorized inferences denoted by @xmath1 ( e.g. , behavior - based authentication , speech - to - text translation ) . all other inferences ( possibly unknown ) are considered sensitive by default . alternatively , a subset of sensitive inferences can be specified by the users , which is denoted by @xmath2 ( e.g. , detection of the text entered in keyboard , speaker identity recognition ) . these useful and sensitive inferences represent functions that map the mobile sensor data to their corresponding inference results ( i.e. , labels ) , which is the information we want to share ( for useful inferences ) or protect ( for sensitive inferences ) . outgoing sensor data is initially intercepted by deeprotect , which obfuscates the data before sharing such that @xmath3 can be derived accurately by the app and @xmath4 is kept private . * approach overview * : we now illustrate the working of deeprotect by using case study 3 as an example . figure [ venn_diagram ] shows the venn diagram of information measures in deeprotect . to protect users privacy , it is required that we should not directly share the raw sensor data @xmath5 ( the entire white square shown in figure [ venn_diagram]1 ) with the untrusted third - party apps . in case study 3 , @xmath5 corresponds to the microphone data . our objective is to enable @xmath3 ( speech - to - text translation ) with high accuracy while protecting against @xmath4 ( speaker identity recognition ) . the key idea in deeprotect is to first ignore the information that is orthogonal to authorized ( useful ) inference information . in this case , we only retain information that enable @xmath3 and ignore the remaining information @xmath6 ( shown in yellow in figure [ venn_diagram]2 ) . in the next step , we perturb those information that are correlated to both the useful and sensitive inferences , @xmath7 ( shown in blue in figure [ venn_diagram]3 ) , to satisfy provable privacy guarantees . this helps in selectively hindering the sensitive inferences without severely degrading the useful inferences . for arbitrary useful and sensitive inferences , the three - stage data obfuscation pipeline used in deeprotect is illustrated in figure [ fig : pipeline ] . in the first stage , we extract features from raw sensor data and perform _ data minimization _ @xcite to guarantee that only features that are relevant to the authorized ( useful ) inferences are retained for further processing . specifically , deeprotect modifies an autoencoder network @xcite to _ automatically _ explore the inherent structural characteristics of the sensor data , and learn a sparse feature representation of the data specific to the useful inferences . at this point , some of the extracted features , although highly specific to the useful inferences , might still have some correlation with the sensitive inferences . in the next step of the pipeline , to provide provable privacy guarantees for a single user s mobile sensor stream , we exploit _ local differential privacy _ @xcite to obfuscate the extracted features . note that differential privacy @xcite is typically applied in multi - user settings , and local differential privacy can be applied in single - user settings . specifically , we develop a computationally efficient mechanism that perturbs the sparse features , using either the conventional local differential privacy framework @xcite or our relaxed notion of local differential privacy framework , depending on users specification of sensitive inferences . finally , we reconstruct obfuscated sensor data from the perturbed features such that the existing interfaces for mobile apps can continue to work without any change . * usage mode : * deeprotect supports two usage modes , depending on how sensitive inferences are specified . _ usage mode 1 : _ only the useful inference set @xmath3 is provided and the sensitive inference set @xmath8 is the default set of ` all possible inferences ' . under this mode , the user is interested in rigorous privacy guarantees with respect to the entire set ( possibly unknown ) of inferences . deeprotect first implements deep learning based feature extraction method to raw sensor data and then applies local differential privacy to these extracted features . _ usage mode 2 : _ both the useful and the sensitive inference sets are provided . under this scenario , the user is interested in protecting only a specific set of sensitive inferences ( and not all possible inferences ) . to provide provable privacy guarantees , we develop a relaxed variant of local differential privacy and use it to perturb the extracted features in a fine - grained manner . our theoretical analysis ( section [ sec_thory ] ) and experimental results on real - world datasets ( section [ sec_evaluation ] ) show that both usage modes of our approach significantly outperform previous state - of - the - art with up to 8x improvement . * contributions : * we now summarize our contributions under three main thrusts : 1 . _ provable privacy guarantees in single - user settings : _ we support rigorous local differential privacy guarantees for a single user s mobile sensor data however , this rigorous privacy guarantee can require significant addition of noise impacting application utility . thus , we propose a novel relaxed variant of local differential privacy for providing inference - based access control over the mobile sensor data . our proposed privacy metric satisfies composition properties , and can achieve improved utility by focusing on protecting a specific set of inferences . _ novel perturbation mechanisms for enhanced privacy - utility tradeoffs : _ we propose an effective perturbation mechanism which consists of two key techniques : deep learning based data minimization and feature obfuscation based data perturbation . our first technique automatically extracts features relevant to the useful / authorized inferences by extending the deep learning models for dimensionality reduction , which to our best knowledge is the first such attempt . for the second technique , we propose effective feature perturbation methods , that achieve both conventional and relaxed notions of local differential privacy guarantees . through rigorous theoretical analysis , we demonstrate the advantage of our approach over the state - of - the - art obfuscation mechanisms . implementation and evaluation : _ we implement our system using real - world datasets and mobile apps ( for all the case studies discussed previously ) . for our first feature learning step , we demonstrate that our modified deep learning network outperforms the state - of - the - art feature extraction approaches with up to @xmath9 improvement . for our end - to - end approach combining both feature extraction and perturbation , we demonstrate the advantage of our technique over the state - of - the - art obfuscation mechanisms with up to @xmath0 improvement . we also implement experiments to show the effectiveness of our method in defending against inference attacks leveraging mobile sensor data . we will provide an open source software framework for all the experiments carried out in our study for independent verification and extensions of our results . mockdroid @xcite is a modified version of the android operating system ( os ) with ability to ` mock ' a given resource requested by an app . this approach allows users to revoke access to particular resources at run - time , encouraging users to consider the trade - off between functionality and privacy . appfence @xcite offers two approaches for protecting sensitive data from untrusted apps : shadowing and blocking sensitive data from being exfiltrated off the device . identidroid @xcite is also a customized android os providing different identifications and privileges , to limit the uniqueness of device and user information . the above systems still rely on the user to determine the sensitive sensors and block , mock , or shadow them . ipshield @xcite , provides users the ability to specify their privacy preferences in terms of inferences , however , they only generate binary privacy policies of _ allow _ and _ deny _ for individual sensors . such binary access control over the sensor data may seriously affect the functionality of various apps . deeprotect on the other hand provides the much needed automatic translation from higher - level privacy specifications in terms of inferences to obfuscation of sensor data . another important limitation of the state - of - the - art sppms is that they are often heuristic in nature and fail to provide rigorous privacy guarantees for inference - based access control over sensors in mobile devices . for instance , ipshield @xcite does provide users with ability to configure obfuscation policies for raw sensor data ( through addition of random noise ) , but does not quantify any privacy guarantees for those policies . boxify @xcite provides app sandboxing for untrusted apps and can be used to run an untrusted app in an isolated environment with minimum access privileges . easedroid @xcite uses semi - supervised learning to perform automatic policy analysis and refinement . this could be used to compliment deeprotect by automatically learning the set of inferences that need to be protected and ones that need to be released from access patterns and audit logs . privacy - preserving mobile data aggregation has been studied in @xcite . however , they assume the presence of a multi - user database whereas we aim to protect inference - based privacy for a single user s sensor stream . we surveyed more than @xmath10 research papers published in relevant conferences and journals over the past @xmath11 years to form a knowledge repository of inferences ( see table [ tableknowledge ] in the appendix ) that can be made using a combination of sensors accessed by mobile apps . this table forms the universe of possible inferences over which the set of useful and sensitive inferences in deeprotect are defined . to the best of our knowledge , deeprotect is the first system that provides a novel inference - based access control in mobile sensing applications . furthermore , deeprotect 1 ) provides provable privacy guarantees including both conventional local differential privacy and our relaxed variant of local differential privacy for sensitive sensor data ; 2 ) presents an effective mechanism consisting of two key techniques : deep learning based data minimization and feature obfuscation based data perturbation ; 3 ) outperforms the state - of - the - art methods in multiple mobile sensing datasets and applications . [ sec_threat ] we assume that deeprotect together with the underlying mobile os , sensors , and system services form the trusted domain . our adversary is an untrusted app provider , who publishes apps in the marketplace for advertised useful inferences . the app accesses sensors including ones to which the user has provided explicit permission , and also others for which no permission is required . the app would then send out the collected sensor data to the adversary . our goal is to ensure that the data shared with the app can only be used for deriving the authorized ( useful ) inferences and not any sensitive inferences . we aim to limit / bound the adversary s inference of the user s sensitive information via access to the obfuscated sensor data while achieving two rigorous privacy properties : ( a ) first , we aim to provide rigorous local differential privacy guarantees @xcite , that limit an adversary against all possible inferences ( corresponding to usage mode 1 in section [ sec : intro ] ) ; and ( b ) second , we aim to provide a novel relaxed variant of local differential privacy that limits an adversary against a specified set of sensitive inferences ( corresponding to usage mode 2 in section [ sec : intro ] ) . furthermore , we adapt the popular laplace perturbation mechanism @xcite to construct effective perturbation approaches , in order to provide both conventional and relaxed notions of local differential privacy guarantees for protecting mobile sensor data ( in sections [ sec : usage1 ] , [ sec : usage2 ] ) . it is also interesting to note that a threat model similar to usage mode 2 ( i.e. , protecting against specified inferences ) has been explored in pufferfish privacy @xcite and blowfish privacy @xcite . for example , the set of sensitive inferences computed over sensor data in deeprotect is the set of potential secrets defined in pufferfish and blowfish privacy . blowfish privacy has recently been adopted by the u.s . census bureau demonstrating its usefulness as an enabler for practical deployment . differential privacy ( dp ) @xcite is a rigorous mathematical framework that prevents an attacker from inferring the presence or absence of a particular record in a statistical database . dp randomizes the query results , computed over the _ multi - user _ database , to ensure that the risk to an individual record s privacy does not increase substantially ( bounded by a function of the privacy budget @xmath12 ) as a result of participating in the database . local differential privacy ( ldp ) @xcite , as an adaption of dp , is defined under the setting where the user does not trust anyone ( not even the central data collector ) . local privacy dates back to warner @xcite , who proposed the randomized response method to provide plausible deniability for individuals responding to sensitive surveys . in our setting , we therefore apply ldp to protect a _ single _ user s mobile sensor data while satisfying rigorous privacy guarantees . similar applications have also been explored in google s rappor system @xcite . specifically , @xmath12-ldp is defined as follows : [ dpdef ] _ ( @xmath12-ldp ) @xcite _ a randomized algorithm @xmath13 provides @xmath12-ldp if for any two databases @xmath14 and for any output set @xmath15 , @xmath16 where @xmath17 ( resp.@xmath18 ) is the output of @xmath13 on input @xmath19 ( resp.@xmath20 ) and @xmath12 is the privacy budget . smaller value of @xmath12 corresponds to a higher privacy level . ) .,title="fig:",width=259,height=124 ] . for a mobile sensing system , consisting of @xmath21 sensors across @xmath22 timestamps , the temporal mobile sensor matrix @xmath23 is a real matrix where @xmath24 records the sensing data of the @xmath25-th sensor at the @xmath26-th timestamp . to fully explore the temporal characteristics of the data , we follow common practice @xcite and partition it into a series of segments according to a user - specified time window size @xmath27 as shown in figure [ twodata ] . for the @xmath28-th window , we stack the corresponding columns within it to form a column vector which is denoted by @xmath29 . the temporal mobile sensor matrix can thus be reformulated as @xmath30 $ ] , where @xmath31 . we therefore apply ldp to a single user s mobile sensor data which is segmented according to a user - specified window size . in this way , we can take the temporal dynamics of users mobile sensor data into consideration , while providing rigorous privacy guarantees . from definition [ dpdef ] , we observe that @xmath12-ldp has the same mathematical formulation as @xmath12-dp , except that the neighboring databases in ldp are _ any two possible databases without the constraint of one tuple s difference as in dp @xcite_. thus , when apply ldp to mobile sensing applications , the neighboring databases represent any two segmented sensor data that differ in _ any possible sensor recording at any timestamp within the same window_. a mechanism that provides rigorous ldp can defend against any inference attacks over the obfuscated data , according to the post - processing invariant property @xcite . therefore , we can provide rigorous ldp for users sensor data to defend against all possible inferences over the data ( corresponding to usage mode 1 in section [ sec : intro ] ) . * the need for a relaxed variant : * although we can provide rigorous ldp and defend against all inferences computed over the obfuscated sensor data ( in usage mode 1 ) , in practice , it may suffice to protect a specific subset of sensitive inferences instead of focusing on all possible inferences . therefore , we further consider a scenario where the user aims to defend against a _ specific subset of sensitive inferences _ over the sensor data ( corresponding to usage mode 2 ) . under this threat model , there are novel opportunities to strategically add noise and enhance utility . we thus relax the definition of ldp to provide provable privacy guarantees for a specific subset of sensitive inferences but significantly improve the utility performance ( will discuss in sections [ sec_thory ] , [ sec_evaluation ] ) . pufferfish @xcite and blowfish @xcite privacy frameworks have similarly motivated the threat model of protecting against a specific set of sensitive computations over the data and have been deployed in real - world settings . inspired by these frameworks and definition [ dpdef ] , we define our relaxed neighboring databases as follows : [ neigh_data ] let us represent a set of sensitive inferences as @xmath32 , and let us represent an orthogonal function of @xmath33 as @xmath34 . thus , @xmath35 , where @xmath36 and @xmath37 is the orthogonal component of @xmath32 consisting of all the functions that are orthogonal to @xmath32 . two databases @xmath38 are relaxed neighboring databases , iff . @xmath39 , @xmath40 for any function @xmath34 . we illustrate the relaxed neighboring databases above in figure [ nei ] . for concrete interpretation of the definition , we refer back to case study 3 in section [ sec : intro ] , where the useful inference is _ speech - to - text translation _ and the sensitive inference is _ speaker identity recognition_. intuitively , both the useful and sensitive inferences rely on a limited number of features extracted from the sensor data . the commonly used features for speech - to - text translation are mel - frequency cepstral coefficients ( mfcc ) , spectrogram and n - gram @xcite , while the popular features for speaker identity recognition are mfcc , spectrogram and the voice biometrics of the user ( such as her pitch , accent , etc . ) according to definition [ neigh_data ] , the relaxed neighboring databases there refer to two sensor data streams that correspond to different speakers ( i.e. , different values of mfcc , spectrogram and the voice biometrics ) and have the same values of n - gram ( the orthogonal component ) . therefore , the relaxed neighboring databases in definition [ neigh_data ] focuses on protecting the privacy of sensitive inferences only ( whose values differ in the two databases ) , and not the component of the data that is orthogonal to sensitive inferences ( whose values remain the same in the two databases ) . , which have different sensitive inference values whilst the same values for the orthogonal inferences.,title="fig:",width=259,height=67 ] . * remark on orthogonal functions : * we further illustrate the relaxed neighboring databases in a three - dimensional space ( our definition can be generally applied to any space ) in figure [ illu ] , where there are two functions @xmath41 that are orthogonal with the sensitive inference function @xmath33 , i.e. , @xmath42 . according to definition [ neigh_data ] , we have the following properties for the two relaxed neighboring databases @xmath43 : 1 ) they have the same projection on the direction corresponding to the orthogonal functions , i.e. , @xmath44 ; 2 ) they have different projection on the direction corresponding to the sensitive inference function , i.e. , @xmath45 . inspired by the relaxed neighboring databases in definition [ neigh_data ] , we now formulate a relaxed variant of local differential privacy ( rldp ) which is also an instantiation of the pufferfish and blowfish privacy frameworks . in fact , we can observe that our set of sensitive inferences computed over sensor data is the set of potential secrets defined in pufferfish privacy ( definition 3.4 in @xcite ) and blowfish privacy ( definition 4.2 in @xcite ) . [ ipdef]_(@xmath12-rldp ) _ for a set of sensitive inference functions @xmath46 and a privacy budget @xmath12 , a randomized algorithm @xmath13 satisfies @xmath12-rldp for protecting @xmath46 , if for any two relaxed neighboring databases @xmath38 defined in definition [ neigh_data ] and any output set @xmath47 , @xmath48 smaller values of @xmath12 correspond to higher privacy level . to achieve @xmath12-rldp for protecting sensitive information , it is required that the shared data @xmath49 conditioned on any two different sensitive information @xmath50 ( where @xmath39 in definition [ neigh_data ] ) , is statistically indistinguishable from each other . this in turn guarantees that an adversary gains negligible information about the true sensitive information upon observing the shared data . we re - iterate that unlike traditional ldp frameworks , its relaxed variant focuses only on the privacy of a specific subset of sensitive inferences . a perturbation mechanism developed to satisfy rldp guarantees would thus result in better utility performance than the ldp mechanisms ( will discuss in sections [ sec_thory ] , [ sec_evaluation ] ) . in definition [ neigh_data ] . @xmath41 are orthogonal functions of the sensitive inference function @xmath33.,title="fig:",width=240,height=115 ] . + . * ( sequential composition theorem)*[seq ] let randomized algorithm @xmath51 ( @xmath52 ) each provide @xmath53-rldp under the sensitive inference functions @xmath33 . the sequence of these algorithms @xmath54 provides @xmath55-rldp . * ( parallel composition theorem)*[paral ] let randomized algorithms @xmath51 ( @xmath52 ) provide @xmath53-rldp under the sensitive inference functions @xmath33 and @xmath56 be arbitrary disjoint data set . the sequence of these randomized algorithm @xmath57 provides @xmath58-rldp . in this section , we design effective perturbation mechanisms to achieve @xmath12-ldp and @xmath12-rldp ( corresponding to the two usage modes in section [ sec : intro ] ) for protecting mobile sensor data in a single - user setting . our key insight is to exploit the structural characteristics of mobile sensor data to enhance privacy - utility tradeoffs . before proposing our privacy mechanisms , we first introduce a baseline approach which directly generalizes the traditional dp perturbation mechanisms to achieve @xmath12-ldp . to achieve @xmath12-dp , the laplace perturbation mechanism ( lpm ) @xcite applies noise drawn from a suitable laplace distribution to perturb the query results . more formally , for a query function @xmath59 , lpm computes and outputs @xmath60 , where @xmath61 is the parameter of the laplacian noise and @xmath62 is the global sensitivity of @xmath59 @xcite . when applying ldp ( recall definition [ dpdef ] ) on segmented mobile sensor data @xmath63 , the neighboring databases @xmath64 may differ in all their possible tuples ( i.e. , all sensor recordings across all timestamps within the same window ) , while the neighboring databases in traditional dp frameworks only differ in one tuple . according to the _ composition theorem _ of dp @xcite , we propose a baseline approach to achieve @xmath12-ldp , which inserts a laplacian noise to each temporal sensor data point with the same parameter of @xmath65 is estimated from our collected data serving as a local sensitivity . ] , where @xmath66 is the dimension of each segmented sensor data @xmath63 ( we refer interested readers to @xcite for similar perturbation mechanisms ) . the baseline approach achieves @xmath12-ldp guarantees and the corresponding proof is deferred to the appendix to improve readability . note that the baseline approach introduces noise that is linear in the number of temporal sensor recordings within the time window . this increases the magnitude of noise that needs to be added and thus degrades the usability of the data . therefore , a better alternative to protect sensor data is to build a compact , privacy - preserving synopsis from the data by exploiting its structural characteristics . data minimization @xcite is a fundamental legal instrument that protects privacy by limiting the collection of personal data to the minimum extent necessary for attaining legitimate goals . in our work , we enforce the principle of data minimization and retain only the minimum number of features necessary to enable the authorized inferences . these extracted features may still be correlated with sensitive inferences ; thus we incorporate local differential privacy to obfuscate the extracted features for protecting against sensitive inferences . in this section , we propose deeprotect ( recall its pipeline in figure [ fig : pipeline ] ) which consists of two key steps of 1 ) first extracting features based on the data minimization principle and 2 ) then perturbing these features to achieve both conventional and relaxed notions of ldp guarantees ( corresponding to the two usage modes in section [ sec : intro ] respectively ) . we will demonstrate the significant advantages of our mechanism over the existing privacy methods theoretically as well as using multiple real - world case studies on real datasets . mobile sensor data is usually high - dimensional in nature , which typically exhibits both structure and redundancy , allowing minimization @xcite . this lays the foundation for the first technique in our deeprotect system : deep learning based data minimization . * existing autoencoder models : * deep learning models @xcite learn multi - layer transformations from the input data to the output representations , which is more powerful for feature extraction than hand - crafted shallow models . among the building blocks of these models , autoencoders @xcite automatically extract features in an unsupervised manner by minimizing the reconstruction error between the input and its reconstructed output . a single - layer autoencoder is shown in figure [ auto ] . the encoder function @xmath67 maps the input data @xmath63 to the hidden units ( features ) according to @xmath68 , where @xmath69 is typically a sigmoid function , @xmath70 is a weight matrix and @xmath71 is a bias vector . the decoder function @xmath72 maps these features back to the original input space according to @xmath73 , where @xmath74 is usually the same form as that in the encoder , @xmath75 is a weight matrix and @xmath76 a bias vector . considering the inherent structure of the mobile sensor data @xcite , we aim to learn an effective feature space on which the mobile sensor data would have a succinct representation , and the corresponding objective function is as follows . . @xmath77 where @xmath78 is the reconstruction error between the input data @xmath63 and its reconstructed output @xmath79 ( detailed mathematical formulation for @xmath80 is deferred to the appendix to improve readability ) . existing autoencoders thus aim to minimize @xmath81 with respect to @xmath82 @xcite . * constructing data minimization model : * using a nonlinear encoding function an autoencoder can typically extract better features than previous linear transformation methods @xcite . but , the features learnt are not specific to the useful ( authorized ) inferences . to handle this , we explicitly modify the autoencoder models in eq . [ cost ] by incorporating the useful inferences and other associated constraints as follows . * incorporating useful inferences : * the objective for our data minimization mechanism is to maximize the utility performance with the minimum amount of information , therefore it is important to combine the useful inferences with the autoencoder models to automatically extract features in a supervised manner . _ to the best of our knowledge , this is the first work to modify the deep learning models through incorporating the authorized inferences . _ specifically , we incorporate the minimization of cost function corresponding to the useful inferences to the objective function of existing autoencoders in eq . + we analyze the cost function for each inference using case study 1 in section [ sec : intro ] , where behavior - based authentication was considered as a useful inference and activity mode detection was deemed sensitive . both the useful and sensitive inferences can be mathematically transformed into a classification problem , which can be addressed by machine learning techniques . for instance , by leveraging the popular _ ridge regression _ technique @xcite , we can learn an optimal classifier as follows . @xmath83 where @xmath84 represents the cost function for the useful and sensitive inferences respectively . for behavior - based authentication ( _ useful inference _ ) , the label @xmath85 where @xmath86 represents the legitimate user and @xmath87 represents the adversary . similarly , for activity mode detection ( _ sensitive inference _ ) , @xmath88 where the labels @xmath89 and @xmath90 represent _ walking _ , _ standing still _ and _ moving up or down the stairs _ respectively . the optimal classifier @xmath91 learned in eq . [ krr_obj ] can be utilized to label the newly - coming mobile sensor data for behavior - based authentication or activity mode detection . the cost function in eq . [ krr_obj ] has been popularly used in machine learning community whilst other general models in @xcite can also be explored as potential cost function to reflect the prediction accuracy of the inferences . note that our analysis are not restricted to these settings and can be utilized for arbitrary inference - based mobile applications . * incorporating orthogonality : * in addition , we also aim to learn orthogonal features so that we can deal with each feature independently in our feature perturbation mechanism for achieving enhanced utility performance . therefore , we add another constraint as @xmath92 to the objective function to guarantee the _ orthogonality _ of the features ( similar intuition has also been utilized in @xcite ) . * our new autoencoder model : * incorporating the two constraints ( useful inferences and orthogonality ) into eq . [ cost ] and combining with eq . [ krr_obj ] , we construct the objective function for our data minimization method as @xmath93 where @xmath94 controls the trade - off between the reconstruction loss and the utility penalty , and @xmath92 represents the orthogonality constraint . note that the last two terms in eq . [ final_obj ] correspond to the cost function in ridge regression ( recall eq . [ krr_obj ] ) . although we only consider the cost function of ridge regression to optimize the data minimization process , our algorithm can be generalized to multiple machine learning techniques such as support vector machine @xcite , naive bayesian @xcite and random forests @xcite in a straightforward manner . * model learning and stacking : * to minimize @xmath95 in eq . [ final_obj ] with respect to @xmath96 , we explore the stochastic gradient descent ( sgd ) technique , which has been shown to perform fairly well in practice @xcite . furthermore , we explore multiple hidden layers to stack multiple model units , in order to generate even more compact and higher - level semantic features resulting in better data minimization . to improve readability , we defer all the details about model learning and stacking into the appendix , based on which we summarize our deep learning based data minimization mechanism in algorithm [ alg1 ] . ldp considers the worst - case adversary which can rigorously protect _ against all possible inferences _ computed over the data ( recall definition [ dpdef ] ) . in the absence of a user - specified set of sensitive inferences , or otherwise if the user chooses to operate under the ldp guarantees ( corresponding to usage mode 1 in section [ sec : intro ] ) , we develop our perturbation mechanism through perturbing the features learnt from the deep learning based data minimization step in section [ sec_feature ] . formally , to achieve @xmath12-ldp , deeprotect under usage mode 1 inserts an laplacian noise with parameter @xmath97 to each previously learned feature , where @xmath98 represent the dimension of the segmented sensor data @xmath63 and the features @xmath99 extracted from @xmath63 , respectively . , the number of features , i.e. , @xmath100 , needed to be perturbed is much smaller than the number of raw sampling points , i.e. , @xmath66 , in the baseline approach . ] deeprotect mechanism under usage mode 1 is summarized in algorithm [ alg3 ] , which satisfies rigorous @xmath12-ldp as will be discussed in theorem [ proofprivacy ] . our mechanism is different from the baseline approach in section [ sec_baseline ] because we add laplacian noise after the application of the deep learning based data minimization step , while the baseline approach directly adds laplacian noise to the raw sensor data without the deep learning mechanism . after perturbing these features , we reconstruct the perturbed sensor data according to the decoder function @xmath72 in autoencoder ( recall eq . [ cost ] ) . we will show that our mechanism significantly outperforms the baseline approach ( in sections [ sec_thory ] , [ sec_evaluation ] ) . note that our privacy objective is also different from that in @xcite since they aim to protect each user s training data in the deep learning training stage under the multi - user settings while in contrast we aim to protect the privacy of mobile sensor data stream in single - user settings . + perturbed sensor data @xmath101 ; + * for each * @xmath102 + 1 . extract features @xmath99 from @xmath63 by data minimiza- + -tion mechanism in algorithm [ alg1 ] ; + 2 . obtain perturbed features @xmath103 by inserting noise + of @xmath104 ; + 3 . reconstruct perturbed sensor data @xmath105 ; if the user has specified a subset of sensitive inferences ( corresponding to usage mode 2 in section [ sec : intro ] ) , we develop our perturbation mechanism ( summarized in algorithm [ alg2 ] ) by first computing the relaxed sensitivity @xmath106 , and then inserting an laplacian noise with parameter @xmath107 to each previously learned feature , to satisfy @xmath12-rldp . the detailed process to compute the relaxed sensitivity is as follows . * computing relaxed sensitivity : * similar to _ sensitivity _ in dp @xcite , our _ relaxed sensitivity _ can be computed as @xmath108 where @xmath109 are relaxed neighboring databases in definition [ neigh_data ] , and the denominator is set to make _ relaxed sensitivity _ comparable with traditional _ sensitivity _ computed over neighboring databases that differ in only one tuple . the sensitive inferences can be mathematically transformed into a classification problem , which can be addressed by machine learning techniques . the ridge regression classifier in eq . [ krr_obj ] is a linear function computed over the segmented sensor data @xmath63 , i.e. , @xmath110 . we explain the computation of relaxed sensitivity using this formulation , though our analysis can be generally applied to non - linear situations using kernel - based techniques @xcite . we apply the _ gram - schmidt orthogonalization _ technique @xcite to a matrix @xmath111 $ ] ( where @xmath112 is an _ identity _ matrix and thus @xmath111 $ ] is full - rank ) , in order to obtain orthogonal vectors of @xmath113 as @xmath114 . based on that , we form an orthogonal matrix @xmath115 $ ] . any function @xmath34 that is orthogonal of the sensitive inference function @xmath32 ( recall definition [ neigh_data ] ) can be represented by a linear combination of @xmath114 . for the two relaxed neighboring databases @xmath64 , we have @xmath116 , according to definition [ neigh_data ] . therefore , we know that @xmath117^t=[\gamma , 0,0,\cdots]^t$ ] , i.e. , @xmath118^t$ ] , and the value of @xmath119 is restricted by the range of @xmath120 . since we consider the _ identity _ query ( as we publish the sanitized sensor data to mobile apps ) , we obtain the constraint of @xmath119 as @xmath121^t\|_1=\|{\bm{x}}_{t1}-{\bm{x}}_{t2}\|_1\le\dim({\bm{x}}_t)\delta q$ ] . therefore , we can compute the relaxed sensitivity as follows . @xmath122^t\|_1\le \dim({\bm{x}}_t)\delta q}}\frac{\|{\bm{s}}^{-1}\cdot[\gamma,0,0,\cdots]^t\|_1}{\dim({\bm{x}}_t)}\ ] ] it is worth noting that @xmath123 due to the constraint of relaxed neighboring databases for achieving the same orthogonal inference values ( recall definition [ neigh_data ] ) . therefore , comparing to usage mode 1 , deeprotect under usage mode 2 would add less noise to the features extracted in the data minimization step thus achieving better utility . + perturbed sensor data @xmath101 ; + * for each * @xmath102 + 1 . extract features @xmath99 from @xmath63 by data minimiza- + -tion mechanism in algorithm [ alg1 ] ; + 2 . compute relaxed sensitivity @xmath106 in eq . [ relaxed_sensitivity0 ] ; + 3 . obtain perturbed features @xmath103 by inserting noise + of @xmath124 ; + 4 . reconstruct perturbed sensor data @xmath105 ; our perturbation mechanisms summarized in algorithms [ alg3 ] , [ alg2 ] satisfy @xmath12-ldp , @xmath12-rldp respectively , according to the following theorems . to improve readability , we defer the corresponding proofs to the appendix . [ proofprivacy ] deeprotect under usage mode 1 ( corresponding to algorithm [ alg3 ] ) satisfies @xmath12-ldp . [ proofprivacy2 ] deeprotect under usage mode 2 ( corresponding to algorithm [ alg2 ] ) satisfies @xmath12-rldp . for a perturbation algorithm @xmath125 , let us denote @xmath126 $ ] as the expected error in the release of data @xmath5 , where @xmath127 $ ] is the expectation taken over the randomness of @xmath128 . we quantify the utility advantage of deeprotect over the baseline approach in the following theorems [ the_error ] , [ the_error2 ] ( corresponding to the two usage modes ) , and the detailed proofs are deferred to the appendix . [ the_error ] for deeprotect under usage mode 1 ( corresponding to algorithm [ alg3 ] ) , the expected error @xmath129 is lower than that of the baseline approach in section [ sec_baseline ] by a factor of @xmath130 , where @xmath99 is the feature set extracted from the segmented sensor data @xmath63 by using our deep learning based data minimization approach . [ the_error2 ] for deeprotect under usage mode 2 ( corresponding to algorithm [ alg2 ] ) , the expected error @xmath129 is lower than that of the baseline approach in section [ sec_baseline ] by a factor of @xmath131 , where @xmath132 are the sensitivity and the relaxed sensitivity corresponding to the query function @xmath59 , respectively . therefore , we can see that the utility advantage of our deep learning based data minimization step in section [ sec_feature ] is @xmath130 and of our relaxed variant of local differential privacy in section [ sec : usage2 ] is @xmath133 . in this section , we describe our methodology for collecting sensor data from mobile devices , and the experimental setup ( including system parameters ) for evaluation . in our experiments , we collected data using a google nexus 5 ( with 2.3ghz , krait 400 processor , 16 gb internal storage and 2 gb ram on android 4.4 ) and a moto360 smartwatch ( with omap 3 processor , 4 gb internal storage , 512 mb ram on android wear os ) . on the smartphone , data was captured from the _ accelerometer _ , _ gyroscope _ , _ orientation _ , _ ambient light _ , _ proximity _ , and _ sensors . on the smartwatch , the _ accelerometer _ and _ gyroscope _ sensors were recorded ( recall figure [ twodata ] ) . the sampling rate was fixed at @xmath134 hz . @xmath135 graduate students ( @xmath136 males and @xmath137 females ) in our university were invited to take our smartphone and smartwatch for two weeks and use them in the same way that they would use their personal devices in their daily lives . to obtain the ground - truth information for performance evaluation , we ask the users to record labels for both the _ useful _ and _ sensitive inferences_. the labelled training data is grouped under two different categories : mode - detection data and identity - recognition data . users perform tasks such as walking , enunciating digits or specific alphabets , and the corresponding data segments are then labelled as per the tasks . the mode - detection data , correspond to labelled user activities ( e.g. , accelerometer data segments are marked with labels such as `` walking '' ) , and speech - to - text translation labels ( where audio segments are labelled with the corresponding spoken digit or alphabet ) . the identity - recognition data is used for authentication and speaker identity recognition experiments . the labelled data is generated by associating the identity of a user as label to a mobile device , on first use . we provide provable privacy guarantees for temporal mobile sensor data ( discussed above ) , which is segmented according to the parameter of time window size @xmath138 . for the deep learning based data minimization step ( in section [ sec_feature ] ) , we use 10-fold cross validation to generate the training data and testing data , where @xmath139 of our collected data is used as training data and the remaining @xmath140 is used as testing data . we repeated this process for @xmath141 iterations and reported the averaged results . in our experiments , we used stacked autoencoders ( @xmath142 in algorithm [ alg1 ] ) with two hidden layers comprising of 15 and 7 units respectively . we will show that an autoencoder with only two - hidden layers was able to extract better features than the state - of - the - art techniques . the reduced number of layers ( and units ) in the autoencoder allowed us to train the model using small amount ( 2 weeks ) of labelled data from the user ( note that we are protecting the sensitive inferences for a single user ) . we implemented all the three case studies ( tradeoff between authentication and activity recognition , tradeoff between transportation detection and text recognition , tradeoff between speech translation and speaker identification ) discussed in section [ sec : intro ] on our real - world dataset using the system parameters discussed above . in this section , we experimentally demonstrate the effectiveness of deeprotect using multiple real - world datasets and applications . we first show the advantage of our deep learning based data minimization step ( in section [ sec_feature ] ) over existing feature extraction approaches with up to 2x improvement . next , we show the advantage of our end - to - end approaches combining both feature extraction and perturbation ( in sections [ sec : usage1 ] , [ sec : usage2 ] ) over the baseline approach with up to 8x improvement . we experimentally evaluate our deep learning based data minimization mechanism in section [ sec_feature ] , using the dataset collected for case study 1 ( recall section [ sec : intro ] ) , where the useful inference is the _ behavior - based authentication_. to show the advantage of our method , we further compare it with the state - of - the - art feature extraction approaches . the discrete _ fourier _ transform ( dft ) and discrete cosine transform ( dct ) are two basic transformation techniques in the signal processing community , and the haar basis forms the fundamental wavelet transformation for time - frequency signal analysis . the principal component analysis ( pca ) technique @xcite can also be utilized to reveal hidden structure in the mobile sensor data . furthermore , we also consider blind compressive sensing ( bcs ) as a typical dictionary learning method for comparison @xcite . * higher accuracy * : figure [ feature_utility ] shows the utility - preserving performance under different feature extraction methods . note that @xmath143-axis is in log scale and @xmath144-axis is the accuracy for behavior - based authentication in case study 1 representing the ratio of correctly authenticated users . note that we use three - dimensional _ accelerometer _ measurement for behavior - based authentication and set the window size as @xmath138 ( recall section [ para ] ) , therefore the dimension of each segmented sensor data is @xmath145 . from figure [ feature_utility ] , we can see that 1 ) more features would be beneficial for improving the utility performance since the combination of multiple features would be more accurate and expressive to represent the input data ; 2 ) our method achieves higher accuracy than the state - of - the - art approaches with up to @xmath9 improvement . @xmath146 features are enough for our method to provide good utility performance with @xmath147 accuracy , while the accuracy by using all the @xmath148 features is @xmath149 . in comparison , the maximum accuracy for the baseline approaches is only @xmath150 by using @xmath146 features ; 3 ) our data minimization mechanism significantly benefits from the automatic learning process which explicitly incorporates the useful inference information into deep learning models . based on our analysis above , we constrain the dimension of features extracted by our deep learning based data minimization method to @xmath146 , whose corresponding accuracy is higher than @xmath151 ( @xmath152 ) of the accuracy achieved by using all the features . therefore , we represent each of the @xmath148-dimensional input sensor data @xmath153 with a feature set @xmath99 consisting of only @xmath146 features . * higher informativeness * : we further leverage _ informativeness _ in @xcite as an important metric to evaluate the information - theoretic relationship between each individual feature and the useful inference information . the _ informativeness _ @xmath154 of the @xmath25-th feature @xmath155 , measures the relative mutual information between the feature and the label of the utility function @xmath156 , and is computed as @xmath157 is the random variable taking the @xmath25-th value in each @xmath158 . ] . under the setting of case study 1 , @xmath159 represents the legitimate user and the adversary , respectively . @xmath160 is the entropy of the variable @xmath156 and @xmath161 is the mutual information between the random variables of @xmath155 and @xmath156 . for each feature @xmath155 , this measure of _ informativeness _ takes a value between @xmath162 and @xmath86 , where @xmath162 means that the feature contains no information about the utility label @xmath156 and @xmath86 means that the feature can completely determine the utility label @xmath156 . recall that we set the dimension of features extracted by our method to @xmath146 ( see figure [ feature_utility ] ) . we further evaluate their corresponding _ informativeness _ in figure [ informativeness ] . we can observe that the features learned by our method have much higher _ informativeness _ with up to @xmath9 improvement over previous works . therefore , our proposed method captures more expressive , higher - quality features than the existing state - of - the - art approaches . to demonstrate the effectiveness of our end - to - end perturbation mechanisms combining both feature extraction and perturbation , we again consider the _ behavior - based authentication _ as useful inference and the _ activity mode detection _ as sensitive inference ( recall case study 1 in section [ sec : intro ] ) . we first compare the baseline method with our mechanism under usage mode 1 ( corresponding to algorithm [ alg3 ] ) , since they achieve the same level of privacy guarantees for preventing all possible sensitive inferences . then , we compare the baseline method with our mechanism under usage mode 2 ( corresponding to algorithm [ alg2 ] ) , to verify the effectiveness of our rldp for protecting a specific subset of sensitive inferences . * utility advantage under both usage modes : * figure [ utiltity_epsilon ] shows the utility performance computed over the obfuscated sensor data generated by deeprotect . we can make the following important observations using figure [ acc_epsilon ] : 1 ) deeprotect achieves considerable advantage over the baseline approach with up to @xmath0 improvement in utility . this validates the effectiveness of deeprotect that not only provides rigorous privacy guarantees for protecting sensitive inferences , but also retains the utility of the perturbed data ; 2 ) deeprotect achieves better utility performance under usage mode 2 ( which only considers specific sensitive inferences ) than that under usage mode 1 ( which considers the entire set of sensitive inferences ) ; 3 ) as expected , at higher values of @xmath12 , there is an improvement in utility but at the cost of degradation in privacy ; 4 ) even at moderate value of @xmath163 which is a typical privacy budget in ldp ( similar values can also be found in google s rappor system @xcite ) , the authentication accuracy using deeprotect ( under usage mode 2 ) is close to the noise - free level ( @xmath149 in figure [ feature_utility ] ) . this observation further validates the effectiveness of our mechanisms . note that the neighboring databases in ldp / rldp may differ in all their possible tuples ( instead of differing in only one tuple as in dp ) . thus , a proper privacy budget in ldp / rldp for balancing utility and privacy is usually higher than that of dp . + to investigate the effectiveness of deeprotect on real - world mobile applications , we plot the trade - off between the accuracy of making useful inferences versus the accuracy of making sensitive inferences for all the three case studies ( corresponding to usage mode 2 and thus algorithm [ alg2 ] is applied ) , as shown in figure [ balance123 ] . * utility - privacy tradeoffs for inference - based access control : * in figure [ balance1 ] , we observe that 1 ) deeprotect achieves good inference performance for behavior - based authentication ( useful inference ) , while significantly deteriorating the activity mode detection ( sensitive inference ) . for privacy budget @xmath163 which is a typical privacy budget in ldp ( similar values have also been used in google s rappor system @xcite ) , the accuracy for inferring the useful information ( authentication ) is larger than @xmath164 while the accuracy for inferring sensitive information ( activity modes ) drops to roughly @xmath165 which is equivalent to random guessing users data in our experiments , the accuracy for random guessing is @xmath166 . ] . therefore , deeprotect can achieve practical privacy while only degrading utility performance by @xmath167 . this provides an effective guide for users to choose a proper value of @xmath12 for real world applications ; 2 ) higher levels of perturbation would degrade the inference performance for both the useful and sensitive information ; 3 ) deeprotect is effective for defending against sensitive inference attacks computed over sensor data . these observations demonstrate that deeprotect works well in practice and returns an acceptable utility performance while satisfying provable privacy guarantees . similarly , for the other two case studies , the accuracy of the sensitive inferences degrades at a much faster rate than that of the useful inferences when more noise is added ( corresponding to a smaller privacy budget @xmath12 ) , which further validates the effectiveness of deeprotect . for case study 2 , in figure [ balance2 ] , when @xmath168 , deeprotect achieves good performance for transportation detection ( useful inference ) , while significantly degrading the identification of entered text ( sensitive inference ) . for case study 3 , in figure [ balance3 ] , when @xmath169 , deeprotect achieves good performance for speech translation ( useful inference ) , while preventing the recognition of speaker identity ( sensitive inference ) . depending on the mode the user chooses to operate in deeprotect can provide either the rigorous ldp guarantee that protects all sensitive inferences ( usage mode 1 ) or our rldp that protects a user - chosen subset of inferences ( usage mode 2 ) . while rldp provides weaker privacy guarantees than ldp , it comes with the advantage of stronger utility properties . when applying ldp / rldp on mobile sensor data , the neighboring databases may differ in all their possible tuples ( i.e. , all sensor recordings across all timestamps within the same window ) , while the neighboring databases in traditional dp only differ in one tuple . therefore , a proper privacy budget in ldp / rldp for balancing privacy and utility is usually higher than that of dp @xcite . previous work in @xcite proposed to ignore several records in the database to bound sensitivity in dp for providing better privacy - utility tradeoffs . our approach of using deep learning techniques to extract authorized features is a conceptual improvement over these methods . while deeprotect has been presented in the context of privacy - aware data sharing on mobile phones , the techniques developed are not restricted to the specific setting , and can be applied to other scenarios beyond mobile phones . we will make deeprotect tool available to public as open source software . although our privacy guarantee is limited to segmented sensor data , we do take the temporal dynamics existing in mobile sensor stream into consideration according to a user - specified parameter of window size . similar to any dp - oriented metrics , our privacy guarantees also compose securely , i.e. , retain privacy guarantees even in the presence of multiple independent releases ( recall theorems [ seq ] , [ paral ] ) . while there is a non - trivial utility cost incurred by our mechanisms , deeprotect ( under both usage modes ) significantly outperforms the baseline approach . in addition , our utility performance can be further improved by 1 ) exploiting fine - grained correlation among sensors ( e.g. , among the three dimensions of accelerometer in case study 1 ) according to @xcite ; 2 ) training customized models specific to each individual user that installs deeprotect in phone ; and 3 ) training models based on a larger amount of users data as in @xcite . our perturbation mechanism can also be easily generalized to consider correlation across time windows , according to the composition properties of our privacy metrics ( refer to @xcite and theorems [ seq],[paral ] ) . an important component of our perturbation mechanism is the computation of the feature sensitivity in algorithms [ alg3],[alg2 ] . ways to accurately compute the sensitivity for arbitrary mobile applications , and deal with its possible underestimation would be a challenge we would like to address in the future . we are working towards an implementation of the auto - encoder on the android platform using one of the deep learning packages on android @xcite . the training itself could be done offline and then transferred on the phone to perform data minimization in real time @xcite . our off - line training process requires a small amount of labeled data , for example , in our experiments , we only required 20 users data collected for 2 weeks . any dp - oriented mechanism that adds noise to the data ( or corresponding query ) can risk the problem of breaking the integrity constraints of these sensors . in practice , since our framework targets simple sensors such as accelerometer , gyroscope , etc . , our decoding process is unlikely to break the integrity constraints for these sensors . therefore , apps can continue to use the perturbed sensor data generated by deeprotect without any change to their existing interface . in this paper , we propose deeprotect , a general privacy - preserving framework for inference - based access control of time - series sensor data on mobile devices . deeprotect not only supports conventional ldp guarantees , but also provides a novel relaxed variant of ldp . we further propose effective perturbation mechanisms consisting of two key steps : 1 ) we uniquely explore deep learning techniques to realize data minimization and only retain features relevant to the useful ( authorized ) inferences . this prevents the leakage of any information that is orthogonal to the useful inferences ; and 2 ) to further enhance the privacy of sensitive inferences , we perturb the extracted features , using an effective obfuscation mechanism that ensures both conventional and relaxed ldp guarantees while simultaneously maintaining the utility of the data . finally , for reasons of compatibility with existing third - party apps , we reconstruct the sensor data , from the noisy features before sharing . through theoretical analysis and extensive experiments over multiple real - world datasets , we demonstrate that compared to the state - of - the - art research , deeprotect significantly improves the accuracy for supporting mobile sensing applications while providing provable privacy guarantees . this work has been partly supported by faculty awards from google , intel and nsf . changchang liu is partly supported by ibm phd fellowship . we are very thankful to brendan mcmahan , keith bonawitz , daniel ramage , nina taft from google , and richard chow from intel for useful feedback and discussions . * sequential composition theorem : * for @xmath170 and @xmath171 , let @xmath172 and @xmath173 . for any sequence @xmath174 of outcomes @xmath175 , the probability of output @xmath174 from the sequence of @xmath176 is @xmath177 . applying the definition of rldp for each @xmath51 , we have @xmath178 . * parallel composition theorem : * for @xmath170 and @xmath171 , let @xmath179 and @xmath180 . for any sequence @xmath174 of outcomes @xmath175 , the probability of output @xmath174 from the sequence of @xmath176 is @xmath177 . applying the definition of rldp for each @xmath51 , we have @xmath181 . * proof for theorem [ the_error ] : * first , we derive the variance of a randomized algorithm @xmath13 for our deeprotect under usage mode 1 as @xmath189 . then , we compute the expected error as @xmath190 \le \mathbb{e}[\|{a}({\bm{x}}_t)-\mathbb{e}[{a}({\bm{x}}_t)]\|_1]+\mathbb{e}[\|{a}({\bm{x}}_t)-{\bm{x}}_t\|_1 ] = re\_{error}({a}({\bm{x}}_t))+\sqrt{\mathbb{e}[\|{a}({\bm{x}}_t)-{\bm{x}}_t\|_2 ^ 2 ] } = re\_{error}({a}({\bm{x}}_t))+\sqrt{var({a}({\bm{x}}_t))}$ ] . note that our deep learning based data minimization mechanism would result in a negligible reconstruction error , which makes it fairly applicable in many practical scenarios . it is likely that the reconstruction error @xmath191 is much lower than the perturbation error caused by adding noise . therefore , we approximate the utility performance of deeprotect as @xmath192 ( similar results can also be found in @xcite ) . similarly , we evaluate the expected error for the baseline approach as @xmath193=\mathbb{e}[\|(lap(\delta q/\epsilon))\|_1]=\delta q/\epsilon$ ] . comparing the utility performance for both methods , we know that deeprotect under usage mode 1 reduces the expected error of the baseline approach with a factor of @xmath194 . * proof for theorem [ the_error2 ] : * first , we derive the variance of a randomized algorithm @xmath13 for our deeprotect under usage mode 2 as @xmath190 \le \le \mathbb{e}[\|{a}({\bm{x}}_t)-\mathbb{e}[{a}({\bm{x}}_t)]\|_1]+\mathbb{e}[\|{a}({\bm{x}}_t)-{\bm{x}}_t\|_1 ] = re\_{error}({a}({\bm{x}}_t))+\sqrt{\mathbb{e}[\|{a}({\bm{x}}_t)-{\bm{x}}_t\|_2 ^ 2 ] } = re\_{error}({a}({\bm{x}}_t))+\sqrt{var({a}({\bm{x}}_t))}$ ] . similarly , we have @xmath195 . comparing the utility performance for both methods , we know that deeprotect under usage mode 2 reduces the expected error of the baseline approach with a factor of @xmath196 . [ [ model - learning - and - stacking - in - deep - learning - based - data - minimization ] ] model learning and stacking in deep learning based data minimization ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ * existing autoencoder models : * the mathematical formulations of existing autoencoder models @xcite ( expanding eq . [ cost ] ) is @xmath197 where the encoder function @xmath67 maps the input data @xmath198 to the hidden units ( features ) @xmath199 according to @xmath68 , and the decoder function @xmath72 maps the outputs of the hidden units ( features ) back to the original input space according to @xmath73 . similar to most existing deep learning methods @xcite , we choose the sigmoid function for both the encoding activation function @xmath69 and the decoding activation function @xmath74 , and we focus on the tied weights case in which @xmath200 ( where @xmath201 is the transpose of @xmath202 ) . @xmath203 is a loss function , typically the square loss @xmath204 . the second term in eq . [ cost2 ] is a weight decay term that helps prevent overfitting @xcite . the third term in eq . [ cost2 ] is a sparsity constraint @xmath205 with a pre - determined sparsity parameter @xmath206 , where @xmath207 is the average activation of hidden unit @xmath25 and @xmath208 is the kl divergence between two bernoulli random variables with mean @xmath206 and @xmath209 respectively . * model learning : * the objective for our data minimization mechanism is to minimize @xmath95 in eq . [ final_obj ] with respect to @xmath96 . we explore the stochastic gradient descent ( sgd ) technique to solve this optimization problem , which has been shown to perform fairly well in practice @xcite . note that we use sgd to only solve the convex optimization problem in eq . [ final_obj ] ( without considering the non - convex constraint of @xmath213 ) . to train our neural network , we first initialize each parameter @xmath96 to a small random value near zero , and then apply sgd for iterative optimization . before computing the exact gradient for the objective function in eq . [ final_obj ] with respect to each variable , we first perform the feedforward pass and evaluate @xmath214 . next , we compute the error terms @xmath215 as @xmath216 . finally , the gradient descents as partial derivatives can be computed as @xmath217 . _ update @xmath70 to satisfy orthogonality constraint : _ after obtaining @xmath70 from sgd , we need to modify @xmath70 to satisfy the orthogonality constraint of @xmath213 in eq . [ final_obj ] . this constraint is difficult to solve since it is non - convex and of high computational complexity . our method to overcome this constraint follows the rigorous analysis in @xcite by adopting a simple operation , @xmath218 , which sets the singular values of @xmath70 to be all ones . the above operation is conducted after every sgd step . the overall process for our deep learning based data minimization method is shown in algorithm [ alg1 ] . * model stacking : * although algorithm [ alg1 ] is effective for solving eq . [ final_obj ] , the learnt result heavily relies on the seeds , used to initialize the optimization process . therefore , we use multiple hidden layers to stack the model in order to achieve more stable performance . in other words , our data minimization mechanism can also be used to build a deep network through model stacking . for the first layer in the deep learning model , we find the optimal layer by minimizing the objective function in eq . [ final_obj ] . the representations learned by the first layer are then used as the input of the second layer , and so on so forth . using a stacked deep auto - encoder , we can learn stable and finer - grained features to better represent the raw sensor data .

after several decades during which the study of systems of many interacting particles has focused on matter fluids such as liquid helium , electron gases in solid - state materials , ultracold atom vapors , nuclear - matter fluids , or quark - gluon plasmas in colliders , photon propagation in suitably designed nonlinear optical systems is presently attracting a growing interest as a novel platform to investigate the physics of interacting bose gases , the so - called _ quantum fluids of light _ ( see ref . @xcite for a review ) . the interactions between the photons constituting the fluid of light are mediated by the kerr ( @xmath0 ) optical nonlinearity of the underlying medium . so far , numerous experimental observations have been performed for semiconductor - planar - microcavity geometries , including the demonstration of a bose - einstein - like condensation @xcite , of a superfluid flow @xcite , and of the hydrodynamic nucleation of nonlinear excitations such as solitons @xcite and quantized vortices @xcite in dilute photon gases . in the meanwhile , active investigations have addressed the possibility of realizing systems characterized by very large optical nonlinearities . correspondingly , the induced photon - photon interactions are expected to introduce strong quantum correlations within the fluid of light , generating in turn new quantum phases . in this quest @xcite , researchers have been faced to ( at least ) two main difficulties . on the one hand , obtaining strong enough nonlinearities in scalable systems to study the dynamics of a strongly interacting photon gas in a spatially extended system turns out to be a major experimental challenge . on the other hand , the dynamics of a light field in devices based on cavities is intrinsically a driven - dissipative one , which introduces severe complications in the theoretical description of such systems and which is typically very detrimental for the study of purely quantum features . an alternative platform for studying many - body physics in photon fluids is based on light propagating in a bulk nonlinear ( of kerr type ) optical medium . from elementary classical optics , it is well known that the paraxial propagation of a spectrally narrow beam of light can be described within the so - called paraxial and slowly - varying - envelope approximations ( see , e.g. , refs . @xcite ) by a nonlinear wave equation formally analogous to the gross - pitaevskii equation for the order parameter of a dilute bose - einstein condensate @xcite . as originally pioneered in refs . @xcite , this framework naturally translates upon quantization to a many - body quantum nonlinear schrdinger formalism with the roles of the propagation coordinate and time exchanged . this framework has been used in a number of theoretical works where laser - physics problems have been reformulated in the hydrodynamics language @xcite , including , e.g. , the investigation of superfluid - like behaviors in the flow of a photon fluid @xcite , of nonlinear phenomena with light waves @xcite , and of the so - called acoustic hawking radiation @xcite . from the experimental point of view , numerous works have been devoted to the study of nonlinear features that may appear in these systems , with a special attention dedicated to their relation to hydrodynamics and superfluidity aspects @xcite . a major first step in the very quantum direction of realizing a gas of strongly interacting photons in a propagating geometry has been recently reported using an optically dressed atom gas in the rydberg - eit regime as a bulk nonlinear medium : this has allowed for the experimental observation of strongly repulsing photons @xcite and , soon after , of two - photon molecular bound states @xcite . these remarkable experimental advances call for a theoretical approach that is able to describe in its full generality the many - body dynamics of strongly interacting photons propagating in a cavityless configuration . building atop the pioneering theoretical works @xcite and refs . @xcite , we report here a fully general quantum field theory of the propagating photon fluid . in this approach , the roles played by the optical - axis coordinate @xmath1 and the time parameter @xmath2 are exchanged : light propagation in the @xmath1 direction is naturally described in terms of evolution equations while the @xmath2 direction corresponds to a third spatial dimension in addition to the transverse @xmath3 and @xmath4 directions . in a paraxial - propagation configuration , light diffraction provides an effective mass to the photons in the @xmath5 plane and chromatic dispersion leads to a typically different effective mass in the @xmath2 direction . as usual , the kerr nonlinearity of the medium gives rise to photon - photon interactions . in contrast to microcavity configurations where driving and dissipation play a major role in the photon - fluid dynamics @xcite , in paraxial - propagation geometries , the quantum fluid of light follows a fully conservative hamilton dynamics starting from an initial condition determined by the incident light field and its coherence properties . the quantum state of the optical field at the end of the evolution is experimentally accessible _ via _ a measurement of the statistical properties of the transmitted light emerging from the nonlinear optical medium . as a most remarkable example of application of this formalism , we then study the transmission of a coherent light across a finite slab of weakly nonlinear medium . in this very simple configuration , the photons experience a pair of sudden jumps of the interaction parameter upon crossing the front and the back faces of the nonlinear medium . as a result of these two _ quantum quenches _ , the fluid of light gets excited and its quantum state after the conservative evolution across the nonlinear material can be reconstructed from the statistical properties of the transmitted light . in the weak - nonlinearity regime , the main excitation process consists in the emission of pairs of correlated counterpropagating bogoliubov phonons , which reflects in peculiar features in the intensity distribution and in the near- and the far - field two - body correlation functions . in its simplicity , this example illustrates the power of the conservative propagation dynamics in view of generating , detecting , and manipulating strongly correlated quantum phases of matter in photon gases , as well as of studying quantum dynamical features of many - body systems @xcite . the paper is organized as follows . first of all , in sec . [ sec : classicalwaveequation ] , we review the classical propagation equation of a paraxial beam of light in a cavityless nonlinear optical medium of kerr type . on this basis , we present in sec . [ sec : quantumtheory ] a general theory that makes it possible to describe the evolution of the quantum optical field for generic values of the nonlinear interaction parameter . in sec . [ sec : bogoliubovtheoryofquantumfluctuations ] , we discuss how the relatively small fluctuations superimposing upon a coherent - light field in the weak - nonlinearity regime can be treated within the framework of the bogoliubov theory of dilute bose - einstein condensates . as an application of this formalism , we investigate in sec . [ sec : responsetoquantumquenchesofthekerrnonlinearity ] the two - body quantum correlations resulting from the propagation of a laser beam across a slab of weakly nonlinear medium and interpret their features in terms of a dynamical casimir emission of bogoliubov collective excitations in a temporally modulated quantum fluid of light . finally , in sec . [ sec : conclusionsandoutlooks ] , we draw our conclusions and give prospects to the present work . we consider the propagation of a laser wave in a dispersive and inhomogeneous kerr dielectric for which the ( frequency - dependent ) electric susceptibility reads [ one writes down @xmath6 , with @xmath7 @xmath8 where the homogeneous contribution @xmath9 takes into account the chromatic dispersion of the medium , the linear modulation @xmath10 comes from the existence of spatial inhomogeneities or of an optical confinement , and @xmath11 is the kerr nonlinear shift of the susceptibility , proportional to the local electric intensity @xmath12 , i.e. , to the square modulus of the ( complex representation of the ) electric field . for simplicity s sake , the dielectric is assumed to be devoid of free charge carriers and nonmagnetic . we finally suppose that the optical field maintains its polarization in the course of its propagation in the medium so that a scalar approach is valid . as explained , e.g. , in ref . @xcite , this is possible for paraxial beam of light and provided @xmath13 slowly varies on space scales of the order of the optical wavelength . spin - orbit - coupling effects resulting from significant deviations from the optical axis and significant spatial variations of the electric susceptibility will be subjected to a future publication @xcite . introducing the envelope @xmath14 of the laser - wave electric field @xmath15 oscillating at the angular frequency @xmath16 as @xmath17 where `` c.c . '' stands for `` complex conjugate , '' and making use of the standard paraxial and slowly - varying - envelope approximations ( see , e.g. , refs . @xcite ) , the maxwell equations supplemented by eq . lead to the following classical wave equation for @xmath14 : @xmath18 in this equation , @xmath19 denotes the nabla operator in the @xmath20 plane and the functions @xmath21 and @xmath22 are defined as @xmath23=-\beta_{0}\,[\delta\chi(\mathbf{x}),\chi^{(3)}(\mathbf{x})]/[2\,(1+\chi_{0})]$ ] , where @xmath24 ; the parameters @xmath25 , @xmath26 , and @xmath27 are respectively the propagation constant @xmath28^{1/2}\,\omega / c$ ] ( @xmath29 is the vacuum speed of light ) of the laser wave in the @xmath30 direction , the group velocity @xmath31^{-1}$ ] of the photons in the medium , and the group - velocity dispersion @xmath32 evaluated at the carrier s angular frequency @xmath16 . the hydrodynamic interpretation of the propagation equation is mostly well known in the limiting case of a purely monochromatic wave at @xmath16 @xcite and has offered a transparent physical interpretation to a number of nonlinear - optics experiments @xcite . in this time - independent case , the first and second derivatives of the envelope @xmath33 with respect to @xmath2 vanish , in such a way that the propagation of the optical field decribed by eq . recovers the mean - field dynamics of the order parameter of a dilute two - dimensional bose - einstein condensate , @xmath34 playing the role of some external potential and @xmath35 corresponding to the effective two - dimensional boson - boson interaction constant . on the other hand , only a very few studies so far @xcite have investigated the consequences of this analogy in time - dependent regimes , where , by extension , the propagation of the photon fluid in the positive-@xmath1 direction can be written in terms of a time evolution in a three - dimensional @xmath36 space where the physical time parameter @xmath2 plays the role of a third spatial coordinate in addition to the transverse variables @xmath3 and @xmath4 . in the following , a special attention will be paid to the new hydrodynamic features that originate from this @xmath2 dependence . in order to facilitate @xmath1 ( @xmath2 ) to be viewed as a true time ( space ) parameter , we introduce the following coordinates : @xmath37 respectively homogeneous to a time and a length . in these new variables , the paraxial - propagation equation takes the form of a time - dependent gross - pitaevskii equation : @xmath38 where the electric - field envelope @xmath33 now has to be considered as a function of @xmath39 and @xmath40 and @xmath41 as functions of @xmath42 . this @xmath43 dependence of @xmath40 and @xmath41 corresponds in our language to a temporal dependence , which , as we shall see in the following , opens the way to the study of quantum - quench physics in the framework of paraxial optics . on the other hand , the fact that the medium properties do not depend on the physical time @xmath2 implies that @xmath40 and @xmath41 are independent on the spatial @xmath44 coordinate . in the gross - pitaevskii - like equation , the rigid - drift ( in the @xmath45 direction ) term @xmath46 originates from the group velocity of the photons in the medium and the kinetic operator @xmath47 involves a contribution in the ( actual - time ) @xmath45 direction in addition to the usual one in the @xmath20 plane . the anisotropy of the `` mass '' tensor @xmath48 in eq . is a natural consequence of the different origins of the effective masses @xmath49 , @xmath50 in the ( spatial ) @xmath3 , @xmath4 directions and @xmath51 in the ( temporal ) @xmath44 one : the former are due to diffraction while the latter originates from dispersion ; note that in vacuum , @xmath52 and so @xmath51 is infinite , by definition . when the medium is characterized by an anomalous group - velocity dispersion , that is , when @xmath53 , one has @xmath54 and so the mass matrix @xmath55 is positive in all the directions . in the following ( see sec . [ sec : bogoliubovtheoryofquantumfluctuations ] for a detailed discussion ) , we shall see that the dynamical stability of the fluid of light requires such a negative @xmath56 , but also repulsive photon - photon interactions , which sets @xmath57 and therefore @xmath58 , that is , that the kerr nonlinearity is self - defocusing . in order to describe the quantum features of a light beam propagating in the nonlinear medium considered in sec . [ sec : classicalwaveequation ] , the classical photon field @xmath59 verifying the paraxial wave equation must be replaced with a quantum field operator satisfying suitable boson commutation relations . to this purpose , in the present work , we perform a canonical quantization @xcite of the classical field theory from which eq . may be derived . the procedure first consists in rewriting the evolution equation in lagrangian and then hamiltonian form ( sec . [ subsec : hamiltonianformulation ] ) . section [ subsec : normalizationconstant ] is dedicated to the precise determination of the global multiplicative constant appearing in the lagrangian and the hamiltonian of the paraxial - propagation problem ; while it plays no role at the classical level , it starts having a crucial importance upon quantization . the quantization is finally accomplished in sec . [ subsec : canonicalquantization ] by replacing the conjugate fields of the classical hamiltonian theory with quantum field operators obeying equal-@xmath43 bosonic commutation relations , standardly deduced from the canonical poisson - bracket relations . of course , similar quantized wave equations describing paraxial light propagation in nonlinear optical media have been considered in the past ( see , e.g. , refs . @xcite ) and applied to concrete problems related to quantum soliton propagation @xcite : as a crucial addition to these works , we do not restrict our attention to one - dimensional fiber geometries for which only the @xmath45 coordinate matters @xcite , but fully take into account the dynamics of the optical field in the transverse @xmath5 plane . furthermore , an explicit expression for the normalization constant appearing in the boson commutators is provided . in contrast to works on the quantization of paraxial electromagnetic fields @xcite , our approach is able to naturally describe spatiotemporal light propagation and to include the possible existence of spatial inhomogeneities and/or of an optical confinement as well as of an optical nonlinearity . as we shall see later , all these features will be crucial for the theory to be applicable to the problem considered in sec . [ sec : responsetoquantumquenchesofthekerrnonlinearity ] , which requires an accurate description of the peculiar spatial , angular , and spectral correlations displayed by the laser beam emerging from the back face of the nonlinear medium . at this stage , the reader who is just interested in knowing the basic elements structuring the quantum theory in order to apply the latter to concrete quantum - optics problems can skip secs . [ subsec : hamiltonianformulation ] and [ subsec : normalizationconstant ] and go directly to sec . [ subsec : canonicalquantization ] , where one finds the commutation relations of the quantum field operator associated to the classical photon field @xmath59 [ eqs . ] , the many - body quantum hamiltonian of the paraxial - propagation problem [ eq . ] , and the corresponding heisenberg evolution equation [ eq . ] . let us introduce the lagrangian @xmath60=\int d\mathbf{r}\;\mathcal{l}(\mathbf{r},\tau),\ ] ] with the lagrangian density @xmath61.\right.\end{aligned}\ ] ] it is immediate to notice that the euler - lagrange equations of motion @xmath62 and @xmath63 deduced from eqs . and coincide with the evolution equation and its complex conjugate , respectively . the global normalization factor @xmath64 in the definition of the lagrangian density @xmath65 will be rigorously determined in sec . [ subsec : normalizationconstant ] on the basis of microscopic calculations . it plays obviously no role at the classical level but is on the contrary crucial to derive the exact commutation relations of the quantum field operators . equation being a first - order differential equation in @xmath43 , the data of the electric - field envelope @xmath59 at a given time @xmath66 is sufficient to determine the subsequent evolution of the system . as a result , the lagrangian @xmath67 , functional of @xmath68 , @xmath33 , and their derivatives with respect to @xmath43 , contains an overabundant number of dynamical variables . thus , before searching for the conjugate momenta and moving on to the hamiltonian formalism , it is convenient to eliminate the redundant dynamical variables from the lagrangian . following ref . @xcite , one starts by rewriting the lagrangian density @xmath65 as a function of @xmath69 $ ] , @xmath70 $ ] , and their spatiotemporal derivatives . in particular , the term involving @xmath71 and @xmath72 in eq . reads @xmath73 . thus , by adding @xmath74 to @xmath67 , we get a lagrangian @xmath75 which does not depend on @xmath76 and which is by definition equivalent to @xmath67 . its density @xmath77 reads @xmath78+\frac{\partial\mathcal{e}_{1}}{\partial\zeta}\,\mathcal{e}_{2}-\mathcal{e}_{1}\,\frac{\partial\mathcal{e}_{2}}{\partial\zeta}\right . \\ \label{newlagrangiandensity } & \left.-\,u(\mathbf{x}_{\perp},\tau)\,(\mathcal{e}_{1}^{2}+\mathcal{e}_{2}^{2})-\frac{g(\mathbf{x}_{\perp},\tau)}{2}\,(\mathcal{e}_{1}^{2}+\mathcal{e}_{2}^{2})^{2}\bigg\}.\right.\end{aligned}\ ] ] by means of the euler - lagrange equation relating to @xmath79 , that is , @xmath80 , one may express @xmath79 as a function of @xmath81 and its time derivative . inserting this expression of @xmath79 into @xmath75 , we finally obtain a lagrangian @xmath82 which only involves the dynamical variable @xmath83 and its time derivative @xmath84 . now let us denote by @xmath85 the conjugate momentum of @xmath81 . by definition , @xmath86}{\delta[\partial_{\tau}\mathcal{e}_{1}(\mathbf{r},\tau)]}.\ ] ] along the extremal `` path , '' @xmath80 , and as a consequence , @xmath87=\delta l'(\tau)/\delta[\partial_{\tau}\mathcal{e}_{1}(\mathbf{r},\tau)]$ ] . according to eq . , this yields @xmath88 as conjugated fields , @xmath81 and @xmath85 obey the canonical relations [ poissonbracketrelations ] @xmath89 where @xmath90 is the equal-@xmath43 poisson bracket . we are now able to write the hamiltonian of the system , @xmath91 in eq . , the hamiltonian @xcite density @xmath92 is by definition obtained by legendre transforming the lagrangian density @xmath93 , @xmath94 , which , in the light of eqs . and and recalling that @xmath95 , leads to @xmath96.\right.\end{aligned}\ ] ] before moving on to the canonical quantization of the classical field theory ( [ lagrangiandensity ] , [ hamiltoniandensity ] ) , let us determine the multiplicative constant @xmath64 in terms of the optical parameters of the electromagnetic wave . @xmath64 was introduced as a global factor in eq . to ensure that @xmath65 has the dimension of an energy per unit volume , as it has to be according to eq . . while such a normalization constant plays no role at the classical level , since it cancels out in the euler - lagrange equations of motion , it becomes important upon quantization as it determines the actual spacing between the energy levels of the system . the role of the quantized action in the bohr - sommerfeld quantization rules is the most well - known examples of such a dependence . a possible strategy to fix @xmath64 is to determine the total energy @xmath97 carried by the laser wave of complex electric field @xmath98 [ @xmath99 ; see eq . ] in the simple case where @xmath100 and @xmath101 , and compare the latter with the well - known formula of classical electrodynamics for the time - averaged energy of a quasimonochromatic electromagnetic field propagating through a dispersive , homogeneous , linear , nonmagnetic medium , that is ( see , e.g. , ref . @xcite ) , @xmath102}{d\omega}(\omega_{0})\,|\tilde{e}|^{2}+\frac{|\tilde{b}|^{2}}{\mu_{0}}\bigg\}.\ ] ] in eq . , @xmath103 $ ] denotes the ( frequency - dependent ) permittivity of the medium , @xmath104 being the one of free space , @xmath105 , where @xmath106 is the slowly varying envelope of the magnetic field ( deduced , e.g. , from maxwell - faraday s law ) , and @xmath107 is the vacuum permeability . in order to get an expression for the physical energy @xmath97 in our lagrangian formalism , one has to perform a legendre transformation of the lagrangian with respect to the actual time coordinate , @xmath2 , instead of the `` propagation one , '' @xmath43 . as the jacobian determinant @xmath108 and since the action @xmath109=\int d\tau~l(\tau)=\int d\mathbf{r}\,d\tau~\mathcal{l}(\mathbf{r},\tau)$ ] of the optical system can be alternatively defined in the @xmath110 coordinates as @xmath109=\int dt~\tilde{l}(t)=\int d\mathbf{x}\,dt~\tilde{\mathcal{l}}(\mathbf{x},t)$ ] , the lagrangian density stays invariant under the coordinate transformations : @xmath111 . thus , one finds from eq . that the conjugate momentum @xmath112 [ @xmath113 of @xmath14 [ @xmath114 in the @xmath110 coordinates is given by [ conjugatemomentumtilde ] @xmath115}{\delta[\partial_{t}\mathcal{e}(\mathbf{x},t ) ] } \\ \label{conjugatemomentumtilde2 } & = \mathcal{n}\,\bigg[\frac{i}{2\,v_{0}}\,\mathcal{e}^{\ast}(\mathbf{x},t)+\frac{d_{0}}{2}\,\frac{\partial\mathcal{e}^{\ast}}{\partial t}(\mathbf{x},t)\bigg],\end{aligned}\ ] ] as a consequence of which the hamiltonian density @xmath116-\tilde{\mathcal{l}}(\mathbf{x},t)$ ] associated to @xmath117 in the @xmath110 coordinates reads when @xmath100 and @xmath101as @xmath118.\ ] ] making the substitution @xmath119 into eq . , one deduces the hamiltonian @xmath120 as a functional of the total electric field @xmath121 : @xmath122.\right.\end{aligned}\ ] ] the temporal evolution of the envelope @xmath14 is given by the hamilton equation of motion @xmath123 , @xmath124 being the equal-@xmath2 poisson bracket not to be confused with the equal-@xmath43 poisson bracket @xmath90 defined in sec . [ subsec : hamiltonianformulation ] . inserting @xmath119 into this evolution equation , one ends up with the following equation for the total field @xmath121 : [ hamiltonequationofmotion ] @xmath125 where , making use of @xmath126 , @xmath127 is the hamiltonian encoding the time evolution of the total electric field @xmath121 . by means of eqs . and , we can write it in the form @xmath128 where @xmath129.\right.\end{aligned}\ ] ] within the framework of the paraxial and slowly - varying - envelope approximations , one can check that @xmath130 defined in eq . gives a negligible contribution to the total energy @xmath97 of the electromagnetic wave . by comparing eqs . and , one finally obtains the normalization constant @xmath64 in terms of the angular frequency @xmath16 and the propagation constant @xmath131 of the laser in the medium : @xmath132 brief discussions on the constant normalizing the lagrangian and the hamiltonian in one - dimensional fiber geometries were given in ref . @xcite and , in a bit more detailed way , in ref . @xcite . a physical interpretation of expression will be given in the next section . in order to carry out the canonical quantization of the ( classical ) hamiltonian theory presented in sec . [ subsec : hamiltonianformulation ] , one replaces the conjugated fields @xmath81 and @xmath85 with quantum field operators @xmath133 and @xmath134 ( by choice , in the heisenberg picture ) and the poisson bracket @xmath90 with the commutator @xmath135/(i\,\hbar)$ ] , where @xmath136 is the reduced planck constant . on doing so , become [ canonicalcommutationrelations ] @xmath137=i\,\hbar\;\delta(\mathbf{r}-\mathbf{r}'),\right . \\ \label{canonicalcommutationrelations2 } & \left.[\hat{\mathcal{e}}_{1}(\mathbf{r},\tau),\hat{\mathcal{e}}_{1}(\mathbf{r}',\tau)]=0,\right . \\ \label{canonicalcommutationrelations3 } \text{and}\quad&\left.[\hat{\pi}(\mathbf{r},\tau),\hat{\pi}(\mathbf{r}',\tau)]=0,\right.\end{aligned}\ ] ] from which and thanks to eq . one deduces the following equal - time , that is , equal-@xmath138 , commutation relations : [ commutationrelations ] @xmath139=\frac{\hbar\,v_{0}}{\mathcal{n}}\,\delta(\mathbf{r}-\mathbf{r}')\right . \\ \label{commutationrelations2 } \text{and}\quad&\left.[\hat{\mathcal{e}}(\mathbf{r},\tau),\hat{\mathcal{e}}(\mathbf{r}',\tau)]=0,\right.\end{aligned}\ ] ] @xmath140 being the hermitian conjugate of @xmath141 : @xmath142 . it is worth noting that , instead of reducing the overabundant number of dynamical variables in the lagrangian @xmath67 in order to avoid dealing with fields which are not independent from each other in the transition from the lagrangian formalism to the hamiltonian one and in the canonical quantization _ via _ the poisson bracket , we could also have implemented the more basic and robust dirac - bergmann quantization procedure @xcite to pass from the classical theory to the quantum one , as recently used in ref . @xcite . as detailed in footnote @xcite , dirac s procedure produces the same equal - time commutation relations for the quantum field operators as the ones deduced from the canonical quantization method adopted in this work , which fully validates our approach . from the definition @xmath143 , one notes that the commutation rules involve quantum field operators at the same `` propagation '' time @xmath138 but at different physical times @xmath2 and @xmath144 . such a writing is based on the coordinate transformation @xmath145 $ ] , which is of course legitimate only in the paraxial and slowly - varying - envelope approximations and if back - scattered waves are assumed not to exist , i.e. , if light propagation is assumed to occur only in the positive-@xmath1 direction . such a quantization procedure imposing equal-@xmath1 and different-@xmath2 ( instead of equal-@xmath2 and different-@xmath1 ) canonical commutation relations was pioneered in refs . @xcite and critically discussed in ref . it is also important to insist on the fact that the slow spatiotemporal variation condition that is assumed for the classical electric - field envelope @xmath14 directly applies to its quantum counterpart @xmath141 : a discussion of such a condition in a quantum framework is illustrated in refs . @xcite . by inserting the explicit expression of the normalization parameter @xmath64 into the commutation rule , one easily gets that the multiplicative constant in the right - hand side of eq . reads @xmath146 $ ] , yielding a simple interpretation of @xmath64 in terms of the energy density of a single - photon wavepacket of total energy @xmath147 . in the light of eqs . and , the quantized hamiltonian of the system reads , after considering the normal order , @xmath148.\right.\end{aligned}\ ] ] equation corresponds to the many - body quantum hamiltonian describing the evolution in time @xmath138 of a many - photon laser beam propagating through the bulk inhomogeneous and nonlinear optical medium of electric susceptibility . in the dielectric , the position of a point is referenced by the coordinates @xmath20 and @xmath44 . the two first contributions are the kinetic terms in the transverse @xmath5 plane and in the @xmath45 direction with different effective masses ( as discussed in sec . [ sec : classicalwaveequation ] ) , the second line describes the rigid global drift along the @xmath45 axis due the group velocity of light in the medium , and the two last terms respectively account for the spatial modulation of the electric susceptibility and for the two - photon interactions mediated by the kerr nonlinearity of the dielectric . in the theory of ultracold bose fluids , contact interactions are usually considered in place of the actual but much more complicated two - body interactions . this approximation is very helpful in simplifying the many - body quantum problem and is well accurate as long as the inter - particle distance is much larger than the range of the boson - boson interactions ( see ref . @xcite ) . here , the assumed local form of the kerr optical nonlinearity automatically leads to contact - like interactions between the photons of the light beam , that is , no dilutness condition for the photon gas is required to get the four - field interaction term @xmath149 in eq . . since no hypothesis is made on the intensity of the photon - photon interaction parameter @xmath41 , the many - body hamiltonian can describe a quantum fluid of weakly interacting photons , that is , in the gross - pitaevskii / bogoliubov regime , as well as a strongly interacting one . in what follows , we will focus our attention on the weak - nonlinearity regime , in which the quantum fluctuations of the fluid of light can be described within the framework of the well - known bogoliubov theory of dilute bose - einstein condensates ( see , e.g. , refs . motivated by the intense experimental investigations that are presently in progress @xcite , the strong - interaction regime will be the subject of future works @xcite . for the sake of completeness , it is useful to explicitly write the evolution equation of the operator @xmath141 . in the heisenberg picture , it is obtained from the system s hamiltonian as @xmath150 $ ] . using eq . and taking advantage of the commutation relations , this gives @xmath151 equation , which governs the time evolution of @xmath141 in the @xmath152 space , is simply the quantized version of the classical equation . as originally pointed out in ref . @xcite for a one - dimensional waveguide geometry , it has the form of a quantum nonlinear schrdinger equation . in an illuminated dielectric medium devoid of free charges as the one considered in this paper , quantum noise of the electromagnetic field only arises from the quantum uncertainty of the optical field , that is , in more physical terms , from the discreteness of the photon . in the case of a strongly coherent light beam propagating across a weakly nonlinear three - dimensional bulk medium , quantum noise is typically small and can be described in terms of weak - amplitude quantum fluctuations oscillating on top of a strongly classical wave . mutuating well - known results from the theory of weakly interacting ultracold atomic gases @xcite , one may develop a bogoliubov - like theory based on an expansion of the envelope operator @xmath141 of the form @xmath153 in this expression , the classical field @xmath59 , which satisfies the gross - pitaevskii - like equation , corresponds to the coherent component of the electric - field envelope and @xmath154 is a small quantum correction to @xmath59 . as the whole quantum nature of the optical field is captured in the fluctuation operator @xmath154 , the equal-@xmath43 commutation relations then totally transfer to the latter , giving [ commutationrelationsfluctuation ] @xmath155=\frac{\hbar\,v_{0}}{\mathcal{n}}\,\delta(\mathbf{r}-\mathbf{r}')\right . \\ \label{commutationrelationsfluctuation2 } \text{and}\quad&\left.[\delta\hat{\mathcal{e}}(\mathbf{r},\tau),\delta\hat{\mathcal{e}}(\mathbf{r}',\tau)]=0.\right.\end{aligned}\ ] ] linearizing the heisenberg equation of motion with respect to @xmath154 and its hermitian conjugate , one readily gets the so - called bogoliubov - de gennes equation @xmath156 which encodes the time evolution of the quantum fluctuation @xmath154 in the three - dimensional @xmath152 space and is in fact the heart of the bogoliubov approach . with respect to similar ( classical ) equations considered in the literature @xcite , this equation explicitly includes the @xmath44 dependence of the field and the corresponding effective mass , given by the group - velocity dispersion @xmath157 at @xmath158 . in sec . [ subsec : spatiallyhomogeneoussystem ] , we review the bogoliubov theory of linearized fluctuations in position- and time - independent configurations . although well established and known in the context of matter fluids , it is important to quickly review it within the nonlinear - propagating - geometry context because ( i ) of the nontrivial role of the effective mass @xmath159 in the dynamics of the luminous fluid in the temporal @xmath44 direction and ( ii ) of the existence of on - going experiments which aim at probing the phononic part of the bogoliubov excitation spectrum of a propagating fluid of light @xcite . in the simplest case where the classical background field @xmath59 is at some time @xmath160 homogeneous in all the ( @xmath3 , @xmath4 , and @xmath45 ) directions , @xmath161 with @xmath162 , and when the nonlinear optical medium is spatially homogeneous with a constant kerr coefficient @xmath0 , [ uniformsituation ] @xmath163 analytical solutions for the classical wave equation and the bogoliubov - de gennes equation are available ( see , e.g. , refs . @xcite ) . in such a configuration , the electric - field envelope @xmath59 follows a simple harmonic evolution with a linearly - evolving ( in time @xmath43 ) global phase : @xmath164 , where @xmath165 . in the theory of dilute bose - einstein condensates , the wavenumber @xmath166 correponds to the chemical potential of the bose gas @xcite . in the homogeneous situation , the elementary excitations of the fluid of light are plane waves of wavevector @xmath167 in the @xmath168 space , the solution of the bogoliubov - de gennes equation obeying the mode expansion @xmath169.\right.\end{aligned}\ ] ] the mode operators @xmath170 satisfy the same-@xmath43 commutation relations [ commutationbogoliubovoperators ] @xmath171=(2\pi)^{3}\,\frac{\hbar\,v_{0}}{\mathcal{n}}\,\delta(\mathbf{q}-\mathbf{q}')\right . \\ \label{commutationbogoliubovoperators2 } \text{and}\quad&\left.[\hat{b}(\mathbf{q},\tau),\hat{b}(\mathbf{q}',\tau)]=0,\right.\end{aligned}\ ] ] and evolve harmonically as @xmath172,\ ] ] with a frequency @xmath173 $ ] determined by the @xmath174-dependent wavenumber @xmath175},\ ] ] where @xmath176 the so - called bogoliubov amplitudes @xmath177 and @xmath178 [ @xmath179 and @xmath43 independent in the situation ] are finally given by @xmath180 by definition , they satisfy the normalization condition @xmath181 . because of the photon - photon interactions ( @xmath182 ) , @xmath178 is nonzero : this indicates that the ground state of the bogoliubov theory differs from the trivial vacuum without particles in the @xmath183 modes @xcite . modulo the @xmath184 contribution to the oscillation pulsation @xmath185 , originating from the presence of the drift term @xmath186 in the bogoliubov - de gennes equation , the wavenumber @xmath187 defined in eq . corresponds to the well - known bogoliubov dispersion relation for the elementary excitations propagating on top of a uniform dilute bose - einstein condensate at rest , the quadratic function @xmath188 given by eq . playing the role of the single - particle kinetic energy of matter - wave superfluids . note that this result stems from the conservative nature of the considered dynamics and is in contrast to the rich variety of dispersions predicted for driven - dissipative fluids of light in microcavity architectures @xcite . the graphical representation of @xmath187 is given in fig . [ bogoliubovlaw ] for repulsive photon - photon interactions , i.e. , for @xmath189 , in the two cases @xmath190 . in the anomalous - dispersion case , that is , when @xmath53 , the photon effective mass @xmath159 in the @xmath45 direction is as the photon effective mass @xmath191 in the transverse @xmath5 plane positive , and the `` kinetic energy '' @xmath188 stays as a consequence positive for any @xmath174 . if one also considers a self - defocusing nonlinearity , that is , if @xmath192 , the bogoliubov dispersion relation @xmath187 never acquires an imaginary part , which means that the photon - photon collision processes mediated by the underlying nonlinear medium do not give rise to unstable behaviors in the photon fluid . ( in units of @xmath166 ) of the elementary excitations propagating on top of a homogeneous quantum fluid of weakly interacting photons ( with @xmath192 ) at rest . in the anomalous - dispersion case [ @xmath53 , panels ( a ) and ( b ) ] , @xmath187 corresponds to the well - known bogoliubov law , linear at small @xmath174 s ( which is indicated by the black straight lines ) and quadratic at large @xmath174 s ( black parabola ) . in the normal - dispersion case [ @xmath193 , panels ( a ) and ( c ) ] , the effective mass @xmath51 in the @xmath45 direction is negative and @xmath187 has an imaginary component [ cf . panel ( c ) ] : in this case , the quantum fluid of light is dynamically unstable . ] in the case mentioned above where @xmath53 and @xmath192 , at low wavevectors , that is , when @xmath194 , where @xmath195 and @xmath196 are the healing lengths @xcite of the photon fluid in the @xmath20 plane and in the @xmath45 direction , respectively , the bogoliubov dispersion relation @xmath187 is phonon - like in all the directions [ as shown by the black straight lines in panels ( a ) and ( b ) of fig . [ bogoliubovlaw ] ] , but with different `` sound velocities '' @xcite @xmath197 and @xmath198 : @xmath199 a pump - probe measurement of @xmath200 was recently reported in ref . another experiment aiming at measuring @xmath201 in a one - dimensional waveguide configuration is presently in progress @xcite . in the opposite limit , that is , when @xmath202 , the dispersion @xmath187 of the elementary excitations takes a single - particle - like shape [ as illustrated by the black parabola in panels ( a ) and ( b ) of fig . [ bogoliubovlaw ] ] : @xmath203 in the present paraxial - optics context , the hartree interaction term @xmath166 corresponds to the usual modification of the propagation constant due to the kerr nonlinearity of the underlying medium . the situation is a bit more complicated in the case of a normal dispersion , that is , when @xmath193 : as presented in panel ( c ) of fig . [ bogoliubovlaw ] , there exists in that case a range of wavevectors @xmath174 in the @xmath45 direction for which the dispersion law @xmath187 is purely imaginary ; this signals that the quantum fluid of light is dynamically unstable against the formation of a train of pulses ( see , e.g. , refs . @xcite ) . to complete the picture , we can also mention that for focusing nonlinearities ( @xmath204 ) the low - wavevector modes in the transverse @xmath5 plane are unstable for all @xmath56 : this gives rise to a modulational instability which , in the optical language , goes under the name of filamentation instability ( see , e.g. , refs . @xcite ) . on the other hand , in this case , the instability in the temporal direction only occurs for a anomalous group - velocity dispersion ( @xmath53 ) . as a first example of application of the general quantum formalism developed in secs . [ sec : quantumtheory ] and [ sec : bogoliubovtheoryofquantumfluctuations ] , we investigate in the present section the propagation of a wide laser beam across a slab of kerr medium immersed in vacuum ( see fig . [ experimentalsetup ] ) . having in mind the reformulation of the propagation of the optical field along the @xmath1 axis in terms of a temporal evolution in time @xmath138 , it is straightforward to see how such a configuration constitutes a very simple realization of a pair of quantum quenches of the nonlinear interaction parameter in the system s hamiltonian . as the optical nonlinearity is nonzero only inside the kerr material , the first ( second ) quench occurs at the entrance ( exit ) face of the dielectric , where the value of the photon - photon interaction constant suddenly jumps from @xmath205 to @xmath206 ( from @xmath206 to @xmath205 ) . while the present work focuses on the weak - nonlinearity regime , regime within which the quantum fluctuations of the fluid of light are accurately described by the bogoliubov theory of linearized fluctuations reviewed in the previous section , application of the general theory to the strongly interacting regimes experimentally realized in refs . @xcite is definitely possible and will be subjected to future works @xcite . given the extreme simplicity of the proposed setup as compared to corresponding experiments using ultracold atomic vapors @xcite , this study demonstrates the promise of nonlinear optical systems to investigate different features of conservative quantum dynamics , including the response to quenches in the hamiltonian s parameters and the subsequent thermalization dynamics @xcite : in contrast to microcavity systems where cavity losses play a crucial role in the dynamics and quickly wash out quantum correlations @xcite , in the present propagating geometry , losses can be made arbitrarily small simply by choosing a suitable transparent dielectric medium . building on decades of expertise in quantum - optics experiments , we discuss below how a detailed information on the response of the photon fluid to the quenches of the interaction parameter can be extracted from the statistical properties of the light emerging from the slab of nonlinear material , in particular in the real - space intensity - correlation signal ( sec . [ subsubsec : correlationsinpositionspace ] ) and in the far - field angular distribution and two - body correlations of the transmitted light ( sec . [ subsubsec : correlationsinmomentumspace ] ) . an intuitive interpretation of our predictions in terms of a dynamical casimir emission of bogoliubov waves on top of a temporally modulated quantum gas of weakly interacting photons is provided and explained in detail , and the analogies and differences with standard four - wave - mixing experiments are discussed . while working with spatially finite , e.g. , gaussian , beams and extending the whole theory to the general case where @xmath207 and @xmath208 depends on @xmath209 [ cf . eq . ] is computationally demanding but conceptually straightforward , for the sake of simplicity we will restrict our attention to homogeneous geometries : in an actual experimental implementation , this assumption requires working with continuous - wave fields with a wide top - hat spatial profile propagating in a homogeneous medium , so that a uniform bulk region can be identified in the fluid of light , as done in the recent experiment @xcite . , constant kerr coefficient @xmath0 , and finite thickness @xmath210 in the @xmath1 direction is immersed in vacuum and illuminated by a wide monochromatic - plane - wave laser beam parallel to the @xmath1 axis . in such a configuration , the photon - photon interaction constant undergoes a sudden gate - shaped modulation along the optical axis . the resulting quantum correlations appearing in the transmitted light reveal the spontaneous emission of correlated counterpropagating excitations in the photon fluid . ] as sketched in fig . [ experimentalsetup ] , the slab of nonlinear material consists in a homogeneous dielectric layer of linear electric susceptibility @xmath9 and uniform kerr coefficient @xmath0 . it is assumed to be infinite in the @xmath3 and @xmath4 directions and to have a uniform thickness @xmath210 in the @xmath1 one . the @xmath211 ( @xmath212 ) surface defines the front ( back ) interface between the dielectric and vacuum . the monochromatic laser beam of angular frequency @xmath16 which illuminates perpendicularly the medium is supposed to have a wide uniform profile in the transverse @xmath5 plane so that it can be legitimately seen as an infinite plane wave . the propagation occurs along the increasing-@xmath1 axis and the propagation constant in vacuum is denoted as @xmath213 . at the angular frequency @xmath16 , the material is supposed to have an anomalous group - velocity dispersion ( @xmath53 ) and a negative kerr coefficient [ @xmath58 , i.e. , @xmath192 ; see eq . ] , so that the fluid of light is dynamically stable inside the nonlinear medium ( see sec . [ subsec : spatiallyhomogeneoussystem ] ) . back - propagating light waves originating from reflection on the interfaces separating the kerr material from vacuum would spoil the reformulation of the laser propagation in the positive-@xmath1 direction in terms of an effective time evolution . in order to avoid dealing with their existence , we assume that the @xmath214 surfaces of the nonlinear - dielectric layer are treated with a perfect antireflection coating . since its characteristic thickness in the @xmath1 direction ( of the order of a few optical wavelengths ) is typically much shorter than all the other lengths of the problem , its effect on light transmission can be summarized as a simple boundary condition guaranteeing the conservation of the energy - flux density of the electromagnetic wave the so - called poynting vector at the @xmath214 interfaces of the dielectic slab . mathematically speaking , this may be explicitly expressed as @xmath215 where @xmath216 is an arbitrarily - small positive parameter . equation allows for matching the quantum fluctuations of the optical field in vacuum ( @xmath217 ) to the ones propagating in the bulk nonlinear medium ( @xmath218 ) ; see sec . [ subsubsec : matching ] . outside the kerr layer ( when @xmath217 ) , that is , in vacuum , the ( quantized ) electric field @xmath219 of the laser wave admits the usual fourier expansion ( see , e.g. , ref . @xcite ) @xmath220}\,\hat{\alpha}(\mathbf{k})+\mathrm{h.c.},\ ] ] where `` h.c . '' stands for `` hermitian conjugate . '' in this equation , @xmath221 denotes the photon dispersion law in vacuum , ( we recall that ) @xmath104 is the vacuum permittivity , and the @xmath222 s [ @xmath223 s ] are the photon destruction ( creation ) operators in the state of wavevector @xmath224 , subject to the boson commutation relations [ photoncommutationrelations ] @xmath225=(2\pi)^{3}\,\delta(\mathbf{k}-\mathbf{k}')\right . \\ \label{photoncommutationrelations2 } \text{and}\quad&\left.[\hat{\alpha}(\mathbf{k}),\hat{\alpha}(\mathbf{k}')]=0.\right.\end{aligned}\ ] ] in vacuum , the second - quantized version of eq . reads @xmath226 , where @xmath213 is the propagation constant of the light wave along the positive-@xmath1 direction in free space . thus , by identification , one deduces from eq . that the envelope @xmath227 of the electric field in vacuum is given by @xmath228}{\varepsilon_{0}}}\right . \\ \label{envelopevacuum } & \left.\times\,e^{i[\mathbf{k}_{\perp}\cdot\mathbf{x}_{\perp}+\delta k_{z}z-\delta\omega(\mathbf{k}_{\perp},\delta k_{z})t]}\,\hat{\alpha}(\mathbf{k}_{\perp},\delta k_{z}),\right.\end{aligned}\ ] ] equation in which we made the variable change @xmath229 and in which we defined @xmath230 and @xmath231 . the @xmath232 s obey the commutation rules , assuming the exchanges @xmath233 and @xmath234 . furthermore , @xmath235 , @xmath236 , and @xmath237 in eq . are related through @xmath238 which is obtained by expanding at leading order the vacuum dispersion law @xmath221 around @xmath239 , that is , within the framework of the paraxial ( @xmath240 ) and slowly - varying - envelope ( @xmath241 and @xmath242 ) approximations considered in this paper . taking advantage of the dispersion law , one can transform the integration over @xmath243 in eq . into an integration over @xmath244 , yielding , at leading order in the paraxial and slowly - varying - envelope approximations , @xmath245}\,\hat{\alpha}[\mathbf{k}_{\perp},\delta k_{z}(\mathbf{k}_{\perp},\delta\omega)].\right.\end{aligned}\ ] ] in this way , the expression of the electric - field envelope outside the nonlinear medium is easily inserted in the @xmath246 mapping language discussed in the previous sections . defining in the integral [ timeevolutionreformulation ] @xmath247,\right.\end{aligned}\ ] ] where @xmath248 denotes the vacuum value of the normalization constant , one easily shows that eq . can be put into the form [ fluctuationvacuumbis ] @xmath249 where , introducing the input / output quantum mode operators @xmath250\,h/2\}\,\hat{a}(\mathbf{q})$ ] , @xmath251\right . \\ \label{ladderoperatorsinout } & \left.\times\,\hat{a}_{\left(\begin{smallmatrix}\mathrm{in}\\\mathrm{out}\end{smallmatrix}\right)}(\mathbf{q}).\right.\end{aligned}\ ] ] by construction , the ladder operators @xmath252 satisfy the equal-@xmath43 commutation relations [ use eqs . and ] [ photoncommutationrelationsbis ] @xmath253=(2\pi)^{3}\,\frac{\hbar\,c}{\mathcal{n}_{\mathrm{vac}}}\,\delta(\mathbf{q}-\mathbf{q}')\right . \\ \label{photoncommutationrelationsbis2 } \text{and}\quad&\left.[\hat{a}(\mathbf{q},\tau),\hat{a}(\mathbf{q}',\tau)]=0.\right.\end{aligned}\ ] ] in order to keep some homogeneity in the notations used outside and inside the nonlinear material , the coordinate referencing a point along the @xmath2 axis is conveniently defined in vacuum in the same way as in the kerr layer : @xmath44 [ eq . ] , where @xmath254 denotes the group velocity in the kerr medium ( i.e. , in the @xmath218 region ) at the laser s angular frequency @xmath16 . finally , note that the @xmath255 s [ @xmath256 s ] are related to the usual input ( output ) photon operators @xmath257 s [ @xmath258 s ] in the @xmath259 ( @xmath260 ) vacuum through eq . . in the specific configuration investigated here , with a coherent monochromatic - plane - wave incident light beam , the optical field in vacuum can be expanded as @xmath261 i.e. , as a classical background field @xmath262 ( with @xmath263 ) plus a small quantum - fluctuation term @xmath154 . the latter may be decomposed according to the same plane - wave expansion as the one , eq . , of the total electric - field envelope : the assumption of having a purely coherent incident beam translates into the condition that all the incident modes other than the coherent - pump one are in their vacuum state @xmath264 , by definition such that @xmath265 , which correspondingly reads in terms of the @xmath255 operators as @xmath266 . inside the nonlinear medium ( when @xmath218 ) , the ( quantized ) electric - field envelope @xmath141 evolves in time @xmath138 in the @xmath143 space according to the heisenberg equation of motion with the prescription . within the bogoliubov weakly nonlinear regime , the background electric - field envelope @xmath59 is of constant amplitude @xmath33 and of global phase @xmath165 ( see sec . [ subsec : spatiallyhomogeneoussystem ] ) , where @xmath267 is the time which ( naturally ) initializes the evolution of the optical field in the medium . from the continuity of the poynting vector at the entrance face of the nonlinear material , i.e. , from eq . for @xmath211 [ @xmath268 , one finds @xmath269 on the other hand , from the continuity of the poynting vector at the exit face of the medium , i.e. , from eq . for @xmath212 [ @xmath270 , one gets that the phase @xmath271 of the optical field in the @xmath260 vacuum is locked at , using eqs . , [ continuityconditionsbackground2 ] @xmath272 finally , the small quantum fluctuation @xmath154 which superimposes upon the mean - field solution @xmath164 is described by eqs . , with @xmath33 and @xmath273 satisfying the constraints . the relation between the fluctuations after ( @xmath260 ) and before ( @xmath259 ) the gate - shaped modulation of the optical nonlinearity along the radiation axis are obtained by matching eqs . and ( [ fluctuationvacuumbis ] , [ lasercontribution ] ) at the @xmath214 interfaces using the continuity equation . at the entrance face of the dielectric plate , that is , at @xmath211 [ @xmath268 , one has @xmath274\,e^{-i\beta_{0}h/2},\right.\end{aligned}\ ] ] in such a way that , taking advantage of @xmath181 and of the second of eqs . , @xmath275.\right.\end{aligned}\ ] ] at the back face of the kerr layer , that is , at @xmath212 [ @xmath270 , the continuity of the poynting vector yields @xmath276\,e^{i\beta_{0}h/2},\right.\end{aligned}\ ] ] from which and using @xmath173 $ ] , eq . , and eq . one finally obtains @xmath277 where we have defined [ newbogoliubovamplitudes ] @xmath278 with [ newbogoliubovamplitudesbis ] @xmath279.\right.\end{aligned}\ ] ] from the input - output relation , it is immediate to extract predictions for all the coherence properties of the transmitted light field past the slab of nonlinear medium . as already mentioned in the end of sec . [ subsubsec : outside ] , our assumption of a perfectly coherent pump incident on the dielectric corresponds to having all the fluctuation modes of the photon field in the vacuum state @xmath264 [ @xmath266 ] . as a consequence , denoting @xmath281 the quantum average in the state @xmath264 , one initially has , just before entering the nonlinear material , @xmath282 and , thanks to the commutation relation , @xmath283 while the average light intensity remains equal to the incident one in vacuum , the effect of the quenches of the photon - photon interaction constant along the optical axis is visible in the intensity - intensity correlation function @xmath284 at a fixed point @xmath260 of the laser - beam axis , i.e. , at a fixed time @xmath285 after the quench sequence . in eq . , @xmath286 is the ( quantum ) intensity operator at @xmath287 and the colon symbol denotes the normal - ordering operation . since we are dealing with a strongly coherent light wave , the electric - field envelope @xmath141 at a time @xmath285 after the second quench of the photon - photon interaction constant is given by the expansion , where the weak quantum modulation @xmath154 obeys the plane - wave decomposition . at leading order in the small fluctuations @xmath154 and @xmath288 , the intensity operator @xmath289 may be expressed as [ intensityoperatorbogoliubov ] @xmath290 is a small quantum correction to the background light intensity @xmath291 outside the nonlinear medium . taking advantage of eqs . and , the intensity correlator becomes , at leading order in the expansion , @xmath292 physically , the second term in the right - hand side of eq . is the shot - noise term originating from the discreteness , i.e. , the quantum nature , of the photon in vacuum ; mathematically , it is the consequence of the normal order considered in eq . . straightforward manipulations using eqs . , , , , , , and lead to @xmath293\right . \\ \label{intensitycorrelations } & \left.+\,\mathcal{v}_{\mathbf{q}}\exp\bigg[i\,\frac{\mathbf{q}_{\perp}^{2}}{2\,k_{0}}\,\big(v_{0}\,\tau-\tfrac{h}{2}\big)\bigg]\bigg|^{2}-1\bigg\}.\right.\end{aligned}\ ] ] the physics of these intensity correlations is most transparent at the exit face of the kerr layer , i.e. , at @xmath294 . in this case , the correlation function admits a very simple integral formulation ; manipulating the @xmath295 s and the @xmath296 s given by eqs . and denoting @xmath297 $ ] , one indeed finds [ intensitycorrelationsexitface ] @xmath298\sin^{2}[k_{\mathrm{b}}(\mathbf{q})\,h]\right . \\ \notag & \left.=-\frac{2}{\pi^{2}}\,\frac{2\,v_{0}}{c}\,\frac{\hbar\,\omega_{0}}{\varepsilon_{0}^{\phantom{2}}\,\mathcal{e}_{0}^{2}\,\xi_{\perp}^{2}\,\xi_{\zeta}^{\phantom{2}}}\int_{0}^{\infty}d\kappa\;\mathrm{sinc}(\kappa\,\rho)\right . \\ \label{intensitycorrelationsexitface2 } & \left.\phantom{=}\times\frac{\kappa^{2}}{\kappa^{2}+4}\,\sin^{2}\bigg(\frac{\ell}{2}\,\kappa\,\sqrt{\kappa^{2}+4}\bigg),\right.\end{aligned}\ ] ] where @xmath195 and @xmath196 are the healing lengths in the @xmath20 plane and in the @xmath45 direction in the dielectric , with @xmath299 the internal light intensity , @xmath300 is the thickness of the nonlinear medium in units of @xmath301 , @xmath302 is the so - called sine cardinal function , and [ relativedistance ] @xmath303 is the dimensionless relative spatiotemporal distance , @xmath304 and @xmath305 being naturally measured in units of @xmath306 and @xmath307 , respectively . the graph of the two - body correlator @xmath308 [ eqs . ] as a function of the dimensionless relative distance @xmath309 [ eqs . ] is represented in fig . [ correlations ] . as one can note from this plot , @xmath308 is characterized by a pronounced dip localized around @xmath310 , that is , at equal @xmath311 in the transverse plane and equal time @xmath2 : @xmath312 . this antibunching is a direct consequence of the repulsive character ( @xmath192 ) of the photon - photon interactions in the nonlinear medium and has been widely studied in the context of many - body physics ( see , e.g. , refs . note that the nontrivial spatial structure of @xmath308 shown in fig . [ correlations ] makes the latter conceptually different from the single - mode antibunching behavior that is obtained , e.g. , by resonantly scattering light off a two - level emitter @xcite or by photon blockade in a strongly nonlinear cavity @xcite . ( red curve ) and @xmath313 ( black curve ) . the oscillation pattern appearing after @xmath314 ( indicated by the red dotted line ) or @xmath315 ( black dotted line ) is the signature of the occurrence of a dynamical casimir emission of elementary excitations in the quantum fluid of light . ] the most important feature of @xmath308 is the fringe pattern that is visible at larger @xmath309 s . the latter is concentrated in the @xmath316 region and can be interpreted from simple physical arguments @xcite . when the sudden change of the photon - photon interaction strength from zero to @xmath182 occurs at @xmath211 [ @xmath268 , pairs of correlated bogoliubov waves of opposite wavevectors @xmath317 are spontaneously emitted in the @xmath3 , @xmath4 , and @xmath45 directions . inside the plate [ that is , for @xmath318 , a bogoliubov excitation of wavevector @xmath174 propagates at the group velocity ( @xmath173 $ ] ; see sec . [ subsec : spatiallyhomogeneoussystem ] ) [ groupvelocity ] @xmath319\right . \\ \label{groupvelocity3 } \text{and}\quad&\left.\mathbf{v}_{\zeta}(\mathbf{q})=v_{0}\,\bigg[\frac{\partial k_{\mathrm{b}}}{\partial q_{\zeta}}(\mathbf{q})+1\bigg]\,\mathbf{u}_{\zeta},\right.\end{aligned}\ ] ] @xmath320 ( @xmath321 , @xmath322 ) being a unit vector in the @xmath3 ( @xmath4 , @xmath45 ) direction . thus , as @xmath323 is a growing function of @xmath174 [ see the definition of @xmath187 in eq . ] , the modes forming a pair of entangled excitations of opposite momenta must be separated at a time @xmath324 [ and such that @xmath325 by a distance [ transverseseparation ] @xmath326 in the @xmath20 plane and by a distance [ temporalseparation ] @xmath327 along the @xmath44 direction , where @xmath328 are the `` sound speeds '' ( see sec . [ subsec : spatiallyhomogeneoussystem ] ) in the transverse and temporal directions inside the nonlinear medium . as a result , according to and , the ( dimensionless ) distance @xmath309 separating two correlated counterpropagating modes has to be @xmath329 and so @xmath330 , with @xmath331 at the back face of the nonlinear medium where @xmath294 . at this point , the correlation between counterpropagating bogoliubov excitations generated by the quench of the optical nonlinearity at @xmath211 results in an oscillatory structure in the @xmath308 function that is concentrated in the @xmath330 region , which we easily verify on the examples of fig . [ correlations ] where @xmath332 ( red curve ) and @xmath313 ( black curve ) . the amplitude of the fringe pattern diminishes as the separation distance @xmath309 grows : the correlations between bogoliubov excitations of opposite momenta are naturally all the weaker as the excitations are separated in space and time . this oscillatory behavior of the correlation function in position space can be interpreted as resulting from a dynamical casimir emission @xcite of elementary excitations on top of a quantum fluid of light presenting a quench of the photon - photon interaction constant in time @xmath138 . this has been theoretically studied in ref . @xcite in the case of a weakly interacting bose - einstein condensate presenting a quick time modulation of the atom - atom @xmath333-wave scattering length . from the experimental point of view , the closely related phenomenon of sakharov oscillations @xcite , resulting from the interference of synchronously generated counterpropagating acoustic waves , has been recently observed in the density fluctuations of cesium superfluids after an abrupt quench in time of the atom interaction strength _ via _ feshbach resonances @xcite : this last observation is of particular interest given its connection to the temperature fluctuations in the cosmic microwave background radiation @xcite . the emission of correlated - phonon pairs in the fluid of light due to the two quantum quenches of the photon - photon interaction constant can also be detected in momentum space by taking a far - field and spectrally resolved picture of the light emerging from the back face of the nonlinear medium . in the optical language , this process of emission of photon pairs at wavevectors and frequencies different from the incident ones goes under the name of spontaneous four - wave mixing . a simple measurement that could be done would be the one of the photon momentum distribution @xmath334 at a fixed point @xmath260 of the radiation axis , that is , at a fixed time @xmath285 , as a function of the wavevector @xmath174 . reminding the definition of @xmath174 , i.e. , @xmath335 , this measurement involves ( i ) an angular resolution to isolate the light deflected with a transverse wavevector @xmath235 and ( ii ) a spectral resolution to isolate the angular - frequency component of the transmitted light at @xmath336 . by construction , the population operator @xmath337 in the state of wavevector @xmath174 does not depend on time @xmath43 [ see eq . ] and one simply has @xmath338 for all @xmath285 . thus , using eqs . , one obtains [ momentumdistribution ] @xmath339^{2}\sin^{2}[k_{\mathrm{b}}(\mathbf{q})\,h]\right . \\ \label{momentumdistribution3 } & \left.=a\,v_{0}\,\delta t\,\frac{\hbar\,\omega_{0}}{\varepsilon_{0}}\,\frac{4}{\kappa^{2}\,(\kappa^{2}+4)}\sin^{2}\bigg(\frac{\ell}{2}\,\kappa\,\sqrt{\kappa^{2}+4}\bigg),\right.\end{aligned}\ ] ] where @xmath340 and @xmath341 ( @xmath342 ) empirically denote the typical transverse cross section of the beam of light and the typical duration of the measurement @xcite , respectively , @xmath300 is , as before , the normalized slab thickness , and @xmath343 is the dimensionless total excitation wavenumber , @xmath344 and @xmath184 being naturally measured in units of @xmath345 and @xmath346 , respectively . the plot of the spontaneous - four - wave - mixing intensity distribution @xmath334 is shown in fig . [ occupationnumber ] for @xmath332 . the most noticeable feature of the momentum distribution is its oscillation profile originating from @xmath347 : depending on the value of the wavevector @xmath174 , the two - body emission processes at the front and the back interfaces reveal constructive or destructive interferences at well - defined periods . the latter are determined by the finite thickness @xmath210 of the dielectric layer and the bogoliubov dispersion @xmath187 _ via _ the product @xmath348 : for example , from eq . , the @xmath174 s for which one has destructive interferences , i.e. , zero minima in the momentum distribution , are such that [ periodicity ] @xmath349 where @xmath350 is an integer @xmath351 . the quantum nature of the emitted particles can be assessed _ via _ a measurement of the two - point correlation function in the far - field regime , namely , @xmath352 function which quantifies the correlations between the different angular and spectral components of the four - wave - mixing emission . one can evaluate by means of wick s theorem [ indeed , since the bogoliubov - de gennes equation is linear , the corresponding hamiltonian the so - called bogoliubov - de gennes hamiltonian admits a quadratic dependence on the fluctuations @xmath353 and @xmath354 , and so , on the ladder operators @xmath355 and @xmath356 . using eq . which connects the outgoing excitations to the ingoing ones and eqs . and , we end up finding , for all @xmath285 , @xmath357.\ ] ] . it reveals the occurrence of constructive and destructive interferences at well - defined periods ; the @xmath358 s [ given by eqs . ] correspond to the @xmath174 s for which one has destructive - interference effects . the main plot focuses on the oscillations of the momentum distribution while the inset shows the whole function . ] as expected on general optical grounds and quantitatively predicted by eq . , the two - body correlation signals in reciprocal space have to be concentrated along the ( diagonal ) @xmath359 and ( antidiagonal ) @xmath360 lines . the first one reflects the vacuum fluctuations in the occupation @xmath361 while the second one is a clear signature of a complete correlation between excitations of opposite wavevectors @xmath362 . the thermal analogue of this phenomenon was recently observed in a quasi - one - dimensional atomic bose - einstein condensate whose speed of sound was modulated _ via _ a suitable time modulation of the transverse confinement @xcite ; even though this resulted in the creation of thermal phonons of equal and opposite velocities , temperature was too high to assess the quantum origin of the observed correlations . as a final point , it is important to highlight the novelty of our predictions with respect to standard nonlinear - optics four - wave - mixing effects , which are also described by paraxial wave equations @xcite similar @xcite to the classical counterpart of the bogoliubov - de gennes equation . as typical four - wave - mixing experiments benefit from maximizing the intensity of the four - wave - mixing signal , they are typically implemented in a phase - matched regime where both energy and momentum are conserved in the process : in our language , this requires that there exist @xmath244 pairs for which @xmath363 , which implies that the transverse ( @xmath191 ) and temporal [ @xmath159 ] masses have opposite signs . as we have seen in sec . [ subsec : spatiallyhomogeneoussystem ] , this is often associated to ( absolute or , in some cases , only convective @xcite ) instabilities in the quantum fluid of light . our angle on this physics is significantly different , as we are interested in the response of the photon fluid to quantum quenches . in an optical language , the energy that is deposed in the fluid by the quenches corresponds to the nonconservation of the normal ( i.e. , @xmath1 ) component of the wavevector at the @xmath214 interfaces . while the four - wave - mixing intensity that results from the quantum quenches is substantially weaker than the one expected for a perfectly phase - matched configuration , it encodes a detailed information on the conservative quantum dynamics of the fluid of light in response to the sudden jumps of the interaction parameter , which is the main object of the present work . as a specific and most illuminating example of such features , the evolution of the photon fluid is ruled by the bogoliubov dispersion law @xmath187 defined in eq . , which involves the optical nonlinearity in a very nonperturbative way . the measurement of the spatial @xmath5 component of @xmath187 , based on a bulk nonlinear medium , was recently reported in ref . @xcite ; another experiment in a one - dimensional waveguide geometry is underway to measure the temporal @xmath44 part of the bogoliubov dispersion @xcite . in this article , we have used a general quantum theory describing paraxial light propagation in a bulk @xmath0 medium in terms of quantum nonlinear schrdinger evolution equations as pioneered in refs . @xcite to theoretically investigate the quantum statistical properties of a laser beam emerging from a finite slab of kerr material . by mapping light propagation onto a temporal evolution and by viewing the physical time as an effective spatial coordinate , our predictions may be straightforwardly interpreted in the many - body - physics language as the response of a gas of many interacting photons to a pair of quantum quenches of the system s hamiltonian in the nonlinear interaction parameter . this exemplifies the potential of nonlinear optics in propagating geometry as a novel platform for experimental studies of time - dependent problems in many - body physics and quantum statistical mechanics @xcite . in the first sections of the present paper , we have reviewed the theory underlying the space @xmath365 time mapping and we have extended it to a fully three - dimensional configuration . we have then provided a comprehensive overview of its features and its power from a many - body - physics perspective . in contrast to quantum fluids of light in planar - microcavity geometries that have been considered and widely studied during the last decade @xcite , the evolution of the quantum photon field in a propagating configuration is fully conservative and not subject to the unavoidable radiative ( or nonradiative ) losses inherent to cavity devices : starting from an initial many - body state that is determined by the statistical properties of the incident beam , quantum coherence can be maintained in the quantum fluid of light over macroscopic evolution times . this gives access to a _ plethora _ of many - body dynamical quantum phenomena that can be eventually reconstructed from the statistical properties of the transmitted light using standard quantum - optics techniques . in the second part , as a simplest , yet most remarkable example of application of the quantum formalism , we have studied in a detailed way the response of a spatially homogeneous quantum fluid of light to the pair of quenches of the interaction constant that it feels upon crossing the front and the back faces of a @xmath0 material immersed in vacuum . in the standard optical configuration where the incident light beam is coherent and the kerr medium weakly nonlinear , we have demonstrated that the quantum quenches of the optical nonlinearity lead to the emission of pairs of correlated counterpropagating bogoliubov waves on top of the laser fluid . in particular , we have pointed out how the peculiar quantum features of this process can be experimentally identified in the angular and spectral distribution as well as in the two - body correlation functions of the transmitted light in both the near and the far field . from a general standpoint , our results illustrate the power of well - known nonlinear - optics setups as a workhorse to investigate many - body phenomena presently of great interest in quantum statistical mechanics . while we have so far focused on the configuration of easiest experimental implementation , this work paves the way to many other interesting directions : as a few examples , one can mention the study of the strong quantum fluctuations of the photon - field phase in reduced dimensions @xcite , of the so - called hawking radiation in analog black - hole configurations @xcite , and , on a longer run , of the strong quantum correlations that naturally appear in the presence of significant nonlinearities at the single - photon level @xcite . we acknowledge dimitris g. angelakis , daniele faccio , stefano finazzi , pjotrs grisins , christian miniatura , and luciano vanzo for stimulating discussions . this work was supported by the erc through the qgbe grant , by the eu - fet proactive grant aqus , project no . 640800 , and by the autonomous province of trento , partly through the siquro project ( `` on silicon chip quantum optics for quantum computing and secure communications '' ) . j. kasprzak , m. richard , s. kundermann , a. baas , p. jeambrun , j. m. j. keeling , f. m. marchetti , m. h. szymaska , r. andr , j. l. staehli , v. savona , p. b. littlewood , b. deveaud , and le si dang , _ bose - einstein condensation of exciton polaritons _ , http://www.nature.com/nature/journal/v443/n7110/full/nature05131.html[nature ( london ) * 443 * , 409 ( 2006 ) ] . a. amo , j. lefrre , s. pigeon , c. adrados , c. ciuti , i. carusotto , r. houdr , e. giacobino , and a. bramati , _ superfluidity of polaritons in semiconductor microcavities _ , http://www.nature.com/nphys/journal/v5/n11/full/nphys1364.html[nat . * 5 * , 805 ( 2009 ) ] . a. amo , s. pigeon , d. sanvitto , v. g. sala , r. hivet , i. carusotto , f. pisanello , g. lemnager , r. houdr , e. giacobino , c. ciuti , and a. bramati , _ polariton superfluids reveal quantum hydrodynamic solitons _ , http://www.sciencemag.org/content/332/6034/1167[science * 332 * , 6034 ( 2011 ) ] . g. nardin , g. grosso , y. lger , b. pitka , f. morier - genoud , and b. deveaud - pldran , _ hydrodynamic nucleation of quantized vortex pairs in a polariton quantum fluid _ , http://www.nature.com/nphys/journal/v7/n8/full/nphys1959.html[nat . phys . * 7 * , 635 ( 2011 ) ] . d. sanvitto , s. pigeon , a. amo , d. ballarini , m. de giorgi , i. carusotto , r. hivet , f. pisanello , v. g. sala , p. s. s. guimaraes , r. houdr , e. giacobino , c. ciuti , a. bramati , and g. gigli , _ all - optical control of the quantum flow of a polariton condensate _ , http://www.nature.com/nphoton/journal/v5/n10/full/nphoton.2011.211.html[nat . photonics * 5 * , 610 ( 2011 ) ] . y. lai and h. a. haus , _ quantum theory of solitons in optical fibers . i. time - dependent hartree approximation _ , http://journals.aps.org/pra/abstract/10.1103/physreva.40.844[phys a * 40 * , 844 ( 1989 ) ] . f. p. mattar and j. teichmann , _ fluid formulation of high intensity laser beam propagation using lagrangian coordinates _ , http://www.sciencedirect.com/science/article/pii/0010465581900758[comput . * 22 * , 1 ( 1981 ) ] . g. dekel , v. fleurov , a. soffer , and c. stucchio , _ temporal dynamics of tunneling : hydrodynamic approach _ , http://journals.aps.org/pra/abstract/10.1103/physreva.75.043617[phys . a * 75 * , 043617 ( 2007 ) ] . e. g. khamis , a. gammal , g. a. el , yu . g. gladush , and a. m. kamchatnov , _ nonlinear diffraction of light beams propagating in photorefractive media with embedded reflecting wire _ , http://journals.aps.org/pra/abstract/10.1103/physreva.78.013829[phys . a * 78 * , 013829 ( 2008 ) ] . g. dekel , o. v. farberovich , a. soffer , and v. fleurov , _ nonlinear dynamic phenomena in macroscopic tunneling _ , http://www.sciencedirect.com/science/article/pii/s0167278908002649[physica d * 238 * , 1475 ( 2009 ) ] . e. cohen , s. muenzel , o. v. farberovich , j. fleischer , v. fleurov , and a. soffer , _ jetlike tunneling from a trapped vortex _ , http://journals.aps.org/pra/abstract/10.1103/physreva.88.043833[phys . rev . a * 88 * , 043833 ( 2013 ) ] . v. fleurov and r. schilling , _ regularization of fluctuations near the sonic horizon due to the quantum potential and its influence on hawking radiation a * 85 * , 045602 ( 2012 ) ] . s. bar - ad , r. schilling , and v. fleurov , _ nonlocality and fluctuations near the optical analog of a sonic horizon _ , http://journals.aps.org/pra/abstract/10.1103/physreva.87.013802[phys . a * 87 * , 013802 ( 2013 ) ] . m. vaupel , k. staliunas , and c. o. weiss , _ hydrodynamic phenomena in laser physics : modes with flow and vortices behind an obstacle in an optical channel _ , http://journals.aps.org/pra/abstract/10.1103/physreva.54.880[phys . a * 54 * , 880 ( 1996 ) ] . w. wan , s. muenzel , and j. w. fleischer , _ wave tunneling and hysteresis in nonlinear junctions _ , http://journals.aps.org/prl/abstract/10.1103/physrevlett.104.073903[phys . lett . * 104 * , 073903 ( 2010 ) ] . d. vocke , t. roger , f. marino , e. m. wright , i. carusotto , m. clerici , and d. faccio , _ experimental characterisation of nonlocal photon superfluids _ , https://www.osapublishing.org/optica/abstract.cfm?uri=optica-2-5-484[optica * 2 * , 484 ( 2015 ) ] . t. peyronel , o. firstenberg , qi - yu liang , s. hofferberth , a. v. gorshkov , t. pohl , m. d. lukin , and v. vuleti , _ quantum nonlinear optics with single photons enabled by strongly interacting atoms _ , http://www.nature.com/nature/journal/v488/n7409/full/nature11361.html[nature ( london ) * 488 * , 57 ( 2012 ) ] . o. firstenberg , t. peyronel , qi - yu liang , a. v. gorshkov , m. d. lukin , and v. vuleti , _ attractive photons in a quantum nonlinear medium _ , http://www.nature.com/nature/journal/v502/n7469/full/nature12512.html[nature ( london ) * 502 * , 71 ( 2013 ) ] . e. m. wright , _ quantum theory of soliton propagation in an optical fiber using the hartree approximation _ , http://journals.aps.org/pra/abstract/10.1103/physreva.43.3836[phys . a * 43 * , 3836 ( 1991 ) ] . b. crosignani , p. di porto , and a. treppiedi , _ finite - dimensional hamiltonian approach to the quantum theory of soliton propagation _ , http://iopscience.iop.org/1355-5111/7/2/001/[quantum semiclassic * 7 * , 73 ( 1995 ) ] . a. b. matsko and v. v. kozlov , _ second - quantized models for optical solitons in nonlinear fibers : equal - time versus equal - space commutation relations _ , http://journals.aps.org/pra/abstract/10.1103/physreva.62.033811[phys . a * 62 * , 033811 ( 2000 ) ] . a. polkovnikov , k. sengupta , a. silva , and m. vengalattore , _ colloquium : nonequilibrium dynamics of closed interacting quantum systems _ , http://journals.aps.org/rmp/abstract/10.1103/revmodphys.83.863[rev . * 83 * , 863 ( 2011 ) ] . m. elazar , v. fleurov , and s. bar - ad , _ all - optical event horizon in an optical analog of a laval nozzle _ , http://journals.aps.org/pra/abstract/10.1103/physreva.86.063821[phys . a * 86 * , 063821 ( 2012 ) ] . m. elazar , s. bar - ad , v. fleurov , and r. schilling , _ an all - optical event horizon in an optical analogue of a laval nozzle _ , http://link.springer.com/chapter/10.1007/978-3-319-00266-8_12[lect . notes phys . * 870 * , 275 ( 2013 ) ] . our choice of referring to @xmath366 as a hamiltonian follows from its interpretation as an evolution operator however , as it generates propagation in the @xmath1 direction , some reader may object that it is related to the @xmath1 component of the linear momentum of the underlying field theory . for the sake of transparency , throughout this paper , we shall stick to the first wording . in dirac s approach @xcite , the fields @xmath33 and @xmath68 are canonically conjugated ( in the sense of the same-@xmath43 poisson bracket @xmath90 ) to @xmath367 and @xmath368 , respectively . as a result , the theory is characterized by two primary constraints , @xmath369 and @xmath370 , which may be shown to be of second class . therefore , @xmath90 becomes irrelevant and must be replaced in the evolution equations by the equal-@xmath43 dirac bracket @xmath371 . making use of the explicit definition of @xmath371 in terms of @xmath90 @xcite , straightforward calculations yield @xmath372 and @xmath373 , which gives the quantum equal-@xmath43 commutation relations after the replacement of @xmath33 with @xmath374 , @xmath68 with @xmath375 , and @xmath371 with @xmath135/(i\,\hbar)$ ] . the results deduced from dirac s theory hence exactly coincide with the ones deduced from the canonical quantization method , legitimating our approach . r. ganapathy , b. a. malomed , and k. porsezian , _ modulational instability and generation of pulse trains in asymmetric dual - core nonlinear optical fibers _ , http://www.sciencedirect.com/science/article/pii/s0375960106001940[phys . a * 354 * , 366 ( 2006 ) ] . l. berg , c. goudard , j. schjdt - eriksen , and h. ward , _ filamentation patterns in kerr media vs. beam shape robustness , nonlinear saturation and polarization states _ , http://www.sciencedirect.com/science/article/pii/s0167278902007406[physica d * 176 * , 181 ( 2003 ) ] . i. carusotto , s. fagnocchi , a. recati , r. balbinot , and a. fabbri , _ numerical observation of hawking radiation from acoustic black holes in atomic bose - einstein condensates _ , http://iopscience.iop.org/1367-2630/10/10/103001[new j. phys . * 10 * , 103001 ( 2008 ) ] . p. deuar , a. g. sykes , d. m. gangardt , m. j. davis , p. d. drummond , and k. v. kheruntsyan , _ nonlocal pair correlations in the one - dimensional bose gas at finite temperature _ , http://journals.aps.org/pra/abstract/10.1103/physreva.79.043619[phys . a * 79 * , 043619 ( 2009 ) ] . t. jacqmin , j. armijo , t. berrada , k. v. kheruntsyan , and i. bouchoule , _ sub - poissonian fluctuations in a 1d bose gas : from the quantum quasicondensate to the strongly interacting regime _ , http://journals.aps.org/prl/abstract/10.1103/physrevlett.106.230405[phys . lett . * 106 * , 230405 ( 2011 ) ] . larr , a. recati , i. carusotto , and n. pavloff , _ quantum fluctuations around black hole horizons in bose - einstein condensates _ , http://journals.aps.org/pra/abstract/10.1103/physreva.85.013621[phys a * 85 * , 013621 ( 2012 ) ] . p. calabrese and j. cardy , _ time dependence of correlation functions following a quantum quench _ , http://journals.aps.org/prl/abstract/10.1103/physrevlett.96.136801[phys . . lett . * 96 * , 136801 ( 2006 ) ] . g. t. moore , _ quantum theory of the electromagnetic field in a variable - length one - dimensional cavity _ , http://scitation.aip.org/content/aip/journal/jmp/11/9/10.1063/1.1665432[j . phys . * 11 * , 2679 ( 1970 ) ] . s. a. fulling and p. c. w. davies , _ radiation from a moving mirror in two dimensional space - time : conformal anomaly _ , http://rspa.royalsocietypublishing.org/content/348/1654/393[proc . london , ser . a * 348 * , 393 ( 1976 ) ] . i. carusotto , r. balbinot , a. fabbri , and a. recati , _ density correlations and analog dynamical casimir emission of bogoliubov phonons in modulated atomic bose - einstein condensates _ , http://link.springer.com/article/10.1140%2fepjd%2fe2009-00314-3[eur . j. d * 56 * , 391 ( 2010 ) ] . a. d. sakharov , _ the initial stage of an expanding universe and the appearance of a nonuniform distribution of matter _ , zh . eksp . teor . fiz . * 49 * , 345 ( 1965 ) ; http://www.jetp.ac.ru/cgi-bin/e/index/e/22/1/p241?a=list[sov . jetp lett . * 22 * , 241 ( 1966 ) ] . . hung , v. gurarie , and c. chin , _ from cosmology to cold atoms : observation of sakharov oscillations in a quenched atomic superfluid _ , http://www.sciencemag.org/content/341/6151/1213[science * 341 * , 1213 ( 2013 ) ] . to treat the infinite factor @xmath376 appearing in eq . , one replaces the @xmath377 function in momentum space with @xmath378 , where @xmath379 denotes a large cylindrical volume in the @xmath380 space whose cross section @xmath340 in the @xmath5 plane and length @xmath381 in the @xmath2 direction have the natural interpretations given in the main text . the same reasoning holds for eq . . jaskula , g. b. partridge , m. bonneau , r. lopes , j. ruaudel , d. boiron , and c. i. westbrook , _ acoustic analog to the dynamical casimir effect in a bose - einstein condensate _ , http://journals.aps.org/prl/abstract/10.1103/physrevlett.109.220401[phys . lett . * 109 * , 220401 ( 2012 ) ] . q. lin , o. j. painter , and g. p. agrawal , _ nonlinear optical phenomena in silicon waveguides : modeling and applications _ , https://www.osapublishing.org/oe/abstract.cfm?uri=oe-15-25-16604[opt . exp . * 15 * , 16604 ( 2007 ) ] . x. zhang , s. gao , and s. he , _ optimal design of a silicon - on - insulator nanowire waveguide for broadband wavelength conversion _ , http://www.jpier.org/pier/pier.php?paper=08120601[prog . 89 * , 183 ( 2009 ) ] . s. gao , z. li , e .- k . tien , s. he , and o. boyraz , _ performance evaluation of nondegenerate wavelength conversion in a silicon nanowire waveguide _ , http://www.osapublishing.org/jlt/abstract.cfm?uri=jlt-28-21-3079[j . lightwave technol . * 28 * , 3079 ( 2010 ) ] . m. blaauboer , a. g. kofman , a. e. kozhekin , g. kurizki , d. lenstra , and a. lodder , _ superluminal optical phase conjugation : pulse reshaping and instability _ , http://journals.aps.org/pra/abstract/10.1103/physreva.57.4905[phys . a * 57 * , 4905 ( 1998 ) ] . g. izs , m. santagiustina , m. s. miguel , and p. colet , _ pattern formation in the presence of walk - off for a type ii optical parametric oscillator _ , https://www.osapublishing.org/josab/abstract.cfm?uri=josab-16-9-1592[j . b * 9 * , 1592 ( 1999 ) ] . m. blaauboer , d. lenstra , and a. lodder , _ optical josephson effect using phase - conjugation mirrors : an analogy with superconductors _ , http://www.nit.eu / archive?itemid=71&option = com_czasopisma&view = wynikiszukania&jezyk = en&nr_mies=&rok=&autor = blaauboer&tytul= & opis=&slowakl=&limit=50[jtit * 1**2 * , 24 ( 2000 ) ] .

fifteen years , the area of channel coding has been revolutionized by the practical realization of capacity - approaching coding schemes , initiated by the invention of turbo codes and their associated decoding algorithms in 1993 @xcite . a few years after the invention of the turbo coding schemes , researchers became aware that gallager s low - density parity - check ( ldpc ) block codes and message - passing iterative decoding , first introduced in @xcite , were also capable of capacity - approaching performance . the analysis and design of these coding schemes quickly attracted considerable attention in the literature , beginning with the work of wiberg @xcite , mackay and neal @xcite , and many others . an irregular version of ldpc codes was first introduced by luby _ _ in @xcite , and analytical tools were presented in @xcite to obtain performance limits for graph - based message - passing iterative decoding algorithms , such as those suggested by tanner @xcite . for many classes of channels , these tools have been successfully employed to design families of irregular ldpc codes that perform very well near capacity @xcite . moreover , for the binary erasure channel these tools have enabled the design of families of irregular ldpc codes that are not only capacity - approaching but in fact capacity - achieving ( see @xcite and references therein ) . the convolutional counterparts of ldpc block codes are ldpc convolutional codes . analogous to ldpc block codes , ldpc convolutional codes are defined by sparse parity - check matrices , which allow them to be decoded using iterative message - passing algorithms . recent studies have shown that ldpc convolutional codes are suitable for practical implementation in a number of different communication scenarios , including continuous transmission and block transmission in frames of arbitrary size @xcite . two major methods have been proposed in the literature for the construction of ldpc convolutional codes , two methods that in fact started the field of ldpc convolutional codes . the first method was proposed by tanner @xcite ( see also @xcite ) and exploits similarities between quasi - cyclic block codes and time - invariant convolutional codes . the second method was presented by jimnez - feltstr " om and zigangirov in @xcite and relies on a matrix - based unwrapping procedure to obtain the parity - check matrix of a periodically time - varying convolutional code from the parity - check matrix of a block code . @xmath0 the aims of this paper are threefold . first , we show that these two ldpc convolutional code construction methods , once suitably generalized , are in fact tightly connected . we establish this connection with the help of so - called graph covers . a second aim is to discuss a variety of ldpc convolutional code constructions . although the underlying principles are mathematically quite straightforward , it is important to understand how they can be applied to obtain convolutional codes with good performance and attractive encoder and decoder architectures . a third aim is to make progress towards a better understanding of where the observed `` convolutional gain '' comes from , and what its costs are in terms of decoder complexity . the paper is structured as follows . after some notational remarks in section [ sec : notation:1:basic ] , we discuss the basics of ldpc convolutional codes in section [ sec : ldpc : convolutional : codes:1 ] . in particular , in that section we give a first exposition of the ldpc convolutional code construction methods due to tanner and due to jimnez - feltstr " om and zigangirov . in section [ sec : graph : covers:1 ] we discuss two types of graph - cover code constructions and show how they can be used to connect the code construction methods due to tanner and due to jimnez - feltstr " om and zigangirov . based on these insights , section [ sec : variations : unwrapping:1 ] presents a variety of ldpc convolutional code constructions ( along with simulation results ) , and in section [ sec : other : ldpc : code : constructions:1 ] we mention some similarities and differences of these constructions compared to other recent code constructions in the literature . afterwards , in section [ sec : analysis ] we analyze some aspects of the constructed ldpc convolutional codes and discuss some possible reasons for the `` convolutional gain , '' before we conclude the paper in section [ sec : discussion ] . we use the following sets , rings , and fields : @xmath1 is the ring of integers ; @xmath2 is the set of non - negative integers ; @xmath3 is the field of size two ; @xmath4 $ ] is the ring of polynomials with coefficients in @xmath3 and indeterminate @xmath5 ; and @xmath6 / \langle x^r{-}1 \rangle}$ ] is the ring of polynomials in @xmath4 $ ] modulo @xmath7 , where @xmath8 is a positive integer . we also use the notational short - hand @xmath9}$ ] for @xmath6 / \langle x^r{-}1 \rangle}$ ] . by @xmath10 and @xmath9}^n$ ] , we mean , respectively , a row vector over @xmath3 of length @xmath11 and a row vector over @xmath9}$ ] of length @xmath11 . in the following , if @xmath12 is some matrix , then @xmath13_{j , i}$ ] denotes the entry in the @xmath14-th row and @xmath15-th column of @xmath12 . note that we use the convention that indices of vector entries start at @xmath16 ( and not at @xmath17 ) , with a similar convention for row and column indices of matrix entries . ( this comment applies also to semi - infinite matrices , which are defined such that the row and column index sets equal @xmath2 . ) the only exception to this convention are bi - infinite matrices , where the row and column index sets equal @xmath1 . finally , @xmath18 will denote the kronecker product of the matrices @xmath19 and @xmath20 . this section defines ldpc convolutional codes and discusses why they are interesting from an implementation perspective . afterwards , we review two popular methods of obtaining ldpc convolutional codes by unwrapping block codes . later in this paper , namely in section [ sec : graph : covers:1 ] , we will use graph covers to show how these two methods are connected , and in section [ sec : variations : unwrapping:1 ] we will see how these two methods can be implemented and combined to obtain ldpc convolutional codes with very good performance . a semi - infinite binary parity - check matrix as in at the top of this page defines a convolutional code @xmath21 as follows . namely , it is the set of semi - infinite sequences given by @xmath22 where @xmath23 denotes the transpose of a vector or of a matrix . we comment on several important aspects and properties of the code @xmath21 and its parity - check matrix @xmath24 . * if the submatrices @xmath25 , @xmath26 , @xmath27 , have size @xmath28 with @xmath29 , then @xmath21 is said to have ( design ) rate @xmath30 . * the parameter @xmath31 that appears in is called the syndrome former memory . it is an important parameter of @xmath21 because the maximal number of non - zero submatrices per block row of @xmath24 is upper bounded by @xmath32 . * the quantity @xmath33 is called the constraint length of @xmath21 . it measures the maximal width ( in symbols ) of the non - zero area of @xmath24 . gives only an upper bound on the maximal width ( in symbols ) of the non - zero area of @xmath24 , but this upper bound will be good enough for our purposes . ] * we do not require that for a given @xmath34 the submatrices @xmath35 are independent of @xmath36 , and so @xmath21 is in general a _ time - varying _ convolutional code . * if there is a positive integer @xmath37 such that @xmath38 for all @xmath34 and all @xmath39 , then @xmath37 is called the period of @xmath24 , and @xmath21 is _ periodically time - varying_. * if the period @xmath37 equals @xmath17 , then @xmath24 is called _ time - invariant _ , and the parity - check matrix can be simply written as @xmath40 . \label{eq : tipcc } \end{aligned}\ ] ] * if the number of ones in each row and column of @xmath24 is small compared to the constraint length @xmath41 , then @xmath21 is an ldpc convolutional code . * an ldpc convolutional code @xmath21 is called @xmath42-regular if , starting from the zeroth column , @xmath24 has @xmath43 ones in each column , and , starting from the @xmath44-th row , @xmath24 has @xmath45 ones in each row . if , however , there are no integers @xmath31 , @xmath43 , and @xmath45 such that @xmath21 is @xmath42-regular , then @xmath21 is called irregular . of course , there is some ambiguity in the above definition . namely , a periodically time - varying ldpc convolutional code with parameters @xmath37 , @xmath46 , and @xmath47 can also be considered to be a periodically time - varying ldpc convolutional code with parameters @xmath48 , @xmath49 , @xmath50 , and @xmath51 for any integer @xmath52 that divides @xmath37 . in particular , for @xmath53 we consider the code to be a time - invariant ldpc convolutional code with parameters @xmath54 and @xmath55 . an advantage of ldpc convolutional codes compared to their block code counterparts is the so - called `` fast encoding '' property . as a result of the diagonal shape of their parity - check matrices , many ldpc convolutional codes enjoy simple shift register based encoders . even randomly constructed ldpc convolutional codes can be formed in such a way as to achieve this feature without any loss in performance ( see , e.g. , @xcite ) . on the other hand , in order to have a simple encoding procedure for ldpc block codes , either the block code must have some sort of structure @xcite or the parity - check matrix must be changed to a more easily `` encodable '' form @xcite . the difficulty in constructing and decoding ldpc convolutional codes is dealing with the unbounded size of the parity - check matrix . this is overcome at the code construction step by considering only periodically time - varying or time - invariant codes . the code construction problem is therefore reduced to designing just one period of the parity - check matrix . for decoding , the most obvious approach is to terminate the encoder and to employ message - passing iterative decoding based on the complete tanner graph representation of the parity - check matrix of the code . although this would be straightforward to implement using a standard ldpc block code decoder , it would be wasteful of resources , since the resulting ( very large ) block decoder would not be taking advantage of two important aspects of the convolutional structure : namely , that decoding can be done continuously without waiting for an entire terminated block to be received and that the distance between two variable nodes that are connected to the same check node is limited by the size of the syndrome former memory . in order to take advantage of the convolutional nature of the parity - check matrix , a continuous sliding window message - passing iterative decoder that operates on a window of size @xmath56 variable nodes , where @xmath57 is the number of decoding iterations to be performed , can be implemented , similar to a viterbi decoder with finite path memory @xcite . this window size is chosen since , in a single iteration , messages from variable ( or check ) nodes can be passed across a span of only one constraint length . thus , in @xmath57 iterations , messages can propagate only over a window of size @xmath57 constraint length . ( see also the recent paper by papaleo _ _ @xcite , which investigates further reducing the window size for codes operating on a binary erasure channel ( bec ) . ) another simplification is achieved by exploiting the fact that a single decoding iteration of two variable nodes that are at least @xmath32 time units apart can be performed independently , since the corresponding bits can not participate in the same parity - check equation . this allows the parallelization of the @xmath57 iterations by employing @xmath57 _ independent identical processors _ working on different regions of the parity - check matrix simultaneously , resulting in the parallel pipeline decoding architecture introduced in @xcite . the pipeline decoder outputs a continuous stream of decoded data after an initial decoding delay of @xmath58 received symbols . the operation of this decoder on the tanner graph of a simple time - invariant rate-@xmath59 convolutional code with @xmath60 and @xmath61 is illustrated in figure [ fig : decoder ] . is usually much larger than typical values of @xmath41 for `` classical '' convolutional codes . therefore the value @xmath61 of the convolutional code shown in figure [ fig : decoder ] is not typical for the codes considered in this paper . ] although the pipeline decoder is capable of fully parallelizing the iterations by using @xmath57 independent identical processors , employing a large number of hardware processors might not be desirable in some applications . in such cases , fewer processors ( even one processor ) can be scheduled to perform subsets of iterations , resulting in a serial looping architecture @xcite with reduced throughput . this ability to balance the processor load and decoding speed dynamically is especially desirable where very large ldpc convolutional codes must be decoded with limited available on - chip memory . further discussion on the implementation aspects of the pipeline decoder can be found in @xcite . in this subsection we discuss two approaches for deriving convolutional codes from block codes , in particular for deriving ldpc convolutional codes from ldpc block codes . the first technique will be the unwrapping due to tanner and the second will be the unwrapping due to jimnez - feltstrm and zigangirov ( jfz ) . in section [ sec : graph : covers:1 ] we will see , with the help of graph covers , how these two seemingly different unwrapping techniques are connected with each other . the term _ unwrapping _ , in particular unwrapping a quasi - cyclic block code to obtain a time - invariant convolutional code , was first introduced in a paper by tanner @xcite ( see also @xcite ) . that paper describes a link between quasi - cyclic block codes and time - invariant convolutional codes and shows that the free distance of the unwrapped convolutional code can not be smaller than the minimum distance of the underlying quasi - cyclic code . this idea was later extended in @xcite . consider the quasi - cyclic block code @xmath62 defined by the polynomial parity - check matrix @xmath63 of size @xmath64 , i.e. , @xmath65}^n \bigm| { { { { \bm{h}}}}_{\mathrm{qc}}^{(r)}}(x ) \cdot { { { { \bm{v}}}}}(x)^{\mathsf{t}}= { { { \bm{0}}}}^{\mathsf{t}}\bigr\}.\end{aligned}\ ] ] here the polynomial operations are performed modulo @xmath7 . the tanner unwrapping technique is simply based on dropping these modulo computations . more precisely , with a quasi - cyclic block code @xmath62 we associate the convolutional code @xmath66^n \bigm| { { { { \bm{h}}}}_{\mathrm{conv}}}(d ) \cdot { { { { \bm{v}}}}}(d)^{\mathsf{t}}= { { { \bm{0}}}}^{\mathsf{t}}\bigr\}\end{aligned}\ ] ] with polynomial parity - check matrix @xmath67 here the change of indeterminate from @xmath5 to @xmath68 indicates the lack of the modulo @xmath69 operations . ( note that in we assume that the exponents appearing in the polynomials in @xmath63 are between @xmath16 and @xmath70 inclusive . ) it can easily be seen that any codeword @xmath71 in @xmath72 maps to a codeword @xmath73 in @xmath62 through @xmath74 a process which was described in @xcite as the wrapping around of a codeword in the convolutional code into a codeword in the quasi - cyclic code . the inverse process was described as unwrapping . having introduced the unwrapping technique due to tanner , we move on to discuss the unwrapping technique due to jfz @xcite , which is another way to unwrap a block code to obtain a convolutional code . the basic idea is best explained with the help of an example . [ example : unwrap:1 ] consider the parity - check matrix with size @xmath75 , of a rate-@xmath76 block code . as indicated above , we can take a pair of scissors and `` cut '' the parity - check matrix into two pieces , whereby the cutting pattern is such that we repeatedly move @xmath77 units to the right and then @xmath78 unit down . having applied this `` diagonal cut , '' we repeatedly copy and paste the two parts to obtain the bi - infinite matrix shown in figure [ fig : jjj:1 ] . this new matrix can be seen as the parity - check matrix of ( in general ) a periodically time - varying convolutional code ( here the period is @xmath79 ) . it is worth observing that this new matrix has the same row and column weights as the matrix that we started with . rows are reduced . ] this example can be formalized easily . namely , starting with an @xmath64 parity - check matrix @xmath80 of some block code , let @xmath81 . then the `` diagonal cut '' is performed by alternately moving @xmath82 units to the right and then @xmath83 units down ( i.e. , @xmath84 . the resulting convolutional code has rate @xmath85 , syndrome former memory @xmath86 , constraint length @xmath87 , and period @xmath88 . analogous to the comment at the end of section [ sec : definition : ldpc : conv : codes:1 ] , it is also possible to cut the matrix @xmath80 in larger step sizes , e.g. , moving @xmath89 units to the right and @xmath90 units down , for any integer @xmath52 that divides @xmath91 , thereby obtaining a periodically time - varying convolutional code with rate @xmath92 , syndrome former memory @xmath93 , constraint length @xmath94 , and period @xmath95 . ( see also the discussion in section [ sec : time : varying : ldpc : convolutional : codes:1 ] . ) in the rest of this paper , the term `` jfz unwrapping technique '' will also stand for the following generalization of the above procedure . namely , starting with a length-@xmath11 block code @xmath96 defined by some size-@xmath97 parity - check matrix @xmath80 , i.e. , @xmath98 we write @xmath80 as the sum @xmath99 of a collection of matrices @xmath100 . the convolutional code @xmath21 is then defined to be @xmath101 where @xmath102 referring to the notation introduced in section [ sec : definition : ldpc : conv : codes:1 ] , the matrix @xmath24 is the parity - check matrix of a time - invariant convolutional code . however , depending on the decomposition of @xmath80 and the internal structure of the terms in that decomposition , the matrix @xmath24 can also be ( and very often is ) viewed as the parity - check matrix of a time - varying convolutional code with non - trivial period @xmath37 . in order to illustrate the generalization of the jfz unwrapping technique that we have introduced in the last paragraph , observe that decomposing @xmath80 from example [ example : unwrap:1 ] as @xmath103 with yields a convolutional code with parity - check matrix @xmath24 whose bi - infinite version equals the matrix shown in figure [ fig : jjj:1 ] . having formally introduced ldpc convolutional codes in the previous section , we now turn our attention to the main tool of this paper , namely graph covers . [ def : graph : cover:1 ] a _ cover _ of a graph @xmath104 with vertex set @xmath105 and edge set @xmath106 is a graph @xmath107 with vertex set @xmath108 and edge set @xmath109 , along with a surjection @xmath110 which is a graph homomorphism ( i.e. , @xmath111 takes adjacent vertices of @xmath107 to adjacent vertices of @xmath104 ) such that for each vertex @xmath112 and each @xmath113 , the neighborhood @xmath114 of @xmath115 is mapped bijectively to @xmath116 . a cover is called an _ @xmath117-cover _ , where @xmath117 is a positive integer , if @xmath118 for every vertex @xmath119 in @xmath105 . is also known as the degree of the cover . ( not to be confused with the degree of a vertex . ) ] these graph covers will be used for the construction of new tanner graphs from old tanner graphs , in particular for the construction of tanner graphs that represent ldpc convolutional codes . more specifically , this section starts by discussing two simple methods to specify a graph cover , which will be called graph - cover construction 1 ( * gcc1 * ) and graph - cover construction 2 ( * gcc2 * ) . although they yield isomorphic tanner graphs , and therefore equivalent codes , it is convenient to have both methods at hand . as we will see , interesting classes of tanner graphs can be obtained by repeatedly applying these graph - cover constructions , by mixing them , and by suitably shortening the resulting codes . moreover , these two graph - cover constructions will allow us to exhibit a connection between the tanner and the jfz unwrapping techniques . let @xmath120 be an @xmath121 matrix over @xmath2 . with such a matrix we can associate a tanner graph @xmath122 , where we draw @xmath123 variable nodes , @xmath124 check nodes , and where there are @xmath125_{j , i}$ ] edges from the @xmath15-th variable node to the @xmath14-th check node . given the role that the matrix @xmath120 will play subsequently , we follow @xcite and call the matrix @xmath120 a proto - matrix and the corresponding graph @xmath122 a proto - graph . the next definition introduces * gcc1 * and * gcc2 * , two ways to specify graph covers that will be used throughout the rest of the paper .. ] [ def : graph : cover : construction:1 ] for some positive integers @xmath124 and @xmath123 , let @xmath126 be a proto - matrix . we also introduce the following objects : * for some finite set @xmath127 , let @xmath128 be a collection of matrices such that @xmath129 , @xmath130 , and such that @xmath131 . * for some positive integer @xmath8 , let @xmath132 be a collection of size-@xmath133 permutation matrices . i.e. , for every @xmath130 , the matrix @xmath134 is such that it contains one `` @xmath17 '' per column , one `` @xmath17 '' per row , and `` @xmath16 ' 's otherwise . based on the collection of matrices @xmath135 and the collection of matrices @xmath136 , there are two common ways of defining a graph cover of the tanner graph @xmath122 . ( in the following expressions , @xmath137 is the identity matrix of size @xmath138 . ) * * graph - cover construction 1 ( * gcc1 * ) . * consider the intermediary matrix @xmath139 whose tanner graph @xmath140 consists of @xmath8 disconnected copies of @xmath122 . this is an @xmath8-fold cover of @xmath122 , albeit a rather trivial one . in order to obtain an interesting @xmath8-fold graph cover of @xmath120 , for each @xmath130 , we replace @xmath141 by @xmath142 , i.e. , we define @xmath143 * * graph - cover construction 2 ( * gcc2 * ) * consider the intermediary matrix @xmath144 whose tanner graph @xmath145 consists of @xmath8 disconnected copies of @xmath122 . this is an @xmath8-fold cover of @xmath122 , albeit a rather trivial one . in order to obtain an interesting @xmath8-fold graph cover of @xmath120 , for each @xmath130 , we replace @xmath146 by @xmath147 , i.e. , we define @xmath148 if all the matrices @xmath136 are circulant matrices , then the graph covers @xmath149 and @xmath150 will be called cyclic covers of @xmath122 . one can verify that the two graph - cover constructions in definition [ def : graph : cover : construction:1 ] are such that the matrix @xmath151 , after a suitable reordering of the rows and columns , equals the matrix @xmath152 . and @xmath153 are permutation equivalent , i.e. , there is a pair of permutation matrices @xmath154 such that @xmath155 . of course , for this to work , the pair @xmath156 must be independent of @xmath130 , i.e. , dependent only on the size of the matrices @xmath135 and @xmath157 . such a @xmath154 pair can easily be found . ] this implies that @xmath149 and @xmath150 are isomorphic graphs ; nevertheless , it is helpful to define both types of constructions . the following examples will help us to better understand how * gcc1 * and * gcc2 * can be used to obtain interesting classes of tanner graphs , and , in particular , how the resulting graph - cover constructions can be visualized graphically . although these examples are very concrete , they are written such that they can be easily generalized . [ ex : graph : cover : construction:1 ] consider the proto - matrix @xmath158 with @xmath159 and @xmath160 , and whose tanner graph @xmath122 is shown in figure [ fig : graph : cover : constructions:1](a ) . let @xmath161 , and let the collection of matrices @xmath162 be given by @xmath163 , where for each @xmath164 and each @xmath165 the matrix @xmath166 is defined as follows @xmath167_{j',i ' } & { \triangleq}\begin{cases } [ { { { { \bm{a}}}}}]_{j',i ' } & \text{if $ ( j',i ' ) = ( j , i)$ } \\ 0 & \text{otherwise } \end{cases}. \end{aligned}\ ] ] moreover , let @xmath168 , and let the collection of matrices @xmath157 be given by @xmath169 , where @xmath170 , @xmath171 , @xmath172 , @xmath173 , @xmath174 , @xmath175 , and where @xmath176 is an @xmath177 times left - shifted identity matrix of size @xmath178 . * using * gcc1 * , we obtain the matrices @xmath179 { { { { \bm{b}}}}}&= \begin{bmatrix } { { { { \bm{i}}}}}_1 & { { { { \bm{i}}}}}_2 & { { { { \bm{i}}}}}_4 \\ { { { { \bm{i}}}}}_6 & { { { { \bm{i}}}}}_5 & { { { { \bm{i}}}}}_3 \end{bmatrix } , \label{eq : ex : graph : cover : constructions:1:b:1 } \end{aligned}\ ] ] whose tanner graphs @xmath140 and @xmath149 , respectively , are shown in figure [ fig : graph : cover : constructions:1](b ) . * using * gcc2 * , we obtain the matrices @xmath180 \hskip-0.20 cm { \overline{{{{\bm{b}}}}}}&= \left [ \begin{array}{@{\;}c@{\;\;}c@{\;\;}c@{\;\;}c@{\;\;}c@{\;\;}c@{\;\;}c@{\ ; } } { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } & { { { { \bm{a}}}}}_{0,1 } & { { { { \bm{a}}}}}_{0,0 } \\ { { { { \bm{a}}}}}_{0,0 } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } & { { { { \bm{a}}}}}_{0,1 } \\ { { { { \bm{a}}}}}_{0,1 } & { { { { \bm{a}}}}}_{0,0 } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } \\ { { { { \bm{a}}}}}_{1,2 } & { { { { \bm{a}}}}}_{0,1 } & { { { { \bm{a}}}}}_{0,0 } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } \\ { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } & { { { { \bm{a}}}}}_{0,1 } & { { { { \bm{a}}}}}_{0,0 } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } \\ { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } & { { { { \bm{a}}}}}_{0,1 } & { { { { \bm{a}}}}}_{0,0 } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } \\ { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } & { { { { \bm{a}}}}}_{0,1 } & { { { { \bm{a}}}}}_{0,0 } & { { { \bm{0 } } } } \end{array } \right]\!\ ! , \label{eq : ex : graph : cover : constructions:1:ob:1 } \end{aligned}\ ] ] whose tanner graphs @xmath145 and @xmath150 , respectively , are shown in figure [ fig : graph : cover : constructions:1](c ) . note that all the block rows and all the block columns sum ( in @xmath1 ) to @xmath120 . ( this observation holds in general , not just for this example . ) we would like to add two comments with respect to the above example . first , instead of defining @xmath176 to be an @xmath177 times _ left_-shifted identity matrix of size @xmath138 , we could have defined @xmath176 to be an @xmath177 times _ right_-shifted identity matrix of size @xmath181 . compared to the matrices and graphs described above , such a change in definition would yield ( in general ) different matrices but isomorphic graphs . second , we note that * gcc2 * was termed the `` copy - and - permute '' construction by thorpe and his co - workers . this terminology stems from the visual appearance of the procedure : namely , in going from figure [ fig : graph : cover : constructions:1](a ) to figure [ fig : graph : cover : constructions:1](c)(top ) we copy the graph several times , and in going from figure [ fig : graph : cover : constructions:1](c)(top ) to figure [ fig : graph : cover : constructions:1](c)(bottom ) we permute the edges of the graph , where the permutations are done within the sets of edges that have the same pre - image in figure [ fig : graph : cover : constructions:1](a ) . [ rem : graph : cover : construction:1:variation:1 ] consider again the matrices that were constructed in example [ ex : graph : cover : construction:1 ] , in particular the matrix @xmath120 in and its @xmath8-fold cover matrix @xmath151 in . because all matrices in the matrix collection @xmath136 are circulant , @xmath149 represents a cyclic cover of @xmath182 . clearly , when seen over @xmath183 , the matrix @xmath184 is the parity - check matrix of a quasi - cyclic binary linear block code @xmath185 using the well - known isomorphism between the addition and multiplication of circulant matrices over @xmath183 and the addition and multiplication of elements of the ring @xmath9}$ ] , this code can be written equivalently as @xmath186}^{{n_{{{{\bm{a } } } } } } } \bigm| { { { { \bm{h}}}}_{\mathrm{qc}}^{(r)}}(x ) \cdot { { { { \bm{v}}}}}(x)^{\mathsf{t}}= { { { \bm{0}}}}^{\mathsf{t}}\bigr\ } \end{aligned}\ ] ] with @xmath187 as noted above , the graphs @xmath149 and @xmath150 that are constructed in definition [ def : graph : cover : construction:1 ] are isomorphic . applying this observation to example [ ex : graph : cover : construction:1 ] , the matrix @xmath188 with @xmath152 from is therefore the parity - check matrix of a binary linear block code @xmath189 that is equivalent to @xmath62 , i.e. , the codewords of @xmath190 can be obtained from the codewords of @xmath62 by a suitable reordering of the codeword components . in terms of the matrices @xmath191 , which also appear in the matrix @xmath152 in , one can verify that the polynomial parity - check matrix @xmath63 can be written as @xmath192 . besides defining finite graph covers , we can also define infinite graph covers , as illustrated in the following examples . these infinite graph covers will be crucial towards defining tanner graphs of convolutional codes . [ ex : graph : cover : construction:1:infinite:1 ] we continue example [ ex : graph : cover : construction:1 ] . however , besides keeping the proto - matrix @xmath120 and the collection of matrices @xmath193 , we consider a different collection of matrices @xmath194 . namely , we set @xmath195 , @xmath196 , @xmath197 , @xmath198 , @xmath199 , @xmath200 . here @xmath201 is a bi - infinite toeplitz matrix with zeros everywhere except for ones in the @xmath177-th diagonal below the main diagonal , i.e. , @xmath202_{j , i } = 1 $ ] if @xmath203 and @xmath202_{j , i } = 0 $ ] otherwise . e.g. , @xmath204 where for clarity we have underlined the entries of the main diagonal . * using * gcc1 * , we obtain the matrices @xmath205 and @xmath206 the tanner graph @xmath140 , which is depicted in figure [ fig : graph : cover : constructions:1:variation:2](b)(top ) , is similar to the corresponding tanner graph in figure [ fig : graph : cover : constructions:1](b)(top ) , but with bi - infinitely many independent components . analogously , the tanner graph @xmath149 , which is depicted in figure [ fig : graph : cover : constructions:1:variation:2](b)(bottom ) , is similar to the tanner graph shown in figure [ fig : graph : cover : constructions:1](b)(bottom ) , but instead of cyclically wrapped edge connections , the edge connections are infinitely continued on both sides . * using * gcc2 * , we obtain the matrices @xmath207 and @xmath208\!. \label{eq : ex : graph : cover : constructions:1:variation:2:ob:1 } \end{aligned}\ ] ] the tanner graph @xmath145 , which is depicted in figure [ fig : graph : cover : constructions:1:variation:2](c)(top ) , is similar to the corresponding tanner graph in figure [ fig : graph : cover : constructions:1](c)(top ) , but with bi - infinitely many independent components . analogously , the tanner graph @xmath150 , which is depicted in figure [ fig : graph : cover : constructions:1:variation:2](c)(bottom ) , is similar to the tanner graph shown in figure [ fig : graph : cover : constructions:1](c)(bottom ) , but instead of cyclically wrapped edge connections , the edge connections are infinitely continued on both sides . although it is tempting to replace in example [ ex : graph : cover : construction:1:infinite:1 ] the bi - infinite toeplitz matrices @xmath201 ( whose row and column index sets equal @xmath1 ) by semi - infinite toeplitz matrices ( whose row and column index sets equal @xmath2 ) , note that the resulting tanner graphs @xmath149 and @xmath150 would then in general not be graph covers of @xmath122 . this follows from the fact that semi - infinite toeplitz matrices are not permutation matrices ( except for @xmath209 ) , and so some vertex degrees of @xmath149 and @xmath150 would not equal the corresponding vertex degrees in @xmath122 . [ rem : graph : cover : connections:1 ] it turns out that the tanner graphs in figure [ fig : graph : cover : constructions:1:variation:2 ] are infinite graph covers of the tanner graphs in figure [ fig : graph : cover : constructions:1 ] . more precisely , the tanner graphs @xmath140 , @xmath149 , @xmath145 , @xmath150 in figure [ fig : graph : cover : constructions:1:variation:2 ] are graph covers of the corresponding tanner graphs @xmath140 , @xmath149 , @xmath145 , @xmath150 in figure [ fig : graph : cover : constructions:1 ] . for the tanner graphs @xmath140 in figures [ fig : graph : cover : constructions:1](b)(top ) and [ fig : graph : cover : constructions:1:variation:2](b)(top ) and the tanner graphs @xmath145 in figures [ fig : graph : cover : constructions:1](c)(top ) and [ fig : graph : cover : constructions:1:variation:2](c)(top ) , this statement is easily verified by inspection . to verify that the tanner graph @xmath150 in figure [ fig : graph : cover : constructions:1:variation:2](c)(bottom ) is a graph cover of @xmath150 in figure [ fig : graph : cover : constructions:1](c)(bottom ) , we apply * gcc2 * with proto - matrix @xmath120 , with resulting matrix @xmath152 , with the set @xmath127 , with the collection of matrices @xmath210 , and with the collection of permutation matrices @xmath211 as follows . namely , we let the proto - matrix @xmath120 be the matrix from ( there denoted by @xmath152 ) , we let the resulting matrix @xmath152 be the matrix in ( there denoted by @xmath152 ) , we define @xmath212 , we select @xmath213 , \label{eq : rem : graph : cover : connections:1:a:0 } \\[0.25 cm ] { { { { \bm{a}}}}}_1 & = \left [ \begin{array}{@{\;}c@{\;\;}c@{\;\;}c@{\;\;}c@{\;\;}c@{\;\;}c@{\;\;}c@{\ ; } } { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } & { { { { \bm{a}}}}}_{0,1 } & { { { { \bm{a}}}}}_{0,0 } \\ { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } & { { { { \bm{a}}}}}_{0,1 } \\ { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } & { { { { \bm{a}}}}}_{1,2 } \\ { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } & { { { { \bm{a}}}}}_{0,2 } \\ { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } & { { { { \bm{a}}}}}_{1,1 } \\ { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { { \bm{a}}}}}_{1,0 } \\ { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } & { { { \bm{0 } } } } \end{array } \right ] , \label{eq : rem : graph : cover : connections:1:a:1 } \end{aligned}\ ] ] and we select @xmath214 and @xmath215 , where @xmath201 was defined in example [ ex : graph : cover : construction:1:infinite:1 ] . clearly , @xmath216 . would be put in @xmath217 . ] with this we have @xmath218 \!\ ! , \end{aligned}\ ] ] and one can verify that this matrix equals the matrix in ( there denoted by @xmath152 ) , which means that @xmath150 is indeed an infinite cover of @xmath122 . we remark that , interestingly , in this process we have shown how a certain * gcc2 * graph cover of a proto - matrix can be written as a * gcc2 * graph cover of a certain * gcc2 * graph cover of that proto - matrix . finally , a similar argument shows that the tanner graph @xmath149 in figure [ fig : graph : cover : constructions:1:variation:2](b)(top ) is a graph cover of the tanner graph in figure [ fig : graph : cover : constructions:1](b)(top ) , also denoted by @xmath149 . there are many other ways of writing a proto - matrix @xmath120 as a sum of a collection of matrices @xmath135 . the next example discusses two such possibilities . [ ex : graph : cover : construction:2 ] consider the proto - matrix @xmath219 that is shown in figure [ fig : graph : cover : constructions:2](a ) , and that also appeared in example [ example : unwrap:1 ] . its tanner graph @xmath122 is @xmath220-regular , i.e. , all variable nodes have degree @xmath221 and all check nodes have degree @xmath222 . let @xmath223 , and consider the collection of matrices @xmath136 with @xmath214 and @xmath215 , where the matrices @xmath209 and @xmath224 are defined as in example [ ex : graph : cover : construction:1:infinite:1 ] . in the following , we look at two different choices of the collection of matrices @xmath162 . * figure [ fig : graph : cover : constructions:2](c ) shows a typical part of the matrix @xmath152 that is obtained when * gcc2 * is used to construct a graph cover of @xmath120 with the collection of matrices @xmath225 defined as shown in figure [ fig : graph : cover : constructions:2](a ) . * figure [ fig : graph : cover : constructions:2](d ) shows a typical part of the matrix @xmath152 when * gcc2 * is used to construct a graph cover of @xmath120 with the collection of matrices @xmath128 defined as shown in figure [ fig : graph : cover : constructions:2](b ) . overall , because of the choice of the collection @xmath226 , the support of both matrices @xmath152 possesses a banded diagonal structure . moreover , the different choices of the collection @xmath162 leads to a somewhat narrower banded diagonal structure in the first case compared to the second case . the next example makes a crucial observation ; namely , it shows that the above graph - cover constructions can be applied repeatedly to obtain additional interesting classes of tanner graphs . [ ex : graph : cover : construction:3 ] starting with the proto - matrix @xmath227 from example [ ex : graph : cover : construction:1 ] , we consider two iterated graph - cover constructions . in the first case , we apply * gcc1 * and then * gcc2 * , and in the second case we apply * gcc2 * twice . * consider the matrix @xmath151 obtained from the matrix @xmath120 using * gcc1 * , like in example [ ex : graph : cover : construction:1 ] . the resulting matrix @xmath151 is shown in and will be called @xmath228 in this example , since it is considered to be a proto - matrix by itself , cf . figure [ fig : graph : cover : construction:3](a ) . based on the `` cutting line '' shown in figure [ fig : graph : cover : construction:3](a ) , we define the matrices @xmath229 and @xmath230 as follows : the non - zero part of @xmath229 equals the non - zero part of the lower triangular part of @xmath228 and the non - zero part of @xmath230 equals the non - zero part of the upper triangular part of @xmath228 . ( clearly , @xmath231 . ) applying the procedure from example [ ex : graph : cover : construction:2 ] , figure [ fig : graph : cover : construction:3](c ) shows a typical part of the matrix @xmath232 that is obtained when * gcc2 * is used to construct a graph cover of @xmath228 . * consider the graph - cover @xmath152 obtained from @xmath120 using * gcc2 * , like in example [ ex : graph : cover : construction:1 ] . the resulting matrix @xmath152 is shown in and will be called @xmath233 in this example , since it is considered to be a proto - matrix by itself , cf . figure [ fig : graph : cover : construction:3](b ) . based on the `` cutting line '' shown in figure [ fig : graph : cover : construction:3](b ) , we define the matrices @xmath234 and @xmath235 as follows : the non - zero part of @xmath234 equals the non - zero part of the lower triangular part of @xmath233 and the non - zero part of @xmath235 equals the non - zero part of the upper triangular part of @xmath233 . ( clearly , @xmath236 . ) applying the procedure from example [ ex : graph : cover : construction:2 ] , figure [ fig : graph : cover : construction:3](d ) shows a typical part of the matrix @xmath237 that is obtained when * gcc2 * is used to construct a graph cover of @xmath233 . we observe a large difference in the positions of the non - zero entries in @xmath232 and @xmath237 . * in the first case , the two graph - cover constructions are `` incompatible '' and the positions of the non - zero entries in @xmath232 follow a `` non - simple '' or `` pseudo - random '' pattern . as we will see in example [ ex : iterated : gcc : ldpc : code : construction:1 ] with the help of simulation results , such tanner graphs can lead to time - varying ldpc convolutional codes with very good performance . * in the second case , the two graph - cover constructions are `` compatible '' in the sense that @xmath237 can be obtained from the proto - matrix @xmath120 by applying * gcc2 * with suitable matrix collections @xmath135 and @xmath238 . as such , the positions of the non - zero entries of @xmath237 follow a relatively `` simple '' or `` non - random '' pattern , which leads to a time - invariant ldpc convolutional code . the above procedure of obtaining two matrices that add up to a matrix is called `` cutting a matrix '' . actually , we will also use this term if there is no simple cutting line , as in the above examples , and also if the matrix is written as the sum of more than two matrices ( cf . example [ example : unwrap:1 ] and the paragraphs after it ) . in section [ sec : unwrapping:1 ] we introduced two techniques , termed the tanner and the jfz unwrapping techniques , to derive convolutional codes from block codes . in this subsection we revisit these unwrapping techniques . in particular , we show how they can be cast in terms of graph covers and how the two unwrapping techniques are connected . because of the importance of the coding - theoretic notion of shortening @xcite for this subsection , we briefly revisit this concept . let @xmath239 be a parity - check matrix that defines some length-@xmath11 binary code @xmath240 . we say that the length-@xmath241 code @xmath242 is obtained by shortening @xmath240 at position @xmath15 if @xmath243 in terms of parity - check matrices , a possible parity - check matrix @xmath244 of @xmath242 is obtained by deleting the @xmath15-th column of @xmath239 . in terms of tanner graphs , this means that the tanner graph @xmath245 is obtained from @xmath246 by removing the @xmath15-th variable node , along with its incident edges . in the following , we will also use the term `` shortening '' to denote this graph modification procedure . now , to explain the tanner unwrapping technique in terms of graph covers , consider the quasi - cyclic block code @xmath62 defined by the polynomial parity - check matrix @xmath63 of size @xmath121 , i.e. , @xmath65}^{{n_{{{{\bm{a } } } } } } } \bigm| { { { { \bm{h}}}}_{\mathrm{qc}}^{(r)}}(x ) \cdot { { { { \bm{v}}}}}(x)^{\mathsf{t}}= { { { \bm{0}}}}^{\mathsf{t}}\bigr\},\end{aligned}\ ] ] where the polynomial operations are performed modulo @xmath7 ( see also remark [ rem : graph : cover : construction:1:variation:1 ] ) . as already mentioned in section [ sec : unwrapping:1 ] , the tanner unwrapping technique is simply based on dropping these modulo computations . more precisely , with a quasi - cyclic block code @xmath62 , we associate the convolutional code @xmath66^{{n_{{{{\bm{a } } } } } } } \bigm| { { { { \bm{h}}}}_{\mathrm{conv}}}(d ) \cdot { { { { \bm{v}}}}}(d)^{\mathsf{t}}= { { { \bm{0}}}}^{\mathsf{t}}\bigr\}\end{aligned}\ ] ] with polynomial parity - check matrix @xmath247 again , the change of indeterminate from @xmath5 to @xmath68 indicates the lack of modulo @xmath69 operations . in the following we will give , with the help of an example , two interpretations of the tanner unwrapping technique in terms of graph covers . [ ex : tanner : unwrapping:1:rev:1 ] unwrapping the quasi - cyclic block code @xmath62 that was considered in remark [ rem : graph : cover : construction:1:variation:1 ] , we obtain a rate-@xmath59 time - invariant convolutional code @xmath248^{{n_{{{{\bm{a } } } } } } } \bigm| { { { { \bm{h}}}}_{\mathrm{conv}}}(d ) \cdot { { { { \bm{v}}}}}(d)^{\mathsf{t}}= { { { \bm{0}}}}^{\mathsf{t}}\bigr\ } \end{aligned}\ ] ] with polynomial parity - check matrix @xmath249 consider now the infinite graph covers that were constructed in example [ ex : graph : cover : construction:1:infinite:1 ] using * gcc1 * , in particular @xmath149 . let @xmath250 be the set of codewords defined by the tanner graph @xmath149 . then the convolutional code @xmath72 is a shortened version of @xmath251 where all codeword bits corresponding to negative time indices have been shortened . therefore , the tanner graph of @xmath72 is given by the tanner graph in figure [ fig : graph : cover : constructions:1:variation:2](b)(bottom ) , where all bit nodes with negative time indices , along with their incident edges , are removed . clearly , this bit - node and edge removal process implies decreasing the degrees of some check nodes . in fact , some check nodes become obsolete , because their degree is decreased to zero . therefore , one interpretation of the tanner unwrapping technique in terms of graph covers is that the tanner graph of the convolutional code is obtained by taking a suitable graph cover of the same proto - graph that was used to construct the quasi - cyclic ldpc code , along with some suitable shortening . [ ex : tanner : unwrapping:1:rev:2 ] we continue remark [ rem : graph : cover : construction:1:variation:1 ] and example [ ex : tanner : unwrapping:1:rev:1 ] . clearly , in the same way as the block code @xmath190 is equivalent to the block code @xmath62 , we can define a code @xmath21 ( with parity - check matrix @xmath24 ) that is equivalent to @xmath72 . the observations in remark [ rem : graph : cover : connections:1 ] and example [ ex : tanner : unwrapping:1:rev:1 ] can then be used to show that the tanner graph of @xmath24 equals a graph cover of the tanner graph @xmath252 , along with some suitable shortening . therefore , the second interpretation of the tanner unwrapping in terms of graph covers is that the tanner graph of the convolutional code is obtained by taking a suitable graph cover of the tanner graph of the quasi - cyclic code , along with some suitable shortening . now turning our attention to the jfz unwrapping technique , recall from section [ sec : unwrapping:1 ] that this method is based on writing a parity - check matrix @xmath80 of some block code @xmath96 as the sum @xmath253 of a collection of matrices @xmath254 . the convolutional code is then defined to be @xmath255 where @xmath256 . \label{eq : code : c : conv : jfz : unwrapping:1:rev : pcm:1}\end{aligned}\ ] ] with the help of an example , we now explain how the jfz unwrapping technique can be cast in terms of graph - covers . [ ex : jfz : unwrapping:2:rev ] consider the infinite graph covers that were constructed using * gcc2 * in example [ ex : graph : cover : construction:1:infinite:1 ] , in particular @xmath150 . let @xmath257 be the set of valid assignments to the tanner graph @xmath150 . moreover , let @xmath258 , and let @xmath21 be defined as in . then the code @xmath21 is a shortened version of @xmath259 , where all codeword bits corresponding to negative time indices have been shortened . therefore , the tanner graph of @xmath21 is given by the tanner graph in figure [ fig : graph : cover : constructions:1:variation:2](c)(bottom ) , where all the bit nodes with negative time indices are shortened . in order to connect the unwrapping techniques due to tanner and due to jfz , we show now , with the help of an example , that in fact the unwrapping technique due to tanner can be seen as a _ special case _ of the unwrapping technique due to jfz . consider the quasi - cyclic block code defined by the parity - check matrix @xmath260 , where @xmath152 was defined in . applying the jfz unwrapping technique with the matrix decomposition @xmath261 , with @xmath217 defined in and @xmath262 defined in , @xmath24 turns out to equal a submatrix of @xmath152 in , namely the submatrix of @xmath152 where the row and column index set are equal to @xmath2 . however , the code defined by @xmath24 is equivalent to the code defined by the tanner unwrapping technique applied to the quasi - cyclic code defined by @xmath252 . therefore , the unwrapping technique due to jfz is more general . in fact , whereas the tanner unwrapping technique leads to _ time - invariant _ convolutional codes , the unwrapping technique due to jfz can , depending on the parity - check matrix decomposition and the internal structure of the terms in the decomposition , lead to _ time - varying _ convolutional codes with non - trivial period . despite the fact that the unwrapping technique due to tanner is a special case of the unwrapping technique due to jfz , it is nevertheless helpful to have both unwrapping techniques at hand , because sometimes one framework can be more convenient than the other . we will use both perspectives in the next section . we conclude this section with the following remarks . * although most of the examples in this section have regular bit node degree @xmath263 and regular check node degree @xmath221 , there is nothing special about this choice of bit and check node degrees ; any other choice would work equally well . * although all polynomial parity - check matrices that appear in this section contain only monomials , this is not required , i.e. , the developments in this section work equally well for polynomial parity - check matrices containing the zero polynomial , monomials , binomials , trinomials , and so on . * it can easily be verified that if the matrix @xmath120 in definition [ def : graph : cover : construction:1 ] contains only zeros and ones , then the graph covers constructed in * gcc1 * and * gcc2 * never have parallel edges . in particular , if @xmath120 is the parity - check matrix of a block code ( like in most examples in this paper ) , then the constructed graph covers never have parallel edges . + however , if @xmath120 contains entries that are larger than one , then there is the potential for the constructed graph covers to have parallel edges ; if parallel edges really appear depends then critically on the choice of the decomposition @xmath264 and the choice of the permutation matrices @xmath211 . an example of such a case is the tanner graph construction in section [ sec : ldpc : code : construction : lentmaier : kudekar:1 ] , where @xmath265 $ ] and where @xmath135 and @xmath266 are chosen such that parallel edges are avoided in the constructed graph cover . + we note that in the case of iterated graph - cover constructions it can make sense to have parallel edges in the intermediate graph covers . however , in the last graph - cover construction stage , parallel edges are usually avoided , because parallel edges in tanner graphs typically lead to a weakening of the code and/or of the message - passing iterative decoder . although the graph - cover constructions and unwrapping techniques that were discussed in sections [ sec : ldpc : convolutional : codes:1 ] and [ sec : graph : covers:1 ] are mathematically quite straightforward , it is important to understand how they can be applied to obtain ldpc convolutional codes with good performance and attractive encoder and decoder architectures . to that end , this section explores a variety of code design options and comments on some practical issues . it also proposes a new `` random '' unwrapping technique which leads to convolutional codes whose performance compares favorably to other codes with the same parameters . of course , other variations than the ones presented here are possible , in particular , by suitably combining some of the example constructions . the simulation results for the codes in this section plot the decoded bit error rate ( ber ) versus the signal - to - noise ratio ( snr ) @xmath267 and were obtained by assuming bpsk modulation and an additive white gaussian noise channel ( awgnc ) . all decoders were based on the sum - product algorithm @xcite and were allowed a maximum of 100 iterations , with the block code decoders employing a syndrome - check based stopping rule . for comparing the performance of unwrapped convolutional codes with their underlying block codes we will use the following metric . for a convolutional code constructed from an underlying block code , we define its `` convolutional gain '' to be the difference in snr required to achieve a particular ber with the convolutional code compared to achieving the same ber with the block code . the rest of this section is structured as follows . first we discuss the construction of some _ time - invariant _ ldpc convolutional codes based on the tanner unwrapping technique . in this context we make a simple observation about how the syndrome former memory can sometimes be reduced without changing the convolutional code . secondly , we present a construction of _ time - varying _ ldpc convolutional codes based on iterated graph - cover constructions . an important sub - topic here will be an investigation of the influence of the `` diagonal cut '' ( which is used to define a graph cover ) on the decoding performance . in this section we revisit a class of quasi - cyclic ldpc codes and their associated convolutional codes that was studied in @xcite . as we will see , they are instances of the quasi - cyclic code construction in example [ ex : graph : cover : construction:1 ] and remark [ rem : graph : cover : construction:1:variation:1 ] , and the corresponding convolutional code construction based on tanner s unwrapping technique in example [ ex : tanner : unwrapping:1:rev:1 ] . [ ex : tanner : qc : code:1 ] consider the regular proto - matrix @xmath268 with @xmath269 and @xmath270 . we apply * gcc1 * , as in example [ ex : graph : cover : construction:1 ] and remark [ rem : graph : cover : construction:1:variation:1 ] , with an interesting choice of permutation matrices first suggested by tanner @xcite that yields the parity - check matrix @xmath271 where as before @xmath176 is an @xmath177 times left - circularly shifted identity matrix of size @xmath138 and @xmath272 . the corresponding polynomial parity - check is @xmath273 the resulting quasi - cyclic @xmath274-regular ldpc block codes have block length @xmath275 . in particular , for @xmath276 , @xmath277 , and @xmath278 , we obtain codes of length @xmath279 , @xmath280 , and @xmath281 , respectively , whose simulated ber performance results are shown in figure [ fig : sims ] . the choice @xmath276 yields the well - known length-@xmath279 quasi - cyclic block code that was first introduced by tanner @xcite ( see also the discussion in @xcite ) . unwrapping these codes by the tanner unwrapping technique as in example [ ex : tanner : unwrapping:1:rev:1 ] , we obtain a rate-@xmath282 time - invariant convolutional code with @xmath283 defined by the polynomial parity - check matrix @xmath284 its decoding performance is also shown in figure [ fig : sims ] under the label `` @xmath285 time - invariant conv . code with @xmath283 . '' we conclude this example with a few remarks . * figure [ fig : sims ] shows that the convolutional code exhibits a `` convolutional gain '' of between @xmath286 and @xmath287 compared to the @xmath288 $ ] quasi - cyclic ldpc block code at moderate bers and that the gain remains between @xmath289 and @xmath290 at lower bers . * note that the polynomial parity - check matrix @xmath291 that is obtained by the tanner unwrapping technique is independent of the parameter @xmath8 of the polynomial parity - check matrix @xmath63 , as long as @xmath8 is strictly larger than the largest exponent appearing in @xmath63 . moreover , for @xmath292 , the tanner graph of @xmath63 is closely related to the tanner graph of @xmath291 , and so it is not surprising to see that , for larger @xmath8 , the decoding performance of quasi - cyclic ldpc block codes based on @xmath63 tends to the decoding performance of the ldpc convolutional based on @xmath291 , as illustrated by the two curves labeled `` @xmath293 $ ] qc code '' and `` @xmath294 $ ] qc code '' in figure [ fig : sims ] . * the permutation matrices ( more precisely , the circulant matrices ) that were used for constructing the quasi - cyclic codes in this example were _ not _ chosen to optimize the hamming distance or the pseudo - weight properties of the code . in particular , a different choice of circulant matrices may result in better high - snr performance , i.e. , in the so - called `` error floor '' region of the ber curve . for choices of codes with better hamming distance properties , we refer the reader to @xcite . * the remaining curves in figure [ fig : sims ] will be discussed in example [ ex : iterated : gcc : ldpc : code : construction:1 ] . we conclude this subsection with some comments on the syndrome former memory @xmath31 of the convolutional codes obtained by the tanner unwrapping technique , in particular how this syndrome former memory @xmath31 can sometimes be reduced without changing the convolutional code . assume that we have obtained a polynomial parity - check matrix @xmath291 from @xmath63 according to the tanner method . clearly , the syndrome former memory @xmath31 is given by the largest exponent that appears in @xmath291 . in some instances there is a simple way of reducing @xmath31 without changing the convolutional code . namely , if @xmath295 is the _ minimal _ exponent that appears in the polynomials of a given row of @xmath291 , then the polynomials in this row of @xmath291 can be divided by @xmath296 . we illustrate this syndrome former memory reduction for the small convolutional code that appeared in example [ ex : tanner : unwrapping:1:rev:1 ] . [ ex : syndrome : former : memory : reduction:1 ] applying the tanner unwrapping technique to the polynomial parity - check matrix @xmath63 of the quasi - cyclic ldpc code with @xmath297 in remark [ rem : graph : cover : construction:1:variation:1 ] , we obtain @xmath291 of a rate-@xmath59 time - invariant ldpc convolutional code , as shown in example [ ex : tanner : unwrapping:1:rev:1 ] , with syndrome former memory @xmath298 . following the procedure discussed in the paragraph above , the first and second rows of @xmath291 can be divided by @xmath299 and @xmath300 , respectively , to yield an equivalent convolutional code with syndrome former memory @xmath301 and polynomial parity - check matrix @xmath302 figure [ fig : ex : syndrome : former : memory : reduction:1 ] shows parts of the corresponding scalar parity - check matrix @xmath24 for @xmath301 , together with the original scalar parity - check matrix for @xmath298 , and illustrates the equivalence of the two matrices in the sense that only the ordering of the rows is different , which does not affect the corresponding convolutional code . in this example , the order of the even - numbered rows stays the same , while the odd - numbered rows are shifted by four positions . the equivalence of the two parity - check matrices can be seen by noting that the parity - check matrix , outside of the diagonal structure , is filled with zeros . as was seen in example [ ex : graph : cover : construction:3 ] , interesting graph covers can be obtained by combining * gcc1 * with * gcc2 * , or vice - versa . inspired by that example , this subsection considers iterated graph - cover constructions for constructing tanner graphs of ldpc convolutional codes , in particular of time - varying ldpc convolutional codes . [ def : iterated : gcc : ldpc : code : construction:1 ] based on a combination of * gcc1 * , * gcc2 * , and the code - shortening concept introduced in section [ sec : unwrapping:1:rev ] , we propose the following construction of ldpc convolutional codes . 1 . we start with a proto - matrix @xmath120 of size @xmath121 . [ item : iterated : gcc : first : lifting:1 ] we apply * gcc1 * to @xmath120 with finite - size permutation matrices and obtain the matrix @xmath303 . [ item : iterated : gcc : second : lifting:1 ] we apply * gcc2 * to @xmath303 with permutation matrices that are bi - infinite toeplitz matrices and obtain the matrix @xmath304 . [ item : iterated : gcc : tanner : unwrapping:1 ] finally , looking at @xmath304 as the parity - check matrix of a bi - infinite convolutional code , we obtain the parity - check matrix of a convolutional code by shortening the code bit positions corresponding to negative time indices . here , steps [ item : iterated : gcc : second : lifting:1 ] and [ item : iterated : gcc : tanner : unwrapping:1 ] can be seen as an application of the jfz unwrapping method . and ( b ) @xmath301.,scaledwidth=50.0% ] the following example shows how this construction can be applied to obtain ldpc convolutional codes with excellent performance . ( in the example , where suitable , we will refer to the analogous matrices of example [ ex : graph : cover : construction:3 ] and figure [ fig : graph : cover : construction:3 ] that were used to illustrate the iterated graph - cover construction . ) [ ex : iterated : gcc : ldpc : code : construction:1 ] based on definition [ def : iterated : gcc : ldpc : code : construction:1 ] , we construct an ldpc convolutional code by performing the following steps . 1 . we start with the same regular proto - matrix @xmath120 as in example [ ex : tanner : qc : code:1 ] , for which @xmath269 and @xmath270 . we apply * gcc1 * to @xmath120 with permutation matrices that are circulant matrices of size @xmath138 and obtain the parity - check matrix @xmath305 shown in , which is the analogue of @xmath228 in figure [ fig : graph : cover : construction:3](a ) . we apply * gcc2 * to @xmath305 with permutation matrices that are bi - infinite toeplitz matrices and obtain a new parity - check matrix @xmath304 . this is analogous to the transition of the matrix @xmath228 in figure [ fig : graph : cover : construction:3](a ) to the matrix @xmath232 in figure [ fig : graph : cover : construction:3](c ) . the `` diagonal cut '' is obtained by alternately moving @xmath270 units to the right and then @xmath269 units down . 4 . finally , we obtain the desired convolutional code by shortening the code bit positions corresponding to negative time indices . for the choices @xmath276 , @xmath306 , @xmath307 , this construction results in rate-@xmath282 time - varying convolutional codes with syndrome former memory @xmath308 , @xmath309 , @xmath310 , respectively , and with constraint length @xmath311 , @xmath280 , @xmath281 , respectively . the label `` time - varying '' is indeed justified because the convolutional codes constructed here can be expressed in the form of the parity - check matrix in with a suitable choice of syndrome former memory @xmath31 , non - trivial period @xmath37 , and submatrices @xmath312 . the decoding performance of these codes is shown in figure [ fig : sims ] , labeled `` r = 2/5 time - varying conv . code with @xmath313 '' . as originally noted in @xcite , we observe that these three ldpc convolutional codes achieve significantly better performance at a ber of @xmath314 than the other codes shown in this plot , namely with `` convolutional gains '' of @xmath315 for the @xmath316 convolutional code , @xmath317 for the @xmath318 convolutional code , and @xmath319 for the @xmath320 convolutional code , compared to the three respective underlying ldpc block codes . in order to compare these codes based on a given decoding processor ( hardware ) complexity , we consider a block code of length @xmath321 ( see @xcite and @xcite ) . the above time - varying convolutional code for @xmath276 has constraint length @xmath322 , and hence approximately the same processor complexity as the quasi - cyclic block code of length @xmath323 in figure [ fig : sims ] and the time - invariant convolutional code with @xmath283 in figure [ fig : sims ] , but it achieves large gains compared to both of these codes . we note , in addition , that the performance of the time - varying convolutional code with constraint length @xmath320 is quite remarkable , since , at a ber of @xmath324 , it performs within @xmath325 of the iterative decoding threshold of @xmath326 , while having the same processor complexity as a block code of length only @xmath327 . in section [ sec : cost ] , we discuss some possible reasons for these `` convolutional gains , '' along with their associated implementation costs in terms of decoder memory and decoding delay . we make the following observations with respect to the above definition and example . * the ldpc code construction in the above example yields time - varying ldpc convolutional codes with syndrome former memory @xmath328 and period @xmath329 . most importantly , varying @xmath8 in the above construction leads to different ldpc convolutional codes . this is in contrast to the tanner unwrapping technique discussed in section [ sec : varying : circulant : size:1 ] , where the obtained ldpc convolutional code is independent of the parameter @xmath8 , as long as @xmath8 is strictly larger than the largest exponent in @xmath63 . * as mentioned previously in example [ ex : graph : cover : construction:3 ] , the iterated graph - cover construction based on the combination of * gcc1 * and * gcc2 * yields tanner graphs that have a `` pseudo - random '' structure , a structure that seems to be beneficial as indicated by the above simulation results . ( we remark that the improved performance of the time - varying ldpc convolutional codes obtained by unwrapping a randomly constructed ldpc block code was first noted by lentmaier _ _ in @xcite . ) * instead of constructing a first parity - check matrix as in step [ item : iterated : gcc : first : lifting:1 ] of definition [ def : iterated : gcc : ldpc : code : construction:1 ] , one can also start with any other ( randomly or non - randomly constructed , regular or irregular ) parity - check matrix , and still achieve a `` convolutional gain . '' the next example is an illustration of this point . [ example : divsalar : codes : derivatives:1 ] as was done in @xcite , one can replace the parity - check matrix that was constructed in step [ item : iterated : gcc : first : lifting:1 ] of definition [ def : iterated : gcc : ldpc : code : construction:1 ] by an irregular ldpc block code with optimized iterative decoding thresholds . in particular , one can start with the parity - check matrix of the rate-@xmath76 irregular proto - graph - based code from @xcite with an iterative decoding threshold of @xmath330 , and several of its punctured versions . figure [ fig : jplsims ] shows simulation results for the obtained block and convolutional codes . each simulated block code had a block length of about @xmath331 , with code rates ranging from @xmath76 to @xmath332 . we see that `` convolutional gains '' ranging from @xmath333 to @xmath334 at a ber of @xmath324 were obtained . similarly , it was shown in @xcite that an ldpc convolutional code derived from a randomly constructed rate-@xmath76 irregular ldpc block code with block length @xmath335 outperformed the underlying code by almost @xmath336 at a ber of @xmath324 . the degree distribution of the underlying ldpc block code was fully optimized and had an iterative decoding threshold of @xmath337 @xcite . of course , there are other ways of applying the `` diagonal cut '' in step [ item : iterated : gcc : second : lifting:1 ] of example [ ex : iterated : gcc : ldpc : code : construction:1 ] , and so it is natural to investigate the influence of different `` diagonal cuts '' on the decoding performance . we will do this in the next few paragraphs by extending the discussion that was presented right after example [ example : unwrap:1 ] . we start by assuming that the matrix after step [ item : iterated : gcc : first : lifting:1 ] of definition [ def : iterated : gcc : ldpc : code : construction:1 ] has size @xmath338 , and define @xmath339 . then , for any positive integer @xmath52 that divides @xmath340 , we can perform a `` diagonal cut '' where we alternately move @xmath341 units to the right and then @xmath342 units down ( i.e. , @xmath343 . with this , the obtained convolutional code is a periodically time - varying ldpc convolutional code with rate @xmath92 , syndrome former memory @xmath344 , period @xmath345 , and constraint length @xmath346 . ( note that the syndrome former memory @xmath347 depends on @xmath52 , but the constraint length @xmath348 is independent of @xmath52 . ) -regular ldpc block code using different step sizes . ] [ ex : diagonal : cuts : big : ldpc : block : codes:1 ] here we simulate the performance of some ldpc convolutional codes obtained according to the above generalization of the `` diagonal cut . '' namely , we start with a randomly - constructed @xmath220-regular ldpc block code based on a parity - check matrix of size @xmath349 . therefore @xmath350 , @xmath351 , and @xmath352 . ( note that @xmath353 and @xmath354 in this case . ) figure [ fig : results400 ] shows the performance of the resulting family of ldpc convolutional codes , where @xmath52 varies in powers of @xmath263 from @xmath17 to @xmath355 , each with constraint length @xmath356 . we make the following observations . first , the case @xmath357 is not interesting because it results in @xmath358 , i.e. , it is a trivial concatenation of copies of the block code , and so the ber is the same as for the underlying block code . secondly , for all other choices of @xmath52 , the constructed codes perform very similarly , each exhibiting a sizable `` convolutional gain '' compared to the block code , although the syndrome former memory @xmath347 is different in each case . a special case of the above code construction deserves mention . when @xmath359 , i.e. , @xmath360 and @xmath11 are relatively prime , the only possible step size is obtained by choosing @xmath361 , which results in the above - mentioned uninteresting case of trivial concatenations of copies of the block code . however , all - zero columns can be inserted in the parity - check matrix such that a value of @xmath362 is obtained , which allows a step size to be chosen that results in a convolutional code with @xmath363 . the variable nodes corresponding to the all - zero columns are not transmitted , i.e. , they are punctured , so that the rate corresponds to the size of the original parity - check matrix . for the `` diagonal cut '' ldpc convolutional code constructions discussed above , the unwrapped convolutional codes have the minimum possible constraint length @xmath348 , which is equal to the block length of the underlying block code . although this is a desirable property for practical implementation , we do not need to limit ourselves to diagonal cuts in general . inspired by the graph - cover construction of figures [ fig : graph : cover : constructions:2](b ) and [ fig : graph : cover : constructions:2](d ) in example [ ex : graph : cover : construction:2 ] , instead of a `` diagonal cut '' we now consider a `` random cut , '' which we define as a partition of the parity - check matrix into two matrices that add up ( over @xmath1 ) to the parity - check matrix . despite the randomness of this approach , several of the key unwrapping properties of the `` diagonal cut '' are preserved . for example , the computational complexity per decoded bit does not change , since the degree distributions of the resulting codes are all equal . however , the ldpc convolutional codes based on a `` random cut '' typically require larger decoding processor sizes as a result of increased code constraint lengths . -regular ldpc block code using random partitions . ] [ ex : random : cuts : big : ldpc : block : codes:1 ] we continue example [ ex : diagonal : cuts : big : ldpc : block : codes:1 ] ; however , instead of performing `` diagonal cuts , '' we perform `` random cuts . '' figure [ fig : partition ] shows the performance of five such ldpc convolutional codes , each with rate @xmath76 and constraint length @xmath364 , compared to the underlying block code and the ldpc convolutional code constructed in example [ ex : diagonal : cuts : big : ldpc : block : codes:1 ] ( with parameters @xmath365 , @xmath366 , @xmath367 , and @xmath356 ) . we note that the increase in constraint length from @xmath356 to @xmath364 due to the `` random cut '' results in a small additional coding gain in exchange for the larger decoding processor size . finally , we note that , for a size @xmath368 sparse parity - check matrix @xmath369 with @xmath370 nonzero entries , there are a total of @xmath371 possible ways of choosing a random cut . however , due to the sparsity , there are only @xmath372 distinct random cuts , where @xmath373 . in this section we briefly discuss some other graph - cover - based ldpc code constructions proposed in the literature , namely by ivkovic _ et al . _ @xcite , by divsalar _ et al . _ @xcite , by lentmaier _ _ et al.__@xcite , and by kudekar _ et al . _ @xcite . the ldpc code construction by ivkovic _ _ et al.__in @xcite can be seen as an application of the graph - cover construction in figures [ fig : graph : cover : constructions:2](b ) and [ fig : graph : cover : constructions:2](d ) in example [ ex : graph : cover : construction:2 ] . namely , in terms of our notation , ivkovic _ et al . _ @xcite start with a parity - check matrix @xmath239 , choose the set @xmath212 , a collection of zero - one matrices @xmath374 such that @xmath375 , and the collection of permutation matrices @xmath376 , \bigl [ \begin{smallmatrix } 0 & 1 \\ 1 & 0 \end{smallmatrix } \bigr ] \bigr\}$ ] . most importantly , the decomposition of @xmath239 into @xmath377 and @xmath378 is done such that trapping sets that were present in the tanner graph of @xmath239 are not present in the tanner graph of the new parity - check matrix . in addition , ivkovic _ et al . _ give guarantees on the relationship between the minimum hamming distances of the old and new code . one of the ldpc code constructions by divsalar _ _ et al.__in @xcite is the so - called rate-@xmath76 ar4ja ldpc code construction , which was also considered earlier in example [ example : divsalar : codes : derivatives:1 ] . a particularly attractive , from an implementation perspective , version of this code construction is obtained by an iterated graph - cover construction procedure , where each graph - cover construction is based on a cyclic cover , as in the application of * gcc1 * in example [ ex : graph : cover : construction:1 ] . although cyclic covers result in simplified encoding and decoding circuitry , codes based on cyclic covers are known to have the disadvantage that the minimum hamming distance is upper bounded by a number that is a function of the proto - graph structure @xcite . however , because the cyclic cover of a cyclic cover of the proto - graph is _ not necessarily _ a cyclic cover of the proto - graph , such disadvantages are avoided to a certain extent in the ar4ja ldpc code construction . nevertheless , ultimately the minimum hamming distance of such codes will also be upper bounded by some number ; however , these bounds usually become relevant only beyond the code length of interest . the ldpc code constructions by lentmaier _ et al . _ @xcite and kudekar _ et al . _ @xcite can also be seen as iterated graph - cover constructions . we now describe a specific instance of this construction . * it starts with a proto - matrix @xmath379 $ ] . * the first graph - cover construction is very similar to the bi - infinite graph - cover construction in example [ ex : graph : cover : construction:1:infinite:1 ] and figure [ fig : graph : cover : constructions:1:variation:2 ] . namely , in terms of our notation , we define the set @xmath380 , the collection of matrices @xmath135 with @xmath381 $ ] and @xmath382 $ ] , and the collection of permutation matrices @xmath136 with @xmath383 , @xmath384 , @xmath385 , @xmath386 , @xmath387 , @xmath388 , where as before @xmath201 is a bi - infinite toeplitz matrix with zeros everywhere except for ones in the @xmath177-th diagonal below the main diagonal . * the second graph - cover construction is a random graph - cover construction of cover - degree @xmath117 . * the code is shortened . namely , for some positive integer @xmath389 all codeword indices corresponding to values outside the range @xmath390 $ ] are shortened.$ ] , we have opted to present the code construction such that the shortening is done after the two graph - cover construction steps . in this way , the structure of the code construction description matches better the description in definition [ def : iterated : gcc : ldpc : code : construction:1 ] . ] we now point out some differences between this code construction and the ldpc convolutional code construction in definition [ def : iterated : gcc : ldpc : code : construction:1 ] . namely , the ldpc code ensemble constructed above has the following properties . * the first graph - cover construction is based on bi - infinite toeplitz permutation matrices , and the second graph - cover construction is based on finite - size permutation matrices . * the analysis focuses on the case where @xmath117 and @xmath389 go to infinity ( in that order ) , i.e. , for a fixed @xmath389 the parameter @xmath117 tends to infinity . afterwards , @xmath389 tends to infinity . * the number of check nodes with degree smaller than @xmath222 in the tanner graph is proportional to @xmath117 . * in @xcite , for the binary erasure channel , when @xmath117 and @xmath389 go to infinity ( in that order ) , kudekar _ et al . _ prove that the sum - product algorithm decoding threshold for a slight variation of the above - mentioned ensemble of codes equals the maximum a - posteriori decoding threshold for the ensemble of @xmath220-regular ldpc codes . this is a very remarkable property ! ( in @xcite , using density evolution methods , lentmaier _ et al . _ give numerical evidence that this statement might also hold for binary - input output - symmetric channels beyond the binary erasure channel . ) on the other hand , the codes constructed in definition [ def : iterated : gcc : ldpc : code : construction:1 ] have the following properties . ( we assume that the underlying block code is a @xmath220-regular ldpc code . ) * the first graph - cover construction is based on finite - size permutation matrices , and the second graph - cover construction is based on bi - infinite toeplitz permutation matrices . * in a typical application of this construction , @xmath8 is fixed . * the number of check nodes with degree smaller than @xmath222 in the tanner graph of the ldpc convolutional code is proportional to @xmath8 . * for a binary - input output - symmetric channel , the performance of the unterminated ldpc convolutional code under the continuous sliding window sum - product algorithm decoding discussed in section [ sec : implement ] improves with increasing @xmath8 ( see , e.g. , fig . [ fig : sims ] ) , but the ultimate asymptotic threshold of such unterminated decoding is unknown . the differences between these two code families come mainly from the fact that the codes constructed by lentmaier _ et al . _ and kudekar _ et al . _ are essentially block codes , although sophisticated ones , whereas the codes in definition [ def : iterated : gcc : ldpc : code : construction:1 ] are convolutional codes , along with their advantages and disadvantages . in particular , the way the limits of the parameters are taken , there is a significant difference in the fraction of check nodes with degree strictly smaller than @xmath222 . namely , in the case of the codes by lentmaier _ et al . _ and kudekar _ et al . _ this fraction is a fixed non - zero function of @xmath389 ( here we assume fixed @xmath389 and @xmath391 ) , whereas in the case of the codes considered in this paper , this fraction is zero ( here we assume fixed @xmath8 and an unterminated convolutional code ) . we conclude this section with the following remarks . namely , although the convolutional codes in definition [ def : iterated : gcc : ldpc : code : construction:1 ] may not enjoy the same asymptotic thresholds as the block code constructions by lentmaier _ et al . _ and by kudekar _ et al . _ , they lend themselves to a continuous decoding architecture , as described in section [ sec : implement ] , which can be advantageous in certain applications , such as data streaming , without a predetermined frame structure . more importantly , however , it is very encouraging that the simulation results reported in this paper indicate that sizable `` convolutional gains '' are already visible for very reasonable constraint / code lengths . in the next section we discuss some possible reasons for these gains . finally , it is worth noting that , as the block lengths and associated constraint lengths of the constructions presented in this section become larger , the observed `` convolutional gains '' will become smaller since the block code results will approach their respective thresholds . this section collects some analytical results about ldpc convolutional codes . in particular , we compare the existence / non - existence of cycles in ldpc block and ldpc convolutional codes , we present some properties of pseudo - codewords , and we discuss the mostly moderate cost increase in decoder complexity that is incurred by going from ldpc block to ldpc convolutional codes . it is well known that cycles in the tanner graph representation of a sparse code affect message - passing iterative decoding algorithms , with short cycles generally pushing the performance further away from optimum . ( indeed , attempts to investigate and minimize these effects have been made in @xcite and @xcite , where the authors propose ldpc code construction procedures to maximize the connectivity of short cycles to the rest of the graph , thus also maximizing the independence of the messages flowing through a cycle . ) hence it is common practice to design codes that do not contain short cycles , so as to obtain independent messages in at least the initial iterations of the decoding process . avoiding cycles in tanner graphs also has the benefit of avoiding pseudo - codewords . to see this , let the active part of a pseudo - codeword be defined as the set of bit nodes corresponding to the support of the pseudo - codeword , along with the adjacent edges and check nodes . with this , it holds that the active part of any pseudo - codeword contains at least one cycle and/or at least one bit node of degree one . and so , given that the typical tanner graph under consideration in this paper does not contain bit nodes of degree one , the active part of a pseudo - codeword must contain at least one cycle . therefore , avoiding cycles implicitly means avoiding pseudo - codewords . let @xmath392 and @xmath239 be two parity - check matrices such that @xmath393 is a graph cover of @xmath246 . it is a well - known result that any cycle in @xmath393 can be mapped into a cycle in @xmath246 . this has several consequences . in particular , the girth of @xmath393 is at least as large as the girth of @xmath246 , and more generally , @xmath393 contains fewer short cycles than @xmath246 . for the codes constructed in this paper , this means that the unwrapping process ( from block code to convolutional code ) can `` break '' some cycles in the tanner graph of the block code . we now revisit some codes that where discussed in earlier sections and analyze their graph cycle structure using a brute - force search algorithm . defined in ( * ? ? ? * eq . ( 3.1 ) ) . note that this search technique works only for counting cycles of length smaller than twice the girth of the graph . for searching longer cycles , more sophisticated algorithms are needed . ] note that , in order to accurately compare the graph cycle distributions of two codes with different block / constraint lengths , we compute the total number of cycles of a given cycle length per block / constraint length , and divide this number by the block / constraint length . .average ( per bit node ) number @xmath394 of cycles of length @xmath52 for the tanner graphs of the block codes ( bcs ) of block length @xmath11 and convolutional codes ( ccs ) of constraint length @xmath41 discussed in example [ ex : cyc1 ] . ( all tanner graphs have girth @xmath395 . ) [ cols="<,^,^,^,^ " , ] [ ex : cyc3 ] table [ table : girth : average : cycle : lengths : example:2 ] shows the cycle analysis results for the rate-@xmath76 proto - graph - based codes that were discussed in example [ example : divsalar : codes : derivatives:1 ] and whose ber performance was plotted in figure [ fig : jplsims ] . from examples [ ex : cyc1 ] and [ ex : cyc3 ] , we see that many of the short cycles in the tanner graphs of the ldpc block codes are `` broken '' to yield cycles of larger length in the tanner graphs of the derived ldpc convolutional codes . this section collects some comments concerning the pseudo - codewords of the parity - check matrices under consideration in this paper . we start by observing that many of the statements that were made in @xcite about pseudo - codewords can be extended to the setup of this paper . in particular , if some parity - check matrices @xmath392 and @xmath239 are such that @xmath393 is a graph cover of @xmath246 , then a pseudo - codeword of @xmath392 can be `` wrapped '' to obtain a pseudo - codeword of @xmath239 , as is formalized in the next lemma . let the parity - check matrices @xmath392 and @xmath239 be such that @xmath393 is an @xmath117-fold graph cover of @xmath246 . more precisely , let @xmath396 for some set @xmath127 , for some collection of parity - check matrices @xmath254 such that @xmath397 ( in @xmath1 ) , and for some collection of @xmath398 permutation matrices @xmath266 . moreover , let @xmath399 be the set of column indices of @xmath239 and let @xmath400 with @xmath401 be the set of column indices of @xmath392 . with this , if @xmath402 is a pseudo - codeword of @xmath392 , then @xmath403 with @xmath404 is a pseudo - codeword of @xmath239 . ( sketch . ) there are different ways to verify this statement . one approach is to show that , based on the fact that @xmath405 satisfies the inequalities that define the fundamental polytope of @xmath392 @xcite , @xmath406 satisfies the inequalities that define the fundamental polytope of @xmath239 . ( we omit the details . ) another approach is to use the fact that pseudo - codewords with rational entries are given by suitable projections of codewords in graph covers @xcite . so , for every pseudo - codeword @xmath405 of @xmath392 with rational entries , there is some graph cover of @xmath393 with a codeword in it , which , when projected down to @xmath393 , gives @xmath405 . however , that graph cover of @xmath393 is also a graph cover of @xmath246 , and so this codeword , when projected down to @xmath246 , gives @xmath406 as defined in . ( we omit the details ; see @xcite for a similar , but less general , result . ) one can then proceed as in @xcite and show that the awgnc , the bsc , and the bec pseudo - weights @xcite of @xmath405 will be at least as large as the corresponding pseudo - weights of @xmath406 . as a corollary , the minimum awgnc , bsc , and bec pseudo - weights of @xmath392 are , respectively , at least as large as the corresponding minimum pseudo - weights of @xmath239 . similar results can also be obtained for the minimum hamming distance . because the high - snr behavior of linear programming decoding is dominated by the minimum pseudo - weight of the relevant parity - check matrix , the high - snr behavior of linear programming decoding of the code defined by @xmath392 is at least as good as the high - snr behavior of linear programming decoding of the code defined by @xmath239 . in general , because of the observations made in section [ sec : cycleanalysis ] about the `` breaking '' of cycles and the fact that the active part of a pseudo - codeword must contain at least one cycle , it follows that the unwrapping process is beneficial for the pseudo - codeword properties of an unwrapped code , in the sense that many pseudo - codewords that exist in the base code do not map to pseudo - codewords in the unwrapped code . it is an intriguing challenge to better understand this process and its influence on the low - to - medium snr behavior of linear programming and message - passing iterative decoders , in particular , to arrive at a better analytical explanation of the significant gains that are visible in the simulation plots that were shown in section [ sec : variations : unwrapping:1 ] . to this end , the results of @xcite and @xcite with respect to some related code families ( see the discussion in section [ sec : other : ldpc : code : constructions:1 ] ) will be very helpful , since they indicate that some of the features of the fundamental polytope deserve further analysis . in this subsection , we investigate the cost of the convolutional gain by comparing several aspects of decoders for ldpc block and convolutional codes . in particular , we consider the computational complexity , hardware complexity , decoder memory requirements , and decoding delay . more details on the various comparisons described in this section can be found in @xcite . ldpc block code decoders and ldpc convolutional code decoders have the same computational complexity per decoded bit and per iteration since ldpc convolutional codes derived from ldpc block codes have the same node degrees ( row and column weights ) in their tanner graph representations , which determines the number of computations required for message - passing decoding . we adopt the notion of _ processor size _ to characterize the hardware complexity of implementing the decoder . a decoder s processor size is proportional to the maximum number of variable nodes that can participate in a common check equation . this is the block length @xmath11 for a block code , since any two variable nodes in a block can participate in the same check equation . for a convolutional code , this is the constraint length @xmath41 , since no two variable nodes that are more than @xmath41 positions apart can participate in the same check equation . the constraint lengths of the ldpc convolutional codes derived from ldpc block codes of length @xmath11 satisfy @xmath407 . therefore , the convolutional codes have a processor size less than or equal to that of the underlying block code . on the other hand , the fully parallel pipeline decoding architecture penalizes ldpc convolutional codes in terms of decoder memory requirements ( and decoding delay / latency ) as a result of the @xmath57 iterations being multiplexed in space rather than in time . the pipeline decoder architecture of figure [ fig : decoder ] consists of @xmath57 identical processors of size @xmath41 performing @xmath57 decoding iterations simultaneously on independent sections of a decoding window containing @xmath57 constraint lengths of received symbols . this requires @xmath57 times more decoder memory elements than an ldpc block code decoder that employs a single processor of size @xmath408 performing @xmath57 decoding iterations successively on the same block of received symbols . therefore , the decoder memory requirements and the decoding delay of the pipeline decoder are proportional to @xmath409 , whereas the block decoder s memory and delay requirements are only proportional to @xmath11 . another way of comparing the two types of codes , preferred by some researchers , is to equate the block length of a block code to the memory / delay requirements , rather than the processor size , of a convolutional code , i.e. , to set @xmath410 . in this case the block code , now having a block length many times larger than the constraint length of the convolutional code , will typically ( depending on @xmath57 ) outperform the convolutional code , but at a cost of a much larger hardware processor . finally , as noted in section [ sec : ldpc : convolutional : codes:1 ] , the parallel pipeline decoding architecture for ldpc convolutional codes can be replaced by a serial looping decoding architecture , resulting in fewer processors but a reduced throughput along with the same memory and delay requirements . in summary , the convolutional gain achieved by ldpc convolutional codes derived from ldpc block codes comes at the expense of increased decoder memory requirements and decoding delays . although this does not cause problems for some applications that are not delay - sensitive ( e.g. , deep - space communication ) , for other applications that are delay - sensitive ( e.g. , real - time voice / video transmission ) , design specifications may be met by deriving ldpc convolutional codes from shorter ldpc block codes , thus sacrificing some coding gain , but reducing memory and delay requirements , or by employing a reduced window size decoder , as suggested in the recent paper by papaleo _ et al . _ @xcite , with a resulting reduction in the `` convolutional gain . '' in this paper we showed that it is possible to connect two known techniques for deriving ldpc convolutional codes from ldpc block codes , namely the techniques due to tanner and due to jimnez - feltstr " om and zigangirov . this connection was explained with the help of graph covers , which were also used as a tool to present a general approach for constructing interesting classes of ldpc convolutional codes . because it is important to understand how the presented code construction methods can be used and in particular combined we then discussed a variety of ldpc convolutional code constructions , along with their simulated performance results . in the future , it will be worthwhile to extend the presented analytical results , in particular to obtain a better quantitative understanding of the low - to - medium snr behavior of ldpc convolutional codes . in that respect , the insights in the papers by lentmaier _ et al . _ @xcite and kudekar _ et al . _ @xcite on the behavior of related code families will be valuable guidelines for further investigation . the authors would like to thank chris jones , michael lentmaier , david mitchell , michael tanner , and kamil zigangirov for their valuable discussions and comments . we also gratefully acknowledge the constructive comments made by the reviewers . c. berrou , a. glavieux , and p. thitimajshima , `` near shannon limit error correcting coding and decoding : turbo codes , '' in _ proc . ieee international conference on communications _ , geneva , switzerland , may 1993 , pp . 10641070 . s. y. chung , t. j. richardson , and r. l. urbanke , `` analysis of sum - product decoding of low - density parity - check codes using a gaussian approximation , '' _ ieee trans . inf . theory _ 47 , no . 2 , 657670 , feb . 2001 . s. y. chung , g. d. forney , jr . , t. j. richardson , and r. l. urbanke , `` on the design of low - density parity - check codes within 0.0045 db of the shannon limit , '' _ ieee communications letters _ , vol . 5 , no . 2 , pp . 5860 , feb . 2001 . s. bates , d. elliot , and r. swamy , `` termination sequence generation circuits for low - density parity - check convolutional codes , '' _ ieee trans . circuits and systems i _ , vol . 53 , no . 9 , pp . 19091917 , sep . s. bates , z. chen , and x. dong , `` low - density parity check convolutional codes for ethernet networks , '' in _ proc . ieee pacific rim conference on communications , computers and signal processing _ , victoria , bc , canada , aug . 2005 . s. bates , l. gunthorpe , a. e. pusane , z. chen , k. sh . zigangirov , and d. j. costello , jr . , `` decoders for low - density parity - check convolutional codes with large memory , '' in _ proc . 12th nasa symposium on vlsi design _ , coeur dalene , i d , usa , oct . 2005 . r. m. tanner , d. sridhara , a. sridharan , t. e. fuja , and d. j. costello , jr . , `` ldpc block and convolutional codes based on circulant matrices , '' _ ieee trans . inf . theory _ , 50 , no . 12 , pp . 29662984 , dec . 2004 . r. m. tanner , `` on quasi - cyclic repeat - accumulate codes , '' in _ proc . of the 37th allerton conference on communications , control , and computing _ , allerton house , monticello , il , usa , sep . 22 - 24 1999 , pp . 249259 . o. vontobel and r. koetter , `` graph - cover decoding and finite - length analysis of message - passing iterative decoding of ldpc codes , '' _ corr , available online under _ ` http://www.arxiv.org/abs/cs.it/ ` _ _ ` 0512078 ` _ _ , dec . a. e. pusane , k. sh . zigangirov , and d. j. costello , jr . , `` construction of irregular ldpc codes with fast encoding property , '' in _ proc . ieee intl . conference on commun . _ , istanbul , turkey , jun . 1115 , 2006 . m. papaleo , a. r. iyengar , p. h. siegel , j. wolf , and g. corazza , `` windowed erasure decoding of ldpc convolutional codes , '' in _ proc . ieee inform . theory workshop _ , cairo , egypt , jan . 68 2010 , pp . 7882 . s. bates , z. chen , l. gunthorpe , a. e. pusane , k. sh . zigangirov , and d. j. costello , jr . , `` a low - cost serial decoder architecture for low - density parity - check convolutional codes , '' _ ieee trans . circuits and systems i _ , vol . 55 , no . 7 , pp . 19671976 , aug . a. e. pusane , a. jimnez - feltstrm , a. sridharan , m. lentmaier , k. sh . zigangirov , and d. j. costello , jr . , `` implementation aspects of ldpc convolutional codes , '' _ ieee trans . commun . _ , vol . 56 , no . 7 , pp . 10601069 , jul . y. levy and d. j. costello , jr . , `` an algebraic approach to constructing convolutional codes from quasi - cyclic codes , '' in _ coding and quantization _ ( piscataway , nj , 1992 ) , vol . 14 of dimacs ser . discrete math . sci . , pp . 189198 , providence , ri : amer . soc . , 1993 . r. smarandache , a. e. pusane , p. o. vontobel , and d. j. costello , jr . , `` pseudo - codeword performance analysis of ldpc convolutional codes , '' _ ieee trans . inf . theory _ 55 , no . 6 , pp . 25772598 , jun . 2009 . f. hug , i. bocharova , r. johannesson , b. kudryashov , and r. satyukov , `` new low - density parity - check codes with large girth based on hypergraphs , '' in _ proc . ieee intl . symposium on inform . austin , tx , usa , jun . 1318 2010 . a. e. pusane , r. smarandache , p. o. vontobel , and d. j. costello , jr . , `` on deriving good ldpc convolutional codes from qc ldpc block codes , '' in _ proc . ieee intl . symposium on inform . theory _ , nice , france , jun . 2429 , 2007 , pp . 12211225 . d. j. costello , jr . , a. e. pusane , s. bates , and k. sh . zigangirov , `` a comparison between ldpc block and convolutional codes , '' in _ proc . information theory and applications workshop _ , san diego , ca , usa , feb . 610 , 2006 . d. j. costello , jr . , a. e. pusane , c. r. jones , and d. divsalar , `` a comparison of ara- and protograph - based ldpc block and convolutional codes , '' in _ proc . information theory and applications workshop _ , san diego , ca , usa , jan . 29feb . 2 , 2007 . m. lentmaier , d. v. truhachev , and k. sh . zigangirov , `` on the theory of low - density convolutional codes ii , '' _ problems of information transmission ( problemy peredachy informatsii ) _ , vol . 288306 , oct .- dec . d. divsalar , c. r. jones , s. dolinar , and j. thorpe , `` protograph based ldpc codes with minimum distance linearly growing with block size , '' in _ proc . ieee global telecommun . _ , vol . 3 , st louis , mo , usa , nov . 28 dec . 5 2005 . m. ivkovic , s. k. chilappagari , and b. vasic , `` eliminating trapping sets in low - density parity - check codes by using tanner graph covers , '' _ ieee trans . inf . theory _ , 54 , no . 8 , pp . 37633768 , aug . 2008 . m. lentmaier , g. p. fettweis , k. sh . zigangirov , and d. j. costello , jr . , `` approaching capacity with asymptotically regular ldpc codes , '' in _ proc . information theory and applications workshop _ , san diego , ca , usa , feb . 813 2009 . m. lentmaier , d. g. m. mitchell , g. p. fettweis , and d. j. costello , jr . , `` asymptotically regular ldpc codes with linear distance growth and thresholds close to capacity , '' in _ proc . information theory and applications workshop _ , san diego , ca , usa , jan . 31 feb . 5 2010 . r. smarandache and p. o. vontobel , `` quasi - cyclic ldpc codes : influence of proto- and tanner - graph structure on minimum hamming distance upper bounds , '' _ submitted to ieee trans . theory , available online under _ ` http://arxiv.org/abs/0901.4129`_ _ , jan . 2009 . m. lentmaier , a. sridharan , d. j. costello , jr . , and k. s. zigangirov , `` iterative decoding threshold analysis for ldpc convolutional codes , '' _ ieee trans . inf . theory _ 56 , no . 10 , pp . 52745289 , oct . 2010 . g. d. forney , jr . , r. koetter , f. kschischang , and a. reznik , `` on the effective weights of pseudocodewords for codes defined on graphs with cycles , '' in _ codes , systems , and graphical models _ , b. marcus and j. rosenthal , eds.1em plus 0.5em minus 0.4emnew york , usa : springer verlag , 2001 , vol .

globular clusters ( gcs ) have played an important role as tracers of stellar populations under the context that gc is a simple system of stars that can be represented by an assumption of simple stellar population . recently , it has been found that gcs have much more diverse sub - populations than conventionally believed : gcs with multiple stellar populations , faint fuzzy clusters , and extended star clusters ( escs ) , etc . this leads to an issue how these new populations of gcs could challenge the long - standing usage of gcs as tracers of stellar populations and kinematics in galaxies . on the other hand , for escs that are characterized by systematically larger size than typical gcs , this kind of challenge could be another opportunity for a new tracer if escs turn out to be associated with some peculiar properties of galaxies that may allow a peek into their evolutionary history . one example is an argument that escs might be associated with dwarf galaxy accretion events ( e.g. , @xcite ) . however , this requires thorough studies on the physical , chemical , and kinematic properties of esc population . one simple complication to such studies is that escs are usually faint and are discovered in rather distant galaxies , sometimes more than several mpc away from the milky way ( mw ) . this is a serious challenge because spectroscopic observations are required for the reliable determination of physical properties , which is feasible only with nearby star clusters . in this regard , the escs in ngc 6822 , one of the nearest dwarf galaxies in the local group , are ideal targets for the spectroscopic studies . there are at least five escs in ngc 6822 discovered to date @xcite . for a small and isolated dwarf irregular galaxy , ngc 6822 is known to have a relatively large stellar halo populated with intermediate age carbon stars with @xmath10 gyr @xcite as well as red giant stars with @xmath11 gyr @xcite . combined with bright and young stars concentrated on the central body as well as hi disk - like structure @xcite , ngc 6822 has complex stellar structure that extends out to about @xmath12 arcmin , well beyond the extent of a central optical body . this has led to a suggestion that ngc 6822 might have undergone merger - like interactions rather recently , which have significantly increased the star formation rate within the last 3 gyr @xcite . to understand physical conditions and extent of primordial ngc 6822 , we need to concentrate on old stellar components that still pervade throughout the galaxy . the spheroidal or halo of a galaxy is an ideal for this purpose since it is mainly composed of old stars with much less contamination by bright and young stars usually found near the galaxy center . a recent study by @xcite shows that a significant population of intermediate age agb stars is found out to 28 arcmin , corresponding to 4 kpc , from the center of ngc 6822 in projected distance , which is consistent with the results of @xcite and @xcite . however , there is no clear evidence on the boundary of ngc 6822 halo and not much information on the properties of stellar components in the very outer part of its halo . the escs discovered in ngc 6822 have expanded the boundary of its stellar halo even further out to 77 arcmin or 10.5 kpc from the center of galaxy . since escs are star clusters with many observational advantages over individual stars , this provides an excellent opportunity to investigate the physical and kinematic properties of outer stellar components as well as to explore the extent of the galaxy . from the photometric study based on the cfht megacam image data , we have shown in @xcite that those new escs are probably old and metal poor populations . however , the results suffer uncertainties caused by insufficient spatial resolution and shallow photometric depth of the ground - based image data as well as the age - metallicity degeneracy inherent to the photometric cmd method . in this study , we present the first result of the integrated spectroscopic study for four escs among those five discovered in ngc 6822 based on the observations made with gmos at gemini - south . we assume the same distance to each esc as ngc 6822 based on the result of cmd analysis @xcite and adopt @xmath13 , an average of distance moduli measured by using rr ryrae @xcite , cepheid @xcite , and nir trgb @xcite . this translates to about 470 kpc in physical distance to ngc 6822 where @xmath14 corresponds to about 0.132 kpc in projected scale . the target escs in ngc 6822 were discovered in our previous imaging survey of ngc 6822 with cfht megacam and the detailed information on these clusters including the coordinates , magnitude , and size is listed in @xcite . the spectroscopic observations of the four escs in ngc 6822 were made using the gemini multi - object spectrograph ( gmos ) of gemini - south 8.1 m telescope in queue mode operation under the program i d gs-2011b - q-16 ( pi : n. hwang ) between august and october 2011 . we used the gmos in long - slit mode with b600_g5323 grating in which each target is placed at the central 0.75 arcsec wide slit in each exposure . the exposures were taken in multiples of 700 sec ( for c1 ) or 1800 sec ( for c2 , c3 , and c4 ) and the total integrated exposure times run from 2800 sec for c1 to 3 hours for c2/c4 and 4 hours for c3 . we dithered each exposure by @xmath15 arcsec to prevent a few bright stars from dominating the whole spectra , which is necessary since the target escs are partially resolved into their member stars . each exposure was binned by 2 pixels in the spectral direction and by 4 pixels in the spatial direction to improve the signal of the obtained spectra . the wavelength coverage of the spectra is about 39006500 with a spectral resolution of about 0.92 per pixel , safely including major absorption features of interest such as ca h+k , h@xmath16 , mgb , and some iron lines including fe5270 and fe5335 . we also obtained spectra of two old gcs in ngc 6822 , hubble vii ( h vii ) and hubble viii ( h viii ) for the comparison with target escs , and four mw gcs ( ngc 6624 , m12 , m15 , m107 ) for the calibration of lick indices using the same observational settings . the integrated exposure times for these clusters range from 1 min to 1 hour depending on the surface brightness . for the flux calibration , we observed ltt7379 as our spectral standard . each science exposure was preceded and followed by flat field and cuar arc exposures with the same observational settings . the reduction of data was carried out with the gemini package in iraf . first , we used @xmath17 to prepare the raw data for further reduction and @xmath18 to flatfield , to correct overscan , to trim , and to mosaic the science images . the wavelength calibration was established by using @xmath19 task on cuar arc taken together with science data , and then , the derived wavelength solution was applied to the corresponding science data by using @xmath20 . the calibrated science spectra were combined to make a single data and then , the combined data were traced and extracted by using @xmath21 to make one dimensional spectral data . the finally reduced spectra with several major absorption features are shown figure [ spectra ] . we have derived the radial velocities of target clusters by using @xmath22 task in iraf @xmath23 package . the spectra of four mw gcs ( ngc 6624 , m12 , m107 , and m15 ) were used as templates for the correlation with target spectra over the wavelength range of @xmath24 where several prominent absorption lines including h@xmath16 and mgb are available . the radial velocities of mw gcs were taken from the catalog by @xcite ( 2010 version ) and are in the range of -107.0 km / s for m15 to 53.9 km / s for ngc 6624 . each target spectrum was correlated with template spectra of four mw gcs and then the derived velocities were weight - averaged with errors to calculate the final radial velocity . the determined radial velocities range from @xmath0 km / s for c1 to @xmath25 km / s for c4 and are listed in table 1 . for h vii and h viii , the radial velocities are determined to be @xmath26 km / s and @xmath27 km / s , respectively . these velocities of h vii and h viii are in good agreement with the radial velocity @xmath28 km / s of ngc 6822 listed in the nasa / ipac extragalactic database ( ned ) within the measurement errors . the standardized lick index system @xcite is a set of the absorption line strength measures for low - resolution spectra that can be compared to theoretical stellar population models to derive the ages , metallicities , and [ @xmath29/fe ] of target system . there are 25 lick indices , running from cn@xmath30 to tio@xmath31 , of which measurement wavelength range was defined by @xcite . we measured the lick indices in the spectra of four escs and two hubble clusters in ngc 6822 as well as four mw gcs using gonzo , a code for lick index measurement described in @xcite . gonzo measures the line index of each lick feature according to the definition by @xcite . we used the sky spectrum for each target with gonzo to simulate the uncertainty of the lick index measurements through 100 monte carlo realizations . for each realization , gonzo measures the indices again in the simulated spectra , and the 1 standard deviation of the measured index is adopted as the index uncertainty . for the calibration of the measured lick index values , we compared the index measurements of four mw gcs , such as ngc 6624 , m12 , m107 , and m15 with the standard lick index values listed in @xcite and @xcite . then , the zero point offsets between standard lick and the measured indices were calculated in the following way : @xmath32 . we determined the zero point offsets for 24 indices that were measured in more than two mw gc calibrator spectra . the offsets and their rms are listed in table [ tab2 ] . the finally derived 24 lick indices and their errors for four escs and two hubble clusters in ngc 6822 are listesd in tables [ tab3a ] and [ tab3b ] , respectively . for the reliable determination of age and metallicity from the lick indices , we need high s / n for major indices such as h@xmath16 , mg2 , mgb , fe5270 , and fe5335 , where s / n value is measured at the continuum part of the corresponding lick indices . the s / n for those major lick indices is between 15 and 60 for c1c3 , h7 , and h8 . however , for c4 , the s / n values for the same indices are about 514 , which leads to relatively large errors . we have used two methods for the age and metallicity determination of star clusters using lick indices : a model grid method and a @xmath33 minimization method . we have used the lick index model grid method ( the grid method hereafter ) to determine the age and metallicity of target clusters . this method compares the measured and calibrated lick indices with the index grids calculated from the theoretical ssp models . in this study , we use the grids spanned by [ mgfe]@xmath34 versus balmer lines such as h@xmath16 , h@xmath35 , and h@xmath36 lines to estimate the metallicity and age @xcite . the composite index [ mgfe]@xmath34 , defined as [ mgfe]@xmath34 @xmath37 , is a good tracer for metallicity with little sensitivity to [ @xmath29/fe ] , while balmer lines are a good tracer of age @xcite . for the age and metallicity for a star cluster , we adopted a weighted average of values derived from the model grids of [ mgfe]@xmath34 versus h@xmath16 , h@xmath35 , and h@xmath36 lines , respectively @xcite . however , each theoretical model grid still depends on [ @xmath29/fe ] that can be , in principle , derived from the grid of mg2 versus @xmath38 , where @xmath39 , which is also a function of age . we have tried to determine [ @xmath29/fe ] using the grid of mg2 versus @xmath38 based on the technique used by @xcite that iterates between these model grids to finally determine the age , metallicity , and [ @xmath29/fe ] for target clusters . after some exhaustive tests with this approach , we have found that the derived [ @xmath29/fe ] values include rather large uncertainties that do not allow further physical application . therefore , we decided to use the age and metallicity values determined by applying the theoretical model with [ @xmath29/fe]=0.3 , the average value for mw gcs , to the grid of [ mgfe]@xmath34 and balmer lines . instead of using a few major lick indices , we have also applied a ` @xmath33 minimization method ' ( @xmath33 fit method hereafter ) that utilizes all available line indices to fit the model index lines to determine the age and metallicity from the input spectra . this method was suggested by @xcite as a more robust method for determining physical parameters of globular clusters . we applied this method to our target star clusters using the same ssp models with [ @xmath29/fe]=0.3 as used in the grid method . we also tested the fit result by varying [ @xmath29/fe ] values , but could not find any sign of significant improvement . the total number of lines used for each cluster is about 15 with best signal available for the corresponding target . the error of each parameter was estimated by calculating the 1@xmath40 significance level above the minimum point in the @xmath33 distribution . figure [ kin1 ] displays the radial velocities of escs and hubble clusters as well as those of carbon stars along the projected distance from the ngc 6822 center . it is known that those carbon stars exhibit a systematic rotation around the main body of ngc 6822 @xcite . figure [ kin1 ] shows clearly that escs in ngc 6822 do not fit into the scheme of rotation displayed by carbon stars in the halo of ngc 6822 @xcite . it is also evident that the velocities of escs do not show any sign of systematic motion either , while the velocities of two hubble clusters may be consistent with the rotation of carbon stars : the velocities of h vii and h viii are within @xmath41 of the mean velocity of carbon stars . this result suggests that the escs are not kinematically coupled with the intermediate stellar populations and that the esc system are pressure supported , rather than rotation supported . we have determined the mean velocity and velocity dispersion of carbon stars and escs : @xmath42 and @xmath43 km s@xmath1 for carbon stars , and @xmath44 and @xmath45 km s@xmath1 for escs . however , the carbon stars are systematically rotating so that we have derived the rotation velocity of carbon stars as @xmath46 km s@xmath1 arcmin@xmath1 or @xmath47 km s@xmath48 kpc@xmath48 based on the radial velocity data provided by @xcite . then , the rotation corrected velocity dispersion of carbon stars is @xmath49 km s@xmath1 . considering that the systematic radial velocity of ngc 6822 is @xmath28 km / s as listed in ned , the system of escs is moving toward us faster than ngc 6822 itself and carbon stars , while the velocity dispersion of esc system is very similar to that of carbon stars . these results show that esc system has kinematical properties distinct from those of the other stellar components in ngc 6822 including carbon stars . the hi - disk like structure does not have any common kinematic characteristic with esc system in ngc 6822 , either @xcite . considering the absence of any systematic motion , it may be a clue supporting that escs had accreted into the halo of ngc 6822 , which is separated from the galaxy - wide kinematics involving the carbon stars and the hi - disk . we list the measured metallicities and ages of each cluster in ngc 6822 in table [ tab4 ] , and compare those values derived from the grid and @xmath33 fit methods in figure [ grid2chi ] . the upper panel shows that both methods return consistent results on metallicity [ z / h ] . we have derived linear least square fits between two metallicity values and found that the rms values are not greater than @xmath50 . the lower panel also shows that the ages determined by the grid and @xmath33 methods are very well correlated with each other although there are some scatters with large errors . the result of linear least square fit exhibits an offset between two age measurements of about @xmath51 gyr , which is usually smaller than the age errors marked in the plot . figure [ mwgc ] shows the comparison of age and metallicities of mw gcs derived by the grid and @xmath33 fit methods with those values listed in @xcite ( 2010 version ) . we have assumed [ @xmath29/fe]=0.3 , the average value for mwgcs , to calculate the metallicities with the equation [ fe / h ] @xmath52 - 0.94 \times { \rm [ \alpha / fe]}$ ] it is noted that ages are relatively better reproduced by the grid method , while metallicities are better traced by the @xmath33 fit method . this result of comparison also suggests that the derived age of @xmath53 gyr may actually be an indication of 10 gyr or older due to the uncertainties larger than expected by the grid and @xmath33 fit methods themselves . therefore , we adopt the age and metallicity values derived by the @xmath33 fit method for further discussions unless specifically noted otherwise , while we interpret the age with a caution . however , we keep ages and metallicities derived from both methods in table [ tab4 ] and in figure [ agemet1 ] for reference . figure [ agemet1](a ) displays the age distribution of 4 escs and 2 hubble clusters in ngc 6822 , determined by the @xmath33 fit ( filled symbols ) and the grid ( open symbols ) method . the result shows that all star clusters are older than about 8 gyr except for h viii that is estimated to be @xmath54 gyr old . it is noted that the age of h vii is estimated to be about 12 gyr , while c2 is about 13 gyr and c1 is estimated about 10 gyr old . considering that the oldest age that can be measured through the current theoretical model is about 10 14 gyr , and that h vii is known as a genuine old gc in ngc 6822 through the existing spectroscopic study @xcite , this result supports that escs c1 and c2 are almost as old as h vii , and c1 is a genuine old gc with the largest separation of about 77 arcmin or 10.5 kpc from the center of ngc 6822 . c3 is about 7.8 gyr old although c3 is located relatively closer to ngc 6822 center than the other escs and has higher chance of younger stellar populations from the galaxy body . this implies that c3 may be slightly younger than h vii but still old , especially considering the uncertainty in the age determination . the age of c4 is about 9.0 gyr with rather large error of about 3 gyr . this large error may be due to the extended structure with low surface brightness of c4 , resulting in low s / n spectra . combined with its integrated color @xmath55 from @xcite , c4 is probably old with @xmath56 8 10 gyr in spite of the large age error . it is noted that the age of h viii is estimated younger than the other clusters of ngc 6822 investigated in this study : @xmath57 gyr with the @xmath33 fit and @xmath58 gyr with the grid method . the age measurements made by individual h lines return @xmath59 gyr for h@xmath16 , @xmath60 gyr for h@xmath35 , and @xmath61 gyr for h@xmath36 lines . therefore , h viii may be slightly older than indicated by the grid and the @xmath33 fit methods , maybe up to about @xmath62 gyr . however , it still suggests that h viii is younger than the other gcs in ngc 6822 including h vii and four escs . this result on the age of h viii is in qualitative agreement with the result by @xcite that derived @xmath63 gyr based on the optical spectroscopy made with ctio and another report by @xcite that derived @xmath64 gyr based on the optical photometry data obtained with hst . on the other hand , @xcite reports that the age of h viii is about 34 gyr based on the low resolution spectroscopy made with keck , which is much older than the other results mentioned above . figure [ agemet1](b ) displays the radial metallicity distribution of 4 escs and 2 hubble clusters in ngc 6822 , calculated by assuming [ @xmath29/fe]=0.3 , the average value for mwgcs . this shows that escs and h vii are very metal poor with [ fe / h]@xmath65 , ranging from [ fe / h]@xmath66 for c4 to [ fe / h]@xmath67 for c3 . for h viii , the metallicity is estimated to be [ fe / h]@xmath68 . this metallicity for h viii is higher than [ fe / h]@xmath69 derived for h viii by @xcite . however , the metallicity for h vii , [ fe / h]@xmath70 is consistent with the values in the literature , such as [ fe / h]@xmath71 by @xcite and [ fe / h]@xmath72 by @xcite . the metallicity of c1 is [ fe / h]@xmath73 , which suggests that the outermost esc of ngc 6822 is a genuinely metal poor population . the result on the age and metallicity in this study of ngc 6822 star clusters shows that star clusters older than about 8 gyr are metal poor with [ fe / h]@xmath4 , and that the old and metal poor clusters are preferentially located in the outer halo , farther than 15 arcmin or 2 kpc from the center of ngc 6822 except for h vii . combined with the radial distribution of carbon stars shown in figure [ kin1 ] , this suggests that the old stellar halo of ngc 6822 , located roughly more than 20 arcmin or 2.6 kpc away from the galaxy center , is composed of old ( @xmath56 8 gyr ) and metal poor ( [ fe / h]@xmath74 ) stellar populations . compared with this , the inner halo of ngc 6822 is populated with intermediate age ( @xmath75 gyr ) and slightly metal rich ( [ fe / h]@xmath76 ) as well as old ( @xmath77 gyr ) and metal poor ( [ fe / h]@xmath5 ) populations . the existence of an old stellar spheroidal or halo around a dwarf galaxy and its physical properties have been subjects of many studies . for ngc 6822 , an extended structure of old and intermediate age stars has been first detected by @xcite . further investigation by @xcite and @xcite revealed that the spheroidal or halo structure is extended out to , at least , about 30 arcmin from the center of ngc 6822 . the mean metallicity of stars in ngc 6822 is [ fe / h]=@xmath78 based on the near infrared photometry of rgb stars @xcite , which is in good agreement with [ fe / h]=@xmath79 determined by the equivalent widths of caii triplet lines @xcite . however , it is noted by @xcite that the metallicity ranges from -0.5 to -2.0 for stars selected in the central 5 to 11 arcmin region of ngc 6822 . this metallicity spread is also consistent with [ fe / h]@xmath80 for old rr lyrae variables @xcite and [ fe / h]@xmath81 for a - type supergiants @xcite in the central region of ngc 6822 . recently , @xcite show that the global metallicity of stars is [ fe / h]@xmath82 within about 28 arcmin radius from ngc 6822 center , hinting an existence of systematically lower metallicity component in the outer halo . the star clusters of ngc 6822 investigated in this study are old ( @xmath83 gyr ) and metal poor ( [ fe / h]@xmath4 ) except for one cluster : h viii is relatively young ( @xmath84 gyr ) and metal rich ( [ fe / h]@xmath85 ) . it is noted that the metallicities of very old clusters , c1 , c2 , and h vii , range from @xmath86 to @xmath87 , which is in agreement with [ fe / h]@xmath80 for old rr lyrae variables @xcite . this makes escs a good tracer of old stellar population . considering the wide spatial distribution of escs , this result suggests the presence of old and metal poor stellar halo extended out to as far as 77 arcmin or 10.5 kpc ( e.g. , c1 ) from the center of ngc 6822 , one of the largest old stellar halos known for dwarf galaxies in the local group . this extent goes beyond the previously explored area @xcite and covers about three times larger radial distance than observed by @xcite . it is interesting to find out that , for ngc 6822 , the extended old and metal poor population appears to exhibit different kinematics when compared with the intermediate age components . figure [ kin1 ] shows that the radial velocities of escs do not fit into the rotation by carbon stars @xcite . considering that carbon stars with about 1 3 gyr old are concentrated near the galaxy center , the central component or main body of ngc 6822 may be kinematically detached from the extended stellar halo mostly represented by escs . it is also noted that the rotation direction of carbon stars are roughly consistent ( approaching west wing and receding east wing ) with the rotation curve of hi gas overlaid on the main body but extended toward the perpendicular direction to the stellar halo of ngc 6822 @xcite . the distinct kinematic characteristics of different components in ngc 6822 may be the evidence of merging - like events @xcite , combined with different spatial distributions of different populations @xcite . recent discoveries of many escs ( e.g. , @xcite ) have led to inquiries regarding the origin of these escs and how they are made larger than typical gcs . there have been discussions on the possible origins of these escs and some of the proposed scenarios are summarized in @xcite . among them , there are scenarios involving tidally stripped dwarf galaxies , collisions of star clusters , and intrinsically extended star clusters that have survived tidal disruption . it has been also implied that observed escs may be a mixture of heterogeneous populations with different formation mechanisms . the spectra of escs in ngc 6822 confirms that they are old stellar population with low metallicities . the large separation of these escs from the galaxy center also implies that they may be survivors of weak tidal force in the outer halo of ngc 6822 . these observational results are consistent with a theoretical model by @xcite that escs are natural products of early star formation in small dwarf galaxies with moderate to weak tidal forces , which is an ideal condition for escs to survive the tidal disruption . the relative isolation of ngc 6822 itself could have helped the formation and survival of escs in their current locations . even in the mw , it has been noted that large and extended gcs are preferentially found in the outer halo where the distance from the mw center , @xmath88 is larger than 40 kpc @xcite . many of those extended gcs belong to the young halo ( yh ) population , which leads to a suggestion that these extended gcs may have come from the accreted dwarf galaxies @xcite . considering the similar physical properties of escs in ngc 6822 to those extended gcs in mw , this may be independent evidence supporting the hypothesis that extended gcs mostly came from the accreted dwarf galaxies . however , it is also argued that extended gcs in the outer part of mw - like galaxies can be explained by introducing the expansion of initially compact gcs , without invoking accretions of dwarf galaxies @xcite . it is interesting to note that escs in ngc 6822 do not exhibit any clear signature of systematic kinematics unlike the intermediate age carbon stars , mostly concentrated around the central main body of the galaxy . this indicates that escs , old population in the halo , are kinematically detached from the central body of ngc 6822 , composed of young and intermediate age populations as well as a huge disk - like hi structure @xcite . combined with the lack of any systematic kinematics among escs , this may suggest that escs in ngc 6822 have accreted from the external components such as neighboring dwarf galaxies . the external origin of escs also implies that the observed escs may not be necessarily bound to ngc 6822 system . the best example can be the outermost esc , c1 . it should make c1 the nearest intergalactic globular cluster . at the moment , however , there is no clear evidence to confirm or to deny the dynamical containment of c1 or the other escs in ngc 6822 system . even if we are to confirm that the escs are gravitationally bound to ngc 6822 , it does not necessarily mean that they have originally formed in ngc 6822 or that they have accreted from other neighboring systems because the old ages and low metallicities are easily expected in both scenarios . at least , it seems clear that escs are kinematically separated from intermediate stellar structure ( e.g. , carbon stars ) as well as hi - disk like structure with young stars . since old rgb stars exhibit a similar spatial distribution with escs , it may be interesting to investigate the kinematical and chemical properties of old rgb and rr lyrae stars observed near the escs in ngc 6822 halo . with spectra of many such old stars , it will be easier to determine the overall kinematical properties of any halo stellar structures that might be associated with escs . the existence of such stellar structures with distinct kinematical properties , regardless of whether they are gravitationally bound to the host galaxy or not , could be construed as a result of the unrelaxed recent merger - like interactions of ngc 6822 with its neighbors . therefore , it will be very useful to compare the kinematical as well as chemical properties of escs with those of adjacent halo stars in many galaxies to solve the issue of the esc origin as well as the dwarf galaxy accretion , even in the context of evolution of large mw - like galaxies . we have estimated the dynamical mass of ngc 6822 by using the tracer mass estimator " method based on the assumption that the esc system is pressure supported as a whole because no significant signature of rotation is observed . we have adopted the definition of mass estimator defined by @xcite as eq.[eq1 ] combined with eq.[eq2 ] , which is described as a function of line of sight velocities @xmath89 and galactocentric distances @xmath90 of clusters : @xmath91 where @xmath92 the @xmath36 is determined by the tracer population density distribution and the @xmath29 is an indicator of the properties of galaxy potential under consideration . when @xmath93 , which is for the typical stellar spheroidal halo , the value of c is given by @xmath94 we also assume an isothermal potential for ngc 6822 halo , which makes @xmath95 . for comparison , two extreme cases with velocity distribution of tracers are tried : @xmath96 for an isotropic velocity distribution and @xmath97 for tracers on radial orbits and the estimated mass for each case is quoted as an upper and lower limit of the estimated mass hereafter . figure [ kin2 ] displays the estimated mass of ngc 6822 system calculated by assuming each cluster is the outermost tracer . the mass estimated by assuming a circular rotation of carbon stars is also marked for comparision . the estimated mass increases from @xmath98 to @xmath99 as we move from c3 , the innermost esc , to c1 , the outermost esc in ngc 6822 . if we limit the radial range to include only out to c2 , then the mass estimate is @xmath100 . this value is still larger than the mass estimate made with carbon stars under the rotational velocity field , which is @xmath101 . however , it is noted that the mass estimate converges to about @xmath102 between c4 and c1 , that is , between 40 and 77 arcmin from the center of ngc 6822 . we have determined the total dynamical mass of ngc 6822 , m@xmath103 , measured at the location of c1 , about @xmath104 arcmin or 10.5 kpc from the galaxy center . this mass is about twice the mass estimated from the hi rotation measurements combined by the pseudo - isohtermal ( iso ) dark matter model assumption by @xcite , which is @xmath105 enclosed within about r@xmath106 arcmin radius . we have derived the m / l value of ngc 6822 using the newly determined dynamical mass and found that ( m / l)@xmath107 . this m / l value is higher than previously known for ngc 6822 : ( m / l)@xmath108 ( m / l)@xmath109@xcite and ( m / l)@xmath110 ( m / l)@xmath109 @xcite . it is also higher than m / l of other dwarf galaxies in the local group with similar luminosities , i.e. , @xmath111 , as shown in figure 9 of @xcite : m / l is usually lower than 30 for dwarf galaxies with @xmath111 . it is noted that m / l of ngc 6822 is comparable with that of faint dwarf spheroidal galaxies such as ursa minor with @xmath112 and ( m / l)@xmath113 ( m / l)@xmath109 @xcite . this result indicates that ngc 6822 is one of the most dark matter dominated dwarf galaxies in the local group , which is also consistent with the conclusion made by @xcite based on the hi rotation curve . the spectroscopic study of the four escs as well as two hubble clusters in ngc 6822 has shown that the radial velocities of escs range from about @xmath114 km s@xmath1 ( for c1 ) to about @xmath115 km s@xmath1 ( for c4 ) , and the mean radial velocity of escs are @xmath116 km s@xmath1 , approaching toward us faster than ngc 6822 itself ( @xmath117 km s@xmath1 ) as well as system of carbon stars ( @xmath118 km s@xmath1 ) . interestingly , the escs do not display any sign of systematic rotation unlike carbon stars rotating within 20 arcmin from ngc 6822 center . the ages and metallicities derived using the lick indices suggest that the escs are old ( @xmath3 gyr ) and metal poor ( [ fe / h]@xmath4 ) . the metallicity of c1 , the outermost esc of ngc 6822 , is estimated to be [ fe / h]@xmath73 , while the metallicity of c3 , the innermost esc , is [ fe / h]@xmath119 . however , the lowest metallicity of [ fe / h]@xmath120 is expected from c2 and c4 . there is no clear sign of radial metallicity gradient in the star clusters of ngc 6822 , but it is found that both metal poor ( [ fe / h]@xmath5 ) and metal rich ( [ fe / h]@xmath6 ) components co - exist within 15 arcmin or 2 kpc from the center of ngc 6822 , while only metal poor component is observed in the outer halo with @xmath7 arcmin or 2.6 kpc . considering no systematic motion of esc system , we suggest that escs are primordial stellar components with low metallicity that had accreted into the outer halo of ngc 6822 . the extended structure may have originated under the environment of relatively weak tidal force in a small dwarf galaxy , which is consistent with the theoretical predictions . based on the wide spatial distribution of escs , we have determined that the dynamical mass of ngc 6822 is m@xmath103 enclosed within 77 arcmin or 10.5 kpc in projected distance . the m / l ratio of ngc 6822 is @xmath121 , which is much higher than previously known @xcite and about twice higher than estimated from the hi disk rotation curve @xcite . this strongly indicates that ngc 6822 is a highly dark matter dominated system among the other dwarf galaxies in the local group . it is not clear yet whether the high content of dark matter has anything with formation and/or survival of escs in ngc 6822 halo . however , with more kinematical and chemical data for old stars in ngc 6822 halo , we may be able to better understand the origin of escs as well as the evolutionary history of ngc 6822 itself , a small dwarf galaxy with complex structure . the authors are grateful to the anonymous referee for the useful comments and suggestions that helped to improve the original manuscript . this work was supported by the national research foundation of korea ( nrf ) grant funded by the korea government ( mest ) ( no . 2012r1a4a1028713 ) . n.h . was supported in part by korea gmt project operated by korea astronomy and space science institute ( kasi ) . the authors are grateful to thomas puzia for generously providing the gonzo package and to serge demers for radial velocity data of ngc 6822 carbon stars . this research has made use of the nasa / ipac extragalactic database ( ned ) which is operated by the jet propulsion laboratory , california institute of technology , under contract with the national aeronautics and space administration . hwang , n. , et al . 2005 , iau colloq . 198 : near field cosmology with dwarf elliptical galaxies , eds . h. jerjen & b. binggeli ( cambridge : cambridge univ . press ) , 257 hwang , h. s. , lee , m. g. , park , h. s. , et al . 2008 , , 674 , 869 lee , m. g. , & hwang , n. 2005 , iau colloq . 198 : near field cosmology with dwarf elliptical galaxies , eds . h. jerjen & h. binggeli ( cambridge : cambridge univ . press ) , 181 letarte , b. , demers , s. , battinelli , p. , & kunkel , w. e.2002 , , 123 , 832 lc ngc6822c1 & -61.2 @xmath122 20.4 + ngc6822c2 & -105.6 @xmath122 30.7 + ngc6822c3 & -70.9 @xmath122 17.3 + ngc6822c4 & -115.3 @xmath122 57.9 + hubble vii & -64.5 @xmath122 20.6 + hubble viii & -46.9 @xmath122 31.2 + cccc cn@xmath30 & mag & -0.084 & 0.056 + cn@xmath31 & mag & -0.079 & 0.062 + ca4227 & & -0.140 & 0.989 + g4300 & & -2.621 & 1.145 + fe4383 & & 0.843 & 0.863 + ca4455 & & -0.283 & 1.002 + fe4531 & & -0.408 & 1.375 + fe4668 & & -1.886 & 1.932 + h@xmath16 & & 0.573 & 1.028 + fe5015 & & -0.657 & 1.662 + mg@xmath30 & mag & -0.001 & 0.055 + mg@xmath31 & mag & 0.012 & 0.035 + mgb & & 0.091 & 1.071 + fe5270 & & 0.125 & 1.051 + fe5335 & & -0.386 & 1.058 + fe5406 & & -0.171 & 1.006 + fe5709 & & -0.386 & 0.996 + fe5782 & & 0.017 & 0.865 + nad & & 0.622 & 1.109 + tio@xmath30 & mag & 0.000 & 0.035 + h@xmath123 & & 1.697 & 2.794 + h@xmath124 & & -1.338 & 0.937 + h@xmath125 & & 0.910 & 1.985 + h@xmath126 & & 0.659 & 1.763 + lcccccccccccccccccccccccc c1 & -0.019 & -0.068 & 0.411 & 1.902 & 1.240 & 0.273 & 1.103 & -1.408 & 1.763 & 1.541 & 0.006 & 0.064 & 1.176 & 0.914 & 1.426 & -0.559 & 0.718 & 0.037 & 2.319 & 0.008 & 2.346 & 2.896 & 1.663 & 2.390 + c2 & -0.028 & -0.062 & 0.451 & 1.754 & 0.475 & 0.485 & 0.366 & 0.611 & 2.868 & 0.058 & 0.017 & 0.066 & 0.600 & 0.529 & 0.347 & -0.022 & -0.262 & -0.292 & 4.519 & 0.057 & 2.726 & 1.174 & 2.847 & 1.906 + c3 & -0.013 & -0.005 & 1.030 & 4.475 & -0.393 & 0.429 & 2.275 & 0.542 & 2.827 & 3.494 & 0.002 & 0.050 & 0.581 & 1.104 & 1.945 & 0.852 & 0.495 & 0.117 & 0.973 & 0.036 & -1.060 & 0.715 & -0.375 & 1.083 + c4 & -0.069 & -0.211 & -1.567 & 2.421 & -4.056 & 0.084 & 3.584 & -1.098 & 2.387 & -0.540 & 0.002 & 0.056 & -2.364 & 0.833 & 0.039 & 1.629 & 0.270 & -0.113 & 0.954 & 0.098 & 1.469 & 3.233 & -3.239 & 0.777 + h7 & 0.009 & -0.004 & 0.236 & 1.332 & 0.262 & 0.537 & 0.829 & -0.995 & 2.865 & 1.386 & 0.010 & 0.043 & 0.780 & 0.950 & 0.927 & 0.411 & 0.638 & 0.313 & 0.990 & 0.006 & 3.298 & 2.860 & 3.557 & 3.099 + h8 & -0.101 & -0.085 & 0.251 & 2.482 & -2.825 & 1.309 & 2.156 & 0.014 & 5.001 & 2.451 & -0.015 & 0.047 & 0.678 & 1.281 & 1.401 & 0.730 & 0.139 & 0.347 & 2.242 & 0.026 & 7.536 & 7.059 & 5.699 & 5.119 lcccccccccccccccccccccccc c1 & 0.002 & 0.004 & 0.143 & 0.168 & 0.192 & 0.195 & 0.203 & 0.224 & 0.226 & 0.238 & 0.007 & 0.007 & 0.244 & 0.246 & 0.248 & 0.249 & 0.249 & 0.250 & 0.251 & 0.007 & 0.280 & 0.297 & 0.303 & 0.307 + c2 & 0.005 & 0.008 & 0.269 & 0.289 & 0.328 & 0.333 & 0.357 & 0.396 & 0.400 & 0.413 & 0.012 & 0.012 & 0.424 & 0.426 & 0.429 & 0.430 & 0.431 & 0.432 & 0.433 & 0.012 & 0.479 & 0.490 & 0.500 & 0.505 + c3 & 0.004 & 0.006 & 0.221 & 0.240 & 0.274 & 0.280 & 0.294 & 0.318 & 0.320 & 0.329 & 0.010 & 0.010 & 0.343 & 0.345 & 0.347 & 0.348 & 0.349 & 0.350 & 0.351 & 0.011 & 0.387 & 0.401 & 0.415 & 0.420 + c4 & 0.013 & 0.020 & 0.720 & 0.833 & 0.951 & 0.990 & 1.035 & 1.118 & 1.126 & 1.159 & 0.035 & 0.035 & 1.191 & 1.197 & 1.206 & 1.211 & 1.214 & 1.216 & 1.221 & 0.037 & 1.383 & 1.452 & 1.503 & 1.523 + h7 & 0.001 & 0.002 & 0.073 & 0.079 & 0.092 & 0.093 & 0.097 & 0.109 & 0.109 & 0.112 & 0.003 & 0.003 & 0.115 & 0.116 & 0.116 & 0.117 & 0.117 & 0.117 & 0.118 & 0.003 & 0.130 & 0.133 & 0.138 & 0.139 + h8 & 0.005 & 0.007 & 0.249 & 0.268 & 0.322 & 0.327 & 0.343 & 0.373 & 0.375 & 0.389 & 0.010 & 0.010 & 0.403 & 0.406 & 0.410 & 0.411 & 0.413 & 0.413 & 0.415 & 0.011 & 0.459 & 0.474 & 0.484 & 0.489 lcccccc ngc6822c1 & @xmath127 & @xmath128 & -1.76 & @xmath129 & @xmath130 & -2.00 + ngc6822c2 & @xmath131 & @xmath132 & -2.06 & @xmath133 & @xmath134 & -2.53 + ngc6822c3 & @xmath135 & @xmath136 & -1.61 & @xmath137 & @xmath138 & -1.52 + ngc6822c4 & @xmath139 & @xmath140 & -2.44 & @xmath141 & @xmath142 & -2.53 + hubble vii & @xmath143 & @xmath144 & -2.10 & @xmath145 & @xmath146 & -2.34 + hubble viii & @xmath147 & @xmath148 & -1.21 & @xmath149 & @xmath150 & -0.61 +

classical work on variable selection dates back to @xcite , who proposed to choose a model that minimizes the kullback- leibler ( kl ) divergence of the fitted model from the true model , leading to the well - known akaike information criterion ( aic ) . @xcite took a bayesian approach by assuming prior distributions with nonzero probabilities on lower dimensional subspaces . he proposed what is known as the bic method for model selection . other types of @xmath0 penalties include @xmath1 @xcite , aicc @xcite , ric @xcite and ebic @xcite , among others . the @xmath0 regularization has a natural interpretation in the form of best subset selection . it also exhibits good sampling properties @xcite . however , in a high - dimensional setting , the combinatorial problem has np - complexity , which is computationally prohibitive . as a result , numerous attempts have been made to modify the @xmath0 type regularization to alleviate the computational burden . they include bridge regression @xcite , non - negative garrote @xcite , lasso @xcite , scad @xcite , elastic net @xcite , adaptive lasso or alasso @xcite , dantzig selector @xcite , sica @xcite , mcp @xcite , among others . to a certain extent , existing penalties can be classified into one of the following two categories : convex penalty and nonconvex penalty . convex penalties , such as lasso @xcite , can lead to a sparse solution and are stable as the induced optimization problems are convex . nonconvex penalties , such as scad @xcite and mcp @xcite , can on the other hand lead to sparser solutions as well as the so - called oracle properties ( the estimator works as if the identities of nonzero regression coefficients were known beforehand ) . however , the non - convexity of the penalty could make the entire optimization problem nonconvex , which in turn could lead to a local minimizer and the solution may not be as stable as the one if instead a convex penalty is used . therefore , an important issue for nonconvex penalties is a good balance between sparsity and stability . for example , both scad and mcp have an extra tuning parameter which regulates the concavity of the penalty so that , when it exceeds a threshold , the optimization problem becomes convex . it is well known that penalty functions have bayesian interpretation . the classical @xmath2 penalty ( ridge regression ) is equivalent to the bayesian estimator with a normal prior . the @xmath3-type penalties , such as lasso , alasso etc . , also have bayesian counterparts ; cf . @xcite , @xcite and @xcite . @xcite initiated the discussion about the issue of stability in model selection . he demonstrated that many model selection methods are unstable but can be stabilized by perturbing the data and averaging over many predictors . @xcite introduced the random forest , providing a way to stabilize the selection process . @xcite derived theoretical results to analyze the variance reduction effect of bagging in hard decision problems . @xcite proposed a stable selection , which combines the subsampling with high - dimensional variable selection methods . the main objective of the paper is to introduce a family of penalty functions for generalized linear models that can achieve a proper balance between sparsity and stability . because for the generalized linear models , the loss function is often chosen to be the negative log - likelihood , it is conceivable to take into consideration the form of the likelihood in the construction of penalty functions . the bayesian connection to the penalty construction and to the likelihood function make it natural to introduce penalty functions through suitable prior distributions . to this end , we introduce the family of negative absolute priors ( nap ) and use it to develop what to be called likelihood adaptively modified penalties ( lamp ) . we define two types of asymptotic stability that cover a wide range of situations and provide a mathematical framework under which the issue of stability can be studied rigorously . we show that under suitable regularity conditions , the lamp results in an asymptotically stable estimator . the rest of the paper is organized as follows . section 2 introduces the lamp family with motivations from its likelihood and bayesian connections . specific examples are given for the commonly encountered generalized linear models . in section 3 , we introduce two types of asymptotic stability and study the asymptotic properties of lamp family . the choice of the tuning parameters and an efficient algorithm are discussed in section 4 . in section 5 , we present simulation results and applied the proposed method to two real data sets . we conclude with a short discussion in section 6 . all the technical proofs are relegated to the appendix . to introduce our approach , we will first focus on the generalized linear models @xcite . it will be clear that the approach also works for other types of regression models , including various nonlinear regression models . indeed , our simulation results presented in section 5 also include the probit model , which does not fall into the exponential family induced generalized linear models . throughout the paper , we shall use @xmath4 and @xmath5 to denote the iid random variables for @xmath6 , where @xmath7 is the response observation of @xmath8 and @xmath9 is the @xmath10-dimensional covariate observation of @xmath11 , @xmath12 @xmath13 with @xmath14 . let @xmath15 , the @xmath16 matrix of observed covariates . following @xcite , we assume that the conditional density of @xmath7 given covariates @xmath9 has the following form @xmath17,\end{aligned}\ ] ] where @xmath18 is a smooth convex function , @xmath19 or @xmath20 , @xmath21 and @xmath22 is the dispersion parameter . then up to an affine transformation , the log - likelihood function is given by @xmath23.\end{aligned}\ ] ] note that the form of @xmath24 is uniquely determined by @xmath18 and vice versa . for a given @xmath18 , we propose the induced penalty function @xmath25,\ ] ] which contains three parameters @xmath26 , @xmath27 and @xmath28 . the corresponding negative penalized log - likelihood function is @xmath29 because the penalty function defined by ( [ lamp ] ) is likelihood specific , we will call such a penalty likelihood adaptively modified penalty ( lamp ) . we clearly have @xmath30 and @xmath31 . furthermore , taking the first and second derivatives , we get @xmath32 the parameters have clear interpretations : @xmath33 is the usual tuning parameter to control the overall penalty level ; @xmath34 is a location parameter , which may be fixed as a constant ; @xmath35 controls the concavity of the penalty in the same way as @xmath36 in scad ( fan and li , 2001 ) and @xmath37 in mcp ( zhang , 2010 ) . the exponential family assumption given by implies that @xmath38 and @xmath39 . when @xmath8 is nonnegative , @xmath40 . thus @xmath41 is negative and the corresponding lamp must be concave . like many other penalty functions , the family of lamps also has a bayesian interpretation . to see this , we introduce the following prior for any given @xmath18 that defines the exponential family . let @xmath42 , @xmath43 and @xmath44 @xmath45 be constants . define a prior density , to be called negative absolute prior ( nap ) , @xmath46 . \end{aligned}\ ] ] if we choose @xmath47 , @xmath48 and @xmath49 $ ] , then the posterior mode is exactly the minimizer of ( [ eq::lik - pen ] ) the penalty form of lamp . we will show that such a choice will lead to good asymptotic properties . the additive form for the penalty function suggests that the prior must be of the product form . for each @xmath50 , hyperparameter @xmath51 scales @xmath52 while hyperparameter @xmath53 gives the speed of decay . for the @xmath50th parameter @xmath52 , the larger the values of @xmath51 and @xmath53 are , the more information the prior has for @xmath52 . translating it into the penalty function , it means that values of @xmath51 and @xmath53 represent the level of penalty and they can be adjusted separately for each component . the form of nap is to similar to that of a conjugate prior , in that the shape of the prior is adapted to that of the likelihood function . however , nap also differs from the conjugate prior in several aspects : negative when @xmath54 , absolute , from separate dimension , and with a lasso term taken away . unlike conjugate prior , which assumes additional samples , this can be seen as to take away the absolute value of redundant sample information from each dimension . it is worth looking into the commonly encountered generalized linear models and examine properties of the corresponding lamps . + _ linear regression_. in this case , @xmath55 . thus , lamp reduces to the elastic net @xcite . @xmath56 _ logistic regression_. here @xmath57 . consequently , the penalty function @xmath58 , \end{aligned}\ ] ] where @xmath59 . this penalty will be called sigmoid penalty . _ poisson regression_. here @xmath60 and we have @xmath61.\ ] ] this will be called the poisson penalty . _ gamma regression_. for the gamma regression , we have @xmath62 . then , the penalty has the following form @xmath63.\ ] ] _ inverse gaussian regression_. for the inverse gaussian , we have @xmath64 . the resulting penalty @xmath65.\ ] ] _ probit regression_. as we mentioned earlier , lamp approach can also accommodate regression models not in the form of naturally parametrized exponential families . in particular , it is applicable to the probit regression for binary outcomes . in this case , @xmath66 , which leads to the following penalty form @xmath67},\ ] ] where @xmath68 and @xmath69 are respectively distribution and density functions of the standard normal random variable . [ remark::alpha_1_choice ] for the above examples , the effect of tuning parameters in the penalty function varies across settings . for example , @xmath34 does not play a role in the poisson regression . in addition , we have a natural choice for @xmath34 for the penalty functions . specifically , we can choose @xmath70 for the cases of linear , gamma , and inverse gaussian , and @xmath71 for the cases of logistic and probit . , title="fig : " ] , title="fig : " ] as the examples show , the lamp family is fairly rich . they also differ from the commonly used penalties . figure [ fig : penaltyfunctions1 ] contains plots of penalty functions ( left panel ) along with their derivatives ( right panel ) that include lasso , scad , mcp and two members of the lamp family ( sigmoid and poisson ) . here @xmath72 for mcp , @xmath73 for scad , @xmath74 for sigmoid and @xmath75 for poisson penalty , and @xmath76 for sigmoid penalty . the parameters are chosen to keep the maximum concavity of these penalties the same . figure [ fig : penaltyfunctions1 ] shows that sigmoid and poisson penalties lie between mcp and scad when the maximum concavity is the same . also , from the graphs of the derivatives , it is easy to identify that the penalties of the lamp family have continuous derivatives ( actually they have continuous derivatives of any order for most common generalized linear models ) as compared with the discontinuous ones for scad and mcp . it will be shown that this feature can make the optimization problem easier and the estimation more stable . consider a one - dimensional penalized logistic regression problem , @xmath77 $ ] , where derivative of the penalty function @xmath78 if the true parameter @xmath79 , we need @xmath80 , where @xmath81 is a @xmath82-centered ball with radius @xmath83 uniform in @xmath84 , to get @xmath85 , while for @xmath86 , we need @xmath87 . thus , to differentiate a nonzero @xmath82 from 0 , we should make the difference between @xmath88 for @xmath89 and @xmath90 as large as possible . in addition , we need to control the second derivative of the penalty function to make the penalized negative log - likelihood function globally convex . figure [ fig : concavity ] illustrates graphically how such dual objectives can be met for the logistic regression under the sigmoid penalty and mcp . the grey area represents the difference between @xmath91 and @xmath90 for the sigmoid penalty . a subset of the grey area , marked by darker color , represent the corresponding difference under mcp . the reason that mcp tends to have a smaller difference compared with sigmoid penalty is because the second derivative of the penalty needs to be controlled by that of the negative log - likelihood in order to maintain the global convexity . ] recall that the negative log - likelihood considered here is convex , and the maximum likelihood estimation is uniquely defined and stable . by adding a nonconvex penalty , the convexity may be destroyed . to study stability of the penalized mle , it is necessary to study the impact of the penalty on the concavity , especially when @xmath84 is large . for the negative penalized maximum log - likelihood estimation procedure , if nonconvex penalties are used , the solution to the minimization problem may not be unique . therefore , it is natural to study the behavior of the local maximizers in penalized likelihood estimates when the observations are perturbed . here we introduce a new concept of asymptotic stability to describe the asymptotic performances of local minimizers in penalized likelihood estimates . note that even for the negative penalized likelihood estimators with convex penalty where the unique minimizer exists , such asymptotic stability concept is still useful in characterizing the behavior of the global maximizer . suppose we want to minimize with respect to @xmath92 a criterion function @xmath93 , where @xmath94 and @xmath95 is the @xmath50th observation of @xmath96 . denote by @xmath97 and @xmath98 the support for @xmath99 and domain for @xmath92 , respectively . we say that @xmath100 is a local minimizer if there exists a neighborhood within which @xmath101 attains its minimum . more precisely , the set of the local minimizers is defined as @xmath102 throughout the paper , @xmath103 denotes the @xmath2-norm of a vector or matrix ( vectorized ) while @xmath104 denotes the @xmath3 norm . for any @xmath105 let @xmath106 the be @xmath107-centered ball with radius @xmath108 . it is clear from the definition that the set of local maximizers is random . we characterize its asymptotic behavior in terms of whether or not the set converges to a single point as @xmath109 . for a set @xmath110 , define its diameter as @xmath111 . we say that the set of local minimizers of @xmath112 satisfies weak asymptotic stability if @xmath113 @xmath114>\delta\big)=0.\ ] ] the weak asymptotic stability characterizes the asymptotic behavior of local minimizers when the data are perturbed slightly . it shows that for large @xmath84 and small perturbation , the local minimizers stay sufficiently close to each other with high probability . defined below is a stronger version , which guarantees uniqueness of the minimizer . we say that the set of local minimizers of @xmath112 satisfies strong asymptotic stability if @xmath115=0\big)=1.\ ] ] under the weak asymptotic stability , multiple minimizers , though shrinking to @xmath116 , may exist with high probability for any finite @xmath84 . the strong asymptotic stability , on the other hand , entails that for sufficiently large @xmath84 , the probability of having multiple minimizers must converge to @xmath116 , implying that there must be a unique minimizer with high probability . @xmath0 penalties , though may have the weak asymptotic property if we adjust its tuning parameter like aic , will never possess the strong asymptotic property , because then the solution of each submodel with the number of parameters constrained will be a local optimizer , and with no probability all solutions will coincide , as @xmath117 . we now consider the situation in which @xmath93 can be approximated by an i.i.d . sum with a remainder term , i.e. , @xmath118 this general form includes the form of a negative log - likelihood plus a penalty . let @xmath119 . we assume throughout the rest of the paper that @xmath98 is compact with the true parameter value , denoted by @xmath120 , lying in its interior and that @xmath121 is finite . we need the following regularity conditions . 1 . for any @xmath122 is convex as a function @xmath92 , and @xmath123 is strictly convex . 2 . there exists function @xmath124 such that @xmath125 and , for any @xmath126 @xmath127 and @xmath128 , such that @xmath129 @xmath130 4 . there exists a local minimizer of that is a consistent estimator of @xmath131 5 . there exists @xmath132 such that @xmath133 when @xmath134 and @xmath135 are both smooth functions , conditions ( c2 ) and ( c3 ) can be guaranteed by assuming that the derivative of @xmath134 is bounded uniformly and that the derivative of @xmath135 tends to 0 uniformly . condition ( c5 ) is guaranteed by the convexity around the true parameters , uniformly in @xmath84 . the next two lemmas provide sufficient conditions for the two types of asymptotic stability . [ lemma::stability : weak ] if conditions ( c1)-(c3 ) are satisfied , then we have weak asymptotic stability . it is straightforward to verify that for generalized linear models , scad , mcp , sigmoid penalty and poisson penalty all satisfy conditions ( c1)-(c3 ) , leading to the weak asymptotic stability . [ lemma : stability : strong ] if conditions ( c1)-(c5 ) are satisfied , then strong asymptotic stability holds . in this subsection , we study asymptotic properties for the proposed lamp , including parameter estimation consistency , model selection consistency and asymptotic stability . suppose @xmath136 or @xmath137 is the true value of @xmath138 or @xmath139 , @xmath140 . let @xmath141 , the number of signals . without loss of generalization , we assume @xmath142 , where the true value @xmath143 has no zero element , and @xmath144 , which indicates a @xmath145 vector with each element being @xmath116 . denote @xmath146 , @xmath147 , then @xmath148 $ ] is the range of signal level . for any sequences @xmath149 we say @xmath150 if @xmath151 . here , we consider the setting of fixed @xmath152 when @xmath153 though some results may be extended to the case of @xmath154 . we first introduce certain regularity conditions which are needed for establishing asymptotic properties . 1 . @xmath155 where @xmath156 denotes the @xmath157th derivative of @xmath18 , and @xmath158 of a matrix `` @xmath159 '' denotes its minimum eigenvalue . 2 . @xmath160 at @xmath161 and @xmath162 ; @xmath163 is increasing at @xmath164 , where @xmath26 is a constant . 3 . @xmath165 , @xmath166 , and @xmath167 4 . @xmath168 [ ^t]}>.@xmath169 [ theo : lampasymp ] suppose that conditions ( c6)-(c9 ) are satisfied . then the penalized maximum likelihood estimator based on the lamp family is consistent and asymptotically normal , and achieves model selection consistency and strong asymptotic stability . [ lemma : rate ] suppose that ( c6 ) and ( c7 ) are satisfied and @xmath170=x^ul(x)$ ] , where @xmath171 is a constant and @xmath172 is negative and slowly varying at @xmath173 , i.e. , for any @xmath174 , @xmath175=1,$ ] and @xmath176 . then @xmath177 implies ( c8 ) . it can be verified that the condition on @xmath170 $ ] is satisfied for the logistic , poisson and probit regression models . to achieve model selection consistency , we can choose @xmath178 and @xmath179 we now present the corresponding results for the case of sigmoid penalty . here , @xmath180 where @xmath181 $ ] is a slowly varying function . the following conditions simplify ( c6)-(c9 ) . 1 . parameter @xmath183 is a constant ( does not depend on @xmath84 ) and @xmath184 . 3 . @xmath185(1+\rho)/\rho.\ ] ] the following proposition shows that the results of theorem [ theo : lampasymp ] carry over to the penalized logistic regression when ( c6)(c8 ) are assumed . for the penalized maximum likelihood estimator of logistic regression with the sigmoid penalty , we have parameter estimation consistency , model selection consistency and strong asymptotic stability under conditions ( c6)-(c8 ) . an important aspect of the penalized likelihood estimation method is the computational efficiency . for the lasso penalty , @xcite proposed the path - following lars algorithm . in @xcite , the coordinate - wise descent method was proposed . it optimizes a target function with respect to a single parameter at a time , cycling through all parameters until convergence is reached . for non - convex penalties , @xcite used the lqa approximation approach . in @xcite , a local linear approximation type method was proposed for maximizing the penalized likelihood for a broad class of penalty functions . in @xcite , the coordinate - wise descent method was implemented for non - convex penalties as well . @xcite proposed a hybrid approach of newton - raphson and coordinate descent for calculating the approximate path for penalized likelihood estimators with both convex and non - convex penalties . we apply quadratic approximation and use the coordinate decent algorithm similar to @xcite and @xcite . recall that our objective function is the negative penalized log - likelihood @xmath186 following @xcite , let @xmath187 and define the violation function @xmath188 we see that the objective function achieves its maximum value if and only if @xmath189 for all @xmath190 . thus we use @xmath191 as the stop condition of our iteration , with @xmath192 being the chosen tolerance threshold . in each step , we use quadratic approximation to the log - likelihood function @xmath193 where @xmath194 and @xmath195 depend on the current value of @xmath92 . the algorithm is summarized as follows . * algorithm : * set values for @xmath196 . denote by @xmath197 the @xmath198th column of @xmath199 * standardize @xmath200 * initialize @xmath201 . calculate @xmath202 and go to s3 if @xmath203 or else go to s5 . * choose @xmath204 . calculate @xmath205 @xmath206 @xmath207 if @xmath208 let @xmath209 ; else @xmath210 . * if @xmath211 else @xmath212 . calculate @xmath213 and go to s3 if @xmath203 , else go to s5 . * do the transformation of the coefficients @xmath214 due to standardization . for s2 , the initial solution @xmath215 can be taken as the zero solution or the mle or the estimate calculated using a parameter @xmath216 from the previous steps . for s4 , we first carry out the iterations for the variables in the current active set until convergence , then check whether additional variables should join the active set . alternatively , we may speed up the calculation by using warm start " . readers are referred to @xcite and @xcite for details of the strategies to speed up calculation in coordinate descent algorithms . for example , the logistic regression with sigmoid penalty is , @xmath217 @xmath218 where @xmath219,$ ] for @xmath220 we define @xmath221 , and @xmath222 . in s3 , we have the following two different approximation methods for updating . 1 . quadratic approximation from irls @xcite . let @xmath223 @xmath224 @xmath225 @xmath226 and @xmath227 where @xmath228 . 2 . quadratic approximation using taylor expansion . let @xmath229 and @xmath230,$ ] where @xmath231 $ ] and @xmath232 is the component - wise product operator . for an initial estimator @xmath215 , denote by @xmath233 a quadratic approximation of @xmath234 at @xmath215 , i.e. , @xmath235 let @xmath236 be a closed set and the objective function @xmath237 is strictly convex . in addition , assume the quadratic approximation at @xmath215 satisfies @xmath238 for all @xmath239 . then the algorithm constrained in @xmath240 ( minimization within @xmath240 ) converges to the true minimum @xmath241 . in addition , method 1 for the logistic regression satisfies the conditions on quadratic approximation . [ theo : algorithmconverge ] an important issue is how to choose tuning parameters in the penalized likelihood estimation . for the lamp family , there are three tuning parameters , i.e. , @xmath33 , @xmath35 and @xmath34 . our numerical experiences show that the resulting estimator is not sensitive to the choice of @xmath34 . in most cases , we may simply take @xmath70 or @xmath71 , depending on the type of regression ( see remark [ remark::alpha_1_choice ] ) . for @xmath33 and @xmath35 , we recommend using cross - validation or bic , so long as the solutions are stable enough . the @xmath242 package described in @xcite to determine a stable area or perform local diagnosis is recommended . there are two approaches to get a stable area : to control the smoothness of the @xmath33-estimate curve and calculate the smallest eigenvalue of the penalized likelihood at each point of the path as stated in theorem 1 . here , we take the second approach in all numerical analysis . our algorithm differs from @xmath242 in the following two aspects : we use the viol " function as the convergence criteria ; we do not use the adaptive - scale . both algorithms 1 and 2 use the linear approximation ( suppose at @xmath243 ) of the penalty term . @xmath244 for algorithm 1 , from concavity , we have @xmath245 which naturally falls into the mm - algorithm framework . to choose the @xmath246 pair , we use the hybrid approach introduced by @xcite , i.e. , combining bic , cross - validation , and convexity diagnostics . for a path of solutions with a given value of @xmath35 large enough , use aic / bic to select @xmath33 and use the convexity diagnostics to determine the locally convex regions of the solution path . if the chosen solution lies outside the stable region , we lower @xmath35 to make the penalty more convex . after this process is iterated multiple times , we can find a value of @xmath35 that produces a good balance between sparsity and convexity . then we can fix @xmath35 and use bic or cross - validation to choose the best @xmath33 . simulation studies cover logistic , poisson and probit cases . the performance of the lamp family is compared with those of lasso , scad and mcp . particular attention is given to the logistic regression to demonstrate how sparsity and stability are properly balanced . two classification examples from microarray experiments involving cancer patients are presented . we simulate from the logistic regression model with @xmath247 , @xmath248 , @xmath249 , @xmath250 , and @xmath251 , where @xmath252 with @xmath253 . the number of replications is 100 for this and all the subsequent simulations . table 1 reports true positive ( tp ) , false positve ( fp ) , proportion of correct fit ( cf ) , proportion of over fit ( of ) , proportion of under fit ( uf ) , @xmath254 ( l1 loss ) , and @xmath255 ( l2 loss ) . to compare performances among lasso , scad , mcp and the sigmoid penalty , we use @xmath256 to calculate the lasso solution path , @xmath242 to calculate the scad and mcp . for all penalties , we use ebic @xcite to choose the tuning parameter @xmath33 with other parameters fixed . the ebic parameter @xmath257 . from table [ tab : logisticcv ] , it is clear that the sigmoid penalty outperforms scad and mcp in the sense that at a similar level of tp , the sigmoid penalty results in a smaller fp . somewhat surprisingly the lasso has a competitive performance , which may be attributed to the use of the ebic selection criterion . . simulation results for model selection and estimation under different penalties . entries are mean values with the standard error in parentheses . here sig " represents the sigmoid penalty . the number inside the parentheses following by the penalty name represents the concavity parameter . [ tab : logisticcv ] [ cols="<,<,<,<,<,<,<,<",options="header " , ] penalty based regularization methods have received much attention in recent years . this paper proposes a family of penalty functions ( lamp ) that is adaptive and specific to the shapes of the log - likelihood functions . the proposed lamp family is different from the well - known lasso , scad and mcp . it can be argued that the new approach provides a good balance between sparsity and stability , two important aspects in model selection . it is shown that the resulting penalized estimation achieves model selection consistency and strong asymptotic stability , in addition to the usual consistency and asymptotic normality . an important issue is how to choose the three parameters imbedded in a lamp . the `` location '' parameter @xmath34 can be chosen in an obvious way for the standard generalized linear models , while @xmath33 , which represents the penalty level , can be chosen through standard cv or information criteria . for @xmath35 , which controls the concavity level , it is computationally intensive to use cv . it is desirable to develop more effective ways to select @xmath35 . it is also important to study the stability of the solution path . the lamp approach can be modified to handle grouped variable selection . it will also be of interest to develop parallel results for semiparametric regression models . we first state a general result about estimation consistency , model selection consistency and asymptotic normality . consider the penalized log - likelihood function @xmath258 as defined by . let @xmath259 be the true value of @xmath260 . recall @xmath261 is the nonzero part of @xmath262 and @xmath263 . for notational simplicity , let let @xmath264 , @xmath265 be i.i.d . with density @xmath266 that satisfies the following regularity conditions ( see @xcite ) : 1 . @xmath267=0,$ ] @xmath268=- \mathrm{e}_{{\boldsymbol{\theta}}}[\frac{\partial^2\log f({\boldsymbol{z } } , { \boldsymbol{\theta}})}{\partial\theta_j\partial\theta_k}],\]]where @xmath269 ; 2 . @xmath270^{\bigotimes 2}<\infty;\ ] ] 3 . there exist functions @xmath271 , a neighborhood @xmath272 of @xmath120 such that @xmath273 and @xmath274 for all @xmath275 , @xmath190 , @xmath157 and @xmath24 . note that for the generalized linear models , ( c10)-(c12 ) are satisfied under mild assumptions on covariates . @xmath276,1\leq j\leq q}|p'_{\lambda , j}(\beta)| ; \ ; p_{2,n}(\beta)\triangleq\inf_{q < j\leq p}p'_{\lambda , j}(\beta);\ ; p_{2,n}\triangleq p_{2,n}(0)\ ] ] @xmath277,1\leq j\leq q}|p''_{\lambda , j}(\beta)|;\ ] ] @xmath278,1\leq j\leq q}p''_{\lambda , j}(\beta);\ ; { \boldsymbol{\sigma}}_1= \mathrm{diag}\{p''_{\lambda,1}(|\beta_{10}|),\cdots , p''_{\lambda , q}(|\beta_{q0}|)\};\ ] ] @xmath279 let @xmath280 be the fisher information matrix at @xmath120 partitioned by the intercept , the nonzero and zero parts and @xmath281 for @xmath282 , define @xmath283 let @xmath284 be the @xmath173 norm of a matrix @xmath285 . suppose that both @xmath286 and @xmath287 are strictly convex , satisfying ( c10)-(c12 ) , and that @xmath288 are continuous at 0 . assume there exist @xmath289 such that @xmath290 , the k - th derivative of @xmath291 is continuous for @xmath292 $ ] and exists at the limit @xmath293 , @xmath294.\ ] ] @xmath295 for @xmath296 , @xmath297 then , we have the following results . 1 . ( parameter estimation consistency ) if either of the following two conditions , ( 1.a ) and ( 1.b ) , holds , then there exists a consistent local maximizer @xmath298 * @xmath299 as @xmath153 . * @xmath300 and @xmath301 2 . ( model selection consistency ) if any of the following three conditions , ( 2.a)-(2.c ) , holds , then any consistent local maximizer is model selection consistent . * as @xmath302 , @xmath303 and for any @xmath304 , @xmath305 . * there exists @xmath306 such that @xmath307 and @xmath308 ; @xmath309 ; and @xmath310 * for any @xmath311 , @xmath312 3 . ( asymptotic normality ) assume the estimator has the model selection consistency as stated in 2 . if @xmath313 , we have the asymptotic normality for the nonzero part , @xmath314 [ lemma : oracle ] lemma [ lemma : oracle ] covers commonly used penalties including lasso , scad , mcp , adaptive lasso , hard thresholding , bridge and the lamp family . note that the conditions are satisfied for scad , and are slightly weaker than those imposed in @xcite . it is also easy to verify that condition ( 2.a ) is satisfied for scad and mcp , ( 2.b ) for lasso and ( 2.c ) for bridge . suppose contrary to ( 2.a ) , @xmath315 and @xmath316 then , the resulting estimator @xmath317 does not have the model selection consistency . for any @xmath318 we have @xmath319 , @xmath320 . select @xmath321 define function @xmath322 throughout the proof , we use the following conventions : @xmath323 denotes a quantity approaching 0 as @xmath117 and @xmath324 simultaneously ; @xmath325 for approaching 0 as @xmath117 ; @xmath326 for approaching 0 as @xmath324 with @xmath84 fixed , and @xmath327 for a bounded sequence with @xmath328 if a subscript @xmath329 is used , then the convergence ( boundedness ) is in probability . second , from condition ( c1 ) , the @xmath332 are convex . we will show that for any @xmath333 $ ] , @xmath334 , and @xmath320 , @xmath335 since @xmath336 is a local minimum of @xmath337 , there exists @xmath338 such that @xmath339 where the last inequality follows from the convexity of @xmath340 . therefore , @xmath341 which , in view of condition ( c3 ) , implies for @xmath342 . the case of @xmath343 can be proved similarly . third , @xmath344 from the law of large number . note that @xmath345\\\nonumber&+&[m_n({\mathcal{z}}_n,{\boldsymbol{u}}'_{2,n})-f_2({\boldsymbol{u}}'_{2,n } ) ] \\\nonumber & - & [ \mathrm{e}m_n({\mathcal{z}}_n,{\boldsymbol{u}}_{1,n})- m_n({\mathcal{z}}_n,{\boldsymbol{u}}_{1,n})]\\\nonumber&-&[m_n({\mathcal{z}}_n,{\boldsymbol{u}}_{1,n } ) -f_1({\boldsymbol{u}}_{1,n})]\\\nonumber & + & [ f_2({\boldsymbol{u}}'_{2,n})-f_1({\boldsymbol{u}}'_{2,n})]+[f_1({\boldsymbol{u}}'_{2,n})-f_1({\boldsymbol{u}}_{1,n})]\\\nonumber & \geq&o_p(1,n)+o(1,\epsilon;n)o_p(1,n)= o_p(1,\epsilon , n).\nonumber\end{aligned}\ ] ] in other words , for any @xmath346 , @xmath347<-\delta)=0.\ ] ] similarly , @xmath348<-\delta)=0.\ ] ] combine these two together , we have , for any @xmath346 and @xmath349 $ ] , @xmath350<-\delta)=0.\ ] ] define @xmath351,\ ] ] and @xmath352 since @xmath98 is compact , there exists @xmath353 such that @xmath354 exists for any @xmath355 $ ] . since @xmath356 is strictly convex , @xmath357 as @xmath358 and @xmath359 we thus conclude from ( [ eqn : lemma2delta ] ) that for any @xmath346 @xmath360 by the definition of @xmath354 , for @xmath361 , @xmath362 therefore , for any @xmath363 , @xmath364 since @xmath346 implies @xmath365 and @xmath366 implies @xmath367 , @xmath368 can be arbitrarily small . thus , for any @xmath369 , @xmath370 from condition ( c2 ) and similar to the proof of lemma [ lemma::stability : weak ] , @xmath371}{{\boldsymbol{\theta}}_i\in{\boldsymbol{\theta}},i=1,2}}\large\{[r m_n({\mathcal{z}}_n+{\mathcal{e}}_n,{\boldsymbol{\theta}}_1)+(1-r ) m_n({\mathcal{z}}_n+{\mathcal{e}}_n,{\boldsymbol{\theta}}_2 ) \\ & - & m_n({\mathcal{z}}_n+{\mathcal{e}}_n , r{\boldsymbol{\theta}}_1+(1-r){\boldsymbol{\theta}}_2)]\\&- & [ r m_n({\mathcal{z}}_n,{\boldsymbol{\theta}}_1)+(1-r ) m_n({\mathcal{z}}_n,{\boldsymbol{\theta}}_2 ) \\ & - & m_n({\mathcal{z}}_n , r{\boldsymbol{\theta}}_1+(1-r){\boldsymbol{\theta}}_2)]\large\}\\ & = & o_p(1,n)o(1,\epsilon;n).\end{aligned}\ ] ] in view of this and condition ( c5 ) , there exists @xmath372 such that if @xmath373 then @xmath374 on the other hand , ( c4 ) and weak asymptotic stability imply @xmath375 @xmath376 thus , @xmath377 denote by @xmath378 and @xmath100 the minimizers of @xmath379 and @xmath93 over @xmath380 , respectively . we have @xmath381 thus @xmath382 which , along with strict convexity of @xmath379 as in ( [ eqn : lemmastrongstability : strictconvex ] ) , implies @xmath383 combining and , we have @xmath384 since @xmath394 is a slowly - varing function , we have @xmath395 or @xmath396 . thus , a sufficient condition will be @xmath397 . then taking @xmath398-th power on both slides , we have @xmath399^{1/u}\ll \lambda_0 $ ] , which can be guaranteed by the assumption @xmath400 . part 1 ( parameter estimation consistency ) . + it suffices to show that for any @xmath318 there exists @xmath401 such that @xmath402>1-\epsilon.\ ] ] this is because , under ( [ eqn : para - consistency ] ) , there exists a local maximizer @xmath403 such that @xmath404 , where @xmath405 . therefore , consistency holds as @xmath406 . we now deal with the four terms on right - hand side of . from condition ( 1.a ) , the first term is of order @xmath411 and the second term is of order @xmath412 . by the cauchy inequality , @xmath413 since @xmath414 is fixed , it is also of the same or a smaller order compared to the second term . as @xmath415 , @xmath416 in the fourth term vanishes . thus the fourth term is also controlled by the second term . regarding the constant involving @xmath417 , the second term contains @xmath418 , while both the first and third terms contain @xmath419 . consequently , the whole expression is controlled by its second term as long as we choose a sufficiently large @xmath420 , noting that @xmath421 is positive definite . . + first of all , from any of condition ( 2.a ) , ( 2.b ) , and ( 2.c ) , we have for @xmath84 sufficiently large , @xmath422 . assume we have an @xmath423-consistent local minimizer @xmath424 . if the model selection consistency does not hold , then there exists a @xmath425 such that @xmath426 . this will result in contradiction if we can show that there exist @xmath427 and a neighborhood of @xmath428 , @xmath429 , within which the sign of @xmath430 does not change . since @xmath431 is a solution , the sign of left derivative should be different from the sign of right derivative at this value . so the non - zero @xmath431 does not exist . taking the taylor expansion , we get @xmath432 for @xmath433 we see that the third term on the right - hand side of does not depend on @xmath434 and is only related to the sign of @xmath138 , which remains constant in @xmath435 since @xmath436 . so if the first two terms are controlled by the third one , we can derive the sparsity using the method above . the coefficient of the sign function in the third term should be positive to control the direction of the derivative . that is , @xmath437 under condition ( 2.a ) , the orders of the three terms in are @xmath438 @xmath439 where @xmath440 is the sequence given in condition ( 2.a ) . condition ( 2.a ) guarantees that the first two terms are controlled by the third one . likewise , condition ( 2.c ) leads to the same conclusion . since @xmath443 , we have @xmath444 @xmath445 and , for any @xmath446 , @xmath447 the above results , together with @xmath448 imply that @xmath449\\\nonumber & = & \tilde{l}({\boldsymbol{\theta}}_0)+np_{1,n}^2\big[-\frac{1}{2}{\boldsymbol{v}}{\mathcal{r}}{\boldsymbol{v}}^t-\sum_{j=1}^q\mbox{sgn}(\beta_{j,0})\frac{p'_{\lambda , j}(|\beta_{j,0}|)}{p_{1,n } } v_j\\\nonumber & & -\sum_{j = q+1}^p\frac{p'_{\lambda , j}(0)}{p_{1,n}}|v_j|\big](1+o_p(1)).\nonumber\end{aligned}\ ] ] as @xmath450 is an maximizer given by , there exists @xmath451 such that for any given @xmath318 there exists a constant @xmath306 such that @xmath452 from @xmath453(1+o_p(1)),\ ] ] we get @xmath454 thus we have @xmath455 now , we investigate the minimizer of @xmath456 . let @xmath457 where @xmath458 is a @xmath459 vector . we see @xmath460 in probability is a necessary condition for sparsity . from conclusions above , @xmath461\xrightarrow{~p~}0,\end{aligned}\ ] ] where @xmath462 @xmath463 @xmath464 and @xmath465 kkt conditions lead to the following equations . @xmath466 or @xmath467 where @xmath468 and @xmath469 therefore @xmath470 then @xmath471.\end{aligned}\ ] ] from ( [ eqn::conditionsparsity2 ] ) , @xmath472 . so with @xmath473 a sufficient condition for sparsity is that @xmath474 is controlled by @xmath475 coordinatewisely . that is , @xmath476 which is similar to the last equation of ( [ eq::proof - lemma-4:kkt ] ) and equivalent to @xmath477 , which is implied by @xmath478 . furthermore , it can be guaranteed by @xmath479 from , it follows that the first two terms are controlled by the third one . part 3 ( asymptotic normality and oracle property ) . + suppose @xmath480 is the local minimizer , noting that @xmath481 with probability tending to 1 . using parameter estimation consistency and model selection consistency property , for @xmath482 @xmath483 . \nonumber\end{aligned}\ ] ] thus @xmath484 \longrightarrow n\left[0,\left ( \begin{array}{cc } { \mathcal{r}}_{00 } & { \mathcal{r}}_{01 } \\ { \mathcal{r}}_{10 } & { \mathcal{r}}_{11 } \\ \end{array } \right)\right].\end{aligned}\ ] ] if the order of @xmath485 is controlled by @xmath486 , the oracle property holds . given @xmath311 , from @xmath501 we get @xmath502 from @xmath503 we get @xmath504 . together with the smoothness of the function @xmath18 , @xmath505 from condition ( c8 ) , it is obvious that @xmath506 condition ( 2.a ) and conditions for asymptotic normality in lemma [ lemma : oracle ] hold . lastly , ( c5 ) is implied by ( c9 ) . in conclusion , with ( c6)@xmath507(c9 ) , the penalized maximum likelihood estimator based on the lamp family is consistent and asymptotically normal , and achieves model selection consistency and strong asymptotic stability . the idea of the proof is adapted from @xcite and @xcite . for convenience , we define the following notations . @xmath508 where @xmath509 are the iteration times of @xmath510 respectively and @xmath511 . from the concavity of the penalty on the positive part , we have @xmath512 where @xmath513 is a @xmath514-vector with the @xmath190-th element 1 and all the others 0 ; from conditions of the theorem , we see @xmath515 and @xmath516 in the algorithm , @xmath517 then @xmath518 since @xmath519 decreases as iteration continues and has an lower bound , it converges . by monotonicity of @xmath520 ( @xmath521 represents the number of iterations ) , all points @xmath522 are in a compact set @xmath523 it is compact because @xmath524 is continuous and coercive . then there exists a convergent subsequence @xmath525 , @xmath526 @xmath527 such that @xmath528 next let @xmath529 for any @xmath530 , we have : @xmath531 assume @xmath532 . taking limit @xmath533 on both sides of the equation above , we have @xmath534 thus the subgradient of @xmath535 at @xmath536 is 0 , which is exactly the derivative of @xmath524 with respect to @xmath537 at @xmath538 because it can be easily verified that the smooth approximation keeps the first - order derivative the same . from the algorithm and the definition of the `` viol '' function , @xmath539 taking the derivative of both sides of the equation above , the subgradient of @xmath540 at @xmath536 is 0 , which is exactly the same as the partial derivative of @xmath524 with respect to @xmath541 at @xmath542 from strict convexity , @xmath100 is the unique local minimum of @xmath524 . akaike , h. ( 1973 ) . information theory and an extension of the maximum likelihood principle . _ in : _ b.n.petrov and f.csaki , eds . _ second international symposium on information theory . _ budapest : akademiai kiado , pp . 267 - 281 . gordon , g. j. , jensen , r. v. , hsiao , l .- l . , gullans , s. r. , blumenstock , joshua e. , ramaswamy , sridhar , richards , william g. , sugarbaker , d. j. and bueno , r. ( 2002 ) . translation of microarray data into clinically relevant cancer diagnostic tests using gene expression ratios in lung cancer and mesothelioma . _ cancer research _ , * 62 * 4963 - 4967 . munk , a. , bissantz , n. , dumbgen , l. and stratmann , b. ( 2006 ) . convergence analysis of generalized iteratively reweighted least squares algorithms on convex function spaces . _ siam journal on optimization _ , * 19 * , 1828 - 1845 . singh , d. , febbo , p. , ross , k. , jackson , d. , manola , j. , ladd , c. , tamayo , p. , renshaw , a. , damico , a. and richie , j. ( 2002 ) . gene expression correlates of clinical prostate cancer behavior . _ cancer cell _ * 1 * , 203 - 209 .

matter - wave solitons , that can be created in bose - einstein condensates ( becs ) , are a research subject of great interest , both as nonlinear collective excitations in macroscopic quantum matter , and due to the potential they offer in applications to high - precision interferometry , atomic - wave soliton lasers , quantum information processing , and other emerging technologies . dark and bright matter - wave solitons were experimentally created in nearly one - dimensional ( 1d ) trap configurations @xcite , filled with @xmath1rb and @xmath2li condensate , respectively . in these settings , strong radial confinement freezes transverse dynamics of the condensate , keeping atoms ( with mass @xmath3 ) in the ground state of the corresponding 2d harmonic - oscillator potential , @xmath4 , while a weak axial parabolic trap , @xmath5 , allows quasi - free motion of the soliton in the axial direction , @xmath6 . new recent experiments with @xmath7rb @xcite and related numerical simulations @xcite have showed the existence of stable solitary waves in a weakly elongated ( @xmath8 ) trap . these essentially three - dimensional ( 3d ) bright solitons exhibit rich and complex behavior , not present in their nearly-1d counterparts . notable among the new features , are inelastic collisions between solitons , which are strongly sensitive to phase relations between them and their relative speed . adding a periodic optical - lattice ( ol ) potential in the axial direction of the quasi-1d trap makes it possible to create bright matter - wave solitons of the gap type in repulsive condensates @xcite , which is a particular manifestation of very rich physics of bec trapped in ols , as reviewed in refs . @xcite . recently , 2d and 3d localized states in the gross - pitaevskii equation ( gpe , which is the basic mean - field bec model ) , that may be sustained and stabilized by cellular potentials , i.e. , multidimensional ols , have attracted a great deal of interest , see review @xcite . a great challenge to the experiment is creation of multidimensional matter - wave solitons , as well as the making of spatiotemporal solitons in nonlinear optics @xcite . in particular , multi - dimensional solitons trapped in a low - dimensional ol ( i.e. , 1d lattice in the 2d space @xcite , and 2d lattice in the 3d space @xcite ) , have been predicted @xcite . as these solitons keep their mobility in the free direction , the latter settings can be used to test head - on and tangential collisions between solitons . note that the quasi-1d lattice potential can not stabilize 3d solitons @xcite ( this becomes possible if the quasi-1d lattice is combined with the periodic time modulation of the nonlinearity provided by the feshbach - resonance - management technique @xcite ; a general account of the technique of periodic management for solitons can be found in book @xcite ) . the low - dimensional ol can also be implemented as a radial ( axially symmetric ) lattice in both 2d @xcite-@xcite and 3d @xcite settings . in both cases , stable solitons in the self - attractive medium were predicted , in the form of a spot trapped either at the central potential well , or ( in the 2d case ) in a radial potential trough ( additionally , the so - called azimuthons were predicted as azimuthally periodic deformations of vortices in the uniform 2d medium @xcite ; however , they are unstable in the case of the cubic nonlinearity ) . in the latter case , the spot - shaped soliton may run at a constant angular velocity in the trough ( thus being a _ rotary soliton _ the circular motion may be more convenient for the experimental study of mobility and collisions of matter - wave solitons ( and soliton trains ) than the already realized soliton - supporting settings in the cigar - shaped traps @xcite , as in the circular geometry the motion is not affected by the weak longitudinal trap . this setting is also convenient for studies of patterns in self - repulsive becs . in particular , dark solitons were , thus far , created in cigar - shaped traps @xcite , where the background density is essentially non - uniform , as it vanishes at edges of the trap . in a circular trough , pairs of dark solitons can be created without this complication . in fact , both axisymmetric ( azimuthally uniform ) vortices @xcite and dipole and quadrupole states , that may be regarded as stable complexes of two or four dark solitons @xcite , were predicted in the two - dimensional gpe with the repulsive nonlinearity and bessel - lattice radial potential . the stable dipole and quadrupole patterns may also rotate at a constant angular velocity @xcite . radial potentials in the form of @xmath9 or @xmath10 , adopted in refs . @xcite and @xcite , respectively ( @xmath11 and @xmath12 are positive constants , and @xmath13 and @xmath14 are the bessel functions ) , can support patterns of most types considered in those works in the linear limit of the model . indeed , axisymmetric states trapped in the central potential well , as well as azimuthally uniform vortices and dipolar and quadrupolar modes trapped in circular potential troughs , may be found as 2d eigenstates of the quantum - mechanical linear schrdinger equation in such radial potentials . accordingly , nonlinear states with the same symmetries found in refs . @xcite may be considered as continuations of their linear counterparts . an exception is the spot - shaped soliton in a circular trough ( the one that can perform rotary motion along the trough , as mentioned above ) . on the contrary to that , in this paper we report results for the _ radially periodic _ ol potential , @xmath15 , which , obviously , can not support any radially localized state in the linear limit . therefore , all localized patterns reported in this paper are essentially nonlinear objects , and they may be different from their counterparts found in the bessel lattices , as concerns their existence and stability conditions . the solitons that we report in this work may be categorized as solitons of the ordinary type in the model with self - attraction , and _ radial gap solitons _ ( a new species of solitary waves ) , in the case of self - repulsion . note that solitons of the gap type can not exist in the above - mentioned bessel - lattice potentials . as for the attractive model , a somewhat similar problem was considered in refs . @xcite and @xcite , where it was demonstrated that the self - attractive nonlinearity in the 1d and 2d geometry ( the latter setting is axisymmetric ) can turn a quasi - bound state in a potential well dug on top of a potential hill into a true bound state ( including 2d vortices @xcite ) . in the 3d setting , the situation is more complicated , but a similar effect is also possible @xcite . it is relevant to mention that , besides the attractive or repulsive cubic nonlinearity , the radial - lattice potential can also be combined with the self - focusing saturable nonlinearity , characteristic of photorefractive media . it was predicted @xcite and very recently experimentally demonstrated @xcite that stable solitons are possible in this case too ( in fact , ref . @xcite presents the first ever experimental realization of solitons in any radial lattice ) . another very recent experimental result is the observation of light localization , in the form of spot - like and ring - like patterns , in a photorefractive medium equipped with a third - order bessel lattice subjected to azimuthal modulation , which corresponds to a combination of saturable nonlinearity and an effective potential @xmath16 , where @xmath17 is the angular coordinate @xcite . a different experimental setup that also leads to a ring - shaped trap for the bec has been recently reported , based on a specially designed configuration of the external magnetic field that creates a guide for matter waves in the form of a torus @xcite . in parallel , nearly-1d solitons in toroidal traps loaded with self - attractive bec were studied theoretically @xcite ( in the experiment of ref . @xcite , the condensate was self - repulsive ) . the objective of this paper is to investigate various types of matter - wave solitons supported by the periodic radial lattice potential . it should be mentioned that , besides the bec context , the model and results obtained in it may also be relevant to photonic - crystal fibers with a concentric ( rather than usual hexagonal ) structure , similar to that in fibers with the multilayer cladding @xcite . in that case , the soliton will be a self - trapped beam propagating in the fiber . the paper is organized as follows . in section ii we formulate the model based on the respective two - dimensional gpe . section iii is dealing with solitons trapped in the center ( including the radial gap solitons , in the case of the model with repulsion ) , for which results are obtained by means of the variational approximation ( va ; both static and dynamical variants of the va are presented ) , and in direct simulations . section iv treats solitons trapped in annular channels ( potential troughs ) . by means of the va , we derive an effective 1d equation for the attractive model , which predicts a possibility of the existence of stable axisymmetric ring - shaped states , and the onset of their instability against azimuthal modulations at a finite density . exact modulated ( cnoidal - wave ) solutions , which appear above the instability threshold of the uniform state , are found too , as well as exact solutions for azimuthal solitons . direct numerical simulations corroborate the existence of stable uniform and modulated ring - shaped patterns , and reveal the radial gap solitons whose central love is trapped in the annular trough ( in the repulsive version of the model ) . also reported are stable ring - shaped gap solitons carrying a pair of dark solitons on their crests . section v addresses dynamics of moving two - dimensional solitons , trapped in the circular troughs of the attractive model . it is shown that the slow solitons are stable , while fast ones are destroyed due to leakage of matter into the outer trough , under the action of the centrifugal force . head - on collisions between in - phase solitons lead to collapse ( intrinsic blow - up ) , while @xmath0-out - of - phase solitons collide many times , and eventually merge into a single soliton ( without collapsing ) ; the same happens with in - phase solitons colliding tangentially , as they are trapped in adjacent troughs . finally , section vi concludes the paper . the starting point is the standard gpe in a normalized form @xcite , @xmath18 written for the single - atom wave function @xmath19 in polar coordinates @xmath20 . here , @xmath21 and @xmath22 correspond to the self - focusing and defocusing nonlinearity , respectively ( alias negative and positive scattering length of interatomic collisions ) , and , as said above , @xmath23 is the radially periodic potential with strength @xmath11 and wavenumber @xmath24 ( by dint of obvious rescaling , we set @xmath25 ) . in terms of ols , this potential can be created by a cylindrical beam whose amplitude is modulated as @xmath26 ( see further discussion below ) . note that the potential with @xmath27 gives rise to a local minimum at @xmath28 , see an example in fig . [ fig1 ] . together with eq . ( [ gpe2d ] ) , we will use its variational representation . obviously , the equation can be derived from the lagrangian @xmath29 , \label{lagr}\end{aligned}\ ] ] where @xmath30 stands for the complex conjugate . .,width=302,height=151 ] stationary axially symmetric states with chemical potential @xmath31 are looked for as solutions to eq . ( [ gpe2d ] ) in the form of @xmath32 , with a real function @xmath33 obeying the equation@xmath34 for given potential @xmath35 , solutions to eq . ( [ u ] ) form families parameterized by @xmath31 or , alternatively , by the norm,@xmath36 solutions to stationary equation ( [ u ] ) with @xmath21 ( self - attraction ) were obtained by means of a recently developed spectral - renormalization method for finding self - localized states of nonlinear - schrdinger ( nls ) type equations @xcite . independently , the same stationary solutions were also generated by dint of the well - known method which uses the integration of gpe ( [ gpe2d ] ) in imaginary time @xcite . stability of the solitary waves was then tested by adding small perturbations to them and integrating the gpe in real time . a stable perturbed soliton would shed off some radiation , which was absorbed at boundaries of the integration domain , and relax into a slightly different stationary form , corresponding to the smaller norm . in the case of repulsion ( @xmath37 ) , self - localized states were obtained by solving eq . ( [ gpe2d ] ) in real time , with an absorber placed at the boundary of the integration domain . in that case , a localized waveform ( e.g. , gaussian ) with a suitable norm was chosen as the initial condition . in the course of the evolution , excess norm is radiated away with linear waves , which are absorbed at the boundary , and a self - localized state emerges ( see an example in fig . [ fig4 ] below ) . it is obvious that solitons found this way are stable . the above - mentioned radial lattices of the bessel type can be naturally generated by a nondiffracting linear optical beam with the cylindrical symmetry ( bessel beams themselves are usually generated by means of axicoms ) @xcite . on the other hand , the radial potential of the form @xmath38 can be induced , as said above , by a cylindrical beam whose amplitude is modulated as @xmath26 . the modulation can be provided by passing the cylindrical laser beam through a properly shaped plate @xcite . unlike the bessel beam , the one with the @xmath26 transverse modulation will suffer conical diffraction , but it is not a problem for bec experiments , as a tight optical trap created in the transverse direction may readily contain a 2d pancake - shaped configuration of the condensate , of the thickness @xmath39 @xmath31 m @xcite . the diffraction of the paraxial beam , with waist diameter @xmath40 @xmath31 m , relevant to the experiment , on such a short propagation distance is completely negligible , as the respective diffraction length ( alias rayleigh range ) is estimated to be @xmath41 mm , where @xmath42 @xmath31 m is the beam s carrier wavelength . it is known that a radial potential structure can easily trap a soliton in the central potential well @xcite . we start the consideration of the present model with solutions of this type . typical examples of the solitonic shape in the model with self - attraction ( @xmath21 ) are displayed in fig . these solutions were obtained by means of both the above - mentioned spectral - renormalization method @xcite and integration in imaginary time @xcite . , trapped ( in the model with self - attraction ) at the center of the radial structure , @xmath43 , with @xmath44 ( in this case , the norm is @xmath45 ; solid line ) , and @xmath46 ( @xmath47 , dashed line).,width=302,height=226 ] in the repulsive model ( @xmath37 ) , numerical solutions were obtained , as said above , by solving eq . ( [ gpe2d ] ) in real time , in the presence of the boundary absorber . an example of the relaxation of an initial gaussian pulse into a soliton is given in fig . full views of the typical stable solitons trapped at the center of the radial structure in the model with self - attraction and self - repulsion are displayed in fig . [ fig4 ] . ( in the presence of absorbers at the boundary of the integration domain ) from an initial gaussian with norm @xmath48 . the norm of the eventually established state is @xmath49 , i.e. , @xmath50 of the initial value.,width=302,height=151 ] ( with @xmath45 , @xmath51 , and @xmath52 ) . right panel : a localized state in the repulsive bec , with @xmath49 , @xmath53 , amplitude @xmath54 , and @xmath55 , which was generated as shown in fig . [ fig3].,title="fig:",width=151,height=113 ] ( with @xmath45 , @xmath51 , and @xmath52 ) . right panel : a localized state in the repulsive bec , with @xmath49 , @xmath53 , amplitude @xmath54 , and @xmath55 , which was generated as shown in fig . [ fig3].,title="fig:",width=151,height=113 ] as concerns the solitons in the repulsive model ( @xmath37 ) , a typical example of which is displayed in the right panel of fig . [ fig4 ] , they may be clearly classified as solitons of the gap type , since they exist solely due to the interplay of the lattice potential and self - defocusing nonlinearity . the ordinary setting that gives rise to gap solitons is quite similar , with the ol in the 1d , 2d or 3d cartesian coordinates @xcite . slowly decaying fringes , attached to the central core of the soliton and evident in the right panel of fig . [ fig4 ] ( and absent in the ordinary soliton displayed in the left panel ) , is a characteristic feature of gap solitons . thus , this self - trapped state may be naturally called a _ radial gap soliton_. to describe a stationary soliton trapped in the central potential well in an analytical approximation , we adopt a gaussian ansatz , @xmath56 comparison with fig . [ fig2 ] shows that , in the model with attraction ( @xmath21 ) , this ansatz is quite adequate for solitons with a larger norm , trapped in a relatively weak lattice , but it may be inaccurate as an approximation for solitons with smaller @xmath57 trapped in a stronger lattice . the application of the variational approximation ( va ) to stationary equation ( [ u ] ) , with the help of the accordingly modified lagrangian ( [ lagr ] ) , is straightforward ( cf . @xcite , where the same gaussian ansatz was used to predict 2d solitons trapped in quasi-1d and square ols ) . thus we arrive at a set of va - generated equations that relate the soliton s size @xmath58 to its norm and chemical potential : @xmath59 -\frac{1}{a^{2 } } , \label{va1}\ ] ] where @xmath60 , @xmath61 , and @xmath62is the standard error function . the analysis of dependence @xmath63 following from eq . ( [ va1 ] ) with @xmath21 ( self - focusing ) predicts the existence of a family of solitons trapped at the center of the radial lattice , which are stable according to the vakhitov - kolokolov ( vk ) stability criterion @xcite , @xmath64 , as shown in fig . although the vk criterion is only necessary for the stability , as it ignores a possibility of oscillatory instability with complex eigenvalues , direct simulations of eq . ( [ gpe2d ] ) confirm the stability of these solutions ( see below ) . following from variational equations ( [ va1 ] ) , for the self - attraction case , @xmath65 . according to the vk criterion , @xmath66 , stable solutions belong to upper branches of the curves.,width=302,height=226 ] additional evidence for the existence of self - trapped localized states of bec in the present model can be provided by a more general , time - dependent , version of the va . to this end , we define a generalized ansatz [ cf . ( [ ansatz - zentrum ] ) ] @xmath67 , \label{ansatz}\ ] ] where @xmath68is a real radial chirp . applying the standard va procedure @xcite , one arrives at the following evolution equation for the width of the localized state , @xmath69 , \label{att}\ ] ] where an effective nonlinearity strength is @xmath70 and the amplitude is given by @xmath71 equation ( [ att ] ) describes motion of a unit - mass particle with the coordinate @xmath72 in an effective potential @xmath73 in fig . [ fig6 ] the effective potential ( [ pot ] ) is depicted for different values of @xmath74 , for the attractive and repulsive versions of the model . ) for different values of @xmath75 , at fixed strength @xmath76 of the radial optical lattice.,title="fig:",width=151,height=151 ] ) for different values of @xmath75 , at fixed strength @xmath76 of the radial optical lattice.,title="fig:",width=151,height=151 ] solitons in the attractive bec ( @xmath21 ) exist when potential ( [ pot ] ) possesses a local minimum . in particular , for @xmath77 ( see fig . [ fig6 ] ) , the potential minimum exists provided that @xmath78 . for the norm smaller than that corresponding to @xmath79 [ see eq . ( [ chi ] ) ] , the self - attraction is too weak to form a soliton , and the matter - wave pulse spreads out through barriers of the radial potential . on the other hand , in the case of attraction , @xmath74 is also limited from above , by the collapse occurring at a critical value , @xmath80 . in the absence of the lattice ( @xmath81 ) , a numerically exact critical norm is @xmath82 ( it is equal to the norm of the _ townes soliton _ ) @xcite . for any @xmath11 , the va predicts @xmath83 , as the value of coefficient @xmath74 at which the first term in effective potential ( [ pot ] ) vanishes . according to eq . ( [ chi ] ) , this gives @xmath84 @xcite . thus , the va predicts a finite existence region for stable self - trapped localized states in the attractive bec ; for example , it is @xmath85 for @xmath77 . while the upper edge of this region is fixed at @xmath86 , numerical analysis of expression ( [ pot ] ) demonstrates that the lower critical norm necessary for the existence of the self - trapped state almost linearly decreases with the increase of the ol strength @xmath11 . numerically exact solution of the @xmath87-independent version of full gpe ( [ gpe2d ] ) demonstrates that , although the collapse sets in at a critical value of the norm which is slightly higher than @xmath84 , similar to the case of the 2d soliton stabilized by the quasi-1d or square ol @xcite , the dependence of @xmath88 on @xmath11 is weak . the physical mechanism behind the rise of the collapse threshold in a trapping potential is that the confinement allows the soliton to increase the internal quantum pressure " , generated by the dispersion term in gpe ( [ gpe2d ] ) , which counteracts the nonlinear self - focusing of the wave packet . predictions of the va for dynamical regimes can be compared to direct simulations in terms of the time dependence of the amplitude , @xmath89 , in a soliton which was excited by a sudden perturbation [ the variational prediction for @xmath89 is taken as per eq . ( [ ampl ] ) ] . to display an example , we take the stationary solution from fig . [ fig2 ] , corresponding to @xmath77 and @xmath47 , and suddenly increase the lattice strength to @xmath90 . comparison of the variational results with simulations of eq . ( [ gpe2d ] ) is presented in fig . as observed , agreement for absolute values of the amplitude is , at best , qualitative , while the oscillation frequencies coincide much better . ) ( gpe " ) with @xmath65 , in terms of the soliton s amplitude as a function of time . intrinsic vibrations of the soliton were triggered by sudden increase of the lattice strength from @xmath46 to @xmath91 , for the self - trapped state with @xmath47 in fig . [ fig2].,width=226,height=151 ] in the case of the repulsive bec ( @xmath37 ) , the behavior is opposite , as seen in the right panel of fig . [ fig6 ] . in this case , a larger norm for given ol strength @xmath11 causes spreading of the localized state , because the self - defocusing nonlinearity overcomes the confining effect of the trapping potential . simulations of eq . ( [ gpe2d ] ) with @xmath37 show that the wave packet sheds off excessive norm in the form of linear waves , which are absorbed at domain boundaries , and relaxes into a stable shape after the norm becomes adequate . for example , at @xmath92 stable localized states in the repulsive bec exist at @xmath93 @xmath94 [ i.e. , the norm is limited to values @xmath57 @xmath95 , see eq . ( [ chi ] ) ] . the most essential property of radially periodic potentials is that localized states can self - trap not only in the central potential well , but also in remote circular troughs @xcite . if the trough s curvature is not significant , these localized states resemble 2d solitons in the quasi-1d ol reported in ref . @xcite . here , we aim to derive an asymptotic 1d equation for patterns trapped in a circular trough , and obtain its relevant solutions . to this end , we adopt the following ansatz for the solution , @xmath96 where @xmath97 is the coordinate running along the circumference of radius @xmath98 , @xmath99 is a slowly varying complex amplitude , and the transverse variable is @xmath100 , with @xmath101 being a radial potential minimum in a vicinity of which the ring - shaped pattern is trapped . assuming @xmath3 to be large enough , we consider the annular pattern placed in a remote circular trough , which may be treated as a quasi - rectilinear potential trap , with negligible curvature . obviously , the latter condition amounts to @xmath102 [ recall @xmath103 is the radial size in ansatz ( [ ansatz1 ] ) ] . neglecting the curvature in eq . ( [ gpe2d ] ) , we rewrite it as a 2d equation with @xmath6 and @xmath104 treated as local cartesian coordinates , @xmath105 the next step is to reduce eq . ( [ no - curv ] ) to an effectively 1d equation along azimuthal direction @xmath104 . to this end , we note that eq . ( [ no - curv ] ) corresponds to the lagrangian [ cf . ( [ lagr ] ) ] @xmath106.\end{aligned}\ ] ] in this lagrangian , we substitute ansatz ( [ ansatz1 ] ) and integrate the resulting expression in transverse direction @xmath6 . after a straightforward calculation , this yields an effective ( averaged ) lagrangian , @xmath107 , \label{leff}\end{aligned}\ ] ] where @xmath108 . the application of the standard variational procedure to averaged lagrangian ( [ leff ] ) gives rise to a cumbersome system of coupled equations for the complex amplitude @xmath99 and real width @xmath109 . the equations strongly simply if we assume that @xmath110 and @xmath103 are slowly varying functions of @xmath104 , which , in particular , corresponds to long - wave perturbations that account for the onset of the modulational instability ( mi ) , see below . in particular , in the attractive model ( @xmath21 ) one may first adopt the lowest - order approximation , completely neglecting the @xmath104-dependence and assumes time - independent @xmath111 and @xmath103 , which gives rise to a result previously known from the application of the va to the nls equation in 1d with potential @xmath112 @xcite@xmath113 : @xmath114 } . \label{a^2}\ ] ] this relation is cumbersome too , but straightforward analysis demonstrates that it takes a very simple form , @xmath115 , when @xmath11 is both small and large [ in the latter case , scaling @xmath116 is assumed , to lend the last three term in the lagrangian density in eq . ( [ leff ] ) the same order of magnitude ] . to obtain a closed - form lagrangian for the amplitude field @xmath99 in the attractive model , one should express @xmath103 in eq . ( [ leff ] ) in terms of @xmath117 , using relation ( [ a^2 ] ) . this will lead , in the general case , to a messy result ; however , the above - mentioned simplest approximation , @xmath118 , which replaces eq . ( [ a^2 ] ) , yields a tractable expression , @xmath119 \right . \nonumber \\ & & \left . -\left ( 2+i\right ) \frac{\left\vert a_{y}\right\vert ^{2 } } { \left\vert a\right\vert } -2\varepsilon |a| \right\ } \label{lfinal}\end{aligned}\ ] ] [ the last term in the lagrangian density should be dropped if @xmath11 is small ; if @xmath11 is large , the term is obtained within the framework of the above scaling , that assumes small @xmath103 , hence @xmath120 . finally , the action functional , @xmath121 , as defined by lagrangian ( [ lfinal ] ) , gives rise to the euler - lagrange equation , @xmath122 , in the following form : @xmath123 + 2\left\vert \tilde{a}\right\vert ^{2}\tilde{a}=0 , \label{new}\end{aligned}\ ] ] where @xmath124 [ this transformation eliminates a linear term in the equation generated by the last term in lagrangian ( [ lfinal ] ) ] . formally setting @xmath125 in eq . ( [ new ] ) recovers an equation known as a model for the propagation of surface waves on a plasma layer with a sharp boundary @xcite . a class of nonpolynomial generalized nls equations similar to eq . ( [ new ] ) , and some of their fundamental solutions , such as single - soliton ones , were introduced in ref . @xcite ( see also ref . @xcite ) . stationary solutions to eq . ( [ new ] ) with chemical potential @xmath31 are looked for in the ordinary form , @xmath126 , with @xmath127 satisfying equation @xmath128 + 2b^{3}+\mu b=0 $ ] , where @xmath129.@xmath113 this equation has its hamiltonian , @xmath130 setting @xmath131 , one obtains a family of soliton solutions , for any @xmath132 : @xmath133 for @xmath134 and @xmath132 , a family of periodic cnoidal - wave solutions is obtained in the form @xmath135 where @xmath136 is the jacobi s elliptic cosine , with modulus @xmath137 and @xmath138 are three roots of equation @xmath139 [ it is obtained by setting @xmath140 in eq . ( [ h ] ) ] . solutions ( [ cn ] ) exist , for given @xmath132 , in a region of @xmath141 ; as said above , the limit of @xmath131 corresponds to soliton ( [ soliton ] ) , and the opposite limit , @xmath142 cnoidal solutions on the ring of radius @xmath98 must satisfy the corresponding periodic boundary conditions ( b.c . ) , @xmath143 ( for the ordinary 1d nls equations with repulsion and attraction , analytical solutions satisfying the periodic b.c . are collected in refs . @xcite and @xcite , respectively ) . accordingly , parameter @xmath144 in solutions ( [ cn ] ) , ( [ modulus ] ) is not continuous , but rather takes discrete values selected by matching the b.c . to the periodicity of @xmath145 , @xmath146 where @xmath147 is the complete elliptic integral , and @xmath148 is an arbitrary integer . note that eq . ( [ new ] ) features the galilean invariance : if @xmath149 is a solution , then its counterpart in the form of a solution moving at an arbitrary velocity @xmath150 is @xmath151 hence one can boost solitons ( [ soliton ] ) and cnoidal waves ( [ cn ] ) to a ( formally ) arbitrary velocity by means of transformation ( [ galileo ] ) . in fact , the velocity is not arbitrary because the factor @xmath152 also must satisfy the periodic b.c . , which leads to a quantization condition , @xmath153 , with @xmath154 . in the model with attraction ( @xmath21 ) , the existence of uniform ring - shaped solitons trapped in annular troughs of large radii is obvious , while a nontrivial issue is a possibility of finding a stability region for such solutions . in terms of eq . ( [ new ] ) , the ring soliton is represented by a cw ( continuous - wave ) solution with a constant amplitude @xmath155 , which was mentioned above as a limit case of the cnoidal solution family , with @xmath156 . it is well known that , while all cw solutions of nls - like equations with attraction are modulationally unstable [ including nonpolynomial equations similar to eq . ( [ new ] ) @xcite ] , periodic b.c . may stabilize them if the cw amplitude is smaller than a certain critical value . this argument suggests a possibility to find a stability region for the uniform ring solitons in the present model with attraction . a standard analysis of the mi assumes a perturbation of the cw solution , @xmath157 \exp \left ( 2ib_{0}^{2}t+i\chi _ { 1}(y , t)\right ) , \ ] ] with infinitesimal amplitude and phase perturbations @xmath158 and @xmath159 . eigenmodes of the perturbations may be looked for assuming that they are proportional to @xmath160 , where @xmath161 is an arbitrary real wavenumber , and @xmath162 is the corresponding instability growth rate . straightforward calculations demonstrate that the cw solution to eq . ( [ new ] ) is subject to the mi under the condition @xmath163 > 0 . \label{mi}\ ] ] in the annular system that we are dealing with , @xmath161 is subjected to the geometric quantization , the same way as above , i.e. , @xmath164 with integer @xmath148 . therefore , condition ( [ mi ] ) with lowest @xmath165 is not satisfied , making the cw solution _ modulationally stable _ , under the condition @xmath166 in a final form , this stability condition may be expressed in terms of the full norm of the axisymmetric ring soliton . substituting ansatz ( [ ansatz1 ] ) in eq . ( [ norm ] ) , and making use of the assumptions adopted above , one obtains @xmath167 . further , substituting here the above approximation , @xmath168 , yields @xmath169 with regard to this relation , stability condition ( [ nomi ] ) amounts to a limitation on @xmath57 : @xmath170 ( note that this result does not contain any parameter ) . a natural way to look at these stability conditions is to consider a situation with gradually increasing @xmath57 ( or @xmath155 ) . the azimuthally uniform ring solution will loose its stability when the amplitude ( or norm ) attains the critical value defined by eq . ( [ nomi ] ) [ or by eq . ( [ thr ] ) ] . one may expect that the onset of the mi gives rise to a _ bifurcation _ , which creates stable azimuthally modulated solutions , with a modulation depth scaling as @xmath171 for @xmath172 . alternatively , one may fix the cw amplitude , @xmath155 , and gradually increase the trough s radius @xmath98 ; the loss of the cw stability and emergence of the modulated solutions should occur when @xmath98 attains a critical value following from eq . ( [ nomi ] ) , @xmath173 in fact , the modulated states generated by the mi onset should be nothing else but the stationary cnoidal wave ( [ cn ] ) , ( [ modulus ] ) . indeed , with regard to eq . ( [ cw ] ) and the fact that @xmath174 , it follows form matching condition ( [ matching ] ) that , with the increase of @xmath98 for fixed @xmath155 , the cnoidal - wave solution appears , with an infinitely small modulation depth , at @xmath175 , where the critical value @xmath176 is _ precisely the same _ as defined by the onset of the mi [ i.e. , given by eq . ( [ cr ] ) ] . a defect of the above approximate analysis is that , with regard to relation @xmath177 , stability condition ( [ thr ] ) may be cast in the form of @xmath178 , which _ does not _ comply with the underlying low - curvature assumption , @xmath179 . this fact makes the existence of modulationally stable ring - shaped solitons in the present model with attraction doubtful . in direct simulations of eq . ( [ gpe2d ] ) with @xmath21 , we were unable to find modulationally stable annular uniform nonlinear states . in this connection , it is relevant to recall that the linear limit of eq . ( [ gpe2d ] ) with potential @xmath180 does not allow any stationary solution localized in the radial direction , therefore the weak nonlinearity is essential in stabilizing the axisymmetric rings . note that localized ring - shaped states may be possible in the linear limit of the bessel - lattice models considered in refs . @xcite and @xcite , with potentials @xmath181 and @xmath182 , respectively ; however , the azimuthal ( in)stability of such states was not studied in the presence of attraction . on the other hand , _ stable _ annular patterns ( both static ones and persistent ring - shaped breathers ) were recently found in an attractive model with a single circular trough , which was created not by the linear potential , but rather by modulation of the nonlinear coefficient in the radial direction @xcite . in that model , annular states may only be stable if the annulus is _ wide enough _ , with the ratio of its outer and inner radii exceeding @xmath183 . vorticity - carrying rings were found to be unstable in the same setting . in this section we investigate the gap soliton nature and the band structure associated with ring shaped states in terms of the following 2d gpe @xmath184 , with @xmath185 the squared 2d angular momentum . this equation corresponds to the @xmath186 limit ( @xmath187 ) of the spherically symmetric 3d gpe , with @xmath188 denoting spherical coordinates . notice that apart for the factor of two in front of the term @xmath189 , due to spherical rather than cylindrical symmetry , eq . ( [ gpespher ] ) is the same as eq . ( [ gpe2d ] ) . from a physical point of view ( [ gpespher ] ) is appropriate to describe a condensate trapped in the outer part of the conical region created by a cylindrical beam of light passed through a very short focal distance lens . in the following we shall refer to this setting as the `` pizza '' geometry since , in contrast with the pancake geometry , the thickness of the condensate in the axial direction is not uniform and reduces to zero as @xmath190 . this geometry could be implemented by combining a cylindrical beam creating the radial lattice with a blue - detuned conical beam ( the pizza - shaper ) ) with @xmath191 , @xmath192 and @xmath193 . right panel : the same as in the left panel for the repulsive case @xmath194 with @xmath195 and @xmath196.,title="fig:",width=151,height=151 ] ) with @xmath191 , @xmath192 and @xmath193 . right panel : the same as in the left panel for the repulsive case @xmath194 with @xmath195 and @xmath196.,title="fig:",width=151,height=151 ] launched in the opposite direction . notice that at large radial distances ( i.e. in the outer part of the pizza ) the differences between pancake and pizza geometry become negligible especially if the thickness of the border is very small compared to the radius . ring shaped states of eq . ( [ gpespher ] ) located in potential wells far from the center are expected , therefore , to be qualitatively similar to the ones of the pancake setting eq . ( [ gpe2d ] ) . in the following we look for ring gap solitons of eq . ( [ gpespher ] ) in the form of stationary states @xmath197 of the form @xmath198 by substituting eq . ( [ stat ] ) into eq.([gpespher ] ) we obtain the following equation for the radial function @xmath199 @xmath200 where @xmath201 are integer numbers which allow to satisfy the periodic boundary condition in @xmath202 : @xmath203 . equation ( [ radial ] ) can be mapped into a 1d eigenvalue problem by means of the transformation @xmath204 leading to the following reduced radial equation @xmath205\xi = 0 \label{reducedradial}\ ] ] which has the form of a 1d schrdinger equation with periodic potential and nonlinear effective centrifugal barrier : @xmath206 ( here @xmath207 ) . for sufficiently large distances @xmath208 and fixed angular momentum and nonlinearity , this equation reduces to the mathieu equation with the well known band gap structure . the part of the spectrum associated to excitations localized nearby the origin , however , is expected to be strongly influenced by the centrifugal barrier which at short distances enhances the nonlinearity and favors the formation of localized states . to investigate the band structure and the gap solitons of eq . ( [ reducedradial ] ) we use a self - consistent method developed in ref . the gap solitons obtained with the self - consistent method will then be used as initial conditions for time propagation of the radial 2d gpe in eq . ( [ gpe2d ] ) . and corresponding to the energy levels @xmath209 of the left panel of fig . [ fig8 ] , respectively . panels c , d . repulsive radial gap solitons ( thick lines ) trapped in circular troughs of the potential at @xmath210 and corresponding to the energy levels @xmath211 of the right panel of fig . [ fig8 ] , respectively . the thin lines corresponds to the radial potential @xmath212 scaled by a factor @xmath213 in the left panel and by a factor @xmath214 in the right panel , for graphical convenience.,title="fig:",width=132,height=132 ] and corresponding to the energy levels @xmath209 of the left panel of fig . [ fig8 ] , respectively . panels c , d . repulsive radial gap solitons ( thick lines ) trapped in circular troughs of the potential at @xmath210 and corresponding to the energy levels @xmath211 of the right panel of fig . [ fig8 ] , respectively . the thin lines corresponds to the radial potential @xmath212 scaled by a factor @xmath213 in the left panel and by a factor @xmath214 in the right panel , for graphical convenience.,title="fig:",width=132,height=128 ] and corresponding to the energy levels @xmath209 of the left panel of fig . [ fig8 ] , respectively . panels c , d . repulsive radial gap solitons ( thick lines ) trapped in circular troughs of the potential at @xmath210 and corresponding to the energy levels @xmath211 of the right panel of fig . [ fig8 ] , respectively . the thin lines corresponds to the radial potential @xmath212 scaled by a factor @xmath213 in the left panel and by a factor @xmath214 in the right panel , for graphical convenience.,title="fig:",width=132,height=132 ] and corresponding to the energy levels @xmath209 of the left panel of fig . [ fig8 ] , respectively . panels c , d . repulsive radial gap solitons ( thick lines ) trapped in circular troughs of the potential at @xmath210 and corresponding to the energy levels @xmath211 of the right panel of fig . [ fig8 ] , respectively . the thin lines corresponds to the radial potential @xmath212 scaled by a factor @xmath213 in the left panel and by a factor @xmath214 in the right panel , for graphical convenience.,title="fig:",width=132,height=132 ] . bottom panels . 3d view of the gap solitons corresponding to the bound states c ( left panel ) and d ( right panel ) of fig . [ fig8].,title="fig:",width=151,height=113 ] . bottom panels . 3d view of the gap solitons corresponding to the bound states c ( left panel ) and d ( right panel ) of fig . [ fig8].,title="fig:",width=151,height=113 ] . bottom panels . 3d view of the gap solitons corresponding to the bound states c ( left panel ) and d ( right panel ) of fig . [ fig8].,title="fig:",width=151,height=113 ] . bottom panels . 3d view of the gap solitons corresponding to the bound states c ( left panel ) and d ( right panel ) of fig . [ fig8].,title="fig:",width=151,height=113 ] as a result we show the existence of radially symmetric gap solitons which correspond to the ring shaped states discussed before . ) section of @xmath215 obtained from eq . ( [ gpe2d ] ) by taking as initial conditions the localized states depicted in panels a - d of fig . the norm of the attractive states in panels @xmath209 are @xmath216 , @xmath217 , respectively , and correspond to bound states a , b , of fig . repulsive states in panels @xmath211 have norm @xmath218 and @xmath219 , respectively , and correspond to the bound states c , d , of fig . parameters are the same as in corresponding panels of fig . [ fig8 ] . , title="fig:",width=132,height=132 ] ) section of @xmath215 obtained from eq . ( [ gpe2d ] ) by taking as initial conditions the localized states depicted in panels a - d of fig . the norm of the attractive states in panels @xmath209 are @xmath216 , @xmath217 , respectively , and correspond to bound states a , b , of fig . repulsive states in panels @xmath211 have norm @xmath218 and @xmath219 , respectively , and correspond to the bound states c , d , of fig . parameters are the same as in corresponding panels of fig . [ fig8 ] . , title="fig:",width=132,height=132 ] ) section of @xmath215 obtained from eq . ( [ gpe2d ] ) by taking as initial conditions the localized states depicted in panels a - d of fig . the norm of the attractive states in panels @xmath209 are @xmath216 , @xmath217 , respectively , and correspond to bound states a , b , of fig . repulsive states in panels @xmath211 have norm @xmath218 and @xmath219 , respectively , and correspond to the bound states c , d , of fig . parameters are the same as in corresponding panels of fig . [ fig8 ] . , title="fig:",width=132,height=132 ] ) section of @xmath215 obtained from eq . ( [ gpe2d ] ) by taking as initial conditions the localized states depicted in panels a - d of fig . the norm of the attractive states in panels @xmath209 are @xmath216 , @xmath217 , respectively , and correspond to bound states a , b , of fig . repulsive states in panels @xmath211 have norm @xmath218 and @xmath219 , respectively , and correspond to the bound states c , d , of fig . parameters are the same as in corresponding panels of fig . , title="fig:",width=132,height=132 ] in fig . [ fig8 ] we depict the lower energy part of the spectrum obtained from eq . ( [ reducedradial ] ) in the case of attractive and repulsive interactions . we see in both cases the existence of a series of discrete levels located in the gaps between the bands which correspond to radially symmetric localized states . the wave functions of the energy levels labeled a - d in fig . [ fig8 ] are shown in fig . [ fig9 ] . the symmetries of these localized states is similar to the ones of intrinsic localized modes of nonlinear lattices and gap solitons of continuous 1d gpe with a periodic potential . adopting the same terminology introduced for these cases we shall refer to them as the onsite symmetric and the onsite asymmetric ring gap solitons . a 3d plot of these states is reported in fig . [ fig10 ] ( notice that the asymmetric state gives two rings of matter located in the same trough of the potential in the attractive case and many ring oscillations in the repulsive case ) . the stability of these gap solitons has been checked by time propagation under the 2d gpe in eq . ( [ gpe2d ] ) . the results are reported in fig . [ fig11 ] for the @xmath220 section of of @xmath221 as a function of time . we see that for the attractive case the gap solitons are always unstable , the onsite symmetric state becomes modulationally unstable at @xmath222 while the asymmetric state decays into a state localized in the center and a series of ring solitons which expand trough the lattice keeping the same symmetry as the initial state . for the repulsive case we see that while the asymmetric state in the second band gap ( panel d ) is unstable , the onsite symmetric ring gap soliton in the first band gap is quite stable under time evolution . the stability of the onsite symmetric repulsive ring gap solitons and the modulational instability of the repulsive ones has been found also for other parameter values and seems to be a general property of the model . the existence of other types of ring gap solitons ( which resemble intersite symmetric and asymmetric intrinsic localized modes of nonlinear lattices ) is also possible but they look even more unstable under time evolution than the onsite asymmetric ones for both attractive and repulsive cases . in the following section we shall investigate onsite symmetric gap soliton states in more details . generic examples of a ring - shaped soliton in the attractive model , and its counterpart in the repulsive one , found from direct solution of gpe ( [ gpe2d ] ) in , respectively , imaginary and real time , are displayed in fig . [ fig12 ] ( in the latter case , real - time simulations were run with absorbers set at @xmath223 and @xmath28 ) . the latter example represents _ radial gap solitons _ with the core trapped in a remote circular trough ( cf . the right panel in fig . [ fig4 ] that shows a radial gap soliton with the core trapped in the center ) . the fact that the soliton belongs to the gap type is clearly attested to by conspicuous inner and outer radial fringes attached to the core ( the fringes are absent in the ordinary soliton in the left panel of the figure ) . , trapped in a circular trough at @xmath224 . the amplitude of this soliton is @xmath225 , and its chemical potential and norm are @xmath226 and @xmath227 . note that analytical approximation ( [ nfinal ] ) yields , in this case , @xmath228 , which is very close to the numerically found value , and its modulational instability is definitely predicted by eq . ( [ thr ] ) . right panel : a stable ring - shaped soliton ( actually , it is a _ radial gap soliton _ ) in the repulsive model with @xmath55 , trapped at @xmath229 . in the latter case , @xmath230 and @xmath231.,title="fig:",width=151,height=113 ] , trapped in a circular trough at @xmath224 . the amplitude of this soliton is @xmath225 , and its chemical potential and norm are @xmath226 and @xmath227 . note that analytical approximation ( [ nfinal ] ) yields , in this case , @xmath228 , which is very close to the numerically found value , and its modulational instability is definitely predicted by eq . ( [ thr ] ) . right panel : a stable ring - shaped soliton ( actually , it is a _ radial gap soliton _ ) in the repulsive model with @xmath55 , trapped at @xmath229 . in the latter case , @xmath230 and @xmath231.,title="fig:",width=151,height=113 ] the evolution of the unstable axisymmetric ring from fig . [ fig12 ] ( left panel ) , and outcome of the evolution are displayed in fig . [ fig13 ] . it is observed that the system tends to form a stationary necklace - like pattern composed of six peaks towering above the remaining quasi - uniform background . ( left panel ) , triggered by a small azimuthal perturbation , @xmath232 . the evolution is shown in terms of the amplitude , @xmath233 , taken along circumference @xmath234 , with @xmath224 , @xmath235 running from @xmath236 to @xmath237 ( due to the symmetry , only half of the circle is shown ) . ( b ) the field profile formed by @xmath238 , demonstrating the emergence of a regular necklace - shaped pattern.,title="fig:",width=302,height=226 ] * ( a ) * ( left panel ) , triggered by a small azimuthal perturbation , @xmath232 . the evolution is shown in terms of the amplitude , @xmath233 , taken along circumference @xmath234 , with @xmath224 , @xmath235 running from @xmath236 to @xmath237 ( due to the symmetry , only half of the circle is shown ) . ( b ) the field profile formed by @xmath238 , demonstrating the emergence of a regular necklace - shaped pattern.,title="fig:",width=302,height=226]*(b ) * in the repulsive model , a stable ring ( i.e. , the radial gap soliton ) may additionally carry , on its crests , pairs of dark solitons . an example of such a pattern is displayed in fig . [ fig14 ] ( it resembles the dipole - mode ring soliton recently found in the model with repulsion and bessel lattice in ref . @xcite ) . , trapped at @xmath229 . this state has norm @xmath239 and chemical potential @xmath240.,width=302,height=226 ] stable static rings with a density modulation in the azimuthal direction , corresponding to analytical solutions ( [ cn ] ) for the cnoidal waves , have been found too in the attractive model . a typical example of such a pattern is displayed in fig . [ fig15 ] ; in fact , it was generated as an outcome of the evolution of an unstable axisymmetric ring with the same norm . ) . the solution s norm and chemical potential are @xmath241 and @xmath242.,width=302,height=226 ] increase of the norm leads to a transition from weakly modulated stable patterns to deeply modulated ones . an example of such a stable configuration , generated from four gaussian pulses , with the phase shifts @xmath0 between them , by integration of eq . ( [ gpe2d ] ) in imaginary time , is displayed in fig . [ fig16 ] ( note that the total norm of this pattern exceeds that corresponding to fig . [ fig15 ] by a factor of @xmath243 ) . and @xmath244 . the norm of each soliton is @xmath245 , with the phase difference @xmath237 between adjacent ones . the lattice is embedded in an integration domain of size @xmath246 . , width=302,height=226 ] strongly localized solitary waves of the attractive bec , self - trapped in a large - radius annular potential channel , like those shown in fig . [ fig16 ] , can freely move along in the channel . if set in motion with a sufficiently small velocity , the soliton can circulate in the channel indefinitely long , preserving its integrity , as shown in fig . [ fig17 ] . , trapped in the circular trough at @xmath229 , was set in motion in the clockwise direction with velocity @xmath247 . as illustrated by the sequence of snapshots , the circular motion in the trough continues indefinitely long without any visible loss.,width=302,height=302 ] however , when the circulation speed is too large , the centrifugal force acting on the soliton can cause the matter to tunnel into adjacent radial troughs . the loss of the norm ( number of atoms ) caused by the centrifugal tunneling eventually brings the soliton s norm below the threshold value necessary for its existence , which leads to disintegration of the soliton , see fig . [ fig18 ] . ( with initial norm @xmath248 ) is set in motion with a greater velocity , @xmath249 , the centrifugal force gives rise to the underbarrier leakage of atoms into the outer circular troughs . as illustrated by the sequence of snapshots , the decaying soliton can not preserve its shape . the norm remaining within the given radial channel at @xmath250 is @xmath251 , i.e. , @xmath252 of the initial value.,width=302,height=302 ] if the lattice is weak , and the soliton is strongly self - trapped ( with a sufficiently large norm ) , the centrifugal force can also drive it away from the center , across the lattice . in that case ( not shown here ) , the soliton also suffers the radiation loss , and eventually disintegrates . the present model offers unique possibilities for exploring interactions and collisions between matter - wave solitons , as the dynamics in circular channels is free of external perturbations , which are always present in previously reported settings ( in the form of the weak longitudinal trap intended to keep the solitons within a finite spatial domain ) . the same advantage is offered by setups with the toroidal trap @xcite . however , sideline ( tangential ) collisions , which are possible between solitons moving in adjacent potential troughs of the radial lattice , have no counterparts in the toroidal traps . in head - on collisions of solitons in the attractive model ( in the same circular channel ) , two competing processes play major roles . the first is the nonlinear self - focusing , which may give rise to intrinsic collapse of a large - norm lump " temporarily formed by the colliding solitons , and the other is the interference due to coherence of the matter - wave solitons @xcite . the interference pattern , featuring alternating regions of high and low density , can suppress the collapse . the predominance of either mechanism depends on the time of interaction . if it is small ( in the case of large collision speed ) , solitons can pass trough each other without collapse ( although with conspicuous loss , see below , hence the collision can not be termed quasi - elastic ) , even if their total norm exceeds the collapse threshold . on the contrary , slow collisions lead to the onset of collapse . these two possible scenarios of the head - on collisions between in - phase ( mutually attracting ) solitons are shown in figs . [ fig19 ] and [ fig20 ] . figure [ fig19 ] presents an example of a relatively fast collision . at the moment of the full overlap between the solitons , @xmath253 , the interference fringes are evident , and are also seen in the cross section along the circumference of the channel ( lower right panel ) . although the total norm of the colliding solitons is overcritical ( @xmath254 ) , and the collision time ( @xmath255 ) is larger than the collapse time ( which is estimated as @xmath256 ) , the solitons separate without collapse . nevertheless , the collision is essentially inelastic , the loss being additionally enhanced by the matter leakage under the action of the centrifugal force . on the other hand , fig . [ fig20 ] shows that the solitons colliding at a sufficiently small velocity merge together and indeed blow up due to the collapse . and zero phase difference , in the circular trough at @xmath229 . one of the solitons is set in motion ( at @xmath257 ) at @xmath258 with speed @xmath249 in the clockwise direction , while the second one ( at @xmath259 ) stays quiescent . the norm remaining within the given circular channel at @xmath260 is @xmath261 . the soliton which was originally at rest loses less norm than the moving one . , title="fig:",width=151,height=151 ] and zero phase difference , in the circular trough at @xmath229 . one of the solitons is set in motion ( at @xmath257 ) at @xmath258 with speed @xmath249 in the clockwise direction , while the second one ( at @xmath259 ) stays quiescent . the norm remaining within the given circular channel at @xmath260 is @xmath261 . the soliton which was originally at rest loses less norm than the moving one . , title="fig:",width=151,height=151 ] and zero phase difference , in the circular trough at @xmath229 . one of the solitons is set in motion ( at @xmath257 ) at @xmath258 with speed @xmath249 in the clockwise direction , while the second one ( at @xmath259 ) stays quiescent . the norm remaining within the given circular channel at @xmath260 is @xmath261 . the soliton which was originally at rest loses less norm than the moving one . , title="fig:",width=151,height=151 ] and zero phase difference , in the circular trough at @xmath229 . one of the solitons is set in motion ( at @xmath257 ) at @xmath258 with speed @xmath249 in the clockwise direction , while the second one ( at @xmath259 ) stays quiescent . the norm remaining within the given circular channel at @xmath260 is @xmath261 . the soliton which was originally at rest loses less norm than the moving one . , title="fig:",width=151,height=151 ] , in one circular channel ( @xmath229 ) , ends up with collapse . upper panels : density plots at @xmath258 and the moment of collapse ( @xmath262 ) . lower panel : the evolution of the density profiles in the cross section along the circumference of the potential trough . as in fig . [ fig19 ] , the initial norm of each soliton is @xmath263 . , title="fig:",width=151,height=151 ] , in one circular channel ( @xmath229 ) , ends up with collapse . upper panels : density plots at @xmath258 and the moment of collapse ( @xmath262 ) . lower panel : the evolution of the density profiles in the cross section along the circumference of the potential trough . as in fig . [ fig19 ] , the initial norm of each soliton is @xmath263 . , title="fig:",width=151,height=151 ] , in one circular channel ( @xmath229 ) , ends up with collapse . upper panels : density plots at @xmath258 and the moment of collapse ( @xmath262 ) . lower panel : the evolution of the density profiles in the cross section along the circumference of the potential trough . as in fig . [ fig19 ] , the initial norm of each soliton is @xmath263 . , width=302,height=226 ] two @xmath0 out - of - phase solitons colliding in the same potential trough bounce from each other due to repulsion between them . then , repeated bouncing collisions are observed , which is similar to the dynamics in the bessel lattice @xcite , as well as in a quasi-1d ol , equipped with a weak longitudinal trap @xcite . however , in contrast to those works , we have observed that many collisions in the circular trough gradually wash out the phase coherence between the solitons . as a result , the solitons cease to repel each other , and a clear interference pattern disappears . eventually , the originally repelling solitons merge together , shedding the excessive norm away , as shown in fig . [ fig21 ] . the intrinsic collapse _ does not _ occur in this case , as the loss makes the norm of the finally established single soliton smaller than the critical value at the collapse threshold . , but the phase difference of @xmath264 between them . solitons bounce at @xmath265 and @xmath266 . finally , the solitons merge into a single one , without collapse.,width=226,height=226 ] collisions of matter - wave solitons in the attractive bec is accompanied by strong exchange of matter between them . the present setting makes it possible to investigate related phenomena for solitons moving in adjacent channels of the radial potential , and thus colliding tangentially . the result is that two in - phase solitons are unstable against the flow of matter in this setting , and tend to merge ( a similar trend was observed in the model with a quasi-1d lattice in ref . nevertheless , the solitons may avoid collapse and separate , if the collision time is small due to a large relative velocity , as shown in fig . [ fig22 ] . , moving at velocities @xmath267 in adjacent troughs ( at @xmath229 and @xmath268 ) of the radial lattice with @xmath244 . the solitons survive the collision , and remain localized , although the collision is inelastic , giving rise to radiation loss . , title="fig:",width=113,height=113 ] , moving at velocities @xmath267 in adjacent troughs ( at @xmath229 and @xmath268 ) of the radial lattice with @xmath244 . the solitons survive the collision , and remain localized , although the collision is inelastic , giving rise to radiation loss . , title="fig:",width=113,height=113 ] , moving at velocities @xmath267 in adjacent troughs ( at @xmath229 and @xmath268 ) of the radial lattice with @xmath244 . the solitons survive the collision , and remain localized , although the collision is inelastic , giving rise to radiation loss . , title="fig:",width=113,height=113 ] on the other hand , if the solitons are loosely bound ( i.e. their norms are small and/or the radial lattice is weak ) , the slow lateral collision gives rise to merger of the solitons into one , excessive norm being shed off with linear waves , as demonstrated in fig . [ fig23 ] . , but for solitons with smaller norms , @xmath269 , launched to move at velocities @xmath270.,title="fig:",width=113,height=113 ] , but for solitons with smaller norms , @xmath269 , launched to move at velocities @xmath270.,title="fig:",width=113,height=113 ] , but for solitons with smaller norms , @xmath269 , launched to move at velocities @xmath270.,title="fig:",width=113,height=113 ] in this work , we have proposed a new setting to explore two - dimensional ( 2d ) localized states in self - attractive and repulsive becs , in the form of a periodic radial optical lattice . a crucial difference from previously studied 2d models with bessel lattices is the fact that no localized state may exist in the linear limit of the present model , hence all solitary states in it are true solitons " , impossible without the nonlinearity . besides the bec , the model may also apply to spatial solitons ( beams ) in photonic crystal fibers with a circular intrinsic structure . the existence of such objects was demonstrated by means of different variants of the variational approximation , and in direct numerical simulations . we have investigated the localized states trapped in the central potential well and in remote circular potential troughs . in both cases , a new species , namely , stable _ radial gap solitons _ , have been identified in the model with self - repulsion ( in addition , the radial gap soliton trapped in a circular trough can carry stable pairs of dark solitons on its crest ) . in remote troughs , we have investigated ring - shaped patterns delocalized in the azimuthal direction , as well as strongly localized azimuthal solitons ( including moving ones ) . in the attractive model , solutions of both types are described by an effective 1d equation supplemented by the periodic boundary conditions ( it is a nonpolynomial nls equation with the coordinate running along the circumference ) . in particular , a finite threshold of the azimuthal modulational instability ( mi ) of axisymmetric ring - shaped states was predicted , and exact cnoidal - wave solutions , generated by the mi - induced bifurcation were found in the framework of the 1d equation . azimuthal solitons were also found as solutions to the same equation . the existence of stable azimuthally uniform and weakly modulated ring - shaped states was corroborated by direct simulations . dynamics of completely localized solitons circulating in the annular potential troughs was investigated in the attractive model by means of direct simulations . in particular , the solitons with sufficiently small velocities remain stable indefinitely long , while high velocities give rise to leakage of matter into the adjacent ( more remote ) trough under the action of the centrifugal force , which eventually destroys the soliton . collisions between solitary waves running in the same or adjacent circular channels were investigated too . head - on collisions of the in - phase solitons in one trough lead to the collapse ; @xmath0-out of phase solitons bounce from each other many times , but gradually loose their mutual coherence , and eventually merge into a single soliton without collapsing , shedding off excess norm . in - phase solitons colliding in adjacent channels may also merge into a single soliton . b.b.b . is partially supported by the grant no . 20 - 06 from the fund for fundamental research of the uzbek academy of sciences . the work of b.a.m . was supported , in a part , by the center - of - excellence grant no . 8006/03 from the israel science foundation . ms acknowledges financial support from the miur , through the inter - university project:_dynamical properties of bose - einstein condensates in optical lattices _ , prin-2005 . 99 j. denschlag , j. e. simsarian , d. l. feder , c. w. clark , l. a. collins , j. cubizolles , l. deng , e. w. hagley , k. helmerson , w. p. reinhardt , s. l. rolston , b. i. schneider , and w. d. phillips , science * 287 * , 97 ( 2000 ) ; s. burger , k. bongs , s. dettmer , w. ertmer , k. sengstock , a. sanpera , g. v. shlyapnikov , and m. lewenstein , phys . lett . * 83 * , 5198 ( 1999 ) ; b. b. baizakov , m. salerno , and b. a. malomed , in : _ nonlinear waves : classical and quantum aspects _ , ed . by f. kh . abdullaev and v. v. konotop , p. 61 ( kluwer academic publishers : dordrecht , 2004 ) ; also available at http://rsphy2.anu.edu.au/asd124/baizakov_2004_61_nonlinear waves.pdf .

* the main objective of the paper * + the immediate objective of the paper is to present partial - realization theory of linear switched systems . however , the broader goal is to demonstrate that * it is possible to develop partial - realization theory for hybrid systems , * the theory of rational formal power series can be used to obtain partial - realization theory of hybrid systems . the results of the paper just serve as evidence for the above claims . in fact , partial - realization theory can be developed along the same line for other classes of hybrid systems , @xcite . this paper is intended as the first one in a series of papers dealing with partial - realization theory of hybrid systems . * the class of linear switched systems * + linear switched systems have been studied for almost two decades , see @xcite for a survey . their practical and theoretical relevance is widely recognized . _ linear switched systems ( abbreviated by lss ) _ are hybrid systems with external switching only , whose continuous dynamics in each discrete mode is determined by a linear continuous - time system , and whose discrete dynamics is trivial , i.e. any discrete state transition is allowed . the continuous subsystems are defined on the same state- , input- and output - spaces . a discrete state transition occurs if the environment enforces one , i.e. the switching sequence itself functions as an input . the time evolution of a linear switched system can be pictured as follows . a discrete mode is selected and a continuous input is fed into the corresponding linear subsystem . the state and output of the linear switched system is now determined by the time evolution of the chosen linear subsystem . at some point in time , a new discrete mode is selected . then the linear subsystem corresponding to the new discrete mode is launched , using as initial state the state of the previous linear subsystem at the moment when a transition to the new mode occurred . from this point on , the state and output evolve according to the newly selected linear subsystem . * motivation * + we believe that similarly to linear systems , partial - realization theory for lsss will be useful for systems identification and model reduction of lsss and even of more general hybrid systems . we will elaborate on the possible applications in section [ sect : concl ] . * contribution of the paper * + the paper presents the following results . * * partial - realization algorithm and theorem for lsss * + in section [ sect : main_result ] we propose a notion of partial - realization . in theorem [ part_real_lin : theo1 ] and [ part_real_lin : theo1.1 ] we formulate a sufficient condition for existence and minimality of a partial realization . we present two algorithms for computing a partial - realization . in one of them , algorithm [ alg0 ] , a partial - realization is constructed from the columns of a _ finite _ sub - matrix hankel - matrix , the other one , algorithm [ partreal ] , is based on finding a factorization of a _ finite _ sub - matrix of the hankel - matrix . the outcomes of both procedures are isomorphic . the former is potentially useful for theoretical purposes , while the latter one might serve as a basis for subspace identification - like methods . the factorization algorithm was implemented and the implementation is available from the the first - named author on request . * * realization algorithm for lsss * + in theorem [ part_real_lin : theo1 ] and [ part_real_lin : theo2 ] we show that the above partial - realization algorithms return a complete minimal realization of the input - output behavior , provided the rank of the finite sub - matrix of the hankel - matrix equals the rank of the full hankel - matrix . we show that any finite sub - matrix whose size is above a certain lower bound has this property . * * partial - realization theory for families of formal power series * + both the ( partial- ) realization algorithm and the corresponding theoretical results for lsss follow quite easily from analogous statements for families of rational formal series . in appendix [ sect : pow : part ] we present what can be termed as partial - realization theory and algorithms for _ families of formal power series_. although partial - realization theory of a single rational formal power series has been known in various forms @xcite , its extension to families of formal power series appears to be new . as it was already noted , this extension of the classical theory can be then used to develop partial - realization theory for a number of other ( more general ) classes of hybrid systems , @xcite . * formal power series approach * + our main tool is the theory of rational formal power series . recall from @xcite that there is a correspondence between lss realizations and representations of certain rational families of former power series . recall that the hankel - matrix of a family of input - output maps is in fact the hankel - matrix of the corresponding family of formal power series . hence , by formulating a partial - realization theory and algorithms for rational families of formal power series , we immediately obtain partial - realization theory and algorithms for lsss . * related work * + to the best of our knowledge , the only results on partial - realization of linear switched systems is the first - named author s thesis @xcite . in @xcite partial - realization theory of bilinear hybrid and switched systems are announced , but no proofs are provided . the thesis @xcite contains most of the results of this paper . in addition , it also covers partial - realization theory of several other classes of hybrid systems . reference @xcite announces some results on partial - realization theory of stochastic jump - markov systems . realization theory of rational formal power series is a classical topic , see @xcite . it is known to be closely related to realization theory of bilinear and state - affine systems @xcite . in turn , results on partial - realization theory for discrete - time bilinear systems and state - affine systems can be found in @xcite . with respect to @xcite , the main novelty of the paper is the following . 1 . in this paper ( in appendix [ sect : pow : part ] ) partial - realization theory is stated directly for rational formal power series representations , without reformulating it in terms of some particular system class such as bilinear or state - affine systems . 2 . while the classical results @xcite can be thought of as a reformulation of partial - realization theory for a single rational formal power series , here we deal with _ families of formal power series_. 3 . we apply partial - realization theory of rational formal power series to linear switched systems . this represents a novel result in the theory of hybrid systems . we view this as the main contribution of the paper . the statement of the main results on partial - realization theory for families of formal power series was announced @xcite , but no proof was ever presented . the thesis @xcite contains the statement and the proof of the main results on formal power series . * outline of the paper * + the outline of the paper is the following . section [ sect : info ] presents an informal formulation of the partial - realization problem for lsss . section [ sect : num ] discusses a numerical example . it is intended as an accessible explanation of the main results of the paper by means of an example . section [ sect : switch ] presents the definition and some elementary properties of linear switched systems together with the basic notation and terminology . section [ sect : lin_sys : real_theo ] provides a brief overview of the realization theory of linear switched systems . this section is a prerequisite for understanding the main results of the paper . section [ sect : main_result ] presents the formal statement of the main results of the paper . section [ sect : lin_alg ] presents the proof of the results stated in section [ sect : main_result ] . finally , in section [ sect : concl ] we formulate the conclusions of the paper and we discuss potential applications of the presented results . as it was already remarked , partial - realization theory of lsss is based on partial - realization theory of rational formal power series . therefore , in appendix [ sect : pow ] we included a brief overview of the extension of the theory of formal power series to families of formal power series . more precisely , appendix [ sect : pow ] reviews the relevant concepts and results on formal power series from @xcite . in appendix [ sect : pow : part ] we present partial - realization theory of families of formal power series . and algorithms for computing a representation for a family of formal power series . appendix [ sect : pow : part ] is in fact the main technical tool of the paper . it is a prerequisite for understanding the proof of the main results on partial - realization theory of lsss . the goal of this section is to state the partial - realization problem for linear switched systems in an informal way . the next section , section [ sect : num ] , provides a numerical example to illustrate the problem and the solution . we defer the formal problem statement until section [ sect : main_result ] . * brief review of linear - partial realization theory * + the partial - realization problem was originally formulated for linear systems . @xcite as follows . assume that the first @xmath0 markov parameters @xmath1 of an input - output map @xmath2 are specified . * find conditions for existence of a linear system whose first @xmath3 markov - parameters coincide with @xmath4 . such a linear system is called a _ partial - realization of @xmath5_. characterize minimal partial - realizations . find an algorithm for computing a ( minimal ) partial - realization . * find conditions under which the thus obtained linear system is a minimal realization of the input - output map @xmath2 . notice that for discrete - time linear systems , the @xmath6th markov parameter coincides with the output of the system at time @xmath6 for a particular input . hence , the partial - realization problem can be seen as an identification problem . in fact , partial - realization theory of linear systems can be used for systems identification and model reduction . * markov - parameters for lsss * + let @xmath7 be a family of input - output map which map continuous - valued inputs and switching sequences to continuous outputs . that is , the elements of @xmath7 are of the same form as the input - output maps generated by lsss . below we will define _ generalized markov - parameters _ of a family @xmath7 of input - output maps which could potentially be realized by a lss . the generalized markov - parameters of @xmath7 are defined as certain high - order derivatives with respect to the switching times of the elements of @xmath7 . more precisely , the markov - parameters are indexed by triples consisting of the following components * elements of @xmath7 , or pairs @xmath8 where @xmath9 runs through the set of discrete modes and @xmath10 runs through the set @xmath11 , where @xmath12 denotes the number of continuous - valued input channels , and * sequences of discrete modes , and * discrete modes . without going into details , the intuition behind the presented indexing is the following . a markov - parameter of @xmath7 indexed by @xmath13 , a discrete mode @xmath9 , and a sequence of discrete modes @xmath14 , stands for the partial derivative of @xmath2 with respect to the switching times , evaluated at zero , for the following switching scenario . the system goes through the modes @xmath15 and then jumps to @xmath9 . here the continuous input is set to zero . the intuitive meaning of a markov - parameter indexed by a pair @xmath16 , a discrete mode @xmath9 and a sequence of discrete modes @xmath14 is the following . each input - output map @xmath13 can be written as a sum @xmath17 of two maps , where @xmath18 is independent of the continuous input and @xmath19 is common for all the elements of @xmath7 and it is linear in continuous inputs . roughly speaking , @xmath18 accounts for the output from a certain initial state under zero input , and @xmath19 represents the input - output map induced by the zero initial state . the markov - parameter indexed by the pair @xmath16 , discrete mode @xmath9 and sequence @xmath20 is the derivative of @xmath19 with respect to the switching times evaluated at zero , for the switching scenario where the system goes from mode @xmath21 to @xmath22 and ends in @xmath9 , and all the continuous input channels are @xmath23 except the @xmath10th one which is @xmath24 . * partial - realization problem for lsss * + we will refer to a markov parameter indexed by a sequence of discrete modes of length @xmath6 as a markov - parameter of _ order @xmath6_. fix a natural number @xmath25 and let @xmath5 be the collection of all markov - parameters of @xmath7 of order at most @xmath3 . that is , @xmath5 is simply a collection of high - order derivatives of the elements of @xmath7 , such that the degree of derivation is bounded by @xmath3 . a lss @xmath26 is said to be a _ @xmath3-partial realization _ of @xmath7 , if certain products of the matrices of @xmath26 are equal to the corresponding elements of @xmath5 . in other words , @xmath26 is a @xmath3-partial realization of @xmath7 if the input - output maps of @xmath26 and those of @xmath7 have the property that their derivatives corresponding to the markov - parameters of order at most @xmath3 coincide . from realization theory of lsss @xcite it follows that a lss is a realization of @xmath7 if and only if it is a @xmath3-partial realization of @xmath7 for all @xmath0 . the _ partial realization problem _ for lsss can be now stated as follows . * find conditions for existence of a @xmath3-partial realization of @xmath7 by lss . characterize minimal dimensional @xmath3-partial realizations of @xmath7 . find an algorithm for computing a minimal @xmath3-partial lss realizations of @xmath7 . * find conditions under which a minimal @xmath3-partial lss realization of @xmath7 is a complete realization of @xmath7 . the motivation for studying the partial - realization problem for lsss is similar to that of for linear systems , i.e. we expect it to be useful for model reduction and systems identification . in section [ sect : concl ] we will present a more detailed description of the motivation and possible applications . the purpose of this section is to demonstrate the main results of the paper by means of a numerical example . in this section we will tacitly use the notation and terminology of section [ sect : switch ] and section [ sect : main_result ] . consider the linear switched system of the form @xmath27 where @xmath28 is the discrete mode at time @xmath29 , @xmath30 is the continuous - state at @xmath29 , @xmath31 is the scalar output at @xmath29 , and @xmath32 is the scalar input at @xmath29 . the system matrices @xmath33 describing the linear control system residing in a state ( mode ) @xmath34 are of the following form . @xmath35 consider the initial states @xmath36 and @xmath37 . consider the set of input - output maps @xmath38 such that @xmath39 is realized by @xmath26 from the initial state @xmath40 , @xmath41 , i.e. @xmath42 , for @xmath41 , for each continuous - valued input @xmath43 and finite switching sequence @xmath44 . notice that @xmath7 can be realized by the following minimal linear switched system @xmath45 where for each @xmath46 , the matrices @xmath47 are of the following form @xmath48 more precisely , @xmath49 realizes @xmath39 from the initial states @xmath50 , for @xmath41 . here , @xmath51 and @xmath52 . since @xmath7 has a realization by a lss , it is clear that it has a generalized kernel representation and hence its markov - parameters can be defined . in order to give a better intuition on markov - parameters of @xmath7 , we have listed some of them in table [ table1 ] . .[table1 ] markov - parameters of @xmath7 [ cols="<,^,^,>",options="header " , ] consider the upper - left sub - matrix @xmath53 of the hankel - matrix @xmath54 of @xmath7 . recall that @xmath53 is formed by the intersection of the columns of @xmath54 indexed by a sequence of discrete modes of length at most @xmath55 , and by the rows of @xmath54 indexed by a sequence of discrete modes of length at most @xmath56 . from theorem [ part_real_lin : theo1 ] it follows that @xmath57 for all @xmath58 . in fact , it turns out that the hankel - matrix @xmath59 for @xmath60 has already rank @xmath61 . by theorem [ part_real_lin : theo1 ] it means that we can already compute a minimal lss realization of @xmath7 from the generalized markov parameters of @xmath7 of indexed by sequences of discrete modes of length at most @xmath62 . applying algorithm [ partreal ] to @xmath63 yields the following minimal lss realization @xmath64 of @xmath7 . @xmath65 where the system matrices are of the form @xmath66 the lss @xmath64 realizes the input - output map @xmath67 from the initial state @xmath68 , and the map @xmath69 from the initial state @xmath70 . however , it turns out that @xmath71 . that is , the sequence of markov parameters of @xmath7 indexed by sequences of discrete modes of length at most @xmath24 already satisfies the partial realization theorems theorem [ part_real_lin : theo1 ] and theorem [ part_real_lin : theo2 ] . hence , we can apply algorithm [ partreal ] to @xmath72 to obtain a minimal lss @xmath24-partial realization of @xmath7 . however , it can be checked that the thus obtained lss _ is not a lss realization of @xmath7_. indeed , by applying algorithm [ partreal ] to @xmath72 we obtain the following lss realization of @xmath7 @xmath73 where the system matrices @xmath74 , @xmath34 are as follows @xmath75 the lss @xmath76 is a _ partial realization _ of the input - output map @xmath67 from the initial state @xmath77 , and of the input - output map @xmath69 from the initial state @xmath78 . that is , the markov parameters of @xmath38 which are indexed by sequences from the set @xmath79 coincide with those of the input - output maps induced by the initial state @xmath80 , @xmath41 . it is easy to see that @xmath76 is not a realization of the input - output maps @xmath39 from the respective initial states @xmath81 , @xmath82 . that is , the input - output maps induced by the respective states @xmath81 , @xmath41 , do not coincide with the maps @xmath39 . one can either check it by direct calculation , or by using uniqueness of a minimal realization . using the latter approach , it is enough to notice that @xmath49 is a minimal realization of @xmath7 and it is of dimension @xmath61 . since @xmath76 of dimension @xmath83 , hence smaller than the dimension of @xmath49 , and all minimal realizations have to be of the same dimension , it follows that @xmath76 can not be a realization of @xmath7 . in fact , by checking the markov - parameters , one can see that @xmath76 recreates only the markov - parameters indexed by sequences of length at most @xmath24 , but there is a markov - parameter of @xmath7 , indexed by a sequence of length @xmath83 , which is not generated by @xmath76 . that is , it indeed happens that a family of input - output maps generated by a lss satisfies the sufficient conditions for existence of a @xmath3-partial realization for some @xmath3 , but the obtained partial realization is not a complete realization of the family of input - output maps . [ sect : switch ] this section contains the definition of linear switched systems . we will start with fixing notation and terminology which will be used throughout the paper . the notation used in this paper is mostly the standard one used in the field of control theory and formal language theory . in order to make the task of the reader easier , below we will list the most important notational conventions , grouped according to the disciplines . denote by @xmath84 the set @xmath85 of all non - negative reals . the set @xmath84 will be the time - axis of the systems discussed in this paper . for any @xmath86 , denote by @xmath87 the class of piecewise - continuous maps from @xmath84 to @xmath88 . that is , @xmath89 , if @xmath2 has finitely many points of discontinuity on each finite interval @xmath90 $ ] , @xmath91 , and at each point of discontinuity the right- and left - hand side limits exist and they are finite . denote by @xmath92 the set of natural number including @xmath23 . by abuse of notation we will denote any constant function @xmath93 by its value . that is , if @xmath94 for all @xmath91 , then @xmath2 will be denoted by @xmath95 . for any function @xmath96 the range of @xmath96 will be denoted by @xmath97 , i.e. if @xmath98 for some sets @xmath99 and @xmath100 , then @xmath101 . if @xmath102 is a vector space and @xmath103 is a subset of @xmath102 , then @xmath104 denotes the linear span of elements of @xmath103 in @xmath102 . if @xmath105 are vector spaces over @xmath106 , and @xmath107 , @xmath108 are linear maps , then @xmath109 denotes the composition @xmath110 and @xmath111 . let @xmath112 be a smooth map . consider a @xmath6 tuple of natural numbers @xmath113 , where @xmath114 . we will denote @xmath115 the partial derivative of @xmath116 evaluated at zero , such that the order of derivation with respect to the variable @xmath117 is @xmath118 for @xmath119 . that is , @xmath120 for each @xmath121 , @xmath122 denotes the @xmath10th unit vector of @xmath123 , i.e. @xmath124 , where @xmath125 is the kronecker symbol . in this paper we will use the notation of @xcite for matrices indexed by sets other than natural numbers . let @xmath126 and @xmath127 be two arbitrary sets . a ( real ) matrix @xmath128 whose columns are indexed by the elements of @xmath127 and whose rows are indexed by the elements of @xmath126 is simply a map @xmath129 . the set of all such matrices is denoted by @xmath130 . the entry of @xmath128 indexed by the row index @xmath131 and column index @xmath132 is denoted by @xmath133 and it is defined as the value of @xmath128 at @xmath134 , i.e. @xmath135 . the case of usual finite matrices can be recovered by viewing @xmath136 real matrices as matrices from @xmath137 . in the sequel , when referring to the index set of a matrix , we will identify any natural number @xmath138 with the set @xmath139 . in other words , @xmath140 denotes the set of matrices @xmath141 and @xmath142 denotes the set of matrices @xmath143 . for a matrix @xmath144 , the columns of @xmath128 are simply maps of the form @xmath145 and the rows of @xmath128 are maps of the form @xmath146 . the set of maps of the form @xmath145 and @xmath146 will sometimes be denoted by @xmath147 and @xmath148 respectively . furthermore , if @xmath149 ( resp . @xmath150 ) then the value of @xmath96 at @xmath131 ( resp . @xmath132 ) will be denoted by @xmath151 ( resp . @xmath152 ) . the column of @xmath128 indexed by @xmath132 will be denoted by @xmath153 and is defined as @xmath154 , @xmath131 . similarly , the row if @xmath128 indexed by @xmath131 will be denoted by @xmath155 and is defined as @xmath156 for all @xmath132 . if @xmath144 and @xmath157 and @xmath127 is finite , then the product of @xmath128 and @xmath5 is the matrix @xmath158 such that @xmath159 for all @xmath160 . in particular , if @xmath161 and @xmath162 for some natural number @xmath163 , then their product @xmath164 is well - defined and it belongs to @xmath165 . we will identify a map @xmath166 with the matrix @xmath167 defined as @xmath168 for all @xmath132 . hence , for a matrix @xmath144 , the product @xmath169 is defined as the following matrix in @xmath170 ; @xmath171 . in addition , we will occasionally identify the rows of a matrix @xmath144 with matrices @xmath172 . that is , the row @xmath155 of @xmath128 indexed by @xmath173 will be viewed as the matrix @xmath174 . with this identification , the product of the row @xmath155 with @xmath2 is a scalar @xmath175 . notice that here we tacitly assumed that @xmath127 is finite . notice that the set of all maps @xmath147 forms a vector space with respect to point - wise addition and multiplication by scalar . that is , if @xmath176 and @xmath177 , then the linear combination @xmath178 is defined by @xmath179 for all @xmath131 . consider a matrix @xmath144 and recall that its columns are simply elements of @xmath147 . hence , it makes sense to speak of the linear subspace spanned by the columns of a matrix @xmath144 . in the sequel , the _ rank of @xmath128 _ , denoted by @xmath180 will mean the dimension of the linear space spanned by the columns of @xmath128 . if this dimension is not finite , then the rank is taken to be @xmath181 . we will denote by @xmath182 the linear space spanned by the columns of @xmath128 . if @xmath144 and @xmath127 is finite , then @xmath128 can be viewed as a linear map from @xmath148 to @xmath147 , defined by @xmath183 , @xmath131 , for each @xmath166 . if @xmath157 and @xmath56 is finite , then the product @xmath158 corresponds to the linear map @xmath184 obtained by composing the linear map corresponding to @xmath5 with the linear map corresponding to @xmath128 . the notation described below is standard in formal languages and automata theory , see @xcite . consider a finite set @xmath185 which will be called the _ alphabet_. denote by @xmath186 the set of finite sequences of elements of @xmath185 . finite sequences of elements of @xmath185 will be referred to as _ strings _ or _ words _ over the alphabet @xmath185 . for a word @xmath187 , @xmath188 , @xmath189 the length of @xmath190 is denoted by @xmath191 , i.e. @xmath192 . we will denote by @xmath193 the _ empty sequence ( word)_. the length of the empty sequence @xmath193 is zero : @xmath194 . we will denote by @xmath195 the set of of non - empty words over @xmath185 . that is , @xmath196 . consider two words @xmath197 and @xmath198 of the form @xmath199 , and @xmath200 , @xmath201 . define the concatenation @xmath202 of the words @xmath203 and @xmath190 as the the word @xmath204 . in particular , if @xmath205 , i.e. if @xmath206 , then @xmath207 . similarly , if @xmath208 , i.e. @xmath209 , then @xmath210 . if @xmath211 is a word , then @xmath212 denotes the word @xmath213 . the word @xmath214 is just the empty word @xmath193 . below we present the formal definition of lsss . for a more detailed exposition , see @xcite . [ lin_switch : def ] a _ linear switched system _ ( abbreviated by lss ) is a control system @xmath26 of the form @xmath215 * @xmath216 is the continuous state at time @xmath91 , * @xmath217 denotes the continuous input at time @xmath91 , * @xmath218 denotes the discrete mode ( state ) at time @xmath29 , * @xmath219 denotes the continuous output at time @xmath91 . * the state - space is @xmath220 , the input - space is @xmath221 , the output - space is @xmath222 , and @xmath223 is the finite set of discrete modes ( discrete states ) . here @xmath224 are positive integers . * for each discrete mode @xmath225 , the corresponding matrices are of the form @xmath226 , @xmath227 and @xmath228 . we will use @xmath229 as a short - hand notation for lsss of the form ( [ lin_switch0 ] ) . [ switch_sys : not1 ] in the sequel we denote by @xmath230 the space @xmath88 of continuous - valued input , by @xmath231 the space @xmath232 of continuous - valued outputs , and by @xmath223 the set of discrete modes . informally , the system ( [ lin_switch0 ] ) evolves as follows . for a piecewise - constant _ switching signal _ @xmath233 and piecewise - continuous input @xmath234 , the _ state trajectory _ @xmath235 is a continuous piecewise - differentiable function which satisfies the differential equation ( [ lin_switch0 ] ) . the output at time @xmath91 is obtained by applying to @xmath236 the readout map ( matrix ) @xmath237 . below we define state- and output- trajectories more rigorously . to this end , we define the notion of switching sequences . a _ switching sequence _ is a sequence of the form @xmath238 , where @xmath239 are discrete modes and @xmath240 denote the _ switching times _ and @xmath241 . the set of all switching sequences is denoted by @xmath242 . if @xmath206 then we say that @xmath190 is the empty switching sequence and we denote it by @xmath193 . we denote the set of all _ non - empty switching sequences _ by @xmath243 . the interpretation of the sequence @xmath244 is the following . from time instance @xmath23 to time instance @xmath245 the active discrete mode is @xmath246 , i.e. the value of the switching signal is @xmath246 , from @xmath245 to @xmath247 the value of the switching signal is @xmath248 , from @xmath247 to @xmath249 the value of the switching signal is @xmath250 , and so on . that is , the non - negative real @xmath117 indicates the time spent in the discrete mode @xmath251 , for all @xmath252 . _ in this paper the switching sequences are regarded as inputs and we allow any switching sequence to occur . _ [ def : state_traject ] let @xmath43 be a continuous - valued input and let @xmath253 be a non - empty switching sequence . the _ state of @xmath26 reached from the initial state @xmath254 with the inputs @xmath255 and @xmath190 _ is denoted by @xmath256 and it is defined recursively on @xmath6 as follows . * if @xmath257 , then @xmath258 is the solution at time @xmath245 of the differential equation @xmath259 with the initial condition @xmath260 . * if @xmath261 is already defined , then let @xmath256 be the solution at time @xmath262 of the differential equation @xmath263 with the initial condition @xmath264 . that is , the states evolves according to the differential equation determined by the discrete mode . if a switch occurs , then the state at the time of the switch is taken as the initial condition for generating a solution to the differential equation associated with the new discrete mode . as a next step , we will define the output trajectories of lsss . consider a continuous - valued input @xmath43 and a non - empty switching sequence @xmath265 . the _ output generated by the lss @xmath26 if it is started from initial state @xmath254 and fed with the inputs @xmath255 and @xmath190 _ , denoted by @xmath266 , is defined by @xmath267 that is , the current output is obtained from the current continuous state @xmath256 by the application of the readout map @xmath268 associated with the current discrete mode @xmath269 . we define the input - output map of a lss induced by a particular initial state as follows . consider a state @xmath254 of @xmath26 . define the _ input - output map of @xmath26 induced by the state @xmath270 _ as the map @xmath271 such that for all continuous - valued input @xmath43 and for all non - empty switching sequence @xmath44 @xmath272 that is , the value of @xmath273 at @xmath274 equals the output generated by @xmath26 if started from the initial states @xmath270 and fed inputs @xmath255 and @xmath190 . by abuse of notation we will denote @xmath275 by @xmath276 . the goal of this section is to present a brief overview of realization theory of lsss . in subsection [ lin_sys : io : map ] we recall the definition of input - output maps of lsss and the notion of a lss realization . in subsection [ lin_sys : system_theo ] . we review a number of system - theoretic concepts such as observability , semi - reachability , minimality and lss morphisms . in subsection [ lin_sys : real_problem ] the realization problem will be formulated . subsection [ lin_sys : gen_ker ] recalls the concept of generalized kernel representation of a family of input - output maps . subsection [ lin_sys : markov ] presents the definition of markov - parameters for lsss . finally , subsection [ lin_sys : real ] presents the main results on realization theory of lsss . for a more details on the material of this section see @xcite . in this section we define the class of maps which represent potential input - output maps of lsss . in addition , we introduce the notion of a lss realization . [ sect : switch_sys : def-1 ] in this paper , unless stated otherwise , an _ input - output map _ will mean a map of the form @xmath277 . the set of all such maps will be denoted by @xmath278 . a _ family of input - output maps _ is just a ( possibly infinite ) subset of the set of all input - output maps @xmath278 . that is , an input - output map maps continuous - valued inputs and non - empty switching sequences to continuous - valued outputs . in order to formalize the notion of a realization by a lss of a family of input - output maps we will adopt the following formalism . [ switch_sys : real : def1 ] consider a set @xmath279 of input - output maps . the family @xmath7 is said to be _ realized _ by a lss @xmath26 of the form ( [ lin_switch0 ] ) if there exists a map @xmath280 , which maps each input - output map @xmath2 from @xmath7 to a state @xmath281 of @xmath26 , such that @xmath2 equals the input - output map induced by @xmath282 , i.e. @xmath283 one can think of the map @xmath284 as a way to determine the initial state corresponding to each element of @xmath7 . [ switch_sys : real : def2 ] let @xmath7 be a family of input - output maps . we will refer to the pairs @xmath285 , where @xmath26 is a lss of the form ( [ lin_switch0 ] ) and @xmath280 is a map mapping elements of @xmath7 to the states of @xmath26 , as lss _ realizations _ ( _ realizations _ for short ) . the realization @xmath285 is said to be a _ realization of @xmath7 _ , if ( [ sect : problem_form : real0 ] ) holds for all @xmath13 . note that not any realization @xmath285 with @xmath286 is a realization of @xmath7 . the statement that @xmath285 is a _ realization _ only expresses the fact that we associate a state of @xmath26 with each element of @xmath7 . however , we do not yet require that the input - output map induced by a designated state equals the corresponding element of @xmath7 . the latter is required only if we claim that @xmath285 _ is a realization of @xmath7_. the goal of this section is to define system theoretic concepts such as observability , span - reachability , system morphism , dimension and minimality for lsss and for lss realizations . throughout the section @xmath26 denotes a lss of the form ( [ lin_switch0 ] ) . the _ reachable set _ of @xmath26 from a set of initial states @xmath287 is defined as @xmath288 that is , @xmath289 is the set of all those states which are obtained by starting the system from an initial state in @xmath290 , applying some continuous - valued input and some finite switching sequence , and considering the state at the last switching time . the lss @xmath26 is said to be _ reachable _ from @xmath290 if @xmath291 holds . the lss @xmath26 is _ semi - reachable _ from @xmath290 if @xmath102 is the smallest vector space containing @xmath289 . in other words , @xmath26 is semi - reachable from @xmath290 if the linear span of the elements of the reachable set @xmath289 yields the whole state - space @xmath102 . we proceed with defining the notion of observability for lsss . two states @xmath292 of the lss @xmath26 are _ indistinguishable _ if the input - output maps induced by @xmath293 and @xmath294 coincide , i.e. for all @xmath43 and @xmath44 , @xmath295 . the lss @xmath26 is called _ observable _ if it has no pair of distinct indistinguishable states . in other words , observability means that if we pick any two states of the system , then we are able to distinguish between them by feeding a suitable continuous - valued input and a suitable switching sequence and then observing the resulting output . below we will define the notion of dimension for lsss . [ switch_sys : dim : def ] define the dimension of @xmath26 , denoted by @xmath296 , as the dimension @xmath297 of is state - space . now we are ready to define the concept of minimality for lsss . let @xmath7 be a family of input - output maps and let @xmath285 be a lss realization of @xmath7 . @xmath285 is _ a minimal realization of @xmath7 _ , if for any lss realization @xmath298 of @xmath7 , @xmath299 . in simple words , a lss realization is a minimal realization of @xmath7 if it has the smallest dimensional state - space among all the linear switched systems which are realizations of @xmath7 . the notions of observability and semi - reachability can be extended to lss realizations as follows . [ switch_sys : real : def3 ] let @xmath7 be a family of input - output maps and let @xmath280 be a map from @xmath7 to the state - space of @xmath26 . the realization @xmath285 is _ semi - reachable _ , if @xmath26 is semi - reachable from the range @xmath300 of @xmath284 . the realization @xmath285 is _ observable _ , if @xmath26 is observable . as the next step , we will define the notion of a lss morphism . [ sect : problem_form : lin : morphism ] consider a lss @xmath301 of the form ( [ lin_switch0 ] ) and a lss @xmath302 of the form @xmath303 and @xmath302 have the same set of discrete modes . ] . a linear map @xmath304 is said to be a _ lss morphism _ from @xmath305 to @xmath306 , and it is denoted by @xmath307 , if for all discrete modes @xmath225 , @xmath308 the map @xmath5 is called surjective ( injective ) if it is surjective ( injective ) as a linear map . the map @xmath5 is said to be a lss isomorphism , if it is an isomorphism as a linear map . consider two lss realizations @xmath309 and @xmath310 such that the domain of definition of both @xmath311 and @xmath312 is a certain family @xmath7 of input - output maps . a lss morphism @xmath307 is called a _ lss morphism _ from realization @xmath309 to @xmath313 , if @xmath314 holds , or , in other words , if for all @xmath13 , @xmath315 . the fact that @xmath5 is a lss morphism from @xmath316 to @xmath310 will be denoted by @xmath317 . the lsss realizations @xmath309 and @xmath313 are said to be _ algebraically similar _ or _ isomorphic _ if there exists an lss isomorphism @xmath317 . the realization problem for lsss can be formulated as follows . [ lin_switch : real_problem ] find necessary and sufficient conditions for existence of a lss realization for a family of input - output maps @xmath7 . find a characterization of minimal lss realizations of @xmath7 . determine if minimal realizations of @xmath7 are unique in any sense . in this section we will recall the notion of _ generalized kernel representation_. it turns out that a family of input - output maps can be realized by a lss only if it admits a generalized kernel representation . informally , a family @xmath7 of input - output maps has a generalized kernel representation if the following hold . 1 . there exists an input - output map @xmath19 such that for all @xmath13 , @xmath318 for all continuous - valued inputs @xmath255 and switching sequences @xmath190 . each element @xmath2 of @xmath7 is affine in continuous - valued inputs and analytic in switching times for all constant inputs . a good intuition for the notion of generalized kernel representation can be derived by analogy with input - output maps of linear systems . recall from @xcite that an input - output map @xmath319 can be realized by a linear system @xmath320 from the initial state @xmath321 , only if there exist analytic functions @xmath322 and @xmath323 such that @xmath324 more precisely , in this case @xmath325 using the setting above , the maps @xmath326 and @xmath327 to be defined below are analogous to the map @xmath56 and @xmath328 respectively . the formal definition is as follows . [ sect : io : def1 ] consider a family of input - output maps @xmath7 , as defined in definition [ sect : switch_sys : def-1 ] . the family @xmath7 is said to have _ generalized kernel representation _ , if for all input - output maps @xmath13 and for all non - empty sequences of discrete modes @xmath329 , @xmath330 , @xmath189 , there exist functions @xmath331 such that the following holds . 1 . [ gen_ker/1 ] for each word @xmath332 and for each input - output map @xmath13 , the functions @xmath333 and @xmath334 are analytic . [ gen_ker/2 ] for each input - output map @xmath13 and for each ( possibly empty ) sequences @xmath335 , and for each discrete mode @xmath225 , it holds that for all @xmath336 , @xmath337 3 . [ gen_ker/3 ] for each pair of sequences @xmath338 such that and @xmath190 is not the empty word , i.e. @xmath339 , and for each discrete mode @xmath225 , the following holds . for each input - output map @xmath13 and for each @xmath340 , @xmath341 for each pair of words @xmath338 such that both @xmath203 and @xmath190 are not empty , i.e. @xmath342 , @xmath339 , and for each discrete mode @xmath225 , the following holds . for each @xmath340 , @xmath343 4 . [ gen_ker/4 ] for each input - output map @xmath13 , for each non - empty switching sequence @xmath344 , where @xmath345 and @xmath346 , each piecewise - continuous input @xmath43 , the following holds . @xmath347 the reader may view the functions @xmath348 as the part of the output which depends on the initial condition and the functions @xmath334 as functions determining the dependence of the output on the continuous inputs . in fact , let @xmath285 be a lss realization of @xmath7 and assume that @xmath26 is of the form ( [ lin_switch0 ] ) . it is easy to see that for all @xmath13 and for any input @xmath43 and switching sequence @xmath349 , @xmath350 from the equation above it is easy to see that @xmath7 admits a hybrid kernel representation of the form @xmath351 for all non - empty sequences of discrete modes @xmath345 , @xmath352 . next we will define the notion of markov parameters for input - output maps . markov - parameters play a central role in ( partial ) realization theory of lsss . before proceeding further , recall from classical linear systems theory @xcite the notion of markov parameter . consider the linear input - output map of the form ( [ gen_ker_int/0 ] ) and define the markov parameters of this map as derivatives of the maps @xmath322 and @xmath323 ; @xmath353 and @xmath354 for all @xmath241 . in turn , the derivatives of @xmath355 and @xmath356 can be expressed as the derivatives of the input - output map as follows . for a constant input @xmath357 define the map @xmath358 that is , @xmath359 is just the value of the input - output map ( [ gen_ker_int/0 ] ) at time @xmath29 if a constant input @xmath255 is fed in . then @xmath360 and hence @xmath361 . similarly , the @xmath10th column of @xmath362 can be written as @xmath363 where @xmath122 , @xmath364 is the @xmath10th unit vector of @xmath221 . that is , _ the markov - parameters of a linear input - output map are the high - order time derivatives of the output trajectories induced by certain constant inputs , evaluated at zero_. for lsss the markov - parameters are defined in a similar way . before proceeding to the definition , we need the following notation [ sect : io : io_der ] consider an input - output map @xmath2 as in definition [ sect : switch_sys : def-1 ] , a non - empty sequence of discrete modes @xmath365 , @xmath345 , @xmath352 and a continuous - valued input @xmath43 . define the map @xmath366 as follows @xmath367 that is , the values of @xmath368 are obtained from the values of @xmath2 by fixing the a piecewise - continuous input @xmath255 and a sequence of discrete modes @xmath190 and varying the switching times only . [ linswitch_intro : def : markov ] let @xmath7 be a family of input - output maps admitting a generalized kernel representation . the _ markov parameters _ of @xmath7 are the vectors @xmath369 and @xmath370 , defined for all discrete modes @xmath371 , all the sequences of discrete modes @xmath372 ( including the empty sequence ) , all @xmath364 and @xmath13 as follows . if @xmath373 for @xmath241 , @xmath345 , then @xmath374 here @xmath122 is the @xmath10th unit vector of @xmath221 . the vectors @xmath375 correspond to the parameters @xmath376 for linear systems , and @xmath377 corresponds to the @xmath10th column of the @xmath378 component of the markov parameters . as it was shown in @xcite , there is a close relationship between the markov parameters of @xmath7 and products of system matrices . [ sect : io : prop1 ] let @xmath26 be of the form ( [ lin_switch0 ] ) and let @xmath280 . the realization @xmath285 is a realization of @xmath7 , if and only if the markov - parameters of @xmath7 are equal to the following products of the system matrices . @xmath379 for each @xmath380 , for each @xmath345 , @xmath241 , @xmath13 and @xmath364 . similarly to the linear case , the markov - parameters of @xmath7 can also be represented as derivatives of the maps @xmath326 and @xmath327 , see @xcite . the purpose of this section is to present formally the main results on realization theory of lsss . to this end , the notion of the hankel - matrix @xmath54 of @xmath7 is defined . similarly to the linear case , the entries of the hankel - matrix will be formed by the markov parameters . [ main_result : lin : hank : arb : def ] assume that the set of discrete modes @xmath223 has @xmath381 elements and choose an enumeration of @xmath223 @xmath382 recall the notation for infinite matrices presented in section [ sect : prelim : not_matrix ] . define the _ hankel - matrix _ of @xmath7 as the infinite real matrix , columns and rows of which are indexed as follows . the rows of @xmath54 are indexed by pairs @xmath383 where @xmath384 is a finite sequence of discrete modes and @xmath385 . the columns of @xmath54 are indexed by pairs @xmath386 , where @xmath372 is a sequence of discrete modes and @xmath10 is either an element of @xmath7 , or @xmath10 is a pair of the form @xmath387 , where @xmath9 is a discrete mode and @xmath388 . that is , @xmath54 is an infinite real matrix of the form @xmath389 where @xmath390 and @xmath391 . the entries of @xmath54 are defined as follows . fix sequences of discrete modes @xmath392 and fix an element @xmath393 . for any @xmath394 of the form @xmath395 where @xmath396 and @xmath397 , the entry @xmath398 is defined as follows @xmath399 here @xmath400 and @xmath401 denote the @xmath402th element of the vectors @xmath403 and @xmath404 respectively . that is , @xmath54 is constructed from certain high - order derivatives of the input - output maps belonging to @xmath7 . as it was noted in section [ sect : prelim : not_matrix ] , the columns of @xmath54 belong to the vector space of all maps @xmath405 . hence , we can speak of the linear span of the columns of @xmath54 . in addition , according to the convention adopted in section [ sect : prelim : not_matrix ] , the _ rank of @xmath54 _ , denoted by @xmath406 , is the dimension of the linear subspace spanned by the columns of @xmath54 . now we are ready to state the main theorem on the existence of a lss realization for arbitrary switching . [ sect : real : theo2 ] let @xmath7 be a family of input - output maps . then @xmath7 has a realization by a lss if and only if @xmath7 has a generalized kernel representation and the rank of the associated hankel - matrix @xmath54 of @xmath7 is finite , i.e , @xmath407 . next , we state state the main result of on minimality of lsss . [ sect : real : theo3 ] if @xmath285 is a lss realization of @xmath7 , then the following are equivalent . * @xmath285 is a minimal lss realization of @xmath7 . * the realization @xmath285 is semi - reachable and it is observable . * the state - space dimension of @xmath26 equals the rank of the hankel - matrix of @xmath7 , i.e. @xmath408 . in addition , all minimal lss realizations of @xmath7 are isomorphic . the purpose of this section is to present the main results of the paper formally . the outline of this section is as follows . in subsection [ sect : main_result : part_real_problem ] we state the partial - realization problem for lsss formally . in subsection [ sect : main_result : hank ] we define the notion of hankel sub - matrix , which will be needed for the statement of the main results . in subsection [ sect : main_result : alg0 ] we present the theorem characterizing the existence and minimality of partial realizations . in addition , we present a simple algorithm for computing a minimal partial realization . finally , in subsection [ sect : main_result : alg1 ] we present a kalman - ho - like algorithm for computing a minimal partial realization . let @xmath7 be a family of input - output maps admitting a generalized kernel representation . as we saw earlier , the markov - parameters uniquely determine the input - output maps of a linear switched system . in addition , the realization problem for lss can be reduced to the problem of finding a suitable representation of the markov - parameters of the input - output maps . however , in practice we can obtain only markov - parameters only up to some finite order . hence , the _ partial - realization problem _ ( with respect to the number of switches ) arises . in order to present a formal problem formulation , the notion of ( minimal ) partial realization has to be introduced . [ def : part_real ] let @xmath7 be a family of input - output maps admitting a generalized kernel representation . assume that @xmath26 is a lss of the form ( [ lin_switch0 ] ) and let @xmath286 . a realization @xmath285 is said to be an @xmath3-partial realization of @xmath7 , if for all @xmath380 , @xmath13 , @xmath364 , and for all ( possibly empty ) sequences of discrete modes @xmath345 , of length at most @xmath3 , i.e. @xmath409 , the following holds . @xmath410 the lss @xmath26 is said to be an @xmath3-partial realization of @xmath7 , if for there exists a map @xmath280 such that the realization @xmath285 is a @xmath3-partial realization of @xmath7 . the realization @xmath285 is said to be a partial realization of @xmath7 , if it is a @xmath3-partial realization of @xmath7 for some @xmath0 . from lemma [ sect : io : prop1 ] it follows that @xmath285 is a realization of @xmath7 if and only if @xmath285 is an @xmath3-partial realization of @xmath7 for all @xmath0 . note that if @xmath7 and @xmath285 is a @xmath3-partial realization of @xmath7 , then ( [ sect : real : prop1:eq1.1 ] ) holds for _ finitely many _ markov - parameters . in the sequel , we will often refer to the markov - parameters which are indexed by a sequence of discrete modes of length @xmath3 as _ markov - parameters of order @xmath3_. that is , a markov - parameter of order @xmath3 is either of the form @xmath377 or @xmath375 with @xmath411 . with this terminology , a lssis a @xmath3-partial realization of @xmath7 , if it recreates the markov - parameters of @xmath7 of order at most @xmath3 . if @xmath7 is finite , then the set of markov - parameters of @xmath7 of order at most @xmath3 is finite . a lss realization @xmath285 is said to be a minimal @xmath3-partial realization of @xmath7 , if @xmath285 is a @xmath3-partial realization of @xmath7 and for any other @xmath3-partial lss realization @xmath412 of @xmath7 , @xmath413 . in plain words , a minimal @xmath3-partial realization of @xmath7 is a @xmath3-partial realization of @xmath7 with the smallest possible state - space dimension . [ prob : part : real ] let @xmath7 be a family of input - output maps admitting a generalized kernel representation . the partial - realization problem entails the following problems . * find conditions for existence of a @xmath3-partial realization of @xmath7 . formulate an algorithm for computing a @xmath3-partial realization from generalized markov - parameters of @xmath7 of some bounded order . if @xmath7 is finite , this amounts to computing a @xmath3-partial realization from _ finitely many _ markov - parameters . * characterize minimal @xmath3-partial realizations of @xmath7 , find conditions for their existence and uniqueness . * find conditions under which a @xmath3-partial realization of @xmath7 is a complete realization of @xmath7 in the sense of definition [ switch_sys : real : def2 ] . we will devote the remaining part of the section to presenting the solution to the problem formulated above . we will show that it is possible to compute a minimal @xmath3-partial realization from a suitably chosen sub - matrix of the hankel matrix of @xmath7 . in fact , this sub - matrix is formed by markov - parameters of bounded order , and hence it is finite , if @xmath7 is finite . if @xmath3 is large enough , the resulting @xmath3-partial realization will be a complete realization of @xmath7 . in addition , if @xmath7 is finite , then this sub - matrix will be finite as well . in order to state the main results rigorously , we need to define formally . certain finite sub - matrices of the hankel - matrix . to this end , recall from definition [ main_result : lin : hank : arb : def ] the definition the hankel - matrix of @xmath7 and of the sets @xmath414 , @xmath415 and @xmath391 . fix natural numbers @xmath416 . we will use the following notation for sequences of discrete modes of length at most @xmath55 . we will denote by @xmath417 the set of all sequences of discrete modes of length at most @xmath55 , i.e. @xmath418 . notice that @xmath419 is a finite set of cardinality @xmath420 . we will define the sub - matrix of the hankel - matrix @xmath54 indexed by sequences of bounded length as follows . [ sect : main_result : sub - hank : def1 ] let @xmath421 be the sub - matrix of @xmath54 formed by the intersections of all the rows indexed by an index of the form @xmath383 with @xmath131 and @xmath422 and all the columns indexed by an index of the form @xmath386 with @xmath393 and @xmath423 . that is , @xmath424 and for each @xmath425 and @xmath426 , @xmath427 notice that the entries of@xmath421 are made up of markov - parameters of order at most @xmath428 . hence , if @xmath7 is finite , then the set @xmath429 is finite and the matrix @xmath421 is a finite matrix . since the entries of the hankel - matrix were defined in terms of the generalized markov - parameters of @xmath7 , the same is true for the entries of @xmath430 . writing out the definition of the entries of @xmath421 and using the enumeration of @xmath223 fixed in ( [ main_result : lin : hank : arb : def_enum ] ) of definition [ main_result : lin : hank : arb : def ] , we get that for any @xmath394 of the form @xmath431 where @xmath396 and @xmath397 , the entry @xmath432 is defined as follows @xmath433 here @xmath400 and @xmath401 denote the @xmath402th element of the vectors @xmath403 and @xmath404 respectively . that is , if @xmath7 is finite , then @xmath421 represents a finite collection of markov - parameters of @xmath7 . in turn the markov - parameters of @xmath7 are defined via the high - order derivatives of the elements of @xmath7 with respect to the switching times . hence , we get that the entries of @xmath421 are just high - order derivatives with respect to the switching times of the elements of @xmath7 , and the order of these derivatives is bounded by @xmath434 . as we remarked in the introduction , there are essentially two ways to construct a partial lss realization . in algorithm [ alg0 ] we present the first one which is conceptually the simplest one . assume that @xmath435 and let @xmath436 be a linear isomorphism from the column space of @xmath437 to @xmath123 . for each column index @xmath438 , _ denote by @xmath439 the column of @xmath437 indexed by @xmath386_. define the lss realization @xmath440 as follows . @xmath441 is a lss of the form ( [ lin_switch0 ] ) such that * the state - space of @xmath441 is @xmath123 . * for each mode @xmath225 , the matrix @xmath442 satisfies @xmath443 * for each mode @xmath225 , the matrix @xmath228 satisfies @xmath444 for each @xmath445 . * for each mode @xmath225 , the matrix @xmath226 is the solution of the system of linear equations @xmath446 if the system of equations ( [ alg0:eq1 ] ) does not have a solution , then abort . for each input - output map @xmath13 , @xmath447 return @xmath440 , with @xmath448 . informally , the realization @xmath440 returned by algorithm [ alg0 ] is defined on an isomorphic copy of the column space of @xmath437 . the @xmath10th column of the matrix @xmath449 is the isomorphic copy of the column of @xmath437 indexed by @xmath8 . the matrices @xmath450 are such that if they are interpreted as linear maps , then each @xmath450 maps the isomorphic copy of a column of @xmath437 to its first @xmath451 rows , i.e. rows indexed by @xmath452 , where @xmath193 denotes the empty word . the matrices @xmath453 realize a shift on the columns of @xmath437 ; by interpreting @xmath453 as a map on the columns of @xmath437 , @xmath453 maps the column indexed by an index @xmath454 to the column indexed by @xmath455 . finally , the value @xmath456 of the map @xmath457 for the input - output map @xmath13 equals the column of @xmath437 indexed by @xmath458 . notice that algorithm [ alg0 ] may fail to return a realization , as ( [ alg0:eq1 ] ) need not always have a solution . if @xmath7 consists of finitely many input - output maps , then algorithm [ alg0 ] is a indeed an effective procedure and it can be implemented as a numerical algorithm . we state the partial - realization theorem as a theorem providing sufficient conditions for algorithm [ alg0 ] to yield a @xmath459-partial , semi - reachable and observable realization of @xmath7 . [ part_real_lin : theo1 ] let @xmath7 be a family of input - output maps admitting a generalized kernel representation . * * existence and computability of a partial realization * + assume that for some @xmath460 , @xmath461 holds . then algorithm [ alg0 ] returns a realization @xmath462 , and @xmath440 is a semi - reachable and observable @xmath459-partial realization of @xmath7 . * * existence of a complete realization * + if @xmath463 , then ( [ part_real_lin : theo1:eq1 ] ) holds and the @xmath459-partial realization @xmath440 is a minimal realization of @xmath7 . the condition @xmath463 holds for a given @xmath3 , if there exists a lss realization @xmath285 of @xmath7 such that @xmath464 . the proof of theorem [ part_real_lin : theo1 ] is presented in section [ sect : lin_alg ] . theorem [ part_real_lin : theo1 ] implies that if @xmath7 has a realization by a lss , then the parameters of a minimal lss realization of @xmath7 are computable from , and hence are determined by , finitely many time derivatives with respect to the switching times . [ part_real_lin : theo1:remark ] the numerical example in section [ sect : num ] provides an example of a family of input - output maps @xmath7 realizable by a lss , such that @xmath7 has the following property . for some @xmath3 the family @xmath7 satisfies ( [ part_real_lin : theo1:eq1 ] ) but not the condition @xmath463 . that is , there is an @xmath0 such that theorem [ part_real_lin : theo1 ] yields a partial realization of @xmath7 which is not a complete realization . that is , the two statements of theorem [ part_real_lin : theo1 ] indeed describe two separate cases . theorem [ part_real_lin : theo1 ] allows us to formulate the following characterization of minimal partial lss realizations . [ part_real_lin : theo1.1 ] with the notation of theorem [ part_real_lin : theo1 ] , if @xmath7 satisfies ( [ part_real_lin : theo1:eq1 ] ) , then the following holds . 1 . [ part_real_lin : theo1.1:res1 ] a minimal @xmath459 partial realization of @xmath7 exists , in fact , the realization @xmath440 returned by algorithm [ alg0 ] is a minimal @xmath459 partial realization of @xmath7 . [ part_real_lin : theo1.1:res2 ] any minimal @xmath459 partial realization of @xmath7 is semi - reachable and observable and it is of dimension @xmath465 . [ part_real_lin : theo1.1:res3 ] all minimal @xmath459 partial realizations of @xmath7 are isomorphic . the proof of theorem [ part_real_lin : theo1.1 ] is presented in section [ sect : lin_alg ] . the reader might wonder why we talk about @xmath459-partial realizations in theorem [ part_real_lin : theo1 ] and [ part_real_lin : theo1.1 ] . the reason behind it is that the finite hankel - matrix @xmath437 is formed by values of the markov - parameters of order at most @xmath459 . we would like the lss realization obtained from @xmath437 to recreate at least those markov - parameters which are the entries of the matrix @xmath437 . but this means that the lss realization obtained from @xmath437 must be a @xmath459-partial realization of @xmath7 . while algorithm [ alg0 ] is theoretically attractive because of its simplicity , it is not necessarily the most suitable one for numerical implementation . in algorithm [ partreal ] we present an alternative algorithm for computing a semi - reachable and observable @xmath459-partial realization of @xmath7 . the algorithm is based on factorization of the hankel - matrix @xmath466 and it is similar to the kalman - ho algorithm . compute a decomposition of @xmath466 @xmath467 where @xmath468 and @xmath469 are matrices such that @xmath470 . recall from ( [ main_result : lin : hank : arb : def_enum ] ) the enumeration of @xmath223 . for each @xmath225 , @xmath471 , for some @xmath472 define the matrix @xmath473 by @xmath474 where @xmath475 denotes the row of @xmath476 indexed by @xmath6 . for each @xmath225 , define the matrix @xmath477 by @xmath478 where @xmath479 stands for the column of @xmath480 indexed by @xmath481 for @xmath225 and @xmath364 . define the map @xmath482 as @xmath483 where @xmath484 stands for the column of @xmath480 indexed by @xmath458 for @xmath13 . for each @xmath225 let the matrix @xmath485 be the solution of equation @xmath486 where @xmath487 are matrices of the form @xmath488 for all @xmath489 . that is , @xmath490 is obtained from @xmath476 by deleting all the rows indexed by pairs @xmath383 where @xmath203 is a sequence of discrete modes of length @xmath491 . the matrix @xmath492 is the shifted version of @xmath476 , i.e. its row indexed by @xmath383 is the row of @xmath476 indexed by @xmath493 . if there no unique solution to ( [ computepartialreal ] ) then return @xmath494 . otherwise return @xmath495 where @xmath496 is a lss of the form @xmath497 the result of algorithm [ partreal ] is described in the theorem below . [ part_real_lin : theo2 ] with the notation above the following holds . 1 . [ part_real_lin : theo2:res1 ] assume that form some @xmath25 , ( [ part_real_lin : theo1:eq1 ] ) of theorem [ part_real_lin : theo1 ] holds . then algorithm [ partreal ] always returns a lss realization @xmath498 and @xmath498 is a minimal @xmath459-realization of @xmath7 . in fact , @xmath495 is isomorphic to the lss realization @xmath440 of theorem [ part_real_lin : theo1 ] . [ part_real_lin : theo2:res2 ] if for some @xmath499 , @xmath500 , then ( [ part_real_lin : theo1:eq1 ] ) holds , and algorithm [ partreal ] returns a minimal realization @xmath495 of @xmath7 . if there exists a lss realization @xmath285 of @xmath7 , such that @xmath464 , then @xmath463 holds and the realization @xmath498 returned by algorithm [ partreal ] is a minimal realization of @xmath7 . the proof of theorem [ part_real_lin : theo2 ] is presented in section [ sect : lin_alg ] . theorem [ part_real_lin : theo2 ] has immediate consequences for systems identification and model reduction of linear switched systems , by guaranteeing correctness of the realization algorithm described in algorithm [ partreal ] . in section [ sect : concl ] we discuss the potential applications of theorem [ part_real_lin : theo1 ] and theorem [ part_real_lin : theo2 ] in more detail . there are many algorithms to compute a factorization of @xmath466 and algorithm [ partreal ] can work with any of these algorithms . in particular , the factorization of @xmath466 can be done through singular value decomposition . more precisely , assume that @xmath7 is finite , and let @xmath501 be the singular value decomposition of @xmath466 . notice that if @xmath7 is finite , then the index sets of @xmath466 are finite and the singular value decomposition can be computed by enumerating the index sets of @xmath466 and viewing @xmath466 as a usual finite matrix . the choice @xmath502 and @xmath503 yields a decomposition @xmath504 satisfying the conditions of algorithm [ partreal ] . algorithm [ partreal ] has been implemented , and the factorization used in the implementation is precisely the singular value decomposition of @xmath466 . the goal of this section is to present the proof of theorem [ part_real_lin : theo1 ] , [ part_real_lin : theo1.1 ] , and [ part_real_lin : theo2 ] . both proofs rely heavily on the relationship between rational representations and lsss described in @xcite . the outline of the section is the following . in subsection [ sect : lin : real ] we will present a short overview of the relationship between rational representations and lsss . subsection [ sect : lin : part_real ] presents the proof of theorem [ part_real_lin : theo1 ] , [ part_real_lin : theo1.1 ] , and [ part_real_lin : theo2 ] . the reader is advised to review appendix [ sect : pow ] and appendix [ sect : pow : part ] before reading this section . throughout the section we will tacitly use the notation and terminology of appendix [ sect : pow ] and appendix [ sect : pow : part ] . below we will present a brief overview of the relationship between formal power series and their representations and lsss . for a more detailed presentation see @xcite . informally , the relationship is as follows . * let @xmath7 be a family of input - output maps and assume that @xmath7 has a generalized kernel representation . then we can construct a family of formal power series @xmath505 from the generalized markov - parameters of @xmath7 . the details of the construction will be presented below . the family @xmath505 will be referred to as _ the family of formal power series associated with @xmath7_. * if @xmath285 is a lss realization of @xmath7 , then we can construct a rational representation @xmath506 of @xmath505 from the parameters of @xmath285 . the converse is also true ; if @xmath480 is a representation of @xmath505 , then we can construct a realization @xmath507 of @xmath7 from the parameters of @xmath480 . the details of the construction will be presented below . we will call @xmath506 the _ representation associated with @xmath285 _ and we will call @xmath507 the _ realization associated with @xmath480_. we will start with defining the family of formal power series @xmath505 associated with @xmath7 . to this end , recall from definition [ linswitch_intro : def : markov ] the definition of markov - parameters @xmath377 , @xmath375 of @xmath7 , for each pair of discrete modes @xmath380 , input - output map @xmath13 , index @xmath364 , and a word @xmath372 . it is easy to see that the maps @xmath508 and @xmath509 define formal power series @xmath510 and @xmath511 in @xmath512 . recall the enumeration of the set of discrete modes @xmath223 defined in ( [ main_result : lin : hank : arb : def_enum ] ) . that is , @xmath223 assumed to have @xmath381 elements given by the distinct elements @xmath513 . for each discrete mode @xmath225 , index @xmath514 , and input - output map @xmath13 define the formal power series @xmath515 as follows ; for each word @xmath372 let @xmath516 that is , the values of the formal power series @xmath517 are obtained by stacking up the values of @xmath518 for @xmath519 . similarly , the values of @xmath520 are obtained by stacking up the values of @xmath521 for @xmath519 . define the set @xmath522 . define the _ indexed set of formal power series associated with @xmath7 _ as the following family of formal power series @xmath523 [ sect : real : hank ] it is easy to see that the hankel - matrix @xmath524 of the family of formal power series @xmath505 is identical to the hankel - matrix @xmath54 of @xmath7 as defined in definition [ main_result : lin : hank : arb : def ] next we define the rational representation @xmath506 associated with a lss realization @xmath285 . let @xmath26 be a lss of the form ( [ lin_switch0 ] ) and assume that @xmath280 is a map assigning to each element @xmath13 an initial state of @xmath26 . [ sect : real : const1 ] define the _ representation associated with @xmath285 _ as the rational @xmath525 representation @xmath526 the various components of @xmath506 are defined as follows . * * state - space @xmath102 . * the state - space of the rational representation @xmath506 is the same as the state - space of @xmath26 , i.e. @xmath527 . * * alphabet . * the representation @xmath506 is defined over the alphabet which equals the set of discrete modes @xmath223 . * * state - transition maps ( matrix ) @xmath528 . * for each discrete mode @xmath225 , the corresponding state - transition matrix @xmath453 of @xmath506 is identical to the matrix @xmath453 of @xmath26 . * * readout matrix ( map ) @xmath529 . * the readout matrix @xmath530 is obtained by vertically `` stacking up '' the matrices @xmath531 in this order from top to bottom . that is , the @xmath532 matrix @xmath530 is of the form @xmath533 * * initial states*. the indexed set of the initial states of @xmath506 is of the form @xmath534 i.e. it is indexed by the elements of the index set @xmath429 . its elements are defined by @xmath535 if @xmath2 is an element of the family @xmath7 , and @xmath536 , i.e. @xmath537 is the @xmath538th column of @xmath449 , for all @xmath225 and @xmath539 . conversely , below we will construct a lss realization @xmath540 from any rational representation @xmath480 satisfying some mild conditions . we would like to note that these conditions are automatically satisfied by a suitably chosen isomorphic copy of any rational representation of @xmath505 . the definition goes as follows . [ sect : real : const2 ] + consider a @xmath525 representation @xmath480 of the following form @xmath541 and assume that the following holds . 1 . the state - space is of the form @xmath220 for some @xmath542 . does not hold , then replace @xmath480 with the isomorphic copy @xmath543 defined in ( [ repr : euclid : eq1 ] ) , remark [ repr : state_space : rem ] whose state - space is @xmath123 . since @xmath543 and @xmath480 are isomorphic , if @xmath480 is a representation of @xmath505 , then @xmath543 will be a representation of @xmath505 as well . ] the readout map @xmath530 takes its values in @xmath544 . the set of initial states @xmath545 of @xmath480 is indexed by the index set @xmath546 . define the _ lss realization @xmath507 associated with r _ as follows . let @xmath547 be of the form ( [ lin_switch0 ] ) , where * * state - space . * the state - space @xmath123 of @xmath547 is the same as that of @xmath480 . * * system matrices @xmath548 of @xmath547 . * for each discrete mode @xmath225 , the matrix @xmath453 of @xmath547 is identical to the state - transition matrix @xmath453 of @xmath480 . * * system matrices @xmath549 of @xmath547 . * for any discrete state @xmath225 of the form @xmath471 for some @xmath472 and for any @xmath550 , the @xmath538th row of @xmath450 equals the @xmath551-th row of @xmath530 , i.e. @xmath552 * * system matrices @xmath553 of @xmath547 . * for each discrete mode @xmath225 the @xmath136 matrix @xmath449 is obtained as follows ; the @xmath538th column of @xmath449 equals the initial state @xmath554 for all @xmath555 . that is @xmath556 the map @xmath557 assigns to each element @xmath2 of @xmath7 the initial state of @xmath480 indexed by @xmath2 , i.e. @xmath558 the following theorem states the relationship between representations and realizations formally . if we apply construction [ sect : real : const1 ] to @xmath540 then the resulting representation @xmath559 coincides with @xmath480 , i.e. @xmath560 . conversely , if we apply construction [ sect : real : const2 ] , to the representation @xmath506 associated with a lss realization @xmath285 , then we get @xmath285 back , i.e. @xmath561 , @xmath562 . [ sect : real : theo1 ] let @xmath7 be a family of input - output maps and assume that @xmath7 admits a generalized kernel representation . with the notation and assumptions above the following holds . * the realization @xmath285 is a lss realization of @xmath7 if and only if @xmath506 is a rational representation of @xmath505 . * the representation @xmath480 is a representation of @xmath505 if and only if @xmath507 is a realization of @xmath7 . * the realization @xmath285 is semi - reachable if and only if @xmath506 is reachable , and @xmath285 is observable if and only if @xmath506 is observable . similarly , @xmath507 is semi - reachable if and only if @xmath480 is reachable , and @xmath540 is observable if and only if @xmath480 is observable . * if @xmath285 is a minimal realization of @xmath7 , then @xmath506 is a minimal representation of @xmath505 . conversely , if @xmath480 is a minimal representation of @xmath505 , then @xmath507 is a minimal realization of @xmath7 . * the map @xmath563 is a lss morphism if and only if @xmath564 is a representation morphism . coincides with that of @xmath506 and the state - space of @xmath412 coincides with that of @xmath565 . hence @xmath5 can indeed be viewed as a representation morphism , provided that it commutes with the matrices of the representations involved ] in addition , @xmath5 is a representation isomorphism if and only if @xmath5 is a lss isomorphism . conversely , if the representations @xmath480 and @xmath566 satisfy the assumptions of construction [ sect : real : const2 ] , then @xmath567 is a representation morphism if and only if @xmath568 is a lss morphism . moreover , in this case @xmath436 is a representation isomorphism if and only if it is a lss isomorphism . the goal of this section is to present the proof of theorem [ part_real_lin : theo1 ] , theorem [ part_real_lin : theo1.1 ] and theorem [ part_real_lin : theo2 ] . before proceeding to the proof , we will need the following result on the relationship between the partial - realization problem for lsss and the partial - realization problem for rational representations . [ part_real_lin : theo1/2 ] assume that @xmath7 is a family of input - output maps admitting a generalized kernel representation . a lss realization @xmath285 is an @xmath459-partial realization of @xmath7 if and only of the associated representation @xmath506 is a @xmath459-partial representation of the family of formal power series @xmath505 . conversely , a representation @xmath480 is a @xmath459-partial representation of the family @xmath505 of formal power series associated with @xmath7 , if and only if the lss realization @xmath507 associated with @xmath480 is a @xmath459-partial realization of @xmath7 . the second statement of the theorem follows from the first one by noticing that @xmath569 . hence , it is enough to prove the first statement of the theorem . it follows easily from definition [ def : part_real ] and the definition of the formal power series @xmath570 , @xmath13 , @xmath571 , @xmath364 , that @xmath285 is an @xmath459-partial realization of @xmath7 , if and only if for all @xmath572 , @xmath571 , @xmath13 , @xmath364 @xmath573 from the definition of @xmath506 it follows directly that ( [ sect : real : pow : eq3 ] ) is equivalent to @xmath506 being an @xmath459-representation of @xmath505 . now we are ready to present the proof of theorem [ part_real_lin : theo1 ] . first , recall from definition [ part_real_pow : submatrix ] the definition of the matrix @xmath574 for some @xmath575 . it is easy to see that the matrix @xmath53 as defined in definition [ sect : main_result : sub - hank : def1 ] and the matrix @xmath574 coincide , in particular @xmath576 . recall from remark [ sect : real : hank ] that the matrices @xmath54 and @xmath577 coincide and hence their ranks are equal as well . we will prove the two statements of the theorem separately . * proof of existence and computability of a partial realization * + we will apply theorem [ part_real_pow : theo ] to @xmath505 . using the remark above , the condition ( [ part_real_lin : theo1:eq1 ] ) can be rewritten as ( [ part_theo : eq1 ] ) of theorem [ part_real_pow : theo ] . consider the @xmath525 representation @xmath578 defined in theorem [ part_real_pow : theo ] . since ( [ part_theo : eq1 ] ) holds , we get that @xmath579 is well - defined and it is a @xmath459-partial representation of @xmath505 , and @xmath579 is reachable and observable . consider algorithm [ alg0 ] and the isomorphism @xmath436 defined there . it is clear that @xmath436 maps the column space of @xmath580 to @xmath123 with @xmath435 . consider the isomorphic copy @xmath581 of @xmath582 via the morphism @xmath436 , i.e. @xmath583 . obviously , @xmath581 is also a @xmath459-partial representation of @xmath505 , and it is reachable and observable . consider now the lss realization @xmath584 associated with @xmath585 , as defined by construction [ sect : real : const2 ] . notice that if ( [ part_real_lin : theo1:eq1 ] ) holds , then ( [ alg0:eq1 ] ) of algorithm [ alg0 ] has a unique solution . indeed , ( [ alg0:eq1 ] ) can be rewritten as @xmath586 . here @xmath587 and for each @xmath588 , the column of @xmath480 indexed by @xmath386 is the image @xmath589 of the column of @xmath59 indexed by @xmath386 ; the column of @xmath590 indexed by @xmath386 is the image @xmath591 of the column of @xmath437 indexed by @xmath455 . the column space of @xmath480 is the column space of @xmath59 by @xmath436 , hence if ( [ part_real_lin : theo1:eq1 ] ) holds , then @xmath592 . that is , the solution of @xmath593 , if it exists , is unique . hence , if ( [ part_real_lin : theo1:eq1 ] ) holds , then the lss realization @xmath440 is uniquely defined by ( [ alg0:eq1]),([alg0:eq-1]),([alg0:eq-2 ] ) and ( [ alg0:eq-3 ] ) , if it exists . it is easy to see that the lss realization @xmath584 satisfies ( [ alg0:eq1]),([alg0:eq-1]),([alg0:eq-2 ] ) and ( [ alg0:eq-3 ] ) and hence it coincides with @xmath440 . in other words , the lss realization @xmath440 returned by algorithm [ alg0 ] equals the lss realization @xmath594 associated with the isomorphic copy @xmath581 of @xmath582 , i.e. @xmath595 by theorem [ part_real_lin : theo1/2 ] it follows then that @xmath440 is a @xmath459-partial realization of @xmath7 . in addition , from theorem [ sect : real : theo1 ] it follows that @xmath440 is semi - reachable and observable . * proof of existence of a complete realization * + in addition , from theorem [ part_real_pow : theo ] it follows that if @xmath463 , i.e. @xmath596 , then @xmath579 is a minimal representation of @xmath505 , and hence so is @xmath581 . but then by theorem [ sect : real : theo1 ] , @xmath597 is a minimal realization of @xmath505 . finally , assume that there exists a realization @xmath285 of @xmath7 , such that @xmath598 . consider the representation @xmath599 associated with @xmath285 as defined in construction [ sect : real : const1 ] . notice that @xmath600 . then by theorem [ sect : real : theo1 ] , @xmath480 is a representation of @xmath505 . applying theorem [ part_real_pow : theo ] we get that @xmath601 and hence @xmath602 holds , and that @xmath582 , and hence @xmath581 , is a minimal representation of @xmath505 . again , from theorem [ sect : real : theo1 ] it follows then that @xmath603 is a minimal realization of @xmath7 . next , we will present the proof of theorem [ part_real_lin : theo1.1 ] we will prove the statements of theorem [ part_real_lin : theo1.1 ] one by one . notice that if @xmath7 satisfies ( [ part_real_lin : theo1:eq1 ] ) , then the associated family of formal power series @xmath505 satisfies ( [ part_theo : eq1 ] ) . this allows us to use theorem [ part_real_pow : theo2 ] and theorem [ part_real_lin : theo1/2 ] to prove the statements of the theorem . * proof of part [ part_real_lin : theo1.1:res1 ] * + from the proof of theorem [ part_real_lin : theo1 ] it follows that the lss realization @xmath440 is the realization associated with an isomorphic copy of the representation @xmath582 . by the remark above the family @xmath505 associated with @xmath7 satisfies the condition of theorem [ part_real_pow : theo2 ] and hence by theorem [ part_real_pow : theo2 ] the representation @xmath582 is a minimal @xmath459-partial representation of @xmath505 . then theorem [ part_real_lin : theo1/2 ] implies that @xmath440 is a @xmath3-partial realization of @xmath7 and its dimension is @xmath604 . it is easy to see that @xmath440 is a minimal @xmath459-partial realization of @xmath7 . for assume that @xmath285 is a @xmath459-partial realization of @xmath7 . then by theorem [ part_real_lin : theo1/2 ] the associated representation @xmath506 is a @xmath459-partial representation of @xmath505 , and hence , @xmath605 . * proof of part [ part_real_lin : theo1.1:res2 ] * + if @xmath285 is a minimal @xmath459-partial realization of @xmath7 , then from theorem [ part_real_lin : theo1/2 ] it follows that @xmath506 is a minimal @xmath459-partial representation of @xmath505 . indeed , consider a @xmath459-partial representation @xmath480 of @xmath505 . using remark [ repr : state_space : rem ] , we can replace @xmath480 by a suitable isomorphic copy , state - space of which is @xmath123 with @xmath606 . this isomorphic copy will also be a @xmath459-partial representation of @xmath505 . hence , without loss of generality we can assume that the state - space of @xmath480 is @xmath123 for @xmath606 and hence construction [ sect : real : const2 ] can be applied . then it follows from theorem [ part_real_lin : theo1/2 ] that @xmath507 is a @xmath459-partial realization of @xmath7 . hence , @xmath607 by minimality of @xmath285 . but then by theorem [ part_real_pow : theo2 ] the representation @xmath506 has to be reachable , observable , and of dimension @xmath608 . by theorem [ sect : real : theo1 ] , the latter means that @xmath285 is semi - reachable , observable , and of dimension @xmath465 . * proof of part [ part_real_lin : theo1.1:res3 ] * + let @xmath285 and @xmath298 be two minimal @xmath459-partial realizations of @xmath7 . using an argument analogous to the one presented in the proof of part [ part_real_lin : theo1.1:res2 ] , one can show that the associated representations @xmath506 and @xmath609 are minimal @xmath459-partial representations of the family @xmath505 of formal power series associated with @xmath7 . from theorem [ part_real_pow : theo2 ] it follows that @xmath506 and @xmath609 are isomorphic . from theorem [ sect : real : theo1 ] it then follows that @xmath285 and @xmath298 are isomorphic . we will continue with the proof of theorem [ part_real_lin : theo2 ] the proof of the theorem relies on the following observation . assume that algorithm [ partrepr ] is applied to @xmath505 . if we use the same factorization algorithm for factorizing the matrix @xmath610 in both algorithm [ partreal ] and algorithm [ partrepr ] , then the following holds . the equation ( [ computepartialreal ] ) has a unique solution if and only if ( [ computepartialrepr ] ) has a unique solution . in addition , algorithm [ partreal ] returns a lss realization if and only if algorithm [ partrepr ] applied to @xmath610 returns a rational representation @xmath611 , and the realization @xmath612 returned by algorithm [ partreal ] is the lss realization associated with the representation @xmath611 , as defined in construction [ sect : real : const2 ] . that is , @xmath613 . * proof of part [ part_real_lin : theo2:res1 ] * + recall that for all @xmath614 , the matrix @xmath421 is identical to @xmath615 , and that the hankel - matrix @xmath54 is identical to @xmath524 . hence , by applying theorem [ sect : svd : theo1 ] to @xmath616 we get that if ( [ part_real_lin : theo1:eq1 ] ) holds , then algorithm [ partrepr ] returns a @xmath459-partial representation @xmath611 of @xmath505 . then from the discussion above it follows that algorithm [ partreal ] returns a realization @xmath495 and @xmath617 . then from theorem [ part_real_lin : theo1/2 ] it follows that @xmath495 is a @xmath459-partial realization of @xmath7 . moreover , theorem [ sect : svd : theo1 ] implies that in this case @xmath611 is isomorphic to the representation @xmath579 from theorem [ part_real_pow : theo ] . recall from the proof of theorem [ part_real_lin : theo1 ] that the realization @xmath440 from theorem [ part_real_lin : theo1 ] is the realization associated with the isomorphic copy @xmath581 of the representation @xmath579 , as defined in construction [ sect : real : const2 ] . since @xmath582 and @xmath611 are isomorphic , the representations @xmath581 and @xmath611 are isomorphic as well . then by theorem [ sect : real : theo1 ] the realizations @xmath612 and @xmath440 are isomorphic as well . * proof of part [ part_real_lin : theo2:res2 ] * + if @xmath602 , then by theorem [ part_real_lin : theo1 ] , ( [ part_real_lin : theo1:eq1 ] ) holds and @xmath440 is a minimal realization of @xmath7 . similarly , if there exists a lss realization @xmath285 of @xmath7 such that @xmath598 , then by theorem [ part_real_lin : theo1 ] , @xmath602 , and ( [ part_real_lin : theo1:eq1 ] ) holds , and @xmath440 is a minimal realization of @xmath7 . by the preceding argument , in both cases algorithm [ partreal ] returns a realization @xmath618 and @xmath618 is isomorphic to @xmath440 . hence , in both cases the realization @xmath618 returned by algorithm [ partreal ] is a minimal realization of @xmath7 . in this paper we presented basic results on partial - realization theory lsss . we also discussed the algorithmic aspects of the theory and presented two algorithms for computing a partial lss realization . we have also shown that under suitable conditions , the obtained partial realizations are in fact ( complete ) minimal realizations of the input - output maps at hand . the paper represents only the first steps towards partial - realization theory of hybrid systems . a number of problems remains open . in particular , we would like to extend the results of the current paper to other classes of hybrid systems . another important research direction is the improvement of the computational complexity of the presented algorithms . as we mentioned it earlier , the results of the paper can potentially be useful for the solution of a number of problems , namely , for system identification , model reduction and possibly fault detection and computer vision . below we will discuss these potential application domains in more detail . * systems identification algorithms * + as the history of development of system identification demonstrates , partial - realization theory can be used to design algorithms for system identification . in particular , the famous subspace identification algorithms @xcite rely heavily on partial - realization theory . in fact , based on the results of the paper one can derive the following crude identification algorithm . consider a finite family @xmath619 of input - output maps . we would like to find a lss realization of @xmath7 . fix a natural number @xmath3 . the number @xmath3 may be chosen either based on our belief on the dimension of a potential lss realization of @xmath7 , or our ability / willingness to make measurements of the system responses for switching sequences up to length @xmath459 . for each @xmath13 , estimate the first - order time derivatives of @xmath620 and @xmath621 for all sequences of discrete modes @xmath190 of length at most @xmath459 and all discrete modes @xmath371 . use these estimates to construct the markov - parameters of order up to @xmath459 . subsequently , use the thus obtained markov - parameters to construct the hankel - matrices @xmath59 , @xmath466 and @xmath437 . if @xmath622 , then compute the @xmath459 partial realization @xmath623 either by algorithm [ alg0 ] or algorithm [ partreal ] . it follows immediately from theorem [ part_real_lin : theo1 ] and theorem [ part_real_lin : theo2 ] that if @xmath7 can indeed be realized by a lss , then for large enough @xmath3 the procedure above will return a minimal lss realization of @xmath7 . in fact , if @xmath491 is the dimension of a potential lss realization of @xmath7 , then the algorithm above always returns a minimal lss realization of @xmath7 . the algorithm described above has several drawbacks . for example , it is unclear how to obtain in practice the derivatives needed to compute the generalized markov - parameters . another problem that the algorithm does not take into account measurement noise . the authors are hopeful that these shortcoming can be overcome , as similar problems were successfully dealt with in the case of linear systems . however , we would like to remark that turning partial - realization theory to practically usable identification algorithms might take considerable amount of time ; in case of linear systems it took several decades . * spaces of lsss * + partial realization theory can potentially be useful for studying the geometry and topology of the space of lsss . in turn , the latter is useful for parametric system identification , as it helps to formalize the well - known paradigm of system identification : `` find a model of a certain class which matches the measured data best '' . in addition , the geometric insights and the resulting notions of distance between systems can be useful for fault detection and computer vision @xcite . for bilinear systems , the first steps were made in @xcite . for linear systems , the topology and geometry of hankel - matrices were investigated in @xcite . note that spaces of linear systems were also studied using a different approach , namely , by looking at equivalence classes of minimal linear systems under algebraic similarity . this approach resulted in deep results and insights in the geometry of spaces of systems and important algorithms for parametric identification , see @xcite . * model reduction * + the presented results on partial - realization theory can also be used for model reduction . recall from @xcite the moment matching approach to model reduction of linear systems . the core of this approach is to approximate a high - order to system with a lower order one , such that the lower order system is a a partial realization of a finite number of markov - parameter of the original system . for lsss the role of markov - parameters ( or moments ) is played by the generalized markov - parameters . hence , one could try to extend the moment matching approach to lsss and replace a high - order system with a lower order one which is a partial - realization of a finite sequence of generalized markov - parameters . of course , a great number of issues needs to be solved . in particular , obtaining good numerical algorithms will be a challenge . * extension to more general classes of hybrid systems * + we believe that partial - realization theory for lsss can be useful for obtaining realization theory , system identification and model reduction for piecewise - affine hybrid systems with guards . one reason for this is that a piecewise - affine hybrid system with guards can be viewed as a feedback interconnection of a lss with an event generator . the former is just the collection of the continuous subsystems of the piecewise - affine hybrid system , the latter is constructed from the guards of the piecewise - affine hybrid system . understanding the realization theory of one of the components of this feedback loop should help to understand the realization theory of piecewise - affine hybrid systems . the second reason is that lsss form a subclass of piecewise - affine hybrid systems . by identifying each lsss with a piecewise - affine hybrid system whose guards depend only on the choice of inputs . hence , any result on ( partial- ) realization theory of piecewise - affine hybrid systems should be consistent with the corresponding results for lsss . * acknowledgment * the authors thank pieter collins and luc habets for useful discussions and suggestions . 10 athanasios c. antoulas and dan c. sorensen . approximation of large - scale dynamical systems : an overview . 11(5):10931121 , 2001 . numerical analysis and systems theory ( perpignan , 2000 ) . j. berstel and c. reutenauer . . springer - verlag , 1984 . roger w. brockett . the geometry of the partial realization problem . , 17:10481052 , jan . 1978 . christopher i. byrnes and norman e. hurt . on the moduli of linear dynamical systems . 4:83122 , 1979 . christopher i. byrnes and anders lindquist . the stability and instability of partial realizations . , 2(2):99105 , 1982/83 . frank m. callier and charles a. desoer . . springer - verlag , 1991 . h. chen and j.m . new subspace identification method for bilinea systems . technical report cued / finfeng / tr.337 , department of engineering , university of cambridge , 2000 . k. de cock and b. de moor . subspace angles and distances between arma models . in _ proc . of the int . symposium of math . theory of networks and systems _ , 2000 . gianfranco doretto , alessandro chiuso , ying nian wu , and stefano soatto . dynamic textures . , 51(2):91109 , 2003 . samuel eilenberg . . academic press , new york , london , 1974 . w. favoreel , b. de moor , and p. van overschee . subspace identification of bilinear systems subject to white inputs . technical report tr 1996 - 531 , esat , katholieke univesiteit leuven , 1996 . m. fliess . matrices de hankel . , ( 23):197 224 , 1973 . m. fliess . functionnelles causales non linaires et indtermines non commutatives . , ( 109):2 40 , 1981 . f. gcseg and i pek . . akadmiai kiad , budapest , 1972 . dieter gollmann . partial realization by discrete - time internally bilinear systems : an algorithm . in _ mathematical theory of networks and systems , proc . , beer sheva / isr . 1983 _ , 1983 . william b. gragg and anders lindquist . on the partial realization problem . , 50:277319 , 1983 . e. j. grimme . . phd thesis , univ . illinois , urbana champaign , 1997 . e. j. grimme , d. c. sorensen , and p. van dooren . model reduction of state space systems via an implicitly restarted lanczos method . , 12(1 - 2):131 , 1996 . b. hanzon . riemannian geometry on families of linear systems , the deterministic case . technical report 88 - 62 , delft university of technology , faculty of mathematics and informatics , 1988 . b. hanzon . , volume 63 of _ cwi tract_. cwi , amsterdam , 1989 . b. hanzon . on the differentiable manifold of fixed order stable linear systems . , 13:345 352 , 1989 . m. hazewinkel . moduli and canonical forms for linear dynamical systems iii : the algebraic - geometric case . in c. martin and r. hermann , editors , _ the 1976 ames research center ( nasa ) conference on geometric control theory _ , pages 229276 . brookline , mass . : math sci press 1977 . m. hazewinkel . moduli and canonical forms for linear dynamical systems ii : the topological case . , 10:363385 , 1977 . u. helmke and d. hinrichsen . canonical forms and orbit spaces of linear systems . , 3:167 184 , 1986 . u. helmke , d. hinrichsen , and w. manthey . a cell decomposition of the space of real hankel matrices of rank @xmath624 and some applications . , 122/123/124:331355 , 1989 . d. hinrichsen , w. manthey , and d. prtzel - wolters . the bruhat decomposition of finite hankel matrices . , 7(3):173182 , 1986 . diederich hinrichsen and wilfried manthey . on a cell decomposition for hankel matrices and rational functions . , 451:1550 , 1994 . b. l. ho and r. e. kalman . effective construction of linear state - variable models from input / output data . in _ proc . third annual allerton conf . on circuit and system theory _ , pages 449459 . illinois , urbana , ill . , 1965 . a. isidori . direct construction of minimal bilinear realizations from nonlinear input - output maps . , pages 626631 , 1973 . alberto isidori . . springer verlag , 1989 . alberto isidori , paolo dalessandro , and antonio ruberti . realization and structure theory of bilinear dynamical systems . , 12(3 ) , 1974 . nathan jacobson . , volume ii : linear algebra . d. van nostrand company , inc . new york , 1953 . r. e. kalman , p. l. falb , and m. a. arbib . . mcgraw - hill book co. , new york , 1969 . w. kuich and a. salomaa . . springer - verlag , 1986 . daniel liberzon . . birkhuser , boston , 2003 . wilfried manthey and uwe helmke . bruhat canonical form for linear systems . , 425(2 - 3):261282 , 2007 . r. martin . a metric for arma processes . , 48(4):11641170 , 2000 . ralf peeters . . phd thesis , free university , amsterdam , 1994 . mihaly petreczky . realization theory for bilinear hybrid systems . in _ in proceedings 11th ieee international conference on methods and models in automation and robotics _ , mihaly petreczky . realization theory for bilinear switched systems . in _ proc . of 44th ieee conference on decision and control _ , 2005 . mihaly petreczky . realization theory for linear and bilinear hybrid systems . technical report mas - r0502 , cwi , 2005 . mihaly petreczky . realization theory of linear and bilinear switched systems : a formal power series approach . technical report mas - r0403 , cwi , 2005 . to appear in esaim control , optimization and calculus of variations , doi 10.1051/cocv/2010014 and doi 10.1051/cocv/2010015 . mihly petreczky . hybrid formal power series and their application to realization theory of hybrid systems . in _ proceedings 17th international symposium on mathematical theory of networks and systems _ , 2006 . mihaly petreczky . . phd thesis , vrije universiteit , amsterdam , 2006 . available at ` http://www.cwi.nl/~mpetrec ` . mihaly petreczky . realization theory for linear switched systems : formal power series approach . , 56(9 10):588595 , 2007 . mihly petreczky and ren vidal . metrics and topology for nonlinear and hybrid systems . in _ hybrid systems : computation and control _ , pages 459472 , 2007 . mihaly petreczky and rene vidal . realization theory of stochastic jump - markov linear systems . in _ proceedings 46th ieee conference on decision and control _ , 2007 . d. eduardo sontag . a remark on bilinear systems and moduli spaces of instantons . , 9(5):361367 , 1987 . eduardo d. sontag . , volume 13 of _ lecture notes in control and information sciences_. springer verlag , 1979 . eduardo d. sontag . realization theory of discrete - time nonlinear systems : part i the bounded case . , cas-26(4 ) , april 1979 . zhendong sun and shuzhi s. ge . . springer , london , 2005 . a. tether . construction of minimal linear state - variable models from finite input - output data . , 15(4):427 436 , 1970 . van overschee and b. de moor . . kluwer academic publishers , 1996 . s. v. n. vishwanathan , alexander j. smola , and ren vidal . binet - cauchy kernels on dynamical systems and its application to the analysis of dynamic scenes . , 73(1):95119 , 2007 . the section recalls basic results on formal power series . the material of this section is based on the classical theory of formal power series , see @xcite . however , a number of concepts and results are extensions of the standard ones to families of formal power series . the outline of the section is the following . in subsection [ sect : pow : def ] we present the basic concepts for formal power series . in subsection [ sect : pow : rat_repr ] we present the notion of a rational representation of a family of formal power series . in subsection [ sect : pow : main ] we state the main result on existence and minimality of a rational representation of a family of formal power series . let @xmath185 be a finite set , which we will refer to as the alphabet . recall from section [ sect : prelim : lang ] the notion of a finite word over an alphabet and the related notation . a _ formal power series _ @xmath5 with coefficients in @xmath232 is a map @xmath625 there are many ways to give an intuition for the definition of a formal power series . for the purposes of this paper , the most suitable one is to think of a formal power series as an output of a machine defined as follows . the machine reads symbols belonging to @xmath185 from its input tape and writes elements of @xmath232 onto its output tape . we denote by @xmath626 the set of all formal power series with coefficients in @xmath232 . the set of all formal power series over @xmath185 with coefficients in @xmath232 forms a vector space with respect to point - wise addition and multiplication . that is , if @xmath627 are two scalars and @xmath628 are two formal power series , then the linear combination @xmath629 is defined as the formal power series assigning each word @xmath197 the value @xmath630 . in the sequel we will mostly be interested in _ families of formal power series_. let @xmath127 be an arbitrary ( possibly infinite ) set . family of formal power series in @xmath631 indexed by @xmath127 _ is simply a collection @xmath632 of formal power series from @xmath633 indexed by elements of @xmath127 . notice that the definition above does not require @xmath634 , @xmath132 to be all distinct formal power series , i.e. @xmath635 for some indices @xmath636 , @xmath637 is allowed . one can think of a family of formal power series as a family of input - output maps of the machine ( discrete - time system ) described above , realized from a set of initial states indexed by elements of @xmath127 . we would like to point out that the notion of ( family of ) formal power series is a purely formal one , the interpretation of them as input - output maps is just one of the many possible interpretations . above we have defined the notion of a formal power series and a family of formal power series and we have related these notions to input - output maps of some systems . below we will recall the notion of a _ rational representation _ , which can be thought of as a special subclass of these systems . we will also define morphisms between rational representations along with a notion of observability and reachability . [ sect : pow : rat_rep : def0 ] let @xmath127 be an arbitrary set and let @xmath638 . a _ rational representation of type @xmath451-@xmath127 over the alphabet @xmath185 _ is a tuple @xmath639 where * the space @xmath102 is a finite dimensional vector space over @xmath106 , called _ state - space _ of @xmath480 , * for each letter @xmath640 , @xmath641 is a linear map , referred to as the _ state - transition map _ * the map @xmath642 is a linear map , referred to as the _ readout map _ , * the family @xmath643 is a collection of ( not necessarily distinct ) elements of @xmath102 indexed by @xmath127 . if @xmath451 and @xmath127 are clear from the context , then we will refer to @xmath480 simply as a _ rational representation_. the dimension @xmath644 of the state - space is called the _ dimension _ of the representation @xmath480 and it is denoted by @xmath645 . notice that if a basis of @xmath102 is fixed and @xmath646 , then the state - transition maps @xmath647 , @xmath640 and the readout map @xmath648 can be identified with their matrix representations in this basis , and for each @xmath132 , @xmath649 can be identified with the @xmath123 column vector of its coordinates . if @xmath220 , then we will identify the linear maps @xmath647 , @xmath640 and @xmath648 with their matrix representations in the standard orthogonal basis of @xmath123 . in this case we will call them the _ state - transition matrices _ and the _ readout matrix _ respectively . [ sect : pow : rat_rep : def-1 ] let @xmath650 be a family of formal power series indexed by @xmath127 . the representation @xmath480 from ( [ repr : def0 ] ) is said to be a _ representation of @xmath651 _ , if for each index @xmath132 , for all sequences @xmath652 , @xmath241 , @xmath653 we will say that a family of formal power series @xmath632 is _ rational _ , if there exists a representation @xmath480 such that @xmath480 is a representation of @xmath651 . [ repr : not1 ] the following notation will greatly simplify the expressions used for rational representations . let @xmath654 be linear maps and let @xmath655 , @xmath656 , @xmath241 , be a word over @xmath185 . then @xmath657 denotes the composition of the linear maps @xmath658 in that order , that is @xmath659 if @xmath208 is the empty words , then @xmath660 is taken to be the identity map . with the notation above , ( [ repr : def1 ] ) can be rewritten as @xmath661 for all @xmath197 , @xmath132 . notice that if @xmath220 for some @xmath662 , and hence the state - transition maps @xmath647 can be viewed as matrices , then ( [ repr : eq2 ] ) defines a notation for products of the state - transition matrices taken along the word @xmath190 . a representation @xmath663 of @xmath651 is called _ minimal _ , if for each representation @xmath480 of @xmath651 , @xmath664 , i.e. @xmath663 is a rational representation of @xmath651 with the smallest possible state - space dimension . we will continue with presenting the notions of _ observability _ and _ reachability _ for rational representations . define the subspaces @xmath665 and @xmath666 of @xmath102 by @xmath667 that is , the subspace @xmath665 is the linear span of the elements of the state - space of the form @xmath668 , where @xmath190 runs through all words over @xmath185 and @xmath10 runs through all the indices in @xmath127 . the space @xmath666 is the intersection of the null - spaces ( kernels ) of all the linear maps @xmath669 , where @xmath190 runs through the set of all words @xmath186 . [ reach : rep : def ] we will say that the representation @xmath480 is _ reachable _ if @xmath670 . the subspace @xmath665 will be referred to as the _ reachability subspace of @xmath480_. [ obs : rep : def ] we will say that @xmath480 is _ observable _ if @xmath671 , i.e. @xmath666 consists of the zero element only . the subspace @xmath666 will be referred to as the _ observability subspace of @xmath480_. next , we define the notion of morphism between rational representations . let @xmath672 , @xmath673 be two @xmath451-@xmath127 rational representations . a linear map @xmath674 is called a _ representation morphism _ from @xmath480 to @xmath675 and is denoted by @xmath676 if @xmath436 commutes with @xmath647 , @xmath649 and @xmath648 for all @xmath132 , @xmath640 , that is , if the following equalities hold @xmath677 the representation morphism @xmath436 is called _ surjective , injective , isomorphism _ if @xmath436 is a surjective , injective or isomorphism respectively if viewed as a linear map . [ repr : state_space : rem ] if @xmath678 is a representation of @xmath651 , then for any vector space isomorphism @xmath679 , @xmath606 , the tuple @xmath680 where @xmath681 is also a representation of @xmath651 . moreover , for all @xmath640 , @xmath682 can be naturally viewed as a matrix by taking its matrix representation with respect to the natural basis of @xmath123 . similarly , by taking matrix and vector representations of @xmath683 and @xmath684 , @xmath132 in the natural basis of @xmath123 , we can view @xmath683 and @xmath684 , @xmath132 as a @xmath685 matrix and @xmath686 vector respectively . moreover , @xmath687 is a representation isomorphism . that is , we can always replace a representation of @xmath651 with an isomorphic representation , state - space of which is @xmath123 for some @xmath138 , and the parameters of which are matrices and real vectors , as opposed to linear maps and elements of abstract vector spaces . moreover , isomorphisms clearly preserve such properties as observability , reachability , and minimality . the purpose of the section is to state the main results on existence and minimality of representations of families of rational formal power series . we will start by stating the main result on the existence of a rational representation . however , in order to state the main theorem , we need to define the concept of the _ hankel matrix _ of a family of formal power series . let @xmath632 be a family of formal power series . recall the notation of section [ sect : prelim : not_matrix ] . [ sect : pow : hank : def0 ] define the _ hankel - matrix _ of @xmath651 as the following infinite matrix @xmath688 . the rows of @xmath688 are indexed by pairs @xmath383 where @xmath198 is an arbitrary word and @xmath689 . the columns of @xmath688 are indexed by pairs @xmath386 where @xmath197 is a word over @xmath185 and @xmath132 runs through the elements of @xmath127 . the set of all real infinite matrices with the indexing the rows and columns described above will be denoted by @xmath690 where @xmath691 . hence , we can write @xmath692 . the entry of @xmath688 lying on the intersection of the row indexed by @xmath383 and the column indexed by @xmath386 is defined as @xmath693 where @xmath694 denotes the @xmath695 entry of the column vector @xmath696 . according to the convention adopted in section [ sect : prelim : not_matrix ] , we define the _ rank of @xmath688 , denoted by @xmath697 _ , as the dimension of the vector space spanned by the columns of @xmath688 . with the notation above the following theorem holds . [ sect : form : theo1 ] the family @xmath651 is rational , i.e. admits a rational representation , if and only if @xmath698 , i.e. the rank of the hankel - matrix @xmath688 is finite . the proof of the above theorem can be found in @xcite . as the next step we will present below the main result on minimality of rational representations . [ sect : form : theo3 ] let @xmath699 be a family of formal power series . the following are equivalent . * @xmath663 is a minimal representation of @xmath651 . * @xmath663 is reachable and observable . * if @xmath480 is a reachable representation of @xmath651 , then there exists a surjective representation morphism @xmath700 . * @xmath701 . in addition , all minimal representations of @xmath651 are isomorphic . the proof of the above theorem can be found in @xcite . the counterpart of the above two theorems , i.e. theorem [ sect : form : theo1 ] and [ sect : form : theo3 ] , for a single formal power series is a classical result , see @xcite . below we will formulate and solve the counterpart of the partial realization problem for families of formal power series . we use the obtained result to derive a solution to the partial realization problem for linear switched systems formulated in problem [ prob : part : real ] . we start with defining the notion of a @xmath3-partial representation . let @xmath632 be a family of formal power series indexed by @xmath127 . a @xmath127-@xmath451 representation @xmath702 , with @xmath643 is said to be an _ @xmath3-partial representation of @xmath651 _ if for each index @xmath132 and each word @xmath197 of length at most @xmath3 , i.e. @xmath703 , @xmath704 that is , if @xmath480 is a @xmath3-partial representation of @xmath651 , then @xmath480 recreates the values of the elements of @xmath651 for all the words of length at most @xmath3 . now we are ready to formulate the partial - realization problem for formal power series . * find conditions for existence of a @xmath3-partial representation of @xmath651 . formulate an algorithm for computing a @xmath3-partial representation of @xmath651 . * characterize minimal @xmath3-partial representations of @xmath651 , their existence and uniqueness . * find conditions under which the @xmath3-partial representation above becomes a rational representation of @xmath651 in the sense of definition [ sect : pow : rat_rep : def0 ] . we will devote the rest of the section to solving the problem described above . the outline of the section is the following . subsection [ part : form : main_results ] presents the main results on partial - realization theory of families of formal power series , along with a kalman - ho - like algorithm for computing a minimal partial representation . subsection [ app : proof ] presents the proof of the results presented in subsection [ part : form : main_results ] . throughout the section , @xmath651 will denote the family of formal power series @xmath705 . the goal of the section is to present the main results on partial - realization theory of families of formal power series . note that partial - realization theory of a single formal power series is more or less equivalent to partial - realization theory of bilinear and state - affine systems . the latter was already investigated in @xcite . the results to be presented below represent an extension of the results of @xcite . in addition , here we state partial - realization theory directly for formal power series . this is in contrast to @xcite , where the partial - realization problem and solution were stated in terms of state - affine or bilinear systems . the outline of the section is the following . subsection [ part : form : theo ] presents the statement of the main results on partial realization theory for formal power series . subsection [ part : form : svd ] presents an algorithm for computing a partial representation , which is similar to the well - known kalman - ho algorithm for the linear systems . we will start with defining the upper - left block matrix of @xmath688 which is indexed by words over @xmath185 of finite length . more precisely , fix natural numbers @xmath706 and define the following sets @xmath707 intuitively , @xmath708 is a subset of the row indices of the hankel - matrix @xmath688 , made up of indices of the form @xmath383 , where @xmath203 runs through all the words of length at most @xmath128 . similarly , the set @xmath709 is a subset of column indices of @xmath688 , made up of column indices @xmath386 where @xmath190 runs through all the words of length at most @xmath56 . the set @xmath709 is the set of columns indices of the matrix to be defined , and @xmath710 is the set of row indices of this matrix . it turns out that under certain circumstances partial representations not only exist but they also yield a minimal representation of the whole family of formal power series . moreover , such partial representations can be constructed from finite data . 1 . [ part_theo : part1 ] if for some @xmath718 , @xmath719 then there exists a @xmath459-partial representation @xmath579 of @xmath651 , of the form @xmath720 where the parameters of @xmath579 are defined as follows . for each word @xmath197 , @xmath721 , and each index @xmath132 denote by @xmath722 the column of @xmath723 indexed by @xmath386 . with this notation , * @xmath724 denotes the linear space spanned by the columns of @xmath723 . * for each @xmath640 , the linear map @xmath725 has the property that for each word @xmath197 , @xmath703 and each index @xmath132 @xmath726 i.e. @xmath647 maps the column indexed by @xmath727 to the column indexed by @xmath728 . * for each @xmath197 , @xmath721 and @xmath132 , the linear map @xmath729 satisfies , @xmath730 that is , @xmath648 maps each column to the vector in @xmath232 formed by the first @xmath451 entries associated with @xmath205 of the column indexed by @xmath386 . * the set @xmath731 is defined by @xmath732 i.e. @xmath649 is simply the column of @xmath723 indexed by @xmath733 . + in addition , the representation @xmath579 is reachable and observable . [ part_theo : part2 ] if for some @xmath499 @xmath734 then ( [ part_theo : eq1 ] ) holds and the representation @xmath579 from ( [ part_theo : eq3 ] ) is a minimal representation of @xmath651 . [ part_theo : part3 ] if @xmath480 is a representation of @xmath651 , @xmath735 , then ( [ part_theo : eq2 ] ) holds for @xmath3 and the representation @xmath579 from ( [ part_theo : eq3 ] ) exists and it is a minimal representation of @xmath651 . the proof of theorem [ part_real_pow : theo ] is presented in appendix [ app : proof ] . using the results of theorem [ part_real_pow : theo ] we can state the following characterization of minimal partial representations of @xmath651 . 1 . [ part_real_pow : theo2:res1 ] a minimal @xmath459 partial representation of @xmath651 exists , in fact , the representation @xmath582 of theorem [ part_real_pow : theo ] is a minimal @xmath459 partial representation of @xmath651 . [ part_real_pow : theo2:res2 ] any minimal @xmath459 partial representation of @xmath651 is reachable and observable and it is of dimension @xmath736 . [ part_real_pow : theo2:res3 ] all minimal @xmath459 partial representations of @xmath651 are isomorphic . the reader might wonder why we speak of @xmath459-partial representations in theorem [ part_real_pow : theo ] and [ part_real_pow : theo2 ] . the reason behind it is that the finite hankel - matrix @xmath723 is formed by values of the formal power series from @xmath651 for words of length at most @xmath459 . we would like the representation obtained from @xmath723 to recreate at least the entries of the matrix @xmath723 . but this means precisely that the representation obtained from @xmath723 should be a @xmath459-partial representation of @xmath651 . in this section we present an algorithm , described in algorithm [ partrepr ] , which computes a partial representation by factorizing the hankel - matrix . the technique of hankel - matrix factorization has been used in realization theory and systems identification for several decades . it forms the theoretical basis of algorithms for subspace identification , see for example @xcite . throughout the section , @xmath651 stands for the family of formal power series @xmath705 . define the family of vectors @xmath745 , where for each @xmath132 , @xmath746 where @xmath747 stands for the column of @xmath480 indexed by @xmath733 . for each @xmath640 let @xmath748 be the solution of @xmath749 where @xmath750 are matrices of the form @xmath751 for all @xmath752 . if there no unique solution to ( [ computepartialrepr ] ) then return @xmath753 . otherwise return @xmath754 algorithm [ partrepr ] represents an algorithm based on matrix factorization of @xmath737 . in addition , if @xmath3 is large enough , algorithm [ partrepr ] in fact yields a minimal representation of @xmath651 . algorithm [ partrepr ] above may return two different types of data . it returns a rational representation if ( [ computepartialrepr ] ) has a unique solution , and the symbol @xmath753 otherwise . in step [ alg : svd : step1 ] of algorithm [ partrepr ] above one can use any algorithm for computing a factorization . for example , one could use svd decomposition , in which case @xmath755 , and @xmath756 , @xmath757 is a valid choice for decomposition . 1 . [ sect : svd : theo1:part1 ] assume that form some @xmath25 @xmath759 then algorithm [ partrepr ] always returns a formal power series representation @xmath611 and @xmath611 is an @xmath459-partial representation of @xmath651 . in fact , the representation @xmath579 from theorem [ part_real_pow : theo ] and @xmath611 are isomorphic . hence , @xmath760 is a minimal @xmath459-partial representation of @xmath651 , it is reachable and observable . [ sect : svd : theo1:part2 ] if for some @xmath499 , @xmath761 then ( [ sect : svd : theo1:eq1 ] ) holds , and algorithm [ partrepr ] returns a minimal representation @xmath611 of @xmath651 . [ sect : svd : theo1:part3 ] assume @xmath762 , or , equivalently , there exists a representation @xmath480 of @xmath651 , such that @xmath735 . then ( [ sect : svd : theo1:eq2 ] ) holds and the representation @xmath611 returned by algorithm [ partrepr ] is a minimal representation of @xmath651 . although solution to ( [ computepartialrepr ] ) need not always exist , one can always take a matrix @xmath763 , @xmath640 as a solution to the minimization problem @xmath764 . then @xmath763 can be obtained using standard numerical techniques for solving approximation problems . with this modification , the algorithm can be applied even if ( [ sect : svd : theo1:eq1 ] ) fails . however , the representation returned by the modified algorithm need not be a @xmath459-partial representation in this case . the goal of the section is to present the proof of theorem [ part_real_pow : theo ] , [ part_real_pow : theo2 ] and [ sect : svd : theo1 ] . in subsection [ part_pow_pf : prelim ] we will introduce some notation and state some preliminary results , which will be needed for the proof of the theorems . in subsection [ part_pow_pf : pf ] we will present the proof of theorem [ part_real_pow : theo][sect : svd : theo1 ] . finally , in subsection [ part_pow_pf : aux ] we will present the proof of the technical results which are used for the proof of the theorems [ part_real_pow : theo][sect : svd : theo1 ] . to begin with , for the purposes of partial realization theory , we need to recall some basic steps of the proof of theorem [ sect : form : theo1 ] . to this end , we have to introduce additional notation and terminology . let @xmath197 be a word over @xmath186 and let @xmath765 be a formal power series . define the formal power series @xmath766 , called the _ left shift of @xmath5 by @xmath190 _ , as follows ; we require that for all @xmath198 the value of @xmath767 at @xmath203 is as follows @xmath768 i.e. the value of @xmath767 at @xmath203 equals the value of @xmath5 at @xmath769 . notice that for any word @xmath197 of the form @xmath770 , @xmath771 and for any formal power series @xmath772 , the following equality holds @xmath773 moreover , notice that the shift operation is linear , that is , for any @xmath774 , and for any scalars @xmath177 , and for any word @xmath197 , @xmath775 . [ sect : form : def1 ] let @xmath632 be a family of formal power series . define the _ the smallest shift invariant linear space containing all the elements of @xmath651 _ , denoted by @xmath776 , as the following subspace of @xmath633 , @xmath777 that is , @xmath776 is composed of the linear combinations of all formal power series of the form @xmath778 for some @xmath132 and @xmath197 . [ sect : form : def1:rem1 ] it is easy to see that there is one - to - one correspondence between the formal power series @xmath778 and the column of @xmath688 indexed by @xmath386 for any word @xmath197 and index @xmath132 . in particular , it follows that @xmath776 is isomorphic to the span of columns of @xmath688 and hence @xmath779 [ sect : form : theo1:lemma1 ] assume that @xmath780 holds . then a representation @xmath781 of @xmath651 is given by @xmath782 where for each letter @xmath640 , the map @xmath783 is defined as the shift by @xmath784 , i.e. for each @xmath785 , @xmath786 ; the collection @xmath787 is formed by the elements of @xmath651 , that is @xmath788 for each @xmath132 ; the linear map @xmath789 is defined as the evaluation at the empty word , i.e. @xmath790 for all @xmath785 . using the lemmas above the proof of theorem [ sect : form : theo1 ] becomes trivial . for the sake of completeness we present it below . from remark [ sect : form : def1:rem1 ] it follows that @xmath792 . if @xmath793 , then lemma [ sect : form : theo1:lemma1 ] implies that @xmath781 is a well - defined representation of @xmath651 , hence @xmath651 is rational . conversely , if @xmath651 is rational then lemma [ sect : form : theo1:lemma2 ] implies that @xmath794 . we will start with defining the space @xmath797 of formal power series which corresponds to the subset of columns of @xmath688 indexed by indices of the form @xmath386 with @xmath132 and @xmath798 define the vector space @xmath797 as the sub - space of @xmath776 spanned by all formal power series of the form @xmath778 with @xmath132 and @xmath798 , i.e. @xmath799 _ notice that if @xmath127 is finite , then @xmath797 is generated by finitely many elements_. the motivation for considering the space @xmath797 is revealed by the following lemma and its corollary . [ part_pow : lemma1 ] assume that @xmath800 . for any formal power series @xmath785 and any word @xmath197 , there exists scalars @xmath801 , and words @xmath802 such that @xmath803 that is , the shift of @xmath84 by @xmath190 is a linear combination of the shifts of @xmath84 by words of the length at most @xmath804 . the proof of corollary [ part_pow : pf1:col1 ] will be presented in subsection [ part_pow_pf : aux ] . in other words , if @xmath3 is big enough , then @xmath806 generates the whole space @xmath776 . alternatively , stated in the language of hankel - matrices , the columns of @xmath688 indexed by words of length at most @xmath3 span the whole image of @xmath688 . although @xmath797 is generated by finitely many elements , if @xmath127 is finite , its elements are formal power series which contain infinite amount of data . in order to replace @xmath797 by a space generated by finite vectors , we proceed as follows . we will define a linear space @xmath807 which plays a similar role for @xmath808 as @xmath776 does for @xmath688 . similarly to @xmath776 , the elements of @xmath807 are maps mapping words over @xmath185 to vectors in @xmath232 . however , in contrast to @xmath776 , the elements of @xmath807 will be defined only on the words of length at most @xmath128 . similarly to @xmath776 , the space @xmath807 will allow us to prove our main results for the finite hankel sub - matrices @xmath808 in a more intuitive way . as the next step , define the map @xmath812 projecting any formal power series to its restriction to @xmath813 , i.e. for each formal power series @xmath772 , @xmath814 it is easy to see that @xmath815 is a surjective linear map . define the vector space @xmath816 by @xmath817 it is easy to see that @xmath807 is the image of @xmath797 by @xmath815 . the relationship between @xmath807 and @xmath808 can be best described as follows ; the generator set of @xmath807 is formed by the columns of @xmath808 . more precisely , recall from subsection [ sect : prelim : not_matrix ] that @xmath818 denotes the vector space of all maps from @xmath710 to @xmath106 . define the map @xmath819 by @xmath820 where @xmath821 stands for the @xmath394th entry of the vector @xmath822 . it is easy to see that @xmath823 is a linear isomorphism . moreover , @xmath824 . in particular , finally , we will show that if @xmath651 is rational and @xmath128 is big enough , then the restriction of @xmath826 to @xmath776 is injective . it means then that @xmath807 and @xmath797 are isomorphic for all @xmath3 , if @xmath128 is large enough . * proof of part [ part_theo : part1 ] * we will define a @xmath459 partial representation @xmath828 of @xmath651 on the linear space @xmath829 . the representation @xmath828 will be reachable and observable and will satisfy a number of properties . to this end , define the map @xmath830 by requiring that for all @xmath831 and for each word @xmath798 , @xmath832 it is easy to see @xmath833 is a surjective linear map . moreover , @xmath833 maps the sub - space @xmath834 onto @xmath829 . using corollary [ part_pow : pf3:col3 ] , ( [ part_theo : eq1 ] ) can be rewritten as @xmath835 but ( [ part_pow : pf4 ] ) implies that @xmath836 and that the restriction of @xmath833 to @xmath829 is injective . the latter means that the inverse map @xmath837 exists . * for each letter @xmath640 , the linear map @xmath839 is defined as follows . consider the map @xmath840 defined by @xmath841 , for all @xmath842 and @xmath843 . it is easy to see that @xmath844 is a linear map . in addition , if @xmath845 for some @xmath798 , @xmath132 , then @xmath846 . indeed , for any @xmath842 , @xmath847 . hence , because of linearity of @xmath844 , the range of @xmath844 is @xmath848 and thus @xmath844 can be viewed as a map @xmath849 . + then for all @xmath850 , define @xmath851 it is easy to see that @xmath852 is a well - defined linear map and it has the property that it commutes with the shift by @xmath784 of a formal power series in @xmath797 , i.e. for all @xmath853 , @xmath854 here , we used the notation of ( [ part_pow : pf1 ] ) . indeed , notice that @xmath855 . in addition , from the discussion above it follows that @xmath856 . combining this with the definition of @xmath647 yields ( [ part_pow : pf10 ] ) . + in addition , for all @xmath857 , and for all @xmath802 and @xmath640 , @xmath858 * the family @xmath859 is defined as @xmath860 i.e. @xmath861 is just the restriction of @xmath862 to @xmath795 . * the map @xmath863 is defined as @xmath864 i.e. @xmath863 is just the evaluation of the elements of @xmath834 at the empty word @xmath193 . it is easy to see that @xmath828 is a well - defined rational representation . next , we will show that @xmath828 is an @xmath459-partial representation of @xmath651 . to this end , by repeated application of ( [ part_pow : pf10 ] ) and using the equality @xmath865 we get that for any word @xmath866 it holds that @xmath867 using ( [ part_pow : pf8 ] ) and repeatedly applying ( [ part_pow : pf10.1 ] ) yields that for each @xmath842 , @xmath868 by noticing that @xmath869 and that any word @xmath870 in @xmath871 can be represented as the concatenation @xmath872 of a word @xmath866 with a word @xmath842 , we get that for all @xmath873 , @xmath874 that is , @xmath828 is an @xmath459-partial representation of @xmath651 . from ( [ part_pow : pf8 ] ) it is easy to deduce that @xmath828 is reachable . observability of @xmath828 can be derived as follows . from ( [ part_pow : pf10.1 ] ) it follows that for any @xmath875 , @xmath876 for all words @xmath798 . hence , if @xmath877 for all @xmath197 , then for all @xmath798 , @xmath878 , i.e. @xmath879 . that is , @xmath880 . define now the representation @xmath579 as the isomorphic copy of @xmath828 with the isomorphism @xmath881 defined in ( [ part_pow : pf3.1 ] ) . that is , using the notation of the theorem , define the parameters @xmath647 , @xmath640 , @xmath648 and @xmath882 of @xmath579 as follows . @xmath883 since @xmath881 defines a representation isomorphism @xmath884 , it follows that @xmath579 is a @xmath459-rational representation of @xmath651 as well , and it is reachable and observable . in addition , combining ( [ part_pow : pf14 ] ) with ( [ part_pow : pf10],[part_pow : pf11],[part_pow : pf12 ] ) yields ( [ part_theo : eq4],[part_theo : eq5],[part_theo : eq6 ] ) . * proof of part [ part_theo : part2 ] * + recall that @xmath885 for all @xmath886 and @xmath887 . it is easy to see that @xmath888 hence , ( [ part_theo : eq2 ] ) implies that @xmath889 hence , ( [ part_theo : eq1 ] ) holds and by part [ part_theo : part1 ] the representation @xmath582 is well - defined and it is a @xmath459-partial representation of @xmath651 . recall that ( [ part_theo : eq1 ] ) implies that @xmath890 . it is left to show that @xmath579 is a representation of @xmath651 . to this end , recall from ( [ project : def ] ) the definition of the map @xmath891 and recall from ( [ part_pow : pf1 ] ) the definition of the set @xmath797 . since @xmath892 , we get that @xmath893 hence @xmath894 . notice that @xmath891 maps @xmath895 onto @xmath896 for @xmath897 . hence , the map @xmath891 maps @xmath776 onto @xmath865 . that is , the restriction @xmath891 to @xmath776 yields a linear isomorphism @xmath898 . recall from lemma [ sect : form : theo1:lemma1 ] the definition of the free representation @xmath781 of @xmath651 , defined on the space @xmath776 . using ( [ part_pow : pf10]),([part_pow : pf11 ] ) and ( [ part_pow : pf12 ] ) , it is easy to see that the restriction of @xmath891 to @xmath776 in fact yields a representation isomorphism @xmath899 , where @xmath828 is as defined in ( [ part_pow : pf5 ] ) . since , @xmath781 is a minimal representation of @xmath651 , then so is its isomorphic copy @xmath828 . since @xmath579 is merely an isomorphic copy of @xmath828 , the same conclusion holds for @xmath579 . * proof of part [ part_theo : part3 ] * + recall from ( [ part_pow : pf1 ] ) and ( [ part_pow : def2 ] ) the definition of the spaces @xmath797 and @xmath807 for @xmath900 , @xmath901 . from lemma [ sect : form : theo1:lemma2 ] it follows that if @xmath735 , then @xmath902 . then by corollary [ part_pow : pf1:col1 ] , @xmath903 for @xmath904 . moreover , notice that the image of @xmath797 by @xmath815 equals @xmath807 . applying the above observations for @xmath905 and using lemma [ part_pow : lemma3 ] we get that @xmath891 and @xmath906 are injective and hence @xmath907 combining ( [ part_pow : pf2 ] ) with corollary [ part_pow : pf3:col3 ] and remark [ sect : form : def1:rem1 ] we get that ( [ part_theo : eq2 ] ) holds . combining this and part [ part_theo : part2 ] we get that the statement of part [ part_theo : part3 ] holds . we will prove the statements of the theorem one by one . however , before proceeding to the actual proof , we need to introduce some notation . let @xmath480 be @xmath451-@xmath127 rational representation , i.e. @xmath480 is of the form ( [ repr : def0 ] ) . define the family of formal power series _ @xmath908 associated with the representation @xmath480 _ as follows . the family @xmath908 is indexed by elements of @xmath127 , i.e. it can be written as @xmath909 . for each index @xmath132 , and for each word @xmath197 , the value of the formal power series @xmath910 at @xmath190 is defined as @xmath911 . it is easy to see that the rational representation @xmath480 is a representation of @xmath908 , i.e. @xmath908 is rational . in fact , @xmath480 is a representation of the family of formal power series @xmath651 , if and only if @xmath912 , i.e. @xmath913 for all @xmath132 . furthermore , @xmath480 is a @xmath459-partial representation of @xmath651 , if and only if @xmath914 for all @xmath132 and @xmath915 . in addition , if @xmath480 is a @xmath459-partial representation of @xmath651 , then @xmath916 in particular , @xmath917 . now we are ready to proceed to the actual proof of the theorem . * proof of part [ part_real_pow : theo2:res1 ] * + assume that @xmath480 is a @xmath459-partial representation of @xmath651 . from the discussion above it follows that @xmath918 . it is easy to see that @xmath919 . moreover , from theorem [ sect : form : theo3 ] it follows that @xmath920 . that is , the rank of @xmath921 is not greater than the dimension of the representation @xmath480 . but from theorem [ part_real_pow : theo ] it follows the the dimension of the representation @xmath582 is precisely @xmath736 and that @xmath582 is an @xmath459-partial representation of @xmath651 . hence , we get that @xmath922 . in other words , @xmath582 is a minimal @xmath459 representation of @xmath651 . * proof of part [ part_real_pow : theo2:res2 ] * + we have just shown that the rational representation @xmath582 of theorem [ part_real_pow : theo ] is a minimal @xmath459-partial representation of @xmath651 . if @xmath480 is another minimal @xmath459-partial representation of @xmath651 , then the dimension of @xmath480 must be equal to that of @xmath582 , which in turn equals @xmath736 . hence , the second part of the statement is proved . assume now that @xmath480 is a minimal @xmath459-partial representation of @xmath651 . then it means that @xmath923 for each @xmath132 and @xmath915 , where @xmath910 denotes the element of @xmath908 indexed by the index @xmath132 . if @xmath480 is not reachable and observable , then it is not a minimal representation of @xmath908 . from theorem [ sect : form : theo3 ] it follows then that there exists a reachable and observable representation @xmath924 of @xmath925 such that @xmath926 . but if @xmath927 is a representation of @xmath925 , then it is a @xmath459-partial representation of @xmath651 , since , @xmath928 for all @xmath132 and @xmath915 . hence , @xmath480 is not a minimal @xmath459-partial representation of @xmath651 , which is a contradiction . * proof of part [ part_real_pow : theo2:res3 ] * + suppose @xmath480 and @xmath929 are two minimal @xmath459-partial representations of @xmath651 . recall the definition of the families of formal power series @xmath908 and @xmath930 associated with @xmath480 . since @xmath480 and @xmath929 have to be reachable and observable by part [ part_real_pow : theo2:res2 ] of the current theorem , by theorem [ sect : form : theo3 ] the representation @xmath480 ( respectively @xmath929 ) is a minimal representation of @xmath925 ( respectively @xmath930 ) . we will show that @xmath925 and @xmath930 are equal , i.e. @xmath931 for all @xmath132 . this in turn implies that @xmath480 and @xmath929 are both minimal representations of the same family of formal power series , and hence by theorem [ sect : form : theo3 ] they are isomorphic . the proof that @xmath925 and @xmath930 are identical proceeds as follows . recall that @xmath480 and @xmath929 are both of dimension @xmath736 , and recall ( [ part_real_pow : theo2:pfeq1 ] ) . hence , it follows that @xmath932 and @xmath933 . it also follows that @xmath925 and @xmath930 satisfy ( [ part_theo : eq2 ] ) with @xmath651 replaced by @xmath925 and @xmath930 respectively . denote by @xmath934 and @xmath935 the minimal @xmath459-partial representations of @xmath925 and respectively @xmath930 , obtained by applying part [ part_theo : part1 ] of theorem [ part_real_pow : theo ] to @xmath908 and @xmath930 respectively . by part [ part_theo : part2 ] of theorem [ part_real_pow : theo ] the representations @xmath934 and @xmath936 are minimal representations of @xmath925 and @xmath930 respectively . however the construction of @xmath934 and @xmath936 depends only on @xmath937 and @xmath938 . the latter two matrices both coincide with @xmath723 . hence , @xmath934 and @xmath936 are both equal to the partial representation @xmath582 obtained by applying part [ part_theo : part1 ] of theorem [ part_real_pow : theo ] to @xmath651 . that is , rational representation @xmath579 , obtained by applying part [ part_theo : part1 ] of theorem [ part_real_pow : theo ] to @xmath651 , is a minimal representation of both @xmath925 and @xmath930 . this implies that @xmath925 and @xmath930 coincide . * proof of part [ sect : svd : theo1:part1 ] * + we have to show that algorithm [ partrepr ] returns a representation @xmath611 , and @xmath611 is isomorphic to @xmath579 . by theorem [ part_real_pow : theo ] , if ( [ sect : svd : theo1:eq1 ] ) holds then @xmath579 is an @xmath459-representation of @xmath651 . from theorem [ part_real_pow : theo2 ] it then follows that @xmath582 is a minimal @xmath459-partial representation . hence , if @xmath611 is an isomorphic copy of @xmath582 , then we get that @xmath611 is a minimal @xmath459-partial representation of @xmath651 . recall from ( [ part_pow : pf5 ] ) in the proof of theorem [ part_real_pow : theo ] the definition of the representation @xmath828 . using the remark in subsection [ sect : prelim : not_matrix ] , we will interpret the matrix @xmath739 as a linear map @xmath939 defined by @xmath940 for all @xmath941 . using this interpretation , we define the map @xmath942 since @xmath881 defines a representation isomorphism @xmath884 it is enough to show that @xmath943 defines a representation isomorphism @xmath944 and that algorithm [ partrepr ] returns a rational representation if ( [ sect : svd : theo1:eq1 ] ) holds . indeed , in this case @xmath945 defines a representation isomorphism @xmath946 . it is clear that @xmath943 is well defined . indeed , @xmath947 by definition of matrix factorization . moreover , @xmath948 is an injective linear map and @xmath949 is a linear isomorphism . furthermore , since @xmath950 we get that @xmath951 is a linear isomorphism . thus , @xmath952 is a well defined linear isomorphism . it is left to show that @xmath943 is a representation morphism from @xmath611 to @xmath828 . it is easy to see that for all @xmath798 , and @xmath941 . @xmath953 recall from subsection [ sect : prelim : not_matrix ] that @xmath954 stands for the value ( entry ) of @xmath955 corresponding to the index @xmath956 . in particular , for any @xmath957 , @xmath958 where @xmath959 stands for the column of @xmath480 indexed by @xmath960 . indeed , using ( [ sect : svd : theo1:pf : eq2 ] ) we get that for all @xmath798 , @xmath961 but @xmath962 , and hence the right - hand side of ( [ sect : svd : theo1:eq2.0.0 ] ) equals the vector formed by the entries @xmath963 for @xmath964 . but the entry @xmath963 equals the @xmath394th entry of @xmath965 , hence @xmath966 from ( [ sect : svd : theo1:pf : eq2 ] ) it follows that @xmath967 from ( [ sect : svd : theo1:eq2.0 ] ) , it follows that @xmath968 it is left to show that a unique solution to equation ( [ computepartialrepr ] ) exists and @xmath969 for all @xmath640 . first , notice that @xmath970 . thus , @xmath971 , i.e. @xmath972 . hence , if a solution solution ( [ computepartialrepr ] ) exists , then this solution is unique . therefore , if we show that @xmath973 is a solution to ( [ computepartialrepr ] ) then ( [ sect : svd : theo1:eq5 ] ) automatically holds . notice that @xmath974 is a linear map from @xmath975 to @xmath975 and hence it can be identified with its matrix representation . by identifying @xmath974 with its matrix , and using ( [ sect : svd : theo1:eq2.0 ] ) and ( [ part_pow : pf10 ] ) , it follows that for any @xmath957 and @xmath976 , @xmath977 here we used the notation of subsection [ sect : prelim : not_matrix ] to denote rows and columns of matrices . in particular , @xmath978 denotes the column of a matrix @xmath128 indexed by @xmath960 and @xmath979 denotes the row of a matrix @xmath128 indexed by @xmath980 . notice that if @xmath798 , then the row of @xmath490 indexed by @xmath980 equals the row of @xmath476 indexed by @xmath980 . moreover , if @xmath981 , then from ( [ sect : svd : theo1:pf : eq2 ] ) it follows that the row of @xmath982 indexed by @xmath980 equals the @xmath394th entry of @xmath983 . hence , the last expression of ( [ sect : svd : theo1:eq6 ] ) can be rewritten as @xmath984 where @xmath985 denotes the @xmath394th entry of @xmath986 . on the other hand , the row of @xmath987 indexed by @xmath980 equals the row of @xmath476 indexed by @xmath988 , if @xmath798 . hence , @xmath989 combining ( [ sect : svd : theo1:eq6 ] ) , ( [ sect : svd : theo1:eq6.1 ] ) and ( [ sect : svd : theo1:eq7 ] ) we get that @xmath990 . since @xmath991 , i.e the columns of @xmath480 span the whole space @xmath975 , the last equality implies that @xmath974 is a solution to ( [ computepartialrepr ] ) . * proof of part [ sect : svd : theo1:part2 ] * + from theorem [ part_real_pow : theo ] it follows that if ( [ sect : svd : theo1:eq2 ] ) holds , then ( [ sect : svd : theo1:eq1 ] ) holds and @xmath579 is a minimal representation of @xmath651 . by part [ sect : svd : theo1:part1 ] of this theorem , if ( [ sect : svd : theo1:eq1 ] ) holds , then algorithm [ partrepr ] returns a representation @xmath611 and @xmath611 is isomorphic to @xmath579 . hence @xmath611 is a minimal representation too . * proof of part [ sect : svd : theo1:part3 ] * + again , from theorem [ part_real_pow : theo ] it follows that ( [ sect : svd : theo1:eq2 ] ) holds in this case . the statement follows now from ( [ sect : svd : theo1:eq2 ] ) and part [ sect : svd : theo1:part2 ] of this theorem . we conclude the section with presenting the proof of lemma [ part_pow : lemma1 ] and lemma [ part_pow : lemma3 ] used in the proof of theorem [ part_real_pow : theo ] . the proof relies on the following chain of results , which are interesting on their own right . [ part_pow : lemma4 ] let @xmath102 be finite - dimensional vector space , @xmath992 . let @xmath993 , @xmath640 be a family of linear maps . then for each @xmath994 , for each @xmath197 , the vector @xmath995 is a linear combination of the vectors @xmath996 , for finitely many words @xmath802 . if @xmath997 , then the statement of the lemma is trivially true . first we prove the lemma for @xmath411 . assume that @xmath998 , @xmath999 . consider the elements @xmath1000 , @xmath1001 . since every @xmath491 elements of @xmath102 are linearly dependent , we get that there exist @xmath1002 such that @xmath1003 is a linear combination of @xmath1004 , @xmath1005 . then we get that @xmath1006 is a linear combination of vectors of the form @xmath1007 , @xmath1008 . notice that each word @xmath1009 is of length at most @xmath804 . to prove the lemma for arbitrary @xmath1010 we proceed by induction on @xmath1011 . the case of @xmath411 we proved above . assume that the statement of the lemma holds for @xmath1012 . assume now that @xmath190 is of the form @xmath1013 , where @xmath1014 , @xmath1015 and @xmath640 is the last letter of @xmath190 . from the induction hypothesis it follows that @xmath1016 is a linear combination of vectors of the form @xmath996 , @xmath802 . but then @xmath1017 is a linear combination of vectors of the form @xmath1018 . for each @xmath802 , either @xmath1019 or @xmath1020 . in the latter case , by induction hypothesis @xmath1021 is again a linear combination of vectors of the form @xmath1022 , @xmath1023 . altogether , we get that @xmath995 is a linear combination of the vectors @xmath996 , @xmath802 . if @xmath220 and @xmath127 is finite , then the above corollary states that the observability subspace @xmath1028 is the kernel of a finite matrix , and the reachability subspace @xmath1029 is the image of a finite matrix . it is clear that @xmath1030 . we will show that the reverse inclusion holds as well . assume that @xmath1031 is such that @xmath1032 for all @xmath802 . by lemma [ part_pow : lemma4 ] , for each @xmath197 , @xmath1033 is a linear combination of @xmath1034 for @xmath802 . hence , @xmath1035 is a linear combination of @xmath1036 , @xmath802 and the latter implies that @xmath1037 . that is , @xmath1038 . similarly , it is clear from the definition that @xmath1039 . the reverse inclusion follows from lemma [ part_pow : lemma4 ] , according to which for each word @xmath197 , @xmath668 is a linear combination of the vectors @xmath1040 , @xmath802 . hence , we get that @xmath1041 . we will use the fact that @xmath1042 . consider the free representation @xmath1043 of @xmath651 defined in theorem [ sect : form : theo1 ] . notice that for any word @xmath186 , @xmath1044 . apply lemma [ part_pow : lemma4 ] to @xmath776 , @xmath1045 , @xmath640 and @xmath1046 . then we get that for any word @xmath197 , the formal power series @xmath1047 is a linear combination of the formal power series @xmath1048 with @xmath842 . it is easy to see that @xmath815 is a surjective linear map . consider the free realization @xmath1049 of @xmath651 . from theorem [ sect : form : theo3 ] we know that @xmath781 is minimal and therefore it is reachable and observable , i.e. @xmath1050 . from corollary [ part_pow : col1 ] we also know that if @xmath1051 , then @xmath1052 . consider the kernel of @xmath815 . for any formal power series @xmath1053 , @xmath1054 if and only if @xmath1055 for all @xmath1056 . hence , @xmath1057 for each @xmath1056 , i.e. @xmath1058 . thus , @xmath815 is injective .

red blood cells ( rbcs ) have been a model system and an inspiration for the biophysics of the membrane for decades@xcite . among the numerous works in the literature , many studies have been devoted to adhesion of cells or vesicles and capsules on _ flat and solid _ surfaces @xcite , where , in the case of elastic capsules only , a buckling instability may occur due to the interplay between bending and stretching energy@xcite . buckling has also been observed in freestanding capsules due to osmotic pressure@xcite or shear forces by external flow @xcite , and in vesicles due to asymmetric lipid distributions @xcite or shear forces by external flow @xcite . the influence of intercellular adhesion @xcite is less known despite its relevance in dense environments such as blood at physiological hematocrit or even tissues . theoretical studies on aggregation between two rbcs based on minimizing the free energy of the membrane@xcite indicate a significant change of the geometry of contact zones between cells when varying the interaction energy . some of these shapes have been observed experimentally @xcite but not quantified as a function of the dextran or fibrinogen concentrations . a variety of membrane shapes has been predicted theoretically for rbc doublets@xcite , depending mainly on the nondimensionalized reduced adhesion strength @xmath0 and the reduced volume @xmath1 where @xmath2 and @xmath3 are the enclosed volume and surface area of the membrane , a parameter that is indeed physiologically related to the hydration state and age of rbcs and can vary over a significant range . examples include male - female ( convex - concave ) shapes and more or less pronounced sigmoid shapes obtained by varying @xmath4 and @xmath0 as depicted in fig . [ 3dshapes](a ) . the 3d reconstructions of our experimental confocal images show rbc doublets that look very similar to the theoretical predictions , too ( fig.[3dshapes](b ) ) . experimental images also show that the amplitude of the deformation of contact zones depends non monotonically on macromolecular ( dextran ) concentration . there are direct physiological implications of aggregation since rbcs in blood in stasis are known to form aggregates in the form of so called _ rouleaux_. this clustering process is reversible and typical hydrodynamic shear forces in physiological flow are usually sufficient to break _ rouleaux _ , at least in larger vessels . in the 1960 s , fibrinogen could be identified as one of the main plasma proteins causing rbc aggregation or clustering @xcite . it is known that the number and size of aggregates increase with fibrinogen concentration in the physiological range ( from approximately @xmath5 for healthy adults up to @xmath6 in acute inflammatory phases ) . the level of rbc aggregation is used in one of the most fundamental standard hematological blood tests , the erythrocyte sedimentation rate ( esr ) which is carried out worldwide many thousand times each day . the more and the larger the aggregates , the faster the sedimentation of the rbcs : by simply measuring visually the sedimentation front in a standardized glass capillary one gets a robust and quick , if unspecific , indication on the inflammatory state of the patient . of course , the tendency of the rbcs to form aggregates may also increase the risk of thrombosis and cardiovascular diseases , especially in combination with stenosis . the molecular mechanisms of macromolecular induced rbc aggregation have been the subject of controversial studies @xcite . two models have been proposed : one is based on the physical effect of depletion @xcite and the other on physisorption or bridging @xcite . in addition to the specific adhesion mechanism , the strength of rbc aggregation also depends on their physical properties , such as deformability @xcite , surface charge @xcite and reduced volume @xcite . it can also be induced by other macromolecules such as dextran , a widely used molecule in laboratory experiments or as a plasma expander and in veterinary medicine . afm based single cell force spectroscopy , rheology and sedimentation of rbcs @xcite as well as theoretical models @xcite depict that an increasing dextran concentration leads first to an increase of the interaction strength among rbcs up to a specified macromolecule concentration . beyond that maximum the interaction strength decreases again , and cell cell aggregation vanishes . this leads to a characteristic bell shaped adhesion - energy versus concentration curve . for the case of fibrinogen , it was found the interaction energy increases with the concentration , but data from the literature are typically limited to lower concentrations than for the case of dextran @xcite . and adhesion strengths @xmath0 . from left to right : male - female shapes , strong sigmoid shapes and low sigmoid shapes.(b ) 3d - confocal imaging of rbc - doublets . the experimental results qualitatively reproduce the theoretical predictions.[3dshapes ] ] we present here a systematic experimental study in order to analyze quantitatively the evolution of the contact zone as a function of aggregating molecule concentration , in a large domain of parameter space by using either dextran or fibrinogen . we analyze the case of rbc doublets as well as that of larger _ rouleaux_. a systematic numerical investigation supports the experimental results . the numerical results show that in the case of two adhering cells , the interface remains flat up to a critical value of interaction energy above which it bifurcates into a deformed , buckled state . our analytical model reveals a forward bifurcation with critical parameters close to those of the full numerical problem . the analysis of the interfacial shape of two cells _ rouleaux _ , here called doublets , was performed by confocal imaging of doublets during free sedimentation , i.e. while not in contact with surfaces . typical images are shown in fig . [ experimentaldoublets ] for doublets in dextran 70kda solutions ( see s. i. for the fibrinogen case ) . there is a significant evolution of the interfacial shape with polymer concentration : at low and high concentrations the interface is only weakly deformed while in the intermediate range where interaction energy is higher , the interface is strongly bent . qualitatively , three different kinds of interfacial shapes were identified : parachute , male - female and sigmoid . in the sigmoid ( s - shape interface ) , cells were mostly laterally displaced , with a non - monotonous curvature of the contact zone , while in other shapes the two cells had a common symmetry axis . the parachute shape was defined by at least one of the cells having a convex and a concave side . the male - female shape was defined when the interface had a bulge , with rather flat non - interacting large membrane parts . cells with a flat interface were also classified under this category . similar shapes were found for fibrinogen solutions . as can be seen in fig . [ experimentaldoublets](a ) , all shapes can be found for all polymer concentrations , probably due to dispersity of cell properties in the sample , with a probability that depends on concentration . a quantitative analysis of the distribution of the different interfacial shapes for the different concentrations of dextran is shown in fig . [ experimentaldoublets](b ) , revealing an increase of the occurrence of sigmoid shapes in the intermediate range of polymer concentrations , i.e. at larger interaction energy favoring slightly larger contact areas . at low and high concentrations of dextran , doublets are mostly in the parachute ( about 60 % ) and male - female shapes . for fibrinogen ( fig . [ experimentaldoublets](c ) ) , mostly parachute shapes were observed in the physiological range of the protein s concentration . the case of echinocyte rbc doublets , obtained after sedimentation and contact with glass , was also studied ( see s. i. ) . quantification of the interfacial shape deformation was done by fitting it with a heuristically chosen sinus function ( see s.i . ) , whose amplitude was taken as a measure of the interfacial curvature ( fig . s2 ) . a bell shape dependency of the deformation amplitude on polymer concentration is seen ( fig . [ amplitudesdoublets ] ) , similar to the dependency found in afm single cell force spectroscopy @xcite or aggregation index ( see s.i . ) . for the case of dextran , both doublets with discocytic and with echinocytic cells were analysed . the amplitude for discocytic cells can reach much larger values than echinocytes that have a close - to - spherical shape and probably a higher membrane stiffness @xcite . to compute the interaction energies corresponding to the experimental doublets and gain further insight on the variability of shapes we performed numerical simulations ( see s.i . for details of the method ) . based on several successful studies ( especially equilibrium shapes and shape dynamics under external flow ) which have shown that 2d and 3d models capture the same essential features @xcite , we restrict ourselves here to a 2d model . in 2d the notion of in - plane shear elasticity looses its meaning and the vesicles and inextensible capsule model ( often evoked to model rbc ) are equivalent . for reduced areas @xmath7 ( where @xmath3 is the area of the 2d vesicle and @xmath8 its perimeter ) in the range ( @xmath9 ) with different combinations ( cells with same or different reduced areas ) and a membrane stiffness of @xmath10 j , we varied the interaction energy @xmath11 in the range 0 - 8.23 @xmath12j / m@xmath13 . the resulting interfacial shapes for different combinations of reduced area @xmath14 , interaction energy @xmath11 and curvature energy @xmath15 ( fig . [ numericdoublets ] ) were in qualitative agreement with the experimentally obtained shapes , namely male - female , s - shapes and parachute shapes . figure [ numericdoublets](a ) shows the evolution of the deformation as a function of the interaction energy . as for the analysis of experimental data we fit the interfacial shape with a sine function . the mean amplitude of the fitted curve was compared to experimental results ( fig . [ amplitudesdoublets ] ) , by heuristically choosing a correspondence between concentration and interaction energy : we assume a linear increase of the interaction energy with concentration up to 6.9 @xmath12j / m@xmath13 at 60 mg / ml , followed by a symmetric linear decrease , in order to get a good compromise between an agreement on the width and the maximum value of bell shaped curves . this piecewise linear variation of the interaction energy is a simplification which nevertheless yields a qualitative agreement for the overall behavior . interestingly , the numerical results also show that the transition from a flat to a deformed interface only occurs at some finite interaction energy @xmath11 ( fig . [ numericdoublets](b ) ) . in principle , this method could allow the determination of the experimental interaction energy for a given concentration . however , the numerical results depend quite sensitively on the reduced areas @xmath14 of the two cells ( fig . [ numericdoublets](b ) ) , a parameter that varies slightly between cells in experiments , even in the same sample . this dispersity is probably also the reason why the bell shape in fig . [ amplitudesdoublets](a ) of the experimental data does not show this sharp transition . and @xmath16 and variation of the energy [ fig - model ] ] the evolution of the amplitude of contact zone s deformation in fig . [ numericdoublets](b ) suggests that a bifurcation akin to buckling occurs at a critical value of the interaction energy , triggering a change of shape from flat to s - shape . to get some insight into the involved mechanism , this behavior can actually be recovered by an analytical model through a minimization of the energy of the system . let us suppose that the shape of an rbc in a symmetric doublet ( in 2d ) is approximated as shown in fig . [ fig - model](a ) , with two semi - circular caps of radius @xmath17 , a roughly straight portion where curvature is negligible ( length @xmath18 ) and a contact zone with a sinusoidal shape of curvilinear length @xmath19 and amplitude @xmath20 described by the equation : @xmath21 for @xmath22 . the corresponding interaction energy ( per unit length in the third dimension ) is , at order 4 in @xmath20 : @xmath23 where @xmath11 is the interaction energy per unit area . the curvature energy of the system is concentrated in the circular caps @xmath24 and the deformed contact zone @xmath25 given by : @xmath26 the total energy @xmath27 therefore depends on @xmath20 , @xmath17 and @xmath18 which are related through the constant area @xmath3 and constant perimeter @xmath8 constraints . in the following we define a dimensionless energy @xmath28 , a rescaled interaction energy @xmath29 , and rescale all lengths by @xmath8 : @xmath30 and so on , and introduce the reduced area @xmath7 . the total energy @xmath31 can then be derived and expanded at order 4 in @xmath20 ( see supplemental material ) . figure [ fig - model](b ) shows the energy due to deformation for @xmath32 and different interaction energies @xmath33 . below a critical value @xmath34 , the minimum of energy corresponds to @xmath35 ( flat interface ) , while for @xmath36 , the energy is minimal for a finite value of @xmath37 ( see s. i. ) . this defines the bifurcation threshold @xmath34 corresponding to @xmath38 , which is an increasing function of @xmath14 ( fig . [ fig - model](c ) ) : @xmath39 the bifurcation diagram of the deformation amplitude ( fig . [ fig - model](d ) ) reveals a classical supercritical bifurcation , owing to the competition between interaction energy and curvature energy through the constant perimeter and area constraints . orders of magnitude of the bifurcation threshold and amplitude of the deformation are in good agreement with experiments and numerical simulations . we have @xmath40 . if we take @xmath41 m and @xmath42 j as in simulations , the critical interaction energy is @xmath43j / m@xmath13 for @xmath32 and @xmath44j / m@xmath13 for @xmath45 , which is in reasonable agreement with the simulations of fig . [ numericdoublets](b ) , while the amplitude of the deformation @xmath20 saturates at @xmath46 m in this model , rather close to the value @xmath47 m found in the simulations given the simplicity of the model . interestingly , the energy of the equilibrium configuration decreases linearly with @xmath33 as shown in fig . [ fig - model](d ) , with no sign of the shape bifurcation due to the smallness of the non linear terms arising from the deformation in the curvature and interaction energy terms . this result can be of interest in the modeling of clustering dynamics in flow for instance , where the dissociation rate due to hydrodynamic stresses can be assumed to simply scale like @xmath48 despite the possibly complex shapes of contact zones . a characterization of larger clusters was performed with _ rouleaux _ of at least 7 cells . the fast sedimentation and arbitrary orientation of the rouleaux only allowed us to image them when settled on the ( bsa treated ) cover slip . since fibrinogen led to significant inter - connections between _ rouleaux _ which eventually led to a percolated gel like structure at 45% hematocrit , we only focused on rouleaux in linear configuration in dextran solutions ( fig . [ imagelargecluster]a ) . at 10mg / ml the cells in a _ rouleau _ were typically laterally shifted and we did not observe large clusters . at higher concentrations of dextran rouleaux _ were observed and as shown in fig . [ imagelargecluster ] , only the first three interfaces exhibited significant deformation ( fig . [ imagelargecluster]a ) while the interfaces far from the tip were always in a male - female shape with very low deformation , mostly below the experimental resolution limit . again , a large variety of shapes was observed for these first three interfaces ( fig . [ imagelargecluster]b ) . as for doublets , these interfaces can be sorted into different categories and the distribution of shapes is shown separately for the first , second and third interface starting from the doublet tip in fig . [ amplitudeslargecluster]a . unlike the case of doublets , we observed mostly male - female shapes and no parachute shapes at all , likely due to the constraints induced by the next attaching cell . for concentrations in the lower to medium - range , s - shape contact zones exist more frequently for the first contact zone than for the next ones , while male - female shapes are indeed prevalent everywhere except at very low dextran concentrations where rbcs tend to stay very close to their free discoid shape due to the low interaction energy , in agreement with the theoretical results on low - aggregation behavior mentioned in @xcite . the amplitudes of the interfacial deformations ( determined by the same method as for doublets ) are shown in fig . [ amplitudeslargecluster]a . again , a bell shape behavior for all three interfaces was observed . interestingly , the amplitude of the first contact zone was very similar to the values measured for echinocyte doublets ( see s.i . ) , too . indeed , the aggregation into larger clusters gives less degree of freedom to rbcs due to higher constraints on the membrane , leading to a higher effective stiffness of rbcs . numerical simulations were performed for 7 cells and different interaction energies @xmath11 , a constant reduced area @xmath49 and membrane stiffness @xmath10 j ( fig . [ simulationslargecluster ] ) . the results agree qualitatively with the experimental results and the amplitudes of the interfacial deformations of the first and second interface were in good agreement with the experimental data using the same energy scale as for doublets ( see fig . [ amplitudeslargecluster ] ) . however , the amplitudes between experiment and theory differed significantly for the third shape . here the variability of the experimentally observed shapes plays a role , as well as the blurring of the contact zone due to resolution in experiments , which tends to increase the amplitude of the fitted sine function . red blood cells aggregate due to the presence of plasma proteins or biomimicking polymers . this is an example of adhesion between two very soft interfaces , different from the case of adhesion on a solid substrate , where we observe a buckling of the cell cell interface . the nature of this instability is revealed by our numerical simulations and explained by an analytical model . the critical threshold of buckling and the final geometry of the aggregate depends strongly on the physical parameters , i.e. on the reduced volume of the cells . for large reduced volumes that correspond to more spherical cells , the critical interaction energy for buckling increases . in the case of doublets of two cells having identical reduced volumes close to that of a typical rbc , a sigmoid deformation of the interface takes place , while for two different reduced volumes male - female shapes can be obtained due to the asymmetry of properties that forces a different mode . we also characterized long linear clusters experimentally and numerically and have found that only the three interfaces at the two ends are considerably deformed . it would undoubtedly be interesting to check how these contact shapes obtained in quasi - static conditions evolve in pathological conditions , or under flow with the additional constraint of hydrodynamic stresses . human blood withdrawal from healthy donors as well as blood preparation and manipulation were performed according to regulations and protocols that were approved by the ethic commission of the rztekammer des saarlandes ( reference no 24/12 ) . we obtained informed consent from the donors after the nature and possible consequences of the studies were explained . venous blood of three healthy donors ( one female , 26 years , blood group 0 negative ; one male , 43 years , blood group 0 negative ; one male , 27 years , blood group b positive ) was drawn into conventional edta tubes ( s - monovette ; sarstedt , nmbrecht , germany ) and all measurements were performed within 4 hours after the blood had been drawn . after washing the rbcs three times ( 704 g , 3 min ) with phosphate - buffered solution ( pbs , life technologies , waltham , ma , usa ) 0.1 @xmath12l cellmask ( life technologies , waltham , ma , usa ) was added for fluorescence labeling to 200 @xmath12l rbcs in 1ml pbs for 10min . after washing two times ( 704 g , 3min ) the supernatant was removed and the rbcs were resuspended in solutions of dextran ( dextran 70 from leuconostoc mesenteroides , sigma - aldrich , st louis , usa ) or ( washed ) fibrinogen ( fibrinogen from human plasma , sigma - aldrich , st louis , usa ) in pbs . the fibrinogen as delivered contains approximately 60% fibrinogen , 25% sodium chloride and 15% of sodium citrate . in order to remove the salt the solutions were ultrafiltrated with a 10kda filter ( vivaspin turbo 4 , satorius ag , goettingen , germany ) at 3500 g in a centrifuge ( hermle z326k , hermle labortechnik gmbh , wehingen , germany ) for 10 hours by adding 2ml of pbs every hour . the remaining protein solution was freeze - dried and the proteins were resuspended in pbs . fibrinogen was used in the measurements within 24 hours after freeze - drying and the osmolality of this solutions was verified to be in the order of pbs ( + /- 5% ) . cylindrical wells of diameter 5 mm , height 1.6 mm ( @xmath12-slide 18 well - flat from ibidi , munich , germany ) were used as sample chambers . rbcs sitting on the bottom slide of the well changed to echinocytes due to the glass effect . for the measurements in which we wanted to preserve the discoyte shape of sedimented cells , the wells were covered with 1 mg / ml bovine serum albumine ( bsa , sigma - aldrich , st louis , usa ) in pbs for 30 min at 37@xmath50c . the bsa - pbs solution was removed and after a drying time of 10 min at 37@xmath50c the rbc suspensions were added . all observations were performed with the confocal microscope described in the previous section at a temperature of 23 @xmath50c with an hematocrit of 0.2 % for the imaging of rbc doublets and 0.4 % for clusters larger than 6 rbcs . images of doublets of rbc in their normal discoid shape could be recorded in free volume during the slow sedimentation process which prevented any contact with the glass surface of the cover slip . this procedure did allow to study the aggregates without any other influence but we should also note that it was also not possible to investigate the interfacial shape of sedimented cells in discocytic state . in this configuration , they were horizontally oriented with their interface and the rbc membrane was very thin . the optical 2-d sections showed a good signal only when the membrane was oriented along the axis of the point spread function of the laser focus . however , images of sedimented rbcs were also taken on the non - treated cover slips , i.e. for the case of echinocytes which were often oriented with their interface perpendicular to the cover slip , most likely due to their rounder shape . in total , 30 cells for discocytes and 15 for echinocytes were evaluated . for clusters that were larger than 6 cells , measurements could be only be performed on sedimented cells . in this case the cover slip was always treated with bsa to preserve the discocytic shape , and contact zones were always oriented perpendicular to the cover slip , thereby allowing a good characterization . the raw images were contrast enhanced and the position of the interfacial membrane contours was manually determined by visual inspection . the point cloud that indicated the interface was extracted with matlab ( the mathworks gmbh , ismaning , germany ) and stored into a one dimensional array to allow for a fitting procedure with origin ( originlab corporation , northampton , ma , usa ) . the interface was fitted with a heuristically chosen sinus function which yielded reasonable agreement . for sigmoid shapes the wave length of the sinus was left as a free parameter and for the male - female and parachute shape the wave length was limited to a range of @xmath51 around the double of the cell diameter . this accounts only for the images shown in fig.[3dshapes ] . in the preparation step for the wells celltak ( celltak , bc bioscience , san jose , usa ) was added as specified by the manufacturer in order to immobilize the cells on the bsa treated cover slip . a stack of 50 images was taken at a distance of 0.1 @xmath12 m in z - direction , while the x - y area was 12 x 12 @xmath12 m . a surface rendering with imaris ( bitplane ag , zuerich , switzerland ) allowed to look through the cell membranes attached to the bottom onto the contact zones . 10 n. mohandas and p. g. gallagher . red cell membrane : past , present , and future . , 112(10):39393948 , 2008 . u. seifert and r. lipowsky . adhesion of vesicles . , 42:47684771 , oct 1990 . u. s. schwarz and s. a. safran . physics of adherent cells . , 85:13271381 , aug 2013 . s. komura , k. tamura , and t. kato . buckling of spherical shells adhering onto a rigid substrate . , 18:343 , 2005 . s. knoche and j. kierfeld . osmotic buckling of spherical capsules . , 10:83588369 , 2014 . s. s. datta , s .- h . kim , j. paulose , a. abbaspourrad , d. r. nelson , and d. a. weitz . delayed buckling and guided folding of inhomogeneous capsules . , 109:134302 , 2012 . d. barths - biesel . motion and deformation of elastic capsules and vesicles in flow . , 48:2552 , 2016 . woo , , and a. wallqvist . spontaneous buckling of lipid bilayer and vesicle budding induced by antimicrobial peptide magainin 2 : a coarse - grained simulation study . , 115:81228129 , 2011 . v. narsimhan , a. p. spann , and e. s. g. shaqfeh . pearling , wrinkling , and buckling of vesicles in elongational flows . , 777:126 , 2015 . ionic strength dependence of localized contact formation between membranes : nonlinear theory and experiment . , 77(2):817 828 , 1999 . p. ziherl and s. svetina . flat and sigmoidally curved contact zones in vesicle vesicle adhesion . , 104(3):761765 , 2007 . s. svetina and p. ziherl . morphology of small aggregates of red blood cells . , 73:8491 , 2008 . d. tilley , w. t. coakley , r. k. gould , s. e. payne , and l. a. hewison . real time observations of polylysine , dextran and polyethylene glycol induced mutual adhesion of erythrocytes held in suspension in an ultrasonic standing wave field . , 14(8):499507 , 1987 . e. w. merrill , e. r. gilliland , t. s. lee , and e. w. salzman . blood rheology : effect of fibrinogen deduced by addition . , 18(4):437446 , 1966 . r. e. wells , t. h. gawronski , p. j. cox , and r. d. perera . influence of fibrinogen on flow properties of erythrocyte suspensions . , 207(5):10351040 , 1964 . s. chien , s. usami , h. m. taylor , j. l. lundberg , and m. i. gregersen . effects of hematocrit and plasma proteins on human blood rheology at low shear rates . , 21(1):8187 , 1966 . mechanical deformability of biological membranes and the sphering of the erythrocyte . , 13(3):209 217 , 1973 . r. fhraeus . the influence of the rouleau formation of the erythrocytes on the rheology of the blood . , 161(2):151165 , 1958 . e. w. merrill , g. c. cokelet , a. britten , and r. e. wells . non - newtonian rheology of human blood - effect of fibrinogen deduced by `` subtraction '' . , 13(1):4855 , 1963 . a. s. popel and p. c. johnson . microcirculation and hemorheology . , 37:43 , 2005 . j. j. bishop , a. s. popel , m. intaglietta , and p. c. johnson . rheological effects of red blood cell aggregation in the venous network : a review of recent studies . , 38(2 - 3):26374 , 2001 . z. marton , g. kesmarky , j. vekasi , a. cser , r. russai , b. horvath , and k. toth . red blood cell aggregation measurements in whole blood and in fibrinogen solutions by different methods . , 24(2):7583 , 2001 . b. neu and h. j. meiselman . depletion - mediated red blood cell aggregation in polymer solutions . , 83(5):24822490 , 2002 . s. asakura and f. oosawa . on interaction between two bodies immersed in a solution of macromolecules . , ( 22):12551256 , 1954 . h. baumler and e. donath . does dextran indeed significantly increase the surface - potential of human red - blood - cells . , 120(2):113122 , 1987 . e. evans and d. needham . attraction between lipid bilayer membranes in concentrated solutions of nonadsorbing polymers : comparison of mean - field theory with measurements of adhesion energy . , 21(6):18221831 , 1988 . p. steffen , c. verdier , and c. wagner . quantification of depletion - induced adhesion of red blood cells . , 110(1):018102 , 2013 . h. bumler , b. neu , e. donath , and h. kiesewetter . basic phenomena of red blood cell rouleaux formation . , 36(5 - 6):439442 , 1998 . n. maeda , m. seike , s. kume , t. takaku , and t. shiga . fibrinogen - induced erythrocyte aggregation : erythrocyte - binding site in the fibrinogen molecule . , 904(1):8191 , 1987 . d. e. brooks . the effect of neutral polymers on the electrokinetic potential of cells and other charged particles : ii . a model for the effect of adsorbed polymer on the diffuse double layer . , 43(3):687699 , 1973 . s. chien and k. m. jan . red cell aggregation by macromolecules : roles of surface adsorption and electrostatic repulsion . , 1(4 - 5):385409 , 1973 . s. chien . shear dependence of effective cell volume as a determinant of blood viscosity . , 168(3934):977979 , 1970 . jan and s. chien . role of surface electric charge in red blood cell interactions . , 61(5):638654 , 1973 . s. svetina , d. kuzman , r. e. waugh , p. ziherl , and b. zeks . the cooperative role of membrane skeleton and bilayer in the mechanical behaviour of red blood cells . , 62:107113 , 2004 . m. brust , o. aouane , m. thibaud , d. flormann , c. verdier , l. kaestner , m. w. laschke , h. selmi , a. benyoussef , t. podgorski , g. coupier , c. misbah , and c. wagner . the plasma protein fibrinogen stabilizes clusters of red blood cells in microcapillary flows . , 4:4348 , 2014 . d. flormann , e. kuder , p. lipp , c. wagner , and l. kaestner . is there a role of c - reactive protein in red blood cell aggregation ? , 37(4):474482 , 2015 . o. baskurt , m. hardeman , m. rampling , and h. meiselman . . ios press , 2007 . badr kaoui , george biros , and chaouqi misbah . why do red blood cells have asymmetric shapes even in a symmetric flow ? , 103(18):188101 , 2009 . dmitry a. fedosov , matti peltomaki , and gerhard gompper . deformation and dynamics of red blood cells in flow through cylindrical microchannels . , 10:42584267 , 2014 . o. aouane , m. thibaud , a. benyoussef , c. wagner , and c. misbah . vesicle dynamics in a confined poiseuille flow : from steady state to chaos . , 90:033011 , sep 2014 . canham . the minimum energy of bending as a possible explanation of the biconcave shape of the human red blood cell . , 26(1):6181 , 1970 . w. helfrich . elastic properties of lipid bilayers : theory and possible experiments . , 28(11 - 12):693703 , 1973 . g. ghigliotti , t. biben , and c. misbah . rheology of a dilute two - dimensional suspension of vesicles . , 653:489518 , jun 2009 . b. neu and h.j . depletion - mediated red blood cell aggregation in polymer solutions . , 83(5):24822490 , 2002 . yaling liu and wing kam liu . rheology of red blood cell aggregation by computer simulation . , 220(1):139154 , 2006 . m. thibaud and c. misbah . rheology of a vesicle suspension with finite concentration : a numerical study . , 88:062707 , dec 2013 . the research leading to this results has received fundings from the german research foundation ( dfg , sfb 1027 ) , the centre national de la recherche scientifique ( cnrs ) and the german french university ( dfh / ufa ) . t. p. and c. m. acknowledge support from cnes and labex tec21 . d. f. performed experiments and data analysis , o. a. ran numerical simulations , l. k. supervised the experiments , c.r . programmed the analysis software , s.s . compared 3d analysis , c. m. supervised numerical simulations , t. p. designed the research , made the analytical model , supervised data analysis , c. w. designed the research , supervised experiments and numerical simulations . a simple way to quantitatively characterize the clustering of the rbcs is based on the quantification of sedimented aggregates within a certain area on a microscope slide . chien and jan @xcite defined a microscopic aggregation index ( mai ) as the ratio of the total number of cells to the total number of `` particles '' ( individual cells and aggregates ) , representing the average number of cells per particle . here , we investigate rbc aggregation by confocal microscopy and define the mai as the the ratio of the number of aggregated cells at the bottom of the well after sedimentation to the total number of cells , i.e. the fraction of aggregated cells . all observations were performed at a temperature of 23 @xmath50c with a hematocrit of 10 % . this concentration ensured an almost fully covered bottom surface after 5 min of sedimentation for most sample solutions , with an average coverage of @xmath52 cells / mm@xmath13 . imaging was performed using a confocal microscope ( tsc sp5 ii ; leica microsystems , mannheim , germany ) with a laser of 633 nm as the excitation wavelength . rbcs were classified as part of a cluster if the membranes of two cells were optically indistinguishable over a distance of at least 4 pixels , which corresponds to 1 @xmath12 m in our image with a size of 144 x 144 @xmath12m@xmath13 . white - light microscopy was used as a control to exclude any possible influence of cellmask on aggregation ( data not shown ) . fibrinogen and dextran solutions were compared to pbs as negative control . due to the increase of viscosity with increase in dextran or fibrinogen concentration , the sedimentation slows down , which leads to a lower total number of cells per image at higher viscosities in comparison to rbcs in pbs . in dextran solutions , the number of _ rouleaux _ and the mai exhibit a bell shaped curve with a maximum around a concentration of 50 mg / ml and almost no _ rouleaux _ at 120 mg / ml . [ microindex](b ) ) . for fibrinogen , a monotonous increase of the mai is seen up to a concentration of 12mg / ml , which corresponds to the physiological range . both results are reminiscent of the dependencies of interaction energy , in - flow clustering and rheology on dextran concentration obtained in previous studies @xcite . another qualitative observation is that the dextran and fibrinogen induced _ rouleaux _ look slightly different with more interconnections in the fibrinogen case . all measurements of contact zone shapes of rbc clusters were performed at 0.2% v / v hematrocrit . the procedure for extracting interfacial shapes for further analysis is exemplified in fig . [ figcontour ] . the contact zones of rbc clusters are first isolated . the images are then thresholded and skeletonized to extract the interface shape on which a sine function can be fitted to measure the amplitude of deformation . when the doublets were allowed to sediment on the bsa treated cover slip , the cells remained in discocyte state for several minutes but they were mostly oriented with their interface parallel to the coverslip which did prevent a characterization of the thin interfacial membrane that was then oriented perpendicular to the z - axis of the point spread function . this was different for the non - treated cover slips : due to the so called `` glass effect '' , the rbcs transformed into the echinocyte state and we observed only male - female and sigmoid shapes ( fig . [ figechinocytes](a ) ) . a quantitative analysis of the distribution of the different interfacial shapes for the different concentrations of dextran is shown in fig . [ figechinocytes](b ) . for all different cases we observed an increase in the number of sigmoid interfaces in the intermediate range of polymer concentrations , i.e. at larger interaction strengths . the sigmoid shape allows for the largest contact zone , favorized by large interaction strength . at low and high concentrations of dextran we observe only few sigmoid shapes and more parachute than male - female shapes for the cells in discocytic state . with fibrinogen , mostly parachute shapes were observed . this is most likely related to the low ( but physiological ) concentrations of the protein . for cells in echinocyte state no parachute shape was observes which is to be expected in view of the spherical shape of a single echinocyte . based on several successful studies ( especially equilibrium shapes and shape dynamics under external flow ) that have shown that 2d and 3d models capture the same essential features @xcite , have restricted ourselves here to a 2d model . in 2d the notion of in - plane shear elasticity looses its meaning and the vesicles and inextensible capsule model ( often evoked to model rbc ) are equivalent . vesicles consist of closed bilayer membranes enclosing an inner fluid and suspended in an outer fluid . using the typical reduced area ( equivalent of the surface - to - volume ratio in three dimensions ) of the rbc , the minimum energy shape of an unstressed vesicle is the same as for an rbc despite the lack of a cytoskeleton . indeed , the stress free shape is dictated by the curvature energy and the rigidity of the cell . to describe the mechanical behavior of the vesicle , we employ ( for each cell ) the two dimensional helfrich free energy@xcite ( per unit length in the 3rd dimension ) : @xmath53 where @xmath54 is the mean curvature , @xmath55 is the arclength element along the membrane contour , @xmath56 is the membrane bending rigidity modulus , and @xmath57 is the spontaneous curvature that can be neglected in case of symmetrical membranes ; actually in 2d the spontaneous curvature plays no role . in order to fulfill the condition of membrane inextensibility we supplemented the curvature energy with a term @xmath58 , where @xmath59 is a lagrange multiplier that depends on the position along the membrane . the fluid inside and outside the membrane is described by the stokes equations ( the inertial effect is small and is neglected ) with boundary conditions at the membrane and far away from the membrane . far away from the membrane the flow vanishes ( in the absence of any imposed external flow ) . at the membrane we require velocity continuity , and force balance . in addition , the velocity field along the membrane has to be divergence - free in order to enforce membrane incompressibility . this condition represents an implicit equation that allows one to determine the still unknown lagrange multiplier @xmath59 ; the details of how this is achieved in practice can be found in @xcite . the hydrodynamics stress jump across is counterbalanced by the membrane force . this force is composed of the bending force obtained as a functional derivative of the helfrich free energy supplied with a tension - like force to fulfill the condition of inextensibility of the membrane . @xmath60\mathbf{n } - c \zeta \mathbf{n } + \frac{\partial{\zeta}}{\partial{l}}\mathbf{t}\]]where @xmath61 and @xmath62 are the normal and tangent unit vectors . @xmath63 designates the vector position of a given point on a given cell ; the curvature as well as the lagrange multiplier are defined for each cell . in addition to this force we had to consider the force due to cell - cell interaction ( due to the presence of proteins or dextran ) . for this purpose we refer to the work of neu and meiselman @xcite . neu and meiselman have presented a model that takes into account the strong electrostatic repulsion due to the negative charge of rbc membranes , and the weak osmotic forces of attraction due to the depletion effect induced by macromolecules surrounding rbcs @xcite . this model contains several parameters depending on the physiochemical properties and the concentration of the polymers . often , the interaction energy between cells is described by a morse potential ( or alternatively a mie type potential ) , having an exponential form , which has proven to adequately describe experimental data @xcite . as is the case in many physical problems , a morse potential can be well approximated by a lennard - jones potential , and we adopted this form in the present study for practical reasons . the interaction potential per unit ( length of the 1st membrane @xmath64 length of the 2nd membrane @xmath64 length in the 3rd ( invariant ) dimension ) is @xmath65\ ] ] where @xmath66 is a microscopic parameter with dimension [ energy / length@xmath67 . in this formulation , @xmath66 is a microscopic parameter of interaction between surface elements . given the range of the interaction potential ( typically a few times @xmath78 ) , an integration of the potential over planar parallel surfaces of extension much larger than this range must be performed in order to derive the macroscopic interaction energy per unit area @xmath11 . the zero force position is therefore @xmath81 , which is smaller than @xmath82 . the interaction energy per unit surface between infinite plates ( _ i.e. _ whose extension is much larger than @xmath83 ) is then : recently the rbc - rbc adhesion energy has been quantified using the atomic force microscopy - based single cell force spectroscopy @xcite . from these data , in principle , we can determine the adhesion energy @xmath11 and equilibrium distance @xmath83 from which the two parameters entering the lennard - jones potential can be determined , namely the minimal energy @xmath66 and the range of the potential @xmath78 . however , here we chose to adapt the results from the numerical simulations , namely the amplitude of the interfacial deformation , directly to the experimental data and use this to deduce the interaction energy . in this procedure , due to numerical constraints , the zero force length ( i.e. the length corresponding to minimal potential ) was fixed in our simulations to 489 nm . this was probably significantly larger than the experimental value but still in the order of the optical resolution limit . in addition , note that as long as @xmath78 is small compared to the lateral extension of the membrane , the product @xmath85 is the most quantity here . the stokes equations were solved using the boundary integral technique as presented in @xcite @xmath86 dl(\mathbf{x } ) } \\+ \frac{(1-\lambda)}{2 \pi ( 1 + \lambda)}\int\limits_{\sum_{i}{\partial \omega_{i}}}\mathbf{u}(\mathbf{x } ) \cdot \mathbf{t}(\mathbf{x},\mathbf{x_0 } ) \cdot \mathbf{n}(\mathbf{x})dl(\mathbf{x } ) \ ] ] where @xmath87 is the velocity of a vector position @xmath88 belonging to the membrane @xmath89 , @xmath90 is the viscosity contrast between the inner and outer fluids with the respective viscosities @xmath91 and @xmath92 , and @xmath93 and @xmath94 are the green s functions associated with the velocity and the stress fields respectively . to match the average values of a human rbc , the reduced area ( @xmath14 ) of the vesicle typically lies in the range ( @xmath95 ) , the typical radius is @xmath96 , the membrane stiffness is @xmath97 . here we did not investigate any dynamic effects and we therefore fixed the viscosity ratio simply to unity , even if typically a factor of 5 was assumed . we considered doublets and sets of @xmath98 cells , separated by an initial surface - to - surface interdistance of 600 nm . we provide here the details of the analytical model showing buckling of contact zones in symmetric rbc doublets . we suppose that the shape of an rbc in a symmetric doublet ( in 2d ) is approximated as shown in fig . [ fig - model ] , with two semi - circular caps of radius @xmath17 , a roughly straight portion where curvature is negligible ( length @xmath18 ) and a contact zone with a sinusoidal shape of curvilinear length @xmath19 and amplitude @xmath20 described by the equation : in the following we define a dimensionless energy @xmath28 , a rescaled interaction energy @xmath106 , and rescale all lengths by @xmath8 : @xmath30 and so on , and use the definition of the reduced area @xmath7 . the constant perimeter and area constraints can then be written as the following ( dropping the * for simplicity ) : @xmath107 and can be used to eliminate @xmath17 and @xmath18 as in equations [ rexpr ] and [ l1expr ] . if we solve @xmath111 to find the equilibrium configurations we get the two solutions whose expressions are given by [ aeq ] , which defines the bifurcation threshold @xmath34 corresponding to @xmath38 ( eq . [ betac ] )

timing analysis , other than being an important diagnostic tool by itself ( see * ? ? ? * for a review ) , is particularly powerful for dissecting the inner accretion regions in black hole binaries and agn when combined with spectral modeling . good examples of this combination in binaries are studies of the variability of a specific spectral component ( e.g. the iron k@xmath3 line and the reflection component , * ? ? ? * ; * ? ? ? * ) , the study of the time lags between different energy bands and the comparison with the expected behavior of a given spectral component ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , the study of the covariance of the signal at multiple energies and the variability spectrum ( e.g. * ? ? ? * ; * ? ? ? a review of reverberation lag measurements in binaries and agn can be found in @xcite . the _ nuclear spectroscopic telescope array _ ( _ nustar _ ; @xcite ) mission has deployed the first hard x - ray ( above 10kev ) focusing satellite . it features two telescopes , focusing x - rays between 3 and 79kev onto two identical focal planes ( usually called focal plane modules a and b , or fpma and fpmb ) . it has a field of view ( fov ) of 12@xmath412and an angular resolution of 18fwhm ; ( 58hpd ) . these features , together with good spectral resolution in the iron k band ( @xmath00.4kev@6kev ) , have motivated a large number of observations of x - ray binaries . a particularly interesting result was the accurate modeling of the reflection component in several black hole ( bh ) binaries @xcite resulting in improved constraints on black hole spin in several systems . such measurements have also been made by _ nustar_for active galactic nuclei ( agn ; e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? _ nustar _ , _ xmm _ and _ suzaku _ have recently measured reverberation lags in the agn mcg-5 - 23 - 16 in both the iron line and compton reflection hump together for the first time @xcite . from the point of view of timing , _ nustar_has an advantage over other imaging satellites : the time resolution is 10@xmath5 , and so one can in principle study variability over the whole range of interesting frequencies in accreting systems without switching to an observing mode with decreased spectral resolution . the satellite has in fact been used to perform timing analysis on selected targets with very good results : for example , it was used to study a number of accreting and rotation - powered pulsars , permitting the detection of variable cyclotron resonant scattering features @xcite and the discovery of strong hard time lags in the emission of another accreting pulsar ( gs 0834 - 430 ; @xcite ) . _ nustar_was also used to detect the pulsations from a magnetar near the galactic center @xcite and to carry out a detailed timing study of this serendipitous source @xcite . these pulsars are mostly slow rotators , with pulse periods above 1s . aperiodic timing has been measured in faint sources ( e.g. * ? ? ? * ) , or in bright sources but only over the low - frequency part of the power spectrum ( e.g. * ? ? ? * ) . from the point of view of fast ( @xmath6hz ) timing studies of bright sources , however , s default observing mode requires care to analyze properly . each x - ray incident on a focal plane produces a trigger , and the energy signal is read out immediately . the read time is dependent on the number of triggered pixels , and so the dead time for each event is slightly variable ( at the few percent level ) with an average value of @xmath7ms . this dead time , together with additional dead time produced on a separate cadence by housekeeping operations and vetoed events , can be accurately measured and its value over a second is stored in the housekeeping files of the observations . in addition , the live time since the previous x - ray event is recorded in a column in the event lists ( see section sec : deadtime for details ) . for bright sources where the fraction of events vetoed by the anti coincidence system is small , light curves can be accurately produced with any time binning ( see madsen et al 2014 , _ submitted _ , for application of this technique to the crab pulsar ) . however , for timing analysis ( e.g. the production of power density spectra ) , the dead time introduces spurious correlations between event arrival times that are reflected in a distortion of the power spectrum , making it difficult to subtract the poisson noise level above @xmath8hz . this effect is almost completely negligible for count rates up to @xmath9@xmath10 , but it is very pronounced in observations of galactic x - ray binaries during outbursts , with typical incident count rates @xmath11@xmath10over the full band . in this paper , we present a technique to avoid spurious noise introduced by dead time , in particular the distortion of the white noise level in the power spectrum . we show how a slight modification to the methods commonly used for quasi - periodic oscillation ( qpo ) detection and power spectral fitting enables good - quality results at all frequencies . we use these techniques to perform a basic timing analysis of several _ nustar_observations of bh binaries ( bhbs ) , whose spectral analysis is presented in other papers ( grs1915@xmath2105 , @xcite ; cygx-1 , @xcite ; gx339@xmath14 , frst et al . _ in prep . we chose these bhbs as ideal objects with which to demonstrate the timing techniques , since their timing behavior has been extensively studied by other missions and instruments ( see section sec : analysis for details ) . the paper is organized as follows : in section sec : deadtime we discuss details _ nustar _ s dead time , and how it is measured on - board the instrument ; in sections sec : cpds and [ sec : simu ] we describe how to use the cross power density spectrum to obtain a proxy of the pds ; in sections sec : red and [ sec : analysis ] we use these techniques to analyze the data from two black hole x - ray binaries , cyg x-1 and gx 339 - 4 ; and in section sec : concl we provide conclusions . in the appendix we present results from a timing analysis of grs 1915 + 105 , which was observed during the on - orbit commissioning period . this observation suffered from spurious features associated with the cadence of instrument housekeeping functions and so requires a unique analysis not relevant to _ nustar _ science phase observations . dead time in the _ nustar _ instrument is produced when the focal plane module electronics are busy processing an event , when the shield veto prevents the focal plane from triggering ( or interrupts an event that is being processed ) , and for some regular instrument housekeeping functions ( the housekeeping produces less than 1 ms of dead time per second ) . _ has two independent telescopes , and the associated focal plane modules have independent processors , so that the dead time in the two modules is uncorrelated . all of the events processed by the instrument are found in the unfiltered event file . the cleaned event file is a subset of the events in the unfiltered event file where gtis and additional screening have been applied to remove non - science quality events . for the production of light curves and flux measurements the _ nustar _ instrument has two operating modes , one for faint ( @xmath1250 mcrab ) sources , and one that allows highly accurate flux measurements even at very high count rates and on arbitrarily short timescales . the fraction per second of dead time from events that are not processed , like shield vetoes , as well as that due to housekeeping functions is saved in the housekeeping files on a one - second cadence . for light curves binned in 1 s intervals ( or a integer multiple thereof ) multiplying by the live time stored in the housekeeping files correctly accounts for dead time . in addition , the live time since the previous x - ray event is stored in the event lists in the prior " column . in the absence of events vetoed by the active anti - coincidence shield , this column would accurately reflect the true live time . in the operating mode used for most observations taken since in - orbit checkout , the events vetoed by the anti - coincidence shield are not telemetered to the ground to minimize data volume . for this mode adding the prior column does not yield the proper live time since vetoed events are not included . however , for source count rates significantly greater than the veto rate the error is negligible ( for the crab it is only @xmath13% ; see madsen et al . 2014 , _ submitted _ ) . for sources of intermediate brightness , where the veto rate can not be ignored , a fully tested and calibrated instrument mode exists that includes vetoed events in the data stream so that fluxes and light curves can be produced with arbitrary accuracy . for timing analyses that involve power spectral analysis , dead time produces systematic effects even if it is perfectly measured . dead time is classified as either paralyzable or non - paralyzable depending on whether a photon hitting the detector during dead time produces new dead time or not . the _ nustar _ readout architecture produces non - paralyzable dead time . its main effect is on recorded count rates : if we have an incident rate of photons @xmath14 , the detected count rate will be approximately @xmath15 ( see , e.g. , * ? ? ? * and references therein ) , where @xmath16is the dead time produced by each event , assuming that it is constant . but dead time also alters the sensitivity to variable signals , acting as a frequency filter . power density spectra ( pds ) , in particular , are deformed to a `` wavy '' shape that depends on the magnitude of dead time and on count rates ( see figure fig : cpdsvspds ) . power at frequencies slightly above @xmath17 is _ quenched _ , as there is a lack of events whose separation is less than @xmath18 , while there is a relatively higher rate above @xmath18 , and therefore the power at frequencies just below @xmath17 is slightly _ amplified_. these `` waves '' have nodes at @xmath17 and multiples thereof , where the power ( in @xcite normalization ) is equal to 2 , the value that it would have without dead time , and maxima and minima in between are given by the relative contribution of the quenching and amplification . for frequencies @xmath19 , the main effect is a general deficiency of events , and the power has a decreasing level that approaches @xmath20 @xcite . assuming that @xmath16is the same for each event and that only source events contribute to either non - paralyzable or paralyzable dead time , this distortion can be modeled precisely ( see * ? ? ? * ; * ? ? ? * for an exhaustive treatment ) . also , some statistical properties of the pds hold in dead time - affected data . for example , the standard deviation associated with the bin @xmath21 of the pds is always equal to @xmath22 , where @xmath23 is the power in the bin @xmath21 and @xmath24 is the number of averaged pdss ( see figure fig : cpdsvspds ) . in _ nustar_@xmath16is not strictly fixed at the same value for all events , but varies by a few percent depending on the number of pixels that are triggered . for this reason , the models available in literature do not correctly describe the dead time effects for this satellite : the `` wavy '' behavior of the pds shifts slightly , and to fully account for this effect and produce a white - noise subtracted pds , a very precise modeling of the dead time would be required . since at high count rates the `` waves '' can be very prominent , any real variability feature such a qpo can easily be `` hidden '' and difficult to detect . as an additional complication , the models described above assume that dead time is produced completely by the recorded signal . in _ nustar _ , additional dead time comes from events outside the source extraction region , from vetoed events , and from all events discarded for other reasons during the cleaning process in the pipeline ( the step from unfiltered to cleaned event files ) . in the following , we neglect the effect of vetoed events , since their dead time ( @xmath25s ) has characteristic frequency @xmath26khz , much higher than science events , and their total contribution to dead time is small . we instead present a method that permits construction of a proxy of a white - noise subtracted pds , regardless of the count rate and the ratio between source and background ( or spurious ) events . _ nustar_has two completely independent focal plane modules ( each containing four detectors ) that are read out by separate microprocessors . it is therefore possible in principle to obtain the same information given by a pds through the so - called cross power density spectrum ( cpds ; for more details see @xcite ) : instead of considering the pdss in the two individual focal planes : @xmath27 where @xmath28 indicates the fourier transform of the lightcurve detected by the focal plane @xmath21 and @xmath29 is the frequency , one multiplies the complex conjugate of one fourier transform with the other fourier transform : @xmath30 the cpds is often used in other contexts to obtain information on the correlation between the signal in two energy bands . it is a complex quantity : its real part is also called the _ cospectrum _ and gives a measure of the signal that is _ in phase _ between the two channels ; its imaginary part , or _ quadrature spectrum _ , gives instead a measure of the _ off - phase _ signal . therefore , in principle , it should be possible to eliminate all variability that is not related between the two light curves , including the effects of dead time , by just considering the cospectrum ( the real part of the cpds ) . in figure fig : cpdsvspds we show the statistical properties of the cospectrum in the case of pure poisson noise . in both the dead time affected and in the zero - dead time cases the cospectrum mean value is 0 ( in figure fig : cpdsvspds , it has been shifted to 2 for graphical reasons ) . this is a big advantage , as this is independent of whether the dead time is constant or not ( since the distribution of dead time is also independent between the two detectors ) , and therefore it is not necessary to conduct complicated studies of the dead time distribution in order to obtain a white noise - subtracted cospectrum , as opposed to a pds where this procedure would be needed . another important point is that the standard deviation @xmath31 of the cospectrum is linked to the standard deviations ( @xmath32 ) of the pdss of the two light curves . in figure 1 , we show that the ratio between @xmath31 and the geometric average of the standard deviations of the two pdss is close to @xmath33 regardless of the frequency . this is very convenient , as it allows us to assign to the cospectrum bin @xmath34 an uncertainty of @xmath35 , where @xmath36 is the geometric average of the two pdss and @xmath24 is the number of averaged pdss . the argument ( or angle ) of the cpds is called _ phase lag _ ; if we divide it by @xmath37 we obtain the _ time lag _ @xmath38 , that is a measure of the time shift between the two channels . in our case , since we are using two lightcurves in the same energy bands , we expect any time lag to be zero . but time lags between different energy bands have been used as indicators of how a signal produced in an emitting region is reprocessed in other regions : for example , if the disk emission is comptonized in a corona , or when the signal from a region is reflected or propagated in another region , this time delay between the different emissions can be detected through time lags ( see , e.g. , @xcite and references therein ) . we briefly discuss time lags between different energy bands in the following sections , but for the most part our analysis will be performed on two channels in the same energy band , and the only important quantity for our treatment will be the cospectrum . amplitude ( 8% ) at different frequencies . the units in this plot , and in all following power spectra or cospectra if not stated otherwise , are power@xmath4frequency . grey points show the standard , dead time - affected pds . the white noise level to subtract in the dead time - affected datasets was calculated between 10 and 20hz ( minimum between two qpos ) , while in the dead time - free case we subtracted the theoretical level ( 2 in leahy normalization ) . ] in order to provide a fully consistent treatment of dead time effects in the pds and in the cospectrum , we ran a large number of simulations . in each simulation we produced two event series containing variability , one for each fpm , and analyzed the data with cospectrum and pds before and after applying a dead time filter . by doing so , we studied in detail the properties of the cospectrum and compared them to those of the pds . in the following paragraphs we will explain the procedure in more detail , and demonstrate that the cospectrum can be considered a very good proxy of a white noise - subtracted pds , albeit with some corrections to account for in the measured @xmath39 . [ [ light - curve - generation ] ] light curve generation + + + + + + + + + + + + + + + + + + + + + + we used the procedure by @xcite , introduced to astronomers by @xcite , to simulate light curves between two times @xmath40 and @xmath41 , from a number of model pds shapes containing qpos . the sampling frequency of the light curves was at least four times higher than the maximum frequency of the variability components included in the simulation . we normalized the light curves in order to have the desired mean count rate and total @xmath39variability ( 7 - 10% ) . in order to be able to later calculate pdss with a given maximum timescale @xmath42 ( see `` calculation of the pdss and the cospectrum '' below ) , we simulated light curves at least 10 times longer , following the prescriptions in @xcite to avoid aliasing . [ [ event - list - generation ] ] event list generation + + + + + + + + + + + + + + + + + + + + + from every light curve , we generated two event lists , corresponding to the signal from the two focal planes . each event list was produced as follows : first of all , we calculated the number @xmath43 of event times to be generated as a random sample from a poisson distribution centered on the number of total photons expected ( summing up all expected lightcurve counts ) ; then , we generated @xmath43 events by using a _ monte carlo acceptance - rejection method_. this is a classical monte carlo technique ; a more general treatment can be found in most textbooks on monte carlo methods ( e.g. * ? ? ? * ) . in our case , we used the following procedure : ( 1 ) for every event we simulated an event time @xmath44 uniformly distributed between @xmath40 and @xmath41 , and an associated random amplitude ( `` probability '' ) value @xmath45 between 0 and the maximum of the light curve ; ( 2 ) we rejected all @xmath44 whose associated @xmath45 was higher than the light curve at @xmath44 ; to avoid possible spurious effects given by a stepwise model light curve , we used a _ cubic spline _ interpolation to approximate the light curve between bins ; ( 3 ) we sorted the event list by @xmath44 . to simulate the effects of background ( spurious events filtered out by the pipeline , events recorded outside the source regions , etc . ) , we also produced two background series , at constant average flux , one for each simulated source light curve . [ [ dead - time - filtering ] ] dead time filtering + + + + + + + + + + + + + + + + + + + for each event list , we created a corresponding dead time - affected event list by applying a simple dead time filter : for each event , we eliminated all events in the 2.5ms after it . source and background events contributed equally to dead time . we used different versions of the dead time filter by varying slightly the dead time between events ( of @xmath46ms ) . but the pernicious effect of variable dead time is mainly on white noise subtraction . in our case ( figure fig : cpdsvspds ) , the cospectrum allows us to overcome this problem as its white noise is 0 , and we verified that the other effects are not significantly different in the constant and variable dead time cases . in the following , we will treat the case with constant dead time . [ [ calculation - of - the - pdss - and - the - cospectrum ] ] calculation of the pdss and the cospectrum + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + we divided each pair of event lists in segments of length @xmath42 , calculated the pds in each of segments and the cpds from each pair of them . we then averaged the pdss and cpdss from all segments . the cpds white noise level is already 0 . for the pdss , used only in the ideal zero - dead time case , we subtracted the theoretical poisson level ( 2 in leahy normalization ) . we then rebinned the pdss and cpdss , either by a fixed rebin factor or by averaging a larger number of bins at high frequencies , following approximately a geometric progression . we finally multiplied the leahy - normalized pdss by @xmath47 ( where b is the background count rate and s the mean source count rate ) in order to obtain the squared @xmath39normalization often used in literature @xcite . the cpds was instead multiplied by the factor @xmath48 , where bars indicate the _ geometric _ averages of the count rates in each of the two event lists used to calculate it , it s as if each fourier transform contributing to it were multiplied by @xmath49 . since the contribution to the cpds is by two separate fourier transforms , we multiplied by @xmath50 , that is equivalent to the geometric averages described . ] . finally , we calculated the cospectrum by taking the real part of the cpds . as described above , we assigned to each final bin of the cospectrum @xmath34 an uncertainty calculated from the geometrical average of the pdss in the two channels , divided by @xmath51 , where @xmath24 is the number of averaged spectra and @xmath52 is the number of subsequent bins averaged to obtain @xmath34 . [ [ fitting - procedure ] ] fitting procedure + + + + + + + + + + + + + + + + + cospectra do not need poisson noise subtraction ; for pdss , we fitted a constant to the interval outside of the frequency range containing qpos . since in the following paragraphs , in all our examples , we will be showing power spectra obtained by averaging more than 50 pdss , we are in the gaussian regime and a fit with standard @xmath53 minimization routines is appropriate to the precision we are interested in @xcite . then , we fitted the qpos with a lorentzian profile in xspec @xcite . errors were calculated through a monte carlo markov chain , as those intervals where the @xmath53 of the fit with no frozen parameters increased by 1 . according to @xcite , the significance of the detection of qpos is expected to be @xmath54 where @xmath55 is the @xmath39and @xmath56 is the equivalent width of the feature ( for a lorentzian , @xmath57 ) . the exact proportionality factor depends on the definition of significance . in our case , the significance of qpos was defined as the ratio between the amplitude of the lorentzian and its error . the significance calculated in this way gives a value @xmath58 times lower than obtained by using the excess power la @xcite ( a factor 2 is expected due to the fact that the gaussian errors are calculated over @xmath59 , and the excess power over @xmath60 ; see * ? ? ? * ; * ? ? ? * ) , but the trend in equation eq : sig holds provided that the variability is dominated by poisson noise . of a qpo at different peak frequencies , measured with the various techniques and with and without dead time . each point represents a simulated qpo with @xmath39=10% and fwhm=2hz . a total of 281 simulations were used for this plot . ( bottom ) significance measured with each method . the total pds has about twice the significance of the single - module pds in the no - dead time case , as expected , due to double number of photons . the cpds in the no - dead time case is a factor @xmath61 higher than the single pds , and lower by the same amount than the total pds . the dead time - affected cpds , instead , has a much lower level due to the lack of photons . the decrease of significance does not depend on the frequency of the qpo , but only on count rate ( see figure fig : cts ) . ] . a total of 118 simulations were used in this plot . the line shows equation eq : drms . it is not a fit , and describes remarkably well the data . in the bottom panel , we plot the detection significance for all the cases . the lines , again , are not fitted , they just show the theoretical prediction from equation eq : sig . all dead time - free cases are in good agreement with a linear increase with count rate below @xmath62 , above which some curvature appears due to the departure from the quasi - poissonian regime . the dead time - affected case is in good agreement with equation eq : sig if , instead of the incident count rate ( dashed line ) , one uses the _ observed _ count rate ( solid line ; see equation eq : drms ) . ] on the relative contribution of the source in a given energy range to the total count rate . the 132 simulations that compose this plot show how the @xmath39drop is stronger if the source signal dominates the total signal , since the source signal contributes more to the total dead time . the solid black line shows equation eq : drmstot . it is not a fit . ] [ [ first - look ] ] first look + + + + + + + + + + the simulation in figure fig : cpdsvspds shows the comparison between the pds and the cospectrum with and without dead time for pure poisson noise . from these simulations it is clear that the white noise - subtracted pds and the cospectrum are equivalent in the case with no dead time . it is immediately evident that the most problematic effect of dead time , the modulation of the white noise level , disappears in the cospectrum . in the following paragraphs , we investigate the frequency and count rate dependence of these quantities in more detail . [ [ frequency - dependence ] ] frequency dependence + + + + + + + + + + + + + + + + + + + + the general statistical properties of the pds and the cospectrum are also very similar , both in the dead time - affected and in the zero - dead time case . figure fig : cpdsvspds , panel d , shows that the variance of the cospectrum and the pds maintains a constant ratio equal to 2 in both the clean and the dead time - affected datasets . this makes it easy to calculate the variance of the cospectrum values for the subsequent analysis , by simply using the known properties of the pds where the variance is just equal to the square of the power ( in leahy normalization ) . contrary to what might be imagined , it is possible to detect variability even at frequencies that are affected the most by dead time , i.e. those above @xmath17 . figure fig : qpos shows that qpos at all frequencies are detectable , albeit with some modulation of the observed @xmath39 . to measure this change of @xmath39 , we simulated @xmath0500 lightcurves using the method above , each containing a single qpo with frequencies equally distributed between 5 and 1000hz , @xmath39=10% and fwhm=2hz . as explained above , from every light curve we obtained two event lists in order to simulate the signals from the two detectors . we produced for every pair of event lists the cospectrum , the two pdss , and a total pds including the counts from both detectors , both in the zero - dead time and the dead time - affected cases . we then fitted the resulting spectra with a lorentzian model in xspec . figure fig : freq shows the change of the @xmath39measured with the pds and the cospectrum , with and without dead time . the measured @xmath39 in the zero - dead time pdss agrees with the zero - dead time cospectrum , whereas the dead time - affected cospectrum yields a frequency - dependent deviation from the true @xmath39 , with deviation following the same trend as the variance ( see also figure fig : cpdsvspds ) . the detection significance does _ not _ depend on the frequency in any case , with or without dead time . the decrease of the significance is instead driven by the observed count rate , as we discuss shortly . in the no - dead time case , the significance of the single pdss is about half that of the total pds because the significance is directly proportional to the intensity of the signal ( equation eq : sig ) , and in the total pds one uses twice the number of photons . the zero - dead time cospectrum yields instead a significance @xmath61 lower than the total pds , and higher by the same amount than the single - module pds . this is just an effect of the factor @xmath33 between the standard deviations of the cospectrum and the single - module pds . from figure fig : freq , it is clear that it is advantageous to use the total pds for low count rates where dead time is negligible , and the cospectrum otherwise , but with the formulas and the simulations shown above to account for the frequency - dependent distortion of the @xmath39amplitude . in summary , the important point that figure fig : freq makes is that qpos are still detectable at any frequency , even those heavily affected by dead time , albeit with a change of the measured @xmath39that must be taken into account . [ [ count - rate - dependence ] ] count rate dependence + + + + + + + + + + + + + + + + + + + + + we now investigate how the measured @xmath39is influenced by count rate . figure fig : cts and [ fig : ctsback ] show the variation with count rate in the detected @xmath39of a qpo at 30hz , fwhm=2hz and @xmath39=7.5% , in two cases : increasing total count rate , and fixed total ( source + background ) count rate with variable source count rate . for the first case , since @xmath63 and the higher - order corrections are not needed , we use van der klis ( 1989 ; equations 3.8 and 4.8 ) to obtain @xmath64 this relation is plotted with a dashed line in the top panel of figure fig : cts and it is in remarkably good agreement with the simulated data . in general , one would expect the significance of detection in the pds to be proportional to the incident count rate and to the square of the @xmath39(equation eq : sig ) . this condition holds if the qpo can be considered a small disturbance in an otherwise poissonian process , or @xmath65 @xcite . in the bottom panel of figure fig : cts , we fit equation eq : sig below 600@xmath10 , with a multiplicative constant due to the slightly different definition of 1-@xmath66 error that we use ( @xmath67 instead of the @xcite definition ) . the best - fit multiplicative constant , @xmath68 , turns out to be consistent with the factor 2 expected from the fact that we are using gaussian fitting instead of excess power ( see section sec : simutec ) . the departure from the linear condition above @xmath69cts is evident . indeed , it is expected that at count rates above @xmath70 , the significance starts departing from the linear trend . the total pds is visibly more affected because its count rate is double than that of the single - module pds . the significance of detection with the cospectrum is @xmath61 lower than that of the total pds , while it is @xmath61 higher than that of the pds from a single module . the dead time - affected cospectrum , instead , has a large deviation from the linear trend . this is just an effect of the diminishing count rate due to dead time . in fact , what is plotted is the _ incident _ count rate . if one converts it to the _ detected _ count rate , the linear relation between count rate and significance still holds ( solid line ) . the second case ( figure fig : ctsback ) clearly shows a linear decrease of the measured @xmath39as the source gains counts with respect to the background . again , by using van der klis ( 1989 ; equations 3.8 and 4.8 ) , but this time putting the _ total _ count rate ( @xmath71 , where @xmath72 is the non - source count rate ) in the relation between incident and observed count rates , one obtains @xmath73\ ] ] this means that the measure of @xmath39we obtain in our data will be generally affected more if the source signal dominates the background , as is the case in most _ nustar_observations of bright sources . in the examples that we present in the next sections , we make use of monte carlo simulations similar to the ones above to estimate the change of @xmath39at the count rate of the sources we observe . we now demonstrate the methods outlined above by analyzing the power density spectra for a number of bright galactic black hole binaries . a summary of the observations used in this work is listed in table tab : obs . we preprocessed all _ nustar_observations with the nustardas pipeline included in heasoft 6.15 . we produced cleaned event files and calculated good time intervals ( gtis ) . we processed both the unfiltered and the clean event files ( files tagged with `` _ uf '' and `` _ cl '' respectively in the `` event_cl '' directory of _ nustar_data ) . by analyzing the unfiltered and cleaned event files , it is possible to estimate the ratio of `` good '' events to the total of dead time - producing events . we calculated the ratio @xmath74 of `` good '' events over intervals of 100s . we found that in all observations @xmath74 was almost constantly @xmath75 , with some drops due to increased solar activity . we selected the intervals in which this ratio did nt go below a certain threshold ( ranging from 0.9 to 0.92 in different files ) , in order to eliminate possible spurious effects on the observed variability . in the subsequent analysis , we will always consider an additional 10% `` background '' event rate due to unrecorded and spurious events . we selected events from a region around each source depending on how broad the point - spread distribution was ( see the next subsections for details for each source ) . we used the barycorr ftool to calculate event times at the solar system barycenter using the de200 ephemeris and source positions from simbad . additional variability is introduced into the cross - power spectrum in some cases due to mis - matched gti between fpma and fpmb . this mis - match occurs often when the satellite traverses the south atlantic anomaly ( saa ) because the automatic algorithm that turns off the focal plane modules to conserve telemetry kicks in at different thresholds for fpma and fpmb . the parameters that rule this behavior are finely tuned in order to maximize the number of photons available for spectral analysis while avoiding most of the contamination @xcite , but for timing analysis this produces unwanted effects that have to be accounted for . a signature of this variability is a strong time lag between the two detectors in the same energy band ( therefore , unrelated to the physics of the source ) , that disappears when discarding an additional interval at the gti borders or by using a more aggressive filtering during the pipeline process . we decided to use a `` safe '' additional cut of 300s at the borders of all gtis during timing analysis , and verified that it solved the problem in all affected obsids . from each barycentered event list ( section sec : red ) we obtained light curves sampled at @xmath76s ( @xmath77s ) . we extracted separate light curves for each detector , and calculated the common good time intervals ( gtis ) . each light curve was then divided in segments ( the length was chosen in order to optimize the use of gtis ) with the start and stop times locked between the two detectors , and a fourier transform was calculated from each of them . we then calculated the cpdss from each segment by multiplying the fourier transforms as explained in section sec : deadtime , and obtained a single cospectrum from the real part of the average of all cpdss . when needed , we plot the error on the best - fit model as a hatched region around the model line . to calculate these errors , we ran a monte carlo markov chain on xspec to simulate 10000 models around the best fit . we chose all the models having @xmath78 , and we calculated at each energy the maximum deviation from the best fit coming from any of these models . cygx-1is a bright persistent bhb , discovered during one of the earliest rocket flights at the dawn of x - ray astronomy @xcite . as such , it is among the best studied x - ray sources in the galaxy . its spectral evolution is typical for a bh , with the succession of low / hard , high / soft and intermediate states with correspondingly different timing patterns ( see * ? ? ? * for a broad view on bhbs ) . variability in all spectral states has been observed . it shows two dominant lorentzian components in the low / hard state . when the source softens their characteristic frequencies increase and an additional cutoff power law - like component appears . in the soft state the latter dominates the power spectrum ( * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * and references therein ) . time lags have been observed in all spectral states @xcite ; the variability of single spectral components was also investigated , showing that the reflection component shows variability at lower frequencies than the bulk of the emission in the low / hard state @xcite while the @xmath39level and shape of the power spectrum is comparable at all energies in the soft state @xcite . a very thorough investigation of the relation between variability in different energy ranges and spectral states can be found in @xcite . for past studies of the high - energy variability of the source see @xcite . _ nustar_observed cygx-1 in its soft state ( for a full spectral analysis description , see * ? ? ? the spectral analysis shows a strong reflection component , that might produce detectable phase lags . phase lags have been observed with _ rxte_in all spectral states ( @xcite ; see @xcite for a rigorous treatment ) . the pds in this state is often well described by a cutoff power law - like shape almost `` flat '' between @xmath79 and @xmath80hz ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? this is the case in our observations , as shown in figure fig : cyg_pdsvse . there is a clear change in @xmath39variability between the two observations , and a power law - like red noise component appears at the lowest fourier frequencies in only one observation , but for the rest the cospectrum is very well described by a cutoff power law . as shown in figure fig : cts , a drop of @xmath39is expected at higher count rates , and the second observation has indeed a higher count rate than the first . but in figure fig : cyg : rms we show that the expected @xmath39drop due to dead time is far lower than observed . this means that the rms change is largely due to an intrinsic lower variability of the source in the second observation . more interesting is the cospectrum in different energy ranges , shown in figure fig : cyg_pdsvse . there is a very strong change in the shape and normalization of the cospectrum with energy . in general , the @xmath39increases with energy , and it is about an order of magnitude higher above 10 kev than below 5 kev . this is a clear indication that the soft component of the spectrum ( coming from the disk ) is varying less than the reprocessed component ( reflected and comptonized ) . again , in order to evaluate the @xmath39change with energy , it is necessary to apply a correction for dead time . in fact , channels with more counts will have a larger drop of @xmath39than channels with a few counts ( compare to figure fig : ctsback ) . following the procedure in section sec : simu , we simulated the expected relative drop in @xmath39 in all bands . in figure fig : cyg : rms we show the measured @xmath39and the curves showing the expected drop due to count rate . these curves were calculated with the method depicted in figure fig : cts , assuming a total count rate @xmath81 times the source count rate , as the total background plus the rejected events account for about 10% of the contributions to dead time . as can be seen , the @xmath39change between bands is mostly intrinsic , with a drop of only @xmath82% expected for the channels with the highest count rates . our cospectrum results are consistent with the pds results from other works ( e.g. * ? ? ? although we do not find any significant time lags between any of the energy bands , this is mostly due to the fact that our observation is not long enough . we estimate that @xmath83ks should be sufficient to detect time lags similar to those observed in the past in cygx-1 , in this spectral state ( e.g. * ? ? ? * ) . [ fig : cyg : rms ] gx339@xmath14is another well known bhb . discovered in 1973 by _ oso-7 _ @xcite , it has been observed frequently since then . it is a transient system , with long outbursts and periods of quiescence lasting up to several years ( * ? ? ? * and references therein ) . monitored during its outbursts by _ rxte _ , it showed one of the best examples of q - track pattern in the hardness - intensity diagram ( @xcite ; for the q - track pattern see @xcite and @xcite for a comparison with other sources ) . like many accreting bhs , its low / hard state pds is dominated by two or more lorentzian components , whose frequencies are positively correlated to count rate . as the source reaches the intermediate and soft state , the broad lorentzians move further to higher frequencies and various kinds of qpos appear ( see , e.g. , * ? ? ? * ; * ? ? ? * ) , including high frequency qpos . thanks to this wealth of data , it has often been used , like cygx-1 , to test new timing approaches . in its very high state , the signal at low energy ranges is generally coherent , while coherence is lost between low and high energies @xcite . @xcite measured coherence and lags in the hard state of the source , showing that the two lorentzians are likely produced by independent , but internally coherent , processes . @xcite showed that hard time lags in this source are more likely to be produced by propagation of modulations of the accretion flow in the disk rather than from reflection . _ nustar_observed gx339@xmath14four times during the rise of the last outburst . during this outburst , the source remained in a canonical low / hard state without reaching the threshold luminosity that would have marked the transition to the intermediate and soft states . the spectral analysis will be published in a separate publication ( frst et al 2014 , _ in prep . _ ) , but in this paper we will concentrate on the timing properties of the source , analyzed with the methods described above . the whole band and energy - dependent cospectra in the four different observations are shown in figure fig : gx_4obs . the approximate flux of the source increased almost linearly from @xmath82 to @xmath84 in each detector between the first and the last observation . the general shape and the variation of the cospectrum is similar to what has been observed in the past for this source in the hard state ( e.g. * ? ? ? * ; * ? ? ? * ) and for other bhbs @xcite , with two zero - centered lorentzian components dominating the shape ( l@xmath85 and l@xmath86 in @xcite ) , plus occasionally a third lorentzian needed to account for an excess between them . the cospectrum evolves with l@xmath85 consistently increasing its characteristic frequency in subsequent observations ( as luminosity increases ) , while l@xmath86 does not change as much . this is consistent with what was previously reported by @xcite using the pds . the energy dependence of the cospectrum is very weak in the first observation , and tends to increase in the subsequent observations , while the total @xmath39consistently decreases as the count rate increases . in the second , third and fourth observation the high energy cospectrum has consistently a lower @xmath39than the low energy one . again , these cospectra results are consistent with previous pds results ( e.g. * ? ? ? * ) . from figure fig : cts , and the discussion about the much brighter cygx-1 , it is clear that at these frequencies and for count rates changing from 10 to 40@xmath10we do not expect a such a strong change in @xmath39 , which must therefore be intrinsic to the source . a more detailed discussion will be presented in frst et al . , _ in prep._. this paper presents a method to exploit _ nustar _ s two independent detectors to work around the spurious effects resulting from dead time , enabling standard aperiodic timing analysis of x - ray binaries to be applied . as extensively studied in literature ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) , dead time produces a spurious correlation between event times that strongly modifies the shape of the pds . for frequencies @xmath19 , the only visible effect in pdss is a slight decrease of the white noise level . in timing - oriented x - ray missions , dead time is very short ( _ rxte_/pca : 10@xmath87s , @xcite ; _ astrosat_/laxpc : 1035@xmath87s , @xcite ; _ loft _ : @xmath88ns , @xcite ) and the `` interesting '' range of frequencies for accretion ( the highest being the dynamical timescales around neutron stars @xmath89hz ) is well below @xmath17 . in _ @xmath16is around 2.5ms , right in the middle of the frequency range commonly analyzed in bh and ns binaries . moreover , in order to obtain better precision in the measure of photon energies , the dead time is variable and depends on the event grades . due to this , standard techniques for the modeling of the pds shape ( e.g. * ? ? ? * ; * ? ? ? * ) are not applicable . but _ nustar_has two independent focal plane modules that are read out by separate processors , and the dead time is therefore independent . if , instead of taking the pds of each detector , we take the cospectrum ( see section sec : cpds ) of the signals in the two detectors , one is able to cancel out most correlations produced by dead time and obtain a close proxy of a white noise - subtracted pds . we described in sections [ sec : cpds ] and [ sec : simu ] how to calculate the cospectrum and estimate the remaining effects of dead time on the signal , namely the modulation of the variance of data bins , that can not be described using the standard @xcite prescriptions , and the equivalent change of the observed @xmath39at different frequencies and for different count rates . this estimate requires monte carlo simulations tailored to the source that is being analyzed , as described in section sec : simu . we also applied this technique to the analysis of _ nustar_data of two bhbs , cygx-1and gx339@xmath14 . cygx-1was observed in the soft state . we report a strong change of @xmath39with energy , with a significant rise of the variability at higher energies . these cospectrum results are consistent with earlier pds results , and extend the energy coverage with respect to earlier results from _ rxte _ ( * ? ? ? * e.g. ) . comparing these results with the spectral modeling presented by @xcite , it is apparent that the low energies are dominated by the disk emission , while at higher energies it s the reprocessed component ( comptonized and reflected ) that dominates . this is consistent with what is found in many bhbs , and in particular to what has already been reported in this source ( the stable disk and unstable corona picture , e.g. * ? ? ? * ) . in gx339@xmath14 , we observed the rising part of the outburst , in a standard low / hard state . the cospectrum shows that , as the outburst goes on , the @xmath39generally decreases and the characteristic frequency of the first lorentzian component used to described the cospectrum consistently rises . also , this is consistent with what has been observed in the low / hard state of this and other sources using the pds @xcite . we do not find evidence for time lags in the sources analyzed in this paper , due to the short observing time available , but the cross power density spectrum can be used to calculate time lags without major impact from dead time , if data are adequately filtered to avoid intervals with high background activity and in particular at the start and end of each good time interval . the techniques presented in this paper can be used to perform most of the standard timing and spectral timing analysis used in literature , if the change of @xmath39with frequency and count rate is properly taken into account ( section sec : simu ) . for slow timing ( @xmath90 ) , the dead time correction included in the official pipeline is still recommended , because it effectively corrects for all sources of dead time , including shielded events . for low count rates ( @xmath91 ) the standard pds of the sum of the two lightcurves ( what we called the `` total pds '' above ) is still advantageous , because the effects of dead time are negligible and the sensitivity of the total pds , expressed as significance of detection of a standard variability feature , is @xmath61 higher than the cospectrum . as an additional note , this technique can be used in any satellite that has completely independent detectors , and of course by using multiple satellites . _ rxte _ , with its independent pcus , would be a good candidate if not for the presence of so - called very large events ( vle ) , particle events depositing more than 100kev and occurring many times per second ( e.g. * ? ? ? these events often involve more than one pcu , invalidating the assumption of independence . the large area xenon propotional counter ( laxpc ) onboard the upcoming _ astrosat _ mission @xcite will provide three independent detectors . provided that the fraction of vle - equivalent events is negligible , this technique will be applicable to laxpc data . mb and db wish to acknowledge the support from the centre national dtudes spatiales ( cnes ) and the centre national de la recherche scientifique ( cnrs ) . cs acknowledges funding by the german bmwi under dlr grant numbers 50 qr 0801 , 50 qr 0903 , and 50 oo 1111 . thanks stfc for support ( grant reference st / j003697/1 ) . this work was supported under nasa contract no . nng08fd60c , and made use of data from the _ nustar _ mission , a project led by the california institute of technology , managed by the jet propulsion laboratory , and funded by the national aeronautics and space administration . we thank the _ nustar _ operations , software and calibration teams for support with the execution and analysis of these observations . this research has made use of the _ nustar _ data analysis software ( nustardas ) jointly developed by the asi science data center ( asdc , italy ) and the california institute of technology ( usa ) . the timing analysis was executed with matteo bachetti s libraries and tools in python for _ nustar_timing ( malt pynt ) . this code is available upon request . most of the plots were produced with the veusz software by jeremy sanders . the authors wish to thank chris done , denis leahy and tomaso belloni for very insightful discussions . 76 natexlab#1#1 agrawal , p. c. 2006 , asr , 38 , 2989 arnaud , k. a. 1996 , adass , 101 , 17 artigue , r. , barret , d. , lamb , f. k. , lo , k. h. , & miller , m. c. 2013 , arxiv , 1303.0248 , 1303.0248 axelsson , m. , borgonovo , l. , & larsson , s. 2005 , a&a , 438 , 999 bachetti , m. et al . 2013 , apj , 778 , 163 barret , d. , & vaughan , s. 2012 , apj , 746 , 131 belloni , t. , & hasinger , g. 1990 , a&a , 230 , 103 belloni , t. , homan , j. , casella , p. , van der klis , m. , nespoli , e. , lewin , w. h. g. , miller , j. m. , & mendez , m. 2005 , a&a , 440 , 207 bendat , j. s. , & piersol , a. g. 2011 , random data : analysis and measurement procedures ( wiley ) bck , m. et al . 2011 , a&a , 533 , 8 boutelier , m. 2009 , phd thesis , phd thesis , cesr boutelier , m. , barret , d. , & miller , m. c. 2009 , mnras , 399 , 1901 bowyer , s. , byram , e. t. , chubb , t. a. , & friedman , h. 1965 , science , 147 , 394 cabanac , c. , roques , j .- , & jourdain , e. 2011 , apj , 739 , 58 casella , p. , belloni , t. , & stella , l. 2005 , apj , 629 , 403 cassatella , p. , uttley , p. , wilms , j. , & poutanen , j. 2012 , mnras , 422 , 2407 castro - tirado , a. j. , brandt , s. , & lund , n. 1992 , iau circ . , 5590 , 2 churazov , e. , gilfanov , m. , & revnivtsev , m. 2001 , mnras , 321 , 759 cui , w. , zhang , s. n. , jahoda , k. , focke , w. , swank , j. h. , heindl , w. a. , & rothschild , r. e. 1997 , the transparent universe , 382 , 209 davies , r. b. , & harte , d. s. 1987 , biometrika , 74 , 95 dunn , r. j. h. , fender , r. p. , krding , e. g. , belloni , t. , & cabanac , c. 2010 , mnras , 403 , 61 fabian , a. c. et al . 2009 , nat . , 459 , 540 feroci , m. et al . 2012 , experimental astronomy , 34 , 415 foster , r. s. , waltman , e. b. , tavani , m. , harmon , b. a. , zhang , s. n. , paciesas , w. s. , & ghigo , f. d. 1996 , astrophysical journal letters v.467 , 467 , l81 frst , f. et al . 2013 , apj , 779 , 69 . 2014 , apj , 780 , 133 gandhi , p. 2010 , mnras , 407 , 2166 gentle , j. e. 2003 , random number generation and monte carlo methods ( springer ) gilfanov , m. , churazov , e. , & revnivtsev , m. 2000 , mnras , 316 , 923 gleissner , t. , wilms , j. , pottschmidt , k. , uttley , p. , nowak , m. a. , & staubert , r. 2004 , a&a , 414 , 1091 grinberg , v. et al . 2014 , a&a , 565 , 1 harrison , f. a. et al . 2013 , apj , 770 , 103 jahoda , k. , markwardt , c. b. , radeva , y. , rots , a. h. , stark , m. j. , swank , j. h. , strohmayer , t. e. , & zhang , w. 2006 , apj supp . ser . , 163 , 401 jin , c. , done , c. , middleton , m. , & ward , m. 2013 , mnras , 436 , 3173 jones , t. a. , levine , a. m. , morgan , e. h. , & rappaport , s. 2008 , apj , 677 , 1241 kaspi , v. m. et al . 2014 , apj , 786 , 84 krding , e. , & falcke , h. 2004 , a&a , 414 , 795 leahy , d. a. , darbro , w. , elsner , r. f. , weisskopf , m. c. , kahn , s. , sutherland , p. g. , & grindlay , j. e. 1983 , apj , 266 , 160 lewin , w. h. g. , van paradijs , j. , & van der klis , m. 1988 , ssrv , 46 , 273 maccarone , t. j. , & coppi , p. s. 2003 , mnras , 338 , 189 marinucci , a. et al . 2014 , mnras , 440 , 2347 markert , t. h. , canizares , c. r. , clark , g. w. , lewin , w. h. g. , schnopper , h. w. , & sprott , g. f. 1973 , apj , 184 , l67 mcclintock , j. e. , & remillard , r. a. 2006 , in compact stellar x - ray sources , ed . j. e. mcclintock & r. a. remillard , 157213 miller , j. m. et al . 2013 , apjl , 775 , l45 miyamoto , s. , kimura , k. , kitamoto , s. , dotani , t. , & ebisawa , k. 1991 , apj , 383 , 784 miyamoto , s. , & kitamoto , s. 1989 , nat . , 342 , 773 miyasaka , h. et al . 2013 , apj , 775 , 65 morgan , e. h. , remillard , r. a. , & greiner , j. 1997 , apj , 482 , 993 mori , k. et al . 2013 , apjl , 770 , l23 natalucci , l. et al . 2014 , apj , 780 , 63 nowak , m. a. , & vaughan , b. a. 1996 , mnras , 280 , 227 nowak , m. a. , vaughan , b. a. , wilms , j. , dove , j. b. , & begelman , m. c. 1999 , apj , 510 , 874 nowak , m. a. , wilms , j. , & dove , j. b. 1999 , apj , 517 , 355 pahari , m. , neilsen , j. , yadav , j. s. , misra , r. , & uttley , p. 2013 , apj , 778 , 136 papadakis , i. e. , nandra , k. , & kazanas , d. 2001 , apj , 554 , l133 paul , b. , & team , l. 2009 , in astrophysics with all - sky x - ray observations , 362 pottschmidt , k. , wilms , j. , nowak , m. a. , heindl , w. a. , smith , d. m. , & staubert , r. 2000 , a&a , 357 , l17 reig , p. , belloni , t. , van der klis , m. , mendez , m. , kylafis , n. d. , & ford , e. c. 2000 , apj , 541 , 883 remillard , r. a. , muno , m. p. , mcclintock , j. e. , & orosz , j. a. 2003 , aas , 7 , 648 revnivtsev , m. , gilfanov , m. , & churazov , e. 1999 , a&a , 347 , l23 . 2000 , a&a , 363 , 1013 risaliti , g. et al . 2013 , nat . , 494 , 449 strohmayer , t. e. 2001 , apj , 554 , l169 timmer , j. , & koenig , m. 1995 , a&a , 300 , 707 tomsick , j. a. , & kaaret , p. 2001 , apj , 548 , 401 tomsick , j. a. et al . 2014 , apj , 780 , 78 uttley , p. , cackett , e. m. , fabian , a. c. , kara , e. , & wilkins , d. r. 2014 , arxiv , 6575 , 1405.6575 uttley , p. , wilkinson , t. , cassatella , p. , wilms , j. , pottschmidt , k. , hanke , m. , & bck , m. 2011 , mnras let . , 414 , l60 van der klis , m. 1989 , in timing neutron stars : proceedings of the nato advanced study institute on timing neutron stars held april 4 - 15 , 27 vaughan , b. a. , & nowak , m. a. 1997 , apjl , 474 , l43 vaughan , s. 2013 , arxiv , 1309.6435v1 vikhlinin , a. , churazov , e. , & gilfanov , m. 1994 , a&a , 287 , 73 walton , d. j. et al . 2014 , arxiv , 5620 , 1404.5620 weisskopf , m. c. 1985 , in talk presented at the workshop _ time variability in x - ray and gamma - ray sources _ , taos , nm , usa zhang , w. , jahoda , k. , swank , j. h. , morgan , e. h. , & giles , a. b. 1995 , apj , 449 , 930 zoghbi , a. et al . 2014 , apj , 789 , 56 grs1915@xmath2105is a very bright persistent microquasar , discovered in 1992 by _ granat _ @xcite which shows accretion states that depart from the usual behavior of galactic bhs , being associated with very high - eddington fractions . the variability of this source has been extensively studied in the past , in particular for the presence of very strong low frequency ( @xmath120.1hz ) , medium - low frequency ( 0.510hz ) , medium - high frequency ( @xmath040 and 67hz ) and high frequency ( 113 and 165hz ) qpos ( e.g. * ? ? ? * ; * ? ? ? * ; * ? ? ? * ; * ? ? ? * ) . in our observation , the medium - low qpo is the only one observed to very high significance . its behavior has been studied in great detail ( e.g. * ? ? ? * ) and its properties are known up to very high energies thanks to _ rxte_/hexte observations @xcite . this source is discussed as a reference for timing analysis on sources observed very early in the mission . the first version of the flight software , in fact , introduced a large number of spurious modulations in the signal that we will discuss and show a way to work around . during the observatory commissioning phase prior to commencement of science observations , _ nustar_observed the galactic bhb grs1915@xmath2105 in its _ plateau _ state @xcite . the spectral analysis of that observation was presented by @xcite . the version of the flight software that was installed at that time on _ nustar_executed housekeeping operations at regular intervals ( 1 or 8s ) , producing periodic dead time . therefore , the raw power spectrum and the cpds of this observation are plagued by spurious features at 1hz and 0.125hz as well as their harmonics and sub - harmonics ( see figure fig : grscpds ) . in this section we show a technique to clean up the cospectrum and perform a basic analysis of the strong qpo that this source . to `` clean '' the cospectrum from these spurious features , we first produced a cospectrum on a time scale @xmath92 of 128s . the maintenance operations were very regular , and so in this cospectrum the corresponding features were very sharp and limited to a single bin each . we replaced the value of each of these bins with a random value calculated from the mean of the two closest bins and their standard deviation , as shown in figure fig : grscpds . the possible spurious effects of this imperfect statistical treatment of the simulated bin value are negligible due to the small frequency bin ( simulated data are only one bin every @xmath93hz ) and the rebinning applied in the following steps of the analysis . the real and imaginary part were treated independently . from this cleaned cpds we calculated the cospectrum . an example is shown in red in figure fig : grscpds . the cospectrum shows a very strong and variable qpo at @xmath94hz , with @xmath39=10(1)% ( from the best - fit lorentzian amplitude , and corrected for the expected @xmath95% drop at @xmath96 , equation eq : drms ) . this feature has previously been observed in this spectral state ( e.g. * ? ? ? * ) . in figure fig : grscpds we show that with these techniques it is possible in principle to also apply techniques of qpo tracking ( a dynamical cospectrum ) given a sufficient @xmath39 . a hint of the harmonic of this qpo is also observable , both in the single and in the dynamical cospectra . no other significant qpos are detected . this source is known to show high - frequency qpos in _ rxte_observations @xcite . these features are not observed in our dataset . we calculate a rough upper limit of @xmath97% for additional qpos between 50 and 100hz . this upper limit is higher than the observed @xmath39of qpos in this frequency range observed in the past ( e.g. @xmath94% in the full _ rxte_band in @xcite ) , and our non - detection might be largely due to the low statistics of this dataset .

in this paper we develop a globally convergent numerical method for a 1-d inverse medium problem in the frequency domain . the performance of this method is tested on both computationally simulated and experimental data . propagation of electromagnetic waves is used in experiments . a theorem about the global convergence of our method is proved . another name for inverse medium problems is coefficient inverse problems ( cips ) . we model the process as a 1-d cip due to some specifics of our data collection procedure , see subsection 6.2 . this paper is the first one in which the globally convergent method of @xcite is extended to the case of the frequency domain . indeed , the original version of that method works with the laplace transform of the time dependent data . both the 3-d and the 1-d versions of the method of @xcite were verified on experimental data , see @xcite for the 3-d case and karch , ksnf1,ksnf2 for the 1-d case . the experimental data here are the same ones as in @xcite . compared with the previous publications @xcite , two additional difficulties occurring here are : ( 1 ) we now work with complex valued functions instead of real valued ones and , therefore ( 2 ) it is not immediately clear how to deal with the imaginary part of the logarithm of the complex valued solution of the forward problem . on the other hand , we use that logarithm in our numerical procedure . besides , the previous numerical scheme of @xcite is significantly modified here . we call a numerical method for a cip _ globally convergent _ if there is a rigorous guarantee that this method reaches at least one point in a sufficiently small neighborhood of the correct solution without any advanced knowledge about this neighborhood . it is well known that the topic of the global convergence for cips is a highly challenging one . thus , similarly with the above cited publications , we prove the global convergence of our method within the framework of a certain approximation . this is why the term approximate global convergence `` was used in above cited references ( global convergence '' in short ) . that approximation is used only on the first iterative step of our algorithm and it is not used in follow up iterations . besides , this approximation is a quite natural one , since it amounts to taking into account only the first term of a certain asymptotic behavior and truncating the rest . the global convergence property is verified numerically here on both computationally simulated and experimental data . cips are both highly nonlinear and ill - posed . these two factors cause the non - convexity of conventional tikhonov functionals for these problems . it is well known that typically those functionals have multiple local minima and ravines , see , e.g. numerical examples of multiple local minima in scales and on figure 5.3 of @xcite . figure 5.4 of @xcite demonstrates an example of a ravine . hence , there is no guarantee of the convergence of an optimization method to the correct solution , unless the starting point of iterations is located in a sufficiently small neighborhood of that solution . however , it is often unclear in practical scenarios how to reach that neighborhood . in our inverse algorithm we solve a sequence of linear ordinary differential equations of the second order on the interval @xmath0 . the peculiarity here is that we have overdetermined boundary conditions for these equations : we have the solution and its first derivative at @xmath1 and we have the first derivative of the solution at @xmath2 . our attempts to use only boundary conditions at @xmath1 did not lead to good reconstruction results . the same observation is in place in remark 5.1 on page 13 of ksnf1 . thus , similarly with @xcite , we use here the quasi - reversibility method ( qrm ) . the qrm is well suitable to solve pdes with the overdetermined boundary data . to analyze the convergence rate of the qrm , we use a carleman estimate . the qrm was first introduced by lattes and lions @xcite . however , they have not established convergence rates . these rates were first established in @xcite , where it was shown that carleman estimates are the key tool for that goal . a survey of applications of carleman estimates to qrm can be found in @xcite . chapter 6 of @xcite describes the use of the qrm for numerical solutions of cips . we also mention an active work with the qrm of bourgeois and dard , see , e.g. @xcite for some samples of their publications . our experimental data are given in the time domain . to obtain the data in the frequency domain , we apply the fourier transform . previously the same data were treated in works @xcite of this group . in that case the laplace transform was applied . next , the 1-d version of the globally convergent method of @xcite was used . indeed , the original version of this method works with the laplace transform of the time dependent data . a similar 1-d inverse problem in the frequency domain was solved numerically in @xcite by a different method . the experimental data of this publication were collected by the forward looking radar which was built in the us army research laboratory @xcite . the data were collected in the field ( as opposed to a laboratory ) . thus , clutter was present at the measurement site . this certainly adds some additional difficulties to the imaging problem . the goal of this radar is to detect and identify shallow explosives , such as , e.g. improvised explosive devices and antipersonnel land mines . those explosives can be located either on the ground surface or a few centimeters below this surface . this radar provides only a single time dependent curve for a target . therefore , to solve a cip , we have no choice but to model the 3-d process by a 1-d wave - like pde . we model targets as inclusions whose dielectric constants are different from the background . in terms of working with experimental data , the goal in this paper is not to image locations of targets , since this is impossible via solving a cip with our data , see subsection 6.2 for details . in fact , our goal is to image maximal values of dielectric constants of targets . our targets are 3-d objects , while we use a 1-d model : since we measure only one time resolved curve for a target . nevertheless , we show below that our calculated dielectric constants are well within tabulated limits @xcite . of course , an estimate of the dielectric constant is insufficient for the discrimination of explosives from the clutter . on the other hand , the radar community mostly relies on the intensity of the radar image . therefore , we hope that the additional information about values of dielectric constants of targets might lead in the future to designs of algorithms which would better discriminate between explosives and clutter . in the 3-d case the globally convergent method of @xcite works with the data generated either by a single location of the source or by a single direction of the incident plane wave . the second globally convergent method for this type of measurements was developed and tested numerically in @xcite . we also refer to another globally convergent methods for a cips , which is based on the multidimensional version of the gelfand - levitan method . this version was first developed by kabanikhin kab1 and later by kabanikhin and shishlenin @xcite . unlike bk1,bk2,tbkf1,tbkf2 , the technique of @xcite works with multiple directions of incident plane waves . in @xcite the performance of the method of @xcite was compared numerically with the performance of the krein equation @xcite for the same experimental data as ones used here . the krein equation krein is close to the gelfand - levitan equation @xcite . it was shown in @xcite that the performance of the krein equation is inferior to the performance of the method of @xcite for these experimental data . this is because the solution of the krein equation is much more sensitive to the choice of the calibration factor than the solution obtained by the technique of @xcite . on the other hand , it is necessary to apply that factor to those experimental data to make the range of values of the resulting data comparable with the range of values of computationally simulated data . in section 2 we pose forward and inverse problems and analyze some of their features . in section 3 we study a 1-d version of the qrm . in section 4 we present our globally convergent method . in section 5 we prove the global convergence of our method . in section 6 we present our numerical results for both computationally simulated and experimental data . it was shown numerically in @xcite that the component of the electric field , which is incident upon a medium , dominates two other components . it was also shown in @xcite that the propagation of that component is well governed by a wave - like pde . furthermore , this finding was confirmed by imaging from experimental data , see chapter 5 of @xcite and tbkf1,tbkf2 . thus , just as in @xcite , we use a 1-d wave - like pde to model the collection of our experimental data of electromagnetic waves propagation . let @xmath3 be two positive numbers . let the function @xmath4 satisfy the following conditions:@xmath5 , \forall x\in \mathbb{r } , \label{2.1 } \\ c\left ( x\right ) & = & 1+\beta \left ( x\right ) , \beta \left ( x\right ) = 0,\forall x\notin \left ( 0,1\right ) . \label{2.3}\end{aligned}\]]fix the source position @xmath6 . consider the generalized helmholtz equation in the 1-d case,@xmath7let @xmath8 be the solution of the problem ( [ 2.4 ] ) , ( [ 2.6 ] ) for the case @xmath9 then@xmath10the problem , is the forward problem . our interest is in the following inverse problem problem : fix the source position @xmath11 let @xmath12\subset \left ( 0,\infty \right ) $ ] be an interval of frequencies @xmath13 . reconstruct the function @xmath14 assuming that the following function @xmath15 is known @xmath16 . \label{2.7}\ ] ] in this subsection we establish existence and uniqueness of the forward problem . even though these results are likely known , we present them here for reader s convenience . in addition , the techniques using in their proofs help us to verify that the function @xmath17 never vanishes for @xmath18 , which plays an important role in our algorithm of solving the to cip . also , this technique justifies the numerical method for solving the forward problem ( [ 2.4 ] ) , ( [ 2.6 ] ) . computationally simulated data are obtained by numerically solving the problem ( [ 2.4 ] ) , ( [ 2.6 ] ) . [ thm unique forward ] assume that conditions ( [ 2.1])-([2.3 ] ) hold . then for each @xmath19 and for each @xmath6 there exists a single solution @xmath20 of the problem ( [ 2.4 ] ) , ( [ 2.6 ] ) . moreover , the function @xmath21 , called the scattering field , is in @xmath22 . * proof*. we prove uniqueness first . suppose that @xmath23 and @xmath24 are two solutions of the problem ( [ 2.4 ] ) , ( [ 2.6 ] ) . denote @xmath25 then the function @xmath26 satisfies @xmath27since the function @xmath28 outside of the interval @xmath29 then ( [ 2.8 ] ) implies that @xmath30 for @xmath31 this , together with ( [ 2.9 ] ) , yields @xmath32where @xmath33 and @xmath34 are some complex numbers depending on @xmath35 let @xmath36 be an arbitrary number . multiply both sides of ( [ 2.8 ] ) by the function @xmath37 and integrate over the interval @xmath38 using integration by parts . we obtain@xmath39by ( [ 2.11 ] ) @xmath40 and @xmath41 hence , the imaginary part of ( [ 2.12 ] ) , @xmath42 vanishes , so do @xmath43 and @xmath44 . using ( [ 2.11 ] ) , we obtain @xmath45 for @xmath46 due to the classical unique continuation principle , @xmath47 thus , @xmath48 . we now prove existence . consider the 1-d analog of the lippman - schwinger equation with respect to a function @xmath49 , @xmath50fix @xmath6 and @xmath19 . consider equation ( [ 2.13 ] ) only for @xmath51 assume that there exist two functions @xmath52 satisfying ( [ 2.13 ] ) for @xmath53 consider their extensions on the set @xmath54 via the right hand side of ( [ 2.13 ] ) . then so defined functions @xmath52 satisfy ( [ 2.13 ] ) for all @xmath55 @xmath56 hence , both of them are solutions of the problem ( [ 2.4 ] ) , ( [ 2.6 ] ) . hence , the above established uniqueness result implies that @xmath57 consider again the integral equation ( [ 2.13 ] ) for @xmath58 and apply the fredholm alternative . this alternative , combined with the discussion in the previous paragraph , implies that there exists unique solution @xmath59 $ ] of equation ( [ 2.13 ] ) . extending this function for @xmath60 via the right hand side of ( [ 2.13 ] ) , we obtain that there exists unique solution @xmath61 of equation ( [ 2.13 ] ) . furthermore , the function @xmath62 and also this function @xmath49 is the required solution of the problem ( [ 2.4 ] ) , ( [ 2.6 ] ) . @xmath63 below we consider only such a solution @xmath64 of the problem ( [ 2.4 ] ) , ( 2.6 ) that @xmath65 : as in theorem 2.1 . consider the function @xmath67 defined as @xmath68note that by ( [ 2.3 ] ) @xmath69 for @xmath70 .$ ] [ thm assymptotic u ] assume that the function @xmath71 .$ ] then for every pair @xmath72 , the asymptotic behavior of the function @xmath20 is @xmath73 \left ( 1+o\left ( \frac{1}{k}\right ) \right ) , k\rightarrow \infty . \label{201}\]]furthermore , for any finite interval @xmath74 there exists a number @xmath75 such that for all @xmath76 the function @xmath20 can be analytically extended with respect @xmath13 from @xmath77 in the half plane @xmath78 . * proof*. consider the following cauchy problem@xmath79this proof is based on the connection between the solution @xmath64 of the problem ( [ 2.4 ] ) , ( [ 2.6 ] ) and the solution @xmath80 of the problem ( [ 2.15 ] ) , ( [ 2.16 ] ) via the fourier transform with respect to @xmath81 . here and below the fourier transform is understood in terms of distributions , see , e.g. the book @xcite for the fourier transform of distributions . we now obtain a hyperbolic equation with potential from equation ( [ 2.15 ] ) . to do this , we use a well known change of variables @xmath82 @xcite@xmath83denote @xmath84 by ( [ 2.17 ] ) @xmath85 hence , ( [ 2.15 ] ) and ( [ 2.16 ] ) become@xmath86consider now a new function @xmath87 where the function @xmath88 is chosen in such a way that the coefficient at @xmath89 becomes zero after the substitution @xmath90 in equation ( [ 2.18 ] ) . then ( [ 2.18 ] ) and ( [ 2.19 ] ) become@xmath91 it follows from ( [ 2.1 ] ) , ( [ 2.3 ] ) , ( [ 2.17 ] ) and ( [ 2.24 ] ) that the function @xmath92 and has a finite support , @xmath93recall that @xmath67 is the function defined in ( [ 200 ] ) . expression ( [ 2.24 ] ) in the @xmath94coordinate becomes @xmath95 let @xmath96 be the heaviside function . it is well known that the solution @xmath97 of the problem ( [ 2.22 ] ) , ( [ 2.23 ] ) has the following form , see , e.g. chapter 2 in @xcite@xmath98 the backwards substitution @xmath99 transforms the function @xmath100 in the function @xmath101 hence , ( [ 2.250 ] ) implies that the function @xmath102 in ( [ 2.24 ] ) is non positive , @xmath103consider now the operator @xmath104 of the fourier transform,@xmath105 in the sense of distributions we have@xmath106see section 6 in 9 of chapter 2 of @xcite . consider an arbitrary finite interval @xmath107 let @xmath108 be two arbitrary integers such that @xmath109 . since by ( [ 2.25 ] ) the function @xmath102 has a finite support , then ( [ 2.28 ] ) , lemma 6 of chapter 10 of the book @xcite as well as remark 3 after that lemma guarantee that functions @xmath110 decay exponentially with respect to @xmath111 as long as @xmath112 hence , one can apply the operator @xmath104 to the functions @xmath113 in the regular sense . the assertion about the analytic extension follows from and the exponential decay of the functions @xmath110 let @xmath114 the next question is whether the function @xmath115 satisfies analogs of conditions ( [ 2.4 ] ) , ( [ 2.6 ] ) . theorem 3.3 of @xcite and theorem 6 of chapter 9 of @xcite guarantee that the function @xmath115 satisfies the following conditions @xmath116 using ( [ 2.27 ] ) and the integration by parts , we obtain for @xmath66 @xmath117two backwards substitution : @xmath99 via ( [ 2.17 ] ) and @xmath118 imply that conditions ( [ 2.31 ] ) and ( [ 2.32 ] ) for the function @xmath115 turn into conditions ( [ 2.4 ] ) , ( [ 2.6 ] ) for the function @xmath64 . thus , ( [ 2.29 ] ) and ( [ 2.33 ] ) imply ( [ 201 ] ) . @xmath63 since the source position @xmath6 is fixed , we drop everywhere below the dependence on @xmath119 in notations of functions . first , we show that , having the function @xmath15 in ( [ 2.7 ] ) , one can uniquely find the function @xmath120 indeed , for @xmath121 conditions ( 2.4 ) and ( [ 2.6 ] ) become@xmath122@xmath123let @xmath124 then ( [ 2.34 ] ) and ( 2.35 ) imply that@xmath125hence , @xmath126for a complex number @xmath127 . note that by ( [ 2.60 ] ) and ( 2.7 ) @xmath128hence , by and the definition of the function @xmath129 , we have @xmath130 let @xmath131thus , , and ( [ 2.37 ] ) provide us with an additional data @xmath132 to solve our cip , where @xmath133 . \label{2.38}\ ] ] * * theorem 2.3 ( * * uniqueness of our cip**)**. _ let _ @xmath134 _ _ be the function defined in ( [ 200 ] ) . assume that _ _ _ then our cip has at most one solution . _ _ * proof*. by theorem 2.2 one can analytically extend the function @xmath15 from the interval @xmath136 $ ] in the half plane @xmath137 where @xmath138 is a positive number@xmath139 hence , it follows from ( [ 2.37 ] ) and ( [ 2.38 ] ) that we know functions @xmath140 for all @xmath141 consider the inverse fourier transform of the function @xmath64 with respect to @xmath13 , @xmath142 it was established in the proof of theorem 2.2 that this transform can indeed be applied to the function @xmath64 and @xmath143 , where the function @xmath144 is the solution of the problem ( [ 2.15 ] ) , ( [ 2.16 ] ) . functions @xmath145are known . hence , we have obtained the inverse problem for equation ( 2.15 ) with initial conditions ( [ 2.16 ] ) and the data ( [ 2.39 ] ) . it is well known that this inverse problem has at most one solution , see , e.g. chapter 2 of @xcite . @xmath63 as it was pointed out in section 1 , we solve ordinary differential equations with over determined boundary conditions using the qrm . in this section , we develop the qrm for an arbitrary linear ordinary differential equation of the second order with over determined boundary conditions . below for any hilbert space @xmath146 its scalar product is @xmath147 . for convenience , we use in this section the notation @xmath148 " for a generic complex valued function , which is irrelevant to the function @xmath148 in ( [ 2.37 ] ) . let functions @xmath149 and @xmath150 be in @xmath151,\mathbb{c})$ ] and the function @xmath152 be in @xmath153,\mathbb{c})$ ] . let @xmath154 and @xmath155 be complex numbers . in this section we construct an approximate solution of the following problem:@xmath156the qrm for problem ( [ 3.1 ] ) amounts to the minimization of the following functional @xmath157where @xmath158 below in this section we will establish existence and uniqueness of the minimizer of @xmath159 . we will also show how close that minimizer is to the solution of if it exists . we start from the carleman estimate for the operator @xmath160 [ lem car ] for any complex valued function @xmath162 with @xmath163 and for any parameter @xmath164 the following carleman estimate holds @xmath165 , \label{3.10}\]]where @xmath166 is a constant independent of @xmath64 and @xmath167 * proof*. since @xmath168)$ ] is dense in @xmath169 , it is sufficient to prove ( [ 3.10 ] ) only for functions @xmath170 .$ ] moreover , without loss of the generality , we can assume that @xmath64 is real valued . introduce the function @xmath171 we have @xmath172a simple calculation yields @xmath173 ^{2}}{x+1}\geq \frac{4\lambda v^{\prime } \left ( v^{\prime \prime } + \lambda ^{2}v\right ) } { x+1 } \notag \\ & \geq \frac{d}{dx}\left ( \frac{2\lambda \left ( v^{\prime } \right ) ^{2}+2\lambda ^{3}v^{2}}{x+1}\right ) + \frac{2\lambda } { ( x+1)^{2}}(\left ( v^{\prime } \right ) ^{2}+\lambda ^{2}v^{2 } ) . \label{3.3}\end{aligned}\]]one the other hand , we have @xmath174 e^{-2\lambda x}\geq \left ( \frac{1}{3}\left ( u^{\prime } \right ) ^{2}-\frac{\lambda ^{2}}{2}u^{2}\right ) e^{-2\lambda x}. \label{3.44}\]]combine and and then integrate the result . we obtain @xmath175inequality follows . @xmath63 we first establish the existence and uniqueness for the minimizer of @xmath177 [ thm unique ] for every @xmath178 there exists a unique minimizer @xmath179 of the functional @xmath180 furthermore , the following estimate holds@xmath181where the number @xmath182 depends only on @xmath183 and @xmath184 . let @xmath185 be a minimizing sequence of @xmath159 . that means , @xmath186it is not hard to see that @xmath187 is bounded in @xmath188 . in fact , if @xmath187 has a unbounded subsequence then @xmath189without the loss of generality , we can assume that @xmath187 weakly converges in @xmath188 to a function @xmath190 and strongly converges to @xmath190 in @xmath169 . the function @xmath190 belongs to @xmath191 because @xmath191 is close and convex . we have @xmath192the uniqueness of @xmath190 is due to the strict convexity of @xmath193 . inequality can be verified by the fact that @xmath194 where @xmath195 and the function @xmath196 $ ] is such that@xmath197 , \\ 0 & x\in \lbrack 3/4,1].\end{array}\right.\]]@xmath63 let @xmath198 and @xmath199 be the real and imaginary parts respectively of the complex valued function @xmath200 . without confusing , we identify @xmath200 with the pair of real valued functions @xmath201 define @xmath202rewrite @xmath159 as @xmath203 in order to find @xmath190 , we find the zero of the frchet derivative @xmath204 of @xmath205 . since the operators @xmath206 and @xmath207 are linear , @xmath204 is given by @xmath208for all @xmath209 with @xmath210 the existence of a zero of @xmath204 follows from the above existence of @xmath190 and the uniqueness is , again , deduced from the convexity of @xmath159 and @xmath191 . in our computations , we use the finite difference method to approximate the equation @xmath211 , together with the condition @xmath212 , as a linear system for @xmath190 . the minimizer @xmath190 of @xmath159 is called the regularized solution of @xcite . while theorem 3.1 claims the existence and uniqueness of the regularized solution @xmath190 of problem ( [ 3.1 ] ) , we now prove convergence of regularized solutions to the exact solution of this problem , provided that the latter solution exists , see @xcite for the definition of the regularized solution . it is well known that one of concepts of the tikhonov regularization theory is the _ a priori _ assumption about the existence of an exact solution of an ill - posed problem , i.e. solution with noiseless data @xcite . estimate ( [ 3.4 ] ) is valid for the @xmath213norm and it becomes worse as long as @xmath214 however , theorem 3.2 provides an estimate for the @xmath215norm and the latter estimate is not worsening as @xmath214 to prove theorem 3.2 , we use the carleman estimate of subsection 3.1 . suppose that there exists the exact solution @xmath216 of the problem ( 3.1 ) with the exact ( i.e. noiseless ) data @xmath217 , @xmath218 . let the number @xmath219 be the level of the error in the data , i.e.@xmath220 let @xmath179 be the unique minimizer of the functional @xmath221 , which is guaranteed by theorem [ thm unique ] . assume that ( [ 3.12 ] ) holds . then there exists a constant @xmath222 depending only on @xmath223 and @xmath224 such that the following estimate holds@xmath225 \right ) . \label{3.13}\]]in particular , if @xmath226 then the following convergence rate of regularized solutions @xmath190 takes place ( with a different constant @xmath227)@xmath228 \right ) \delta . \label{30}\]][thm convergence of regularized solutions ] * proof*. in this proof , @xmath227 is a generic constant depending only on @xmath223 and @xmath229 let the function @xmath230 $ ] satisfies the following condition @xmath197 , \\ 0 & x\in \lbrack 3/4,1].\end{array}\right.\]]define the error " function @xmath231obviously , @xmath232since @xmath190 is the minimizer of @xmath233 then we have for all @xmath234 with @xmath235 @xmath236since @xmath216 is a solution of , then @xmath237 denoting @xmath238 , using the test function @xmath239 , and using the cauchy - schwarz inequality , we derive in a standard way from ( [ 3.12 ] ) , and that @xmath240on the other hand , lemma [ lem car ] gives @xmath241 .\end{aligned}\]]choosing @xmath242 sufficiently large , we obtain @xmath243combining this and completes the proof . [ lem nonzero ] fix @xmath6 and @xmath19 . let @xmath20 be the solution of the problem ( [ 2.4])-([2.6 ] ) . then @xmath244 in particular , @xmath245,$ ] where @xmath15 is the function defined in ( [ 2.7 ] ) . * proof*. the proof is similar to that of the uniqueness of the problem , in theorem [ thm unique forward ] . @xmath246 below @xmath247 is the function defined in ( [ 2.37 ] ) . since by theorem 2.2 @xmath248 then we can consider @xmath249 since @xmath250 then the natural question is about @xmath251 hence , we consider the asymptotic behavior at @xmath66 of the function @xmath252 . using ( [ 2.60 ] ) and theorem 2.2 , we obtain for @xmath18 @xmath253 \left ( 1+o\left ( \frac{1}{k}\right ) \right ) . \label{4.0}\]]obviously @xmath254 $ ] for sufficiently large @xmath255 hence , we set for sufficiently large @xmath19 and for @xmath256@xmath257the function @xmath258 is defined via ( [ 4.01 ] ) for sufficiently large @xmath259 on the other hand , for not large values of @xmath13 it would be better to work with derivatives of @xmath249 indeed , we would not have problems then with defining @xmath260 hence , taking @xmath261 sufficiently large , we define the function @xmath262 as @xmath263differentiate with respect to @xmath13 . we have @xmath264 multiplying both sides of the equation above by @xmath265 gives @xmath266since @xmath267 , then @xmath268the function @xmath269 , therefore , defines @xmath270 . for each @xmath271 , define @xmath272 it follows from ( [ 2.7 ] ) , ( [ 2.37 ] ) and ( [ 4.0 ] ) that @xmath273 as @xmath274 hence , @xmath275 $ ] for sufficiently large @xmath255 hence , the function @xmath276 can be defined similarly with the function @xmath277 in ( [ 4.2 ] ) . let @xmath278,k\in \lbrack \underline{k},\overline{k } ] . \label{4.1}\]]hence,@xmath279denote@xmath280we call @xmath281 the tail function " and this function is unknown . note that we do not use below the function @xmath281 . rather we use only its @xmath94derivatives . hence , when using these derivatives , we are not concerned with @xmath282 it easily follows from ( [ 2.3 ] ) , ( [ 2.4 ] ) , ( [ 2.60 ] ) , ( [ 2.37 ] ) and ( [ 2.38 ] ) that @xmath283using ( [ def log w ] ) and ( [ 4.3 ] ) , we obtain @xmath284therefore , ( [ 4.1])-([4.40 ] ) imply that @xmath285@xmath286@xmath287 , q_{x}\left ( 0,k\right ) = 0 . \label{4.6}\ ] ] we have obtained an integral differential equation ( [ 4.5 ] ) for the function @xmath288 with the overdetermined boundary data ( [ 4.6 ] ) . the tail function in ( [ 4.5 ] ) is also unknown . hence , to approximate both functions @xmath288 and @xmath281 , we need to use not only conditions ( [ 4.5 ] ) , ( 4.6 ) but something else as well . thus , in our iterative procedure , we solve problem ( [ 4.5 ] ) , ( [ 4.6 ] ) , assuming that @xmath281 is known , and update the function @xmath288 this way . then we update the unknown coefficient @xmath289 next , we solve problem ( [ 4.3 ] ) , ( [ 4.4 ] ) for the function @xmath148 at @xmath290 and update the tail function via ( 4.1 ) , ( [ 4.20 ] ) and ( [ 4.21 ] ) . it is important for the above iterative process to properly choose the initial approximation @xmath292 for the tail function . since we want to construct a globally convergent method , this choice must not use any advanced knowledge of a small neighborhood of the exact solution @xmath293 of our inverse problem . we now describe how do we choose the initial tail @xmath294 it follows from theorem [ thm assymptotic u ] and the definition of @xmath295 via ( [ 4.0])- that there exists a function @xmath296 $ ] such that @xmath297hence , assuming that the number @xmath261 is sufficiently large , we drop terms @xmath298 and @xmath299 in and set@xmath300set @xmath290 in ( [ 4.5 ] ) and ( [ 4.6 ] ) and then substitute ( [ 4.7 ] ) there . we obtain @xmath301 it follows from this , , , and that @xmath302we solve the problem ( [ guess v ] ) , ( [ v0 boundary condition ] ) via the qrm . by the embedding theorem @xmath303 $ ] and @xmath304 } \leq c\left\vert f\right\vert _ { h^{2}\left ( 0,1\right ) } , \forall f\in h^{2}\left ( 0,1\right ) , $ ] where @xmath305 is a generic constant@xmath139 recall that the function @xmath306 in ( [ v0 boundary condition ] ) is linked with the function @xmath15 as in ( [ 2.38 ] ) . thus , theorems [ thm unique ] and [ thm convergence of regularized solutions ] lead to theorem [ thm 4.1 ] . in this theorem , we use the entire interval @xmath307 $ ] rather than just @xmath308 ( in ( [ 4.11 ] ) ) for brevity : since we will use this interval below . [ thm 4.1 ] let @xmath293 satisfying conditions ( 2.1)-([2.3 ] ) be the exact solution of our cip . for @xmath309 let the exact tail @xmath310 have the form ( [ 4.7 ] ) . assume that for @xmath311 $ ] @xmath312where @xmath313 is a sufficiently small number , which characterizes the level of the error in the boundary data . let the function @xmath314 be the approximate solution of the problem ( [ guess v])-([v0 boundary condition ] ) obtained via the qrm with @xmath315 . then there exists a constants @xmath316 depending only on @xmath261 and @xmath317 such that @xmath318 } \leq c\left\vert v_{0,a}\left ( x,\overline{k}\right ) -v^{\ast } \left ( x,\overline{k}\right ) \right\vert _ { h^{2}\left ( 0,1\right ) } \leq c_{3}\delta .\ ] ] [ remark 4.2 ] 1 . @xmath319 2 . @xmath320@xmath321@xmath308 and recalling we conclude that@xmath322@xmath323@xmath317@xmath324@xmath323@xmath317 3 . @xmath325@xmath293 4 . 5 . @xmath323 and @xmath326@xmath281@xmath288 below we consider the error parameter @xmath327 defined as@xmath328 consider a partition of the frequency interval @xmath330 $ ] in @xmath331 subintervals with the step size @xmath332,@xmath333where the number @xmath334 is sufficiently small . we assume that the function @xmath335 is piecewise constant with respect to @xmath13 , @xmath336 for @xmath337 for each @xmath338 and for all @xmath339 define @xmath340hence , by ( [ 4.20])@xmath341 then ( [ 4.5 ] ) and ( [ 4.12])-([4.13 ] ) imply that for all @xmath338 @xmath342choose the step size @xmath332 sufficiently small and ignore terms with @xmath332 and @xmath343 . note that @xmath344 for @xmath345 also , we keep in mind that we will iterate with respect to tail functions for each @xmath346 as well as with respect to @xmath346 . thus , we rewrite the last equation as @xmath347 q_{n , j}^{\prime } = -2k_{n}\left ( -q_{n-1}^{\prime } + v_{n , j}^{\prime } \right ) ^{2}+2i\left ( -q_{n-1}^{\prime } + v_{n , j}^{\prime } \right ) \label{4.15}\]]for all @xmath348 for some @xmath349 . the boundary conditions for @xmath350 in are taken according to those for @xmath288 in . precisely , @xmath351 the procedure to solve the cip is described below : we reconstruct a set @xmath352 of approximations for @xmath353 1 . set @xmath354 find the first approximation @xmath355 for the tail function solving the problem ( [ guess v])-([v0 boundary condition ] ) via the qrm . [ step 2 ] for an integer @xmath356 $ ] , suppose that functions @xmath357 , @xmath358 , @xmath359 are known . therefore , @xmath360 is known . we calculate the function @xmath361 as follows set @xmath362 , @xmath363 @xmath364 . 2 . for @xmath365 : 1 . solve the problem , for @xmath350 via the qrm . 2 . for @xmath366 let @xmath367 due to an analog of . 3 . [ step 2b3]calculate @xmath368 by , which will be explained later . 4 . solve the problem ( [ 4.3 ] ) , ( [ 4.4 ] ) via the qrm with @xmath290 and @xmath369 . let @xmath370 be its solution . next , using ( 4.2 ) and ( [ 4.20 ] ) , set @xmath371 3 . set @xmath372 where @xmath373 3 . [ step 3 ] chose @xmath374 where @xmath375 in the algorithm , all differential equations are solved via the qrm . thus , we keep for those qrm solutions " the same notations for brevity . for simplicity , we assume here that @xmath376 although we also work in one case of experimental data with a non - positive function @xmath377 thus , in the algorithm above , we update the function @xmath378 using ( [ 2.1 ] ) , ( [ 2.3 ] ) , ( [ 4.40 ] ) and ( [ 4.13 ] ) as@xmath379this truncation helps us to get a better accuracy in the reconstructed function @xmath380 . in fact , it follows from ( [ 2.1 ] ) , ( [ 2.3 ] ) , and ( [ 4.2323 ] ) that @xmath381where @xmath382 is the exact solution of the cip . in this section we prove our main result about the global convergence of the algorithm of the previous section . this method actually has the approximately global convergence property , see the third paragraph of section [ sec:1 ] and remarks [ remark 4.2 ] . for brevity we assume in this section that @xmath383 in the above algorithm . in other words , we assume that we do not perform inner iterations . the case @xmath384 can be done similarly . first , we need to introduce some assumptions about the exact solution . everywhere below the superscript @xmath385 " denotes functions which correspond to the exact coefficient @xmath322 . denote @xmath386 then @xmath387 for @xmath388 and for @xmath389 set @xmath390let the function @xmath391 be the same as in ( [ 4.14 ] ) , except that functions @xmath392 are replaced with @xmath393 also , let @xmath394then ( [ 4.40 ] ) and implies that @xmath395also , by ( [ 4.15 ] ) and ( [ 4.16 ] ) we have for @xmath0@xmath396 \left ( q_{n}^{\ast } \right ) ^{\prime } \notag \\ & = & -2k_{n}\left ( -\left ( q_{n-1}^{\ast } \right ) ^{\prime } + \left ( v^{\ast } \right ) ^{\prime } \right ) ^{2}+2i\left ( -\left ( q_{n-1}^{\ast } \right ) ^{\prime } + \left ( v^{\ast } \right ) ^{\prime } \right ) + g_{n}^{\ast } \left ( x\right ) , \label{5.3 } \\ q_{n}^{\ast } \left ( 0\right ) & = & \psi _ { n}^{\ast , 0},\left ( q_{n}^{\ast } \right ) ^{\prime } \left ( 0\right ) = \psi _ { n}^{\ast , 1},\left ( q_{n}^{\ast } \right ) ^{\prime } \left ( 1\right ) = 0 . \notag\end{aligned}\]]since the number @xmath324 characterizes the error in the boundary data and since @xmath397 is the error parameter introduced in ( [ 4.110 ] ) , then , taking into account ( [ 4.16 ] ) , we set@xmath398 in ( [ 5.2 ] ) and ( [ 5.3 ] ) @xmath399 and @xmath400 are error functions , which can be estimated as @xmath401where @xmath402 is a constant . we also assume that @xmath403 } & \leq & m,\left\vert q_{n}^{\ast } \right\vert _ { h^{2}\left ( 0,1\right ) } \leq m,\left\vert \left ( v^{\ast } \right ) ^{\prime } \right\vert _ { c\left [ 0,1\right ] } \leq m,\left\vert \left ( v^{\ast } \right ) ^{\prime \prime } \right\vert _ { l^{2}\left ( 0,1\right ) } \leq m , \label{5.50 } \\ \left\vert \left ( w^{\ast } \right ) ^{\prime } \right\vert _ { c\left [ 0,1\right ] } & \leq & m,\left\vert \left ( w^{\ast } \right ) ^{\prime \prime } \right\vert _ { l^{2}\left ( 0,1\right ) } \leq m. \label{5.501}\end{aligned}\ ] ] * theorem 5.1*. _ let conditions of theorem 4.1 hold . in procedures ( i ) and ( iv ) of the algorithm set in the qrm _ @xmath404 _ _ assume that the number _ _ _ and the number _ _ @xmath261 _ is so large that in ( [ 4.0 ] ) _ @xmath406 _ _ for _ _ @xmath407 _ _ for _ _ @xmath408 _ _ let the function _ _ @xmath409 $ ] _ be the solution of the problem ( [ 4.3 ] ) , ( [ 4.4 ] ) with the exact coefficient _ @xmath410 _ _ and the exact data _ _ @xmath411 _ let @xmath412 be an integer in @xmath413.$ ] then there exists a sufficiently large constant _ @xmath414 _ _ for which estimates ( [ 5.5])-([5.501 ] ) are valid and which also satisfies _ _ @xmath415__as well as a constant _ _ @xmath416 _ _ such that if the error parameter _ _ @xmath327 _ _ is so small that _ _ @xmath417__then for _ _ @xmath418 _ _ the following estimate holds true _ _ @xmath419 * remark 6.1 . * _ thus , this theorem claims that our iteratively found functions _ @xmath361 _ _ are located in a sufficiently small neighborhood of the exact solution _ _ @xmath420 _ _ as long as _ _ .$ ] _ _ since this is achieved without any advanced knowledge of a small neighborhood of the exact solution _ _ @xmath422 _ _ , then theorem 5.1 implies the global convergence of our algorithm , see introduction . on the other hand , this is achieved within the framework of the approximation of subsection 4.2 . hence , to be more precise , this is the approximate global convergence property , see section 1.1.2 of @xcite and section 4 of @xcite for the definition of this property . recall that the number of iterations ( _ _ @xmath412 _ _ in our case ) can be considered sometimes as a regularization parameter in the theory of ill - posed problems @xcite . _ _ * proof*. in addition to ( [ 5.10 ] ) , we will also prove that for @xmath418 @xmath423 } , \left\vert q_{n}-q_{n}^{\ast } \right\vert _ { h^{2}\left ( 0,1\right ) } & \leq & \left ( m_{1}m^{16}\right ) ^{n}\eta < \sqrt{\eta } , \label{5.8 } \\ \left\vert v_{n+1}^{^{\prime } } -\left ( v^{\ast } \right ) ^{\prime } \right\vert _ { c\left [ 0,1\right ] } , \left\vert v_{n+1}^{\prime \prime } -\left ( v^{\ast } \right ) ^{\prime \prime } \right\vert _ { l^{2}\left ( 0,1\right ) } & \leq & \left ( m_{1}m^{16}\right ) ^{n}\eta < \sqrt{\eta } . \label{5.9}\end{aligned}\]]to simplify and shorten the proof , we assume in this proof that we work only with real valued functions . hence , we replace in two terms of ( [ 4.15 ] ) @xmath424 with 1 " and similarly in two terms of ( [ 5.3 ] ) . the case of complex valued functions is very similar . however , it contains some more purely technical details and is , therefore , more space consuming . we use the mathematical induction method . denote@xmath425using theorem 4.1 , we obtain@xmath426 } \leq m\eta < m^{9}\eta , \text { } \left\vert v_{1}^{\prime \prime } -\left ( v^{\ast } \right ) ^{\prime \prime } \right\vert _ { l^{2}\left ( 0,1\right ) } \leq m\eta < m^{9}\eta . \label{5.12}\]]hence , by ( [ 5.50 ] ) and ( [ 5.7])@xmath427 } \leq 2m,\left\vert v_{1}^{\prime \prime } \right\vert _ { l^{2}\left ( 0,1\right ) } \leq 2 m . \label{5.13}\]]following notations of section 3 , denote@xmath428 y^{\prime } , \label{5.14 } \\ l_{n,\ast } \left ( y\right ) & = & y^{\prime \prime } + \left [ 2k_{n}^{2}\left ( -\left ( q_{n-1}^{\prime } \right ) ^{\ast } + v^{\ast } \right ) -2ik_{n}\right ] y^{\prime } . \notag\end{aligned}\]]define @xmath429 since all functions @xmath329 are qrm solutions of corresponding problems with @xmath430 , then , using ( [ 5.3 ] ) and ( [ 4.15 ] ) , we obtain for all functions @xmath431,@xmath432 & = & \left ( -2k_{n}\left ( -q_{n-1}^{\prime } + v_{n}^{\prime } \right ) ^{2}+2\left ( -q_{n-1}^{\prime } + v_{n}^{\prime } \right ) , l_{n}z\right ) , \notag \\ \left ( l_{n,\ast } q_{n}^{\ast } , l_{n}z\right ) + \eta ^{2}\left [ q_{n}^{\ast } , z\right ] & = & \left ( -2k_{n}\left ( -\left ( q_{n-1}^{\ast } \right ) ^{\prime } + \left ( v^{\ast } \right ) ^{\prime } \right ) ^{2},l_{n}z\right ) \label{5.15 } \\ + \left ( 2\left ( -\left ( q_{n-1}^{\ast } \right ) ^{\prime } + \left ( v^{\ast } \right ) ^{\prime } \right ) , l_{n}z\right ) & & + \left ( g_{n}^{\ast } , z\right ) + \eta ^{2}\left [ q_{n}^{\ast } , z\right ] . \notag\end{aligned}\]]subtracting the second equality ( [ 5.15 ] ) from the first one and using ( [ 5.14 ] ) and , we obtain@xmath433 = \left ( \left ( -2k_{n}^{2}\left ( -\widetilde{q}_{n-1}^{\prime } + \widetilde{v}_{n}^{\prime } \right ) \left ( q_{n}^{\ast } \right ) ^{\prime } \right ) , l_{n}z\right)\]]@xmath434 . \notag\end{aligned}\]]in addition , by ( [ 5.4])@xmath435 we now explain the meaning of the constant @xmath436 since the constant @xmath437 in theorem 3.3 depends on @xmath438norms of coefficients of the operator @xmath439 in ( [ 3.1 ] ) , we need to estimate from the above the @xmath438norm of the coefficient of the operator @xmath440 in ( [ 5.14 ] ) . if ( [ 5.8 ] ) is true , then using ( [ 5.50 ] ) and ( [ 5.7 ] ) and noting that by ( [ 5.11 ] ) @xmath441 } \leq \left\vert \widetilde{q}_{n}\right\vert _ { c^{1}\left [ 0,1\right ] } + \left\vert q_{n}^{\ast } \right\vert _ { c^{1}\left [ 0,1\right ] } \leq 1+m\leq 2m,$ ] we obtain@xmath442 } \leq 2m,\left\vert q_{n}\right\vert _ { h^{2}\left ( 0,1\right ) } \leq 2 m . \label{5.18}\]]hence , by ( [ 4.14 ] ) @xmath443 hence , if ( [ 5.9 ] ) is also true , then the coefficient of the operator @xmath440 can be estimated as @xmath444thus , in the case of the operator @xmath440 in the analog of estimate ( [ 30 ] ) for the qrm , the constant @xmath227 should be replaced with another constant @xmath445 first , consider the case @xmath446 and estimate functions @xmath447 in this case @xmath448 and so ( [ 5.19 ] ) is an over - estimate of course . still , to simplify the presentation , we use @xmath449 in this case . estimate first two terms in the right hand side of ( 5.16 ) at @xmath446 . by theorem 4.1 , ( [ 5.5])-([5.6 ] ) , ( [ 5.11 ] ) and ( [ 5.13 ] ) @xmath450@xmath451hence , theorem 3.3 , ( [ 5.16 ] ) , ( [ 5.17 ] ) , ( [ 5.20 ] ) and ( [ 5.21 ] ) lead to@xmath452 } \leq m_{1}\left ( 16\overline{k}^{2}m^{2}\right ) \eta < m_{1}m^{5}\eta . \label{5.22}\ ] ] hence , ( [ 5.8 ] ) is true for @xmath446 . we now estimate derivatives of the function @xmath453 by ( [ 5.13 ] ) and ( [ 5.18 ] ) @xmath454next , by ( [ 5.1 ] ) , ( [ 5.12 ] ) and ( [ 5.22])@xmath455 } \leq \eta \left\vert \widetilde{q}_{1}^{\prime } \right\vert _ { c\left [ 0,1\right ] } + \left\vert \widetilde{v}_{1}^{\prime } \right\vert _ { c\left [ 0,1\right ] } \leq m_{1}m^{5}\eta + m\eta \leq 2m_{1}m^{5}\eta . \label{5.24}\]]we now estimate @xmath456 subtracting ( [ 5.2 ] ) from ( [ 4.40 ] ) and using ( 4.22 ) , ( [ 5.51 ] ) , ( [ 5.6 ] ) , ( [ 5.12 ] ) , ( [ 5.22 ] ) and ( 5.24 ) , we obtain @xmath457 } \label{5.25 } \\ & \leq & m_{1}m^{5}\eta + m\eta + 2m_{1}m^{5}\overline{k}^{2}\left ( 3m+m^{2}\right ) \eta \leq m_{1}m^{8}\eta . \notag\end{aligned}\ ] ] we now estimate functions @xmath458 and @xmath459 let @xmath460 recall that we find the function @xmath461 via solving the problem ( [ 4.3 ] ) , ( [ 4.4 ] ) with @xmath308 using the qrm . also , it follows from ( [ 2.38 ] ) and ( [ 4.11 ] ) that @xmath462 and @xmath463 hence , we obtain similarly with ( [ 5.15 ] ) and ( [ 5.16 ] ) @xmath464 & = & -\left ( \overline{k}^{2}\widetilde{\beta } _ { 1}w^{\ast } , a_{1}z\right ) -\eta ^{2}\left [ w^{\ast } , z\right ] , \forall z\in w_{0 } , \label{5.26 } \\ \left\vert \widetilde{w}_{1}\left ( 0\right ) \right\vert & \leq & \eta , \left\vert \widetilde{w}_{1}^{\prime } \left ( 0\right ) \right\vert \leq m\eta , \widetilde{w}_{1}^{\prime } \left ( 1\right ) = 0 . \notag\end{aligned}\]]the function @xmath465 is the solution of a qrm problem , which is completely similar with ( [ 5.26 ] ) . since by ( 4.2323 ) functions @xmath466 are uniformly bounded for all @xmath346 , @xmath467 then there exists an analog of the constant @xmath227 of theorem 3.2 , which estimates functions @xmath468 for all @xmath346 as solutions of analogs of problems ( [ 5.26 ] ) . hence , we can assume that this constant equals @xmath449 . using theorem 3.2 , ( [ 5.25 ] ) and ( [ 5.26 ] ) , we obtain@xmath469 } , \left\vert \widetilde{w}_{1}\right\vert _ { h^{2}\left ( 0,1\right ) } \leq m_{1}m^{10}\eta . \label{5.27}\]]hence , using ( [ 5.7 ] ) , ( [ 5.27 ] ) and @xmath470 we obtain@xmath471 } , \left\vert w_{1}\right\vert _ { h^{2}\left ( 0,1\right ) } \leq 2m.\ ] ] the final step of the proof for the case @xmath446 is to estimate derivatives of the second tail , i.e. functions @xmath472 by ( [ 4.17])@xmath473estimate the denominator in ( [ 5.29 ] ) . since @xmath474 in ( [ 4.0 ] ) for @xmath475 then ( [ 2.1 ] ) and ( [ 4.0 ] ) imply that @xmath476 for @xmath477 .$ ] hence , using ( [ 5.7 ] ) and ( [ 5.27 ] ) , we obtain @xmath478hence,@xmath479we now estimate from the above the modulus of the nominator in each of two formulas of ( [ 5.29 ] ) . using ( [ 5.501 ] ) , ( [ 5.6 ] ) and ( [ 5.27])-([5.30 ] ) , we obtain for @xmath477 $ ] @xmath480next , @xmath481hence , we obtain similarly with ( [ 5.31 ] ) @xmath482it can be easily derived from ( [ 5.7 ] ) , ( [ 5.31 ] ) and ( [ 5.32 ] ) that @xmath483 } \leq 2m$ ] and @xmath484 thus , in summary ( [ 5.22 ] ) , ( [ 5.25 ] ) , ( [ 5.31 ] ) and ( [ 5.32 ] ) imply that @xmath452 } , \left\vert \widetilde{\beta } _ { 1}\right\vert _ { l^{2}\left ( 0,1\right ) } , \left\vert \widetilde{v}_{2}^{\prime } \right\vert _ { c\left [ 0,1\right ] } , \left\vert \widetilde{v}_{2}^{\prime \prime } \right\vert _ { l^{2}\left ( 0,1\right ) } \leq m_{1}m^{14}\eta . \label{5.33}\]]in other words , estimates ( [ 5.8])-([5.10 ] ) are valid for @xmath446 . assume that they are valid for @xmath485 where @xmath486 . denote @xmath487 then , similarly with the above , one will obtain estimates ( [ 5.33 ] ) where 1 `` in first three terms is replaced with @xmath346 , 2 '' in the fourth and fifth terms is replaced with @xmath488 and the right hand side is @xmath489 @xmath63 in all our computation , @xmath490 and @xmath491 $ ] . we have observed in our computationally simulated data as well as in experimental data that the function @xmath492 becomes very small for @xmath493 on the other hand , the largest values of @xmath494 were observed in some points of the interval @xmath495 $ ] . thus , we assign in all computations @xmath496 we have used @xmath497 actually in all our computations we go along the interval @xmath495 $ ] several times . more precisely , let @xmath498 be the result obtained in step [ step 3 ] of the above globally convergent algorithm . set @xmath499 . next , solve the problem the problem ( [ 4.3 ] ) , ( [ 4.4 ] ) with @xmath290 and @xmath500 . let the function @xmath501 be its solution . then we define derivatives of the new tail function @xmath502 as in ( [ 4.17 ] ) where @xmath503 is replaced with @xmath504 next , we go to step 2 and repeat . the process is repeated @xmath505 times in our computational program . we choose @xmath506 such that@xmath507@xmath508 it is also worth mentioning that in step step 2b3 , after calculating @xmath368 , we replace it by @xmath509where @xmath510 is a small neighborhood of @xmath511 , @xmath339 and @xmath512 is the length of the interval @xmath513 we also use the truncation technique to improve the accuracy of the reconstruction function @xmath514 ( see ) . in this section , we show the numerical reconstruction of the spatially distributed dielectric constant from computationally simulated data . let the function @xmath515 has the form @xmath516let the function @xmath517 be the solution of problem , . as mentioned in the proof of theorem [ thm unique forward ] ( see ) , the function @xmath517 satisfies the lippman - schwinger equation , @xmath518this equation can be approximated as a linear system . we have solved that system numerically to computationally simulate the data . our numerical results are displayed in figures 1 . in the top row @xmath519 and in the bottom row @xmath520 in each figure , we show the true function @xmath515 , the data obtained by solving the forward problem with that true @xmath515 and the solution of the cip . in both cases we had two values of @xmath521 and @xmath522 these figures confirm that both the target / background contrast and the position of the target are computed with small errors . the experimental data were collected by the forward looking radar which was built in the us army research laboratory @xcite . the device consists of two main parts . the first one , emitter , generates the time resolved electric pulses . the emitter sends out only one component of the electric field . the second part involves 16 detectors . these detectors collect the time resolved backscattering electric signal ( voltage ) in the time domain . the same component of the electric field is measured as the one which is generated by the emitter : see our comment about this in the beginning of section 2 . the step size of time is 0.133 nanosecond . the backscattering data in the time domain are collected when the distance between the radar and the target varies from 20 to 8 meters . the average of these data with respect to both the position of the radar and those 16 detectors is the data on which we have tested our algorithm . to identify horizontal coordinates of the position of the target , ground positioning system ( gps ) the error in each of horizontal coordinates does not exceed a few centimeters . when the target is under the ground , the gps provides the distance between the radar and a point on the ground located above the target . as to the depth of a buried target , it is not of a significant interest , since horizontal coordinates are known and it is also known that the depth does not exceed 10 centimeters . we refer to @xcite for more details about the data collection process . publications @xcite contain schematic diagrams of the measurements . our interest is in computing maximal values of dielectric constants of targets . in one target ( plastic cylinder below ) we compute the minimal value of its dielectric constant , since its value was less than the dielectric constant of the ground . for each target , the only information the mathematical team ( mvk , lhn ) had , in addition to just a single experimental curve , was whether it was located in air or below the ground . we calculate @xmath523 the relative spatial dielectric constant of the whole structure including the background ( air or the ground ) and the target . more precisely , @xmath524where @xmath525 is the subinterval of the interval @xmath29 which is occupied by the target . here @xmath526 and @xmath527 are values of the function @xmath528 in target and background respectively . we assume that @xmath529 for each set of experimental data . hence , @xmath530 if the target is located in air the ground was dry sand . it is well known that the dielectric constant of the dry sand varies between 3 and 5 @xcite . hence , in the case of buried targets @xmath531 $ ] . in our mathematical model the time resolved electric signal @xmath532 collected by the detectors satisfies the equations , with @xmath515 being replaced by @xmath523 where the position @xmath119 of the source is actually unknown . the latter is one of the difficulties of working with these data . thus , we set in all our tests @xmath533 which is the same as in @xcite . let @xmath534 be the function @xmath535 which we compute . following @xcite@xmath536 , \\ \min r\left ( x\right ) \text { if } 0<r\left ( x\right ) \leq 1,\forall x\in \left [ 0,1\right ] .\end{array}\right . \label{6.2}\]]since the dielectric constant of air equals 1 , then we have @xmath537 for targets located in air . as to the buried targets , we have developed a procedure of the analysis of the original time resolved data , which provides us with the information on which of two cases ( [ 6.2 ] ) takes place . we refer to case 1 and case 2 on page 2944 of ksnf2 for this procedure . in addition , since we had a significant mismatch of magnitudes of experimental and computationally simulated data , we have multiplied , before computations , our experimental data by the calibration number @xmath538 see @xcite for details of our choice of this number . there is a significant discrepancy between computationally simulated and experimental data , which was noticed in our earlier publications karch , ksnf1,ksnf2 . this discrepancy is evident from , comparison of , e.g. figure 1b with figure 2b and other similar ones . therefore , to at least somehow mitigate this discrepancy , we perform a data pre - processing procedure . besides of the fourier transform of the time resolved data , we multiply them by a calibration factor , truncate a certain part of the data in the frequency domain and shift the data in the frequency domain , see details below . the function @xmath539 which we have studied in the previous sections , is the fourier transform of the time resolved data @xmath540 . the function @xmath541 is called the data in the frequency domain " . observing that @xmath542 is small when @xmath13 belongs to a certain interval , we do not analyze @xmath541 on that interval . rather , we only focus on such a frequency interval which contains the major part of the information . now , to keep the consistency with our study of computationally simulated data , we always force the frequency interval to be @xmath543 $ ] . to do so , we simply shift our data in the frequency domain : compare figures 2b and 2c , figures 2f and 2 g , figures 3b and 3c , figures 3f and 3 g and figures 3j and 3k . we consider two cases : targets in air ( figure [ air ] ) and targets buried about a few centimeters under the ground ( figure [ ground ] ) . we had experimental data for total of five ( 5 ) targets . the reconstructed dielectric constants of these targets are summarized in table 1 . in this table , computed @xmath544 . in tables of dielectric constants of materials , their values are usually given within certain intervals @xcite . now about the intervals of the true @xmath545 in the @xmath546 column of table 1 . in the cases when targets were a wood stake and a plastic cylinder , we have taken those intervals from a published table of dielectric constants @xcite . the interval of the true @xmath547 for the case when the target was bush , was taken from @xcite . as to the metal targets , it was established in @xcite that they can be considered as such targets whose dielectric constants belong to the interval @xmath548.$ ] * * table 1 : * * computed dielectric constants of five targets [ cols="^,^,^,^,^,^",options="header " , ] + in this paper , we have developed a frequency domain analog of the 1-d globally convergent method of @xcite . we have tested this analog on both computationally simulated and time resolved experimental data . the experimental data are the same as ones used in karch , ksnf1,ksnf2 . we have modeled the process of electromagnetic waves propagation by the 1-d wave - like pde . the reason why we have not used a 3-d model , as in , e.g. earlier works of this group on experimental data bk1,tbkf1,tbkf2 , is that we had only one time resolved experimental curve for each of our five targets . our numerical method has the global convergence property . in other words , we have proven a theorem ( theorem 5.1 ) , which claims that we obtain some points in a sufficiently small neighborhood of the exact solution without any advanced knowledge of this neighborhood . our technique heavily relies on the quasi reversibility method ( qrm ) . the proof of the convergence of the qrm is based on a carleman estimate . a significant modification of our technique , as compared with @xcite , is due to two factors . first , we use the fourier transform of time resolved data instead of the laplace transform in @xcite . second , when updating tail functions via ( [ 4.17 ] ) , we solve the problem ( [ 4.3 ] ) , ( [ 4.4 ] ) using the qrm . on the other hand , in @xcite tail functions were updated via solving the laplace transform analog " of the problem ( [ 2.4 ] ) , ( [ 2.6 ] ) as a regular forward problem since the dielectric constants of targets were not measured in experiments , the maximum what we can do to evaluate our results is to compare them with published data in , e.g. @xcite . results of table 1 are close to those obtained in @xcite . one can see in table 1 that our reconstructed values of dielectric constants are well within published limits . we consider the latter as a good result . this is achieved regardless on a significant discrepancy between computationally simulated and experimental data , regardless on a quite approximate nature of our mathematical model and regardless on the presence of clutter at the data collection site . that discrepancy is still very large even after the data pre - processing , as it is evident from comparison of figures 1b , e with figures 2c , g and figures 3c , g , k . besides , the source position @xmath119 was unknown but rather prescribed by ourselves as @xmath549 thus , our results indicate a high degree of stability of our method . l. beilina and m.v . klibanov , a new approximate mathematical model for global convergence for a coefficient inverse problem with backscattering data , _ j. inverse and ill - posed problems _ , 20 , 512 - 565 , 2012 . l. beilina , energy estimates and numerical verification of the stabilized domain decomposition finite element / finite difference approach for the maxwell s system in time domain , _ central european journal of mathematics _ , 11 , 702733 , 2013 . klibanov and j. malinsky , newton - kantorovich method for 3-dimensional potential inverse scattering problem and stability for the hyperbolic cauchy problem with time dependent data , inverse problems , 7 577 - 596 , 1991 . kuzhuget , l. beilina , m.v . klibanov , a. sullivan , l. nguyen and m.a . fiddy , blind backscattering experimental data collected in the field and an approximately globally convergent inverse algorithm , _ inverse problems _ , 28 , 095007 , 2012 . a.v . kuzhuget , l. beilina , m.v . klibanov , a. sullivan , l. nguyen and m.a . fiddy , quantitative image recovery from measured blind backscattered data using a globally convergent inverse method , _ ieee transaction for geoscience and remote sensing , _ 51 , 2937 - 2948 , 2013 . n. nguyen , d. wong , m. ressler , f. koenig , b. stanton , g. smith , j. sichina and k. kappra , obstacle avolidance and concealed target detection using the army research lab ultra - wideband synchronous impulse reconstruction ( uwb sire ) forward imaging radar , _ proc . spie _ * 6553 * 65530h ( 1)-65530h ( 8) , 2007 . n. t. thnh , l. beilina , m. v. klibanov and m. a. fiddy , reconstruction of the refractive index from experimental backscattering data using a globally convergent inverse method , _ siam journal on scientific computing _ , 36 , b273b293 , 2014 . n. t. thnh , l. beilina , m. v. klibanov and m. a. fiddy , imaging of buried objects from experimental backscattering time dependent measurements using a globally convergent inverse algorithm , _ siam j. imaging sciences _ , 8 , 757 - 786 , 2015 .

superwimps , superweakly interacting massive particles produced in the late decays of weakly - interacting massive particles ( wimps ) , are promising non - baryonic dark matter candidates @xcite . well - motivated superwimp candidates are the gravitino in supersymmetric models @xcite and the first excited graviton in universal extra dimension models @xcite . the supersymmetric possibility is realized naturally in supergravity with a gravitino lightest supersymmetric particle ( lsp ) and a slepton or sneutrino next - to - lightest supersymmetric particle ( nlsp ) . ( throughout this paper , `` slepton '' denotes a charged slepton . ) both the gravitino and the nlsp have weak - scale masses . the nlsp freezes out as usual with a relic density near the observed value . however , after time @xmath0 , it decays through @xmath1 the gravitino then inherits much of the relic density of the slepton or sneutrino @xcite , and its relic density is naturally of the right magnitude without the introduction of new scales . this is in contrast to other gravitino dark matter scenarios , where the gravitino is a thermal relic and the desired density is obtained by an appropriately chosen gravitino mass @xmath2 @xcite , or the gravitino is produced during reheating @xcite , where the desired dark matter density is obtained only for a fine - tuned reheat temperature @xmath3 . the decays of eq . ( [ decay ] ) occur well after big bang nucleosynthesis ( bbn ) . an immediate concern , therefore , is that they might destroy the successful light element abundance predictions of bbn . in fact , bbn is not the only worry the planckian spectrum of the cosmic microwave background ( cmb ) , the diffuse photon background , and bounds on late time entropy production from the coincidence between bbn and cmb baryometry all impose significant constraints . as demonstrated in refs . @xcite , however , these constraints on the electromagnetic ( em ) energy released in the decays of eq . ( [ decay ] ) exclude some of the weak scale parameter space , but leave much of it intact . of course , at the border between the excluded and allowed regions , slight deviations from standard cosmological predictions are expected , providing possible signals in future observations . in fact , the existing anomaly in the @xmath4li abundance prediction of standard bbn may already be interpreted as a signal of superwimp dark matter @xcite . such signals are particularly welcome , since superwimp dark matter is so weakly interacting that it is impossible to detect through conventional direct and indirect dark matter searches . the previous work , however , neglected hadronic energy produced in wimp decays . this was natural , as the wimp decays of eq . ( [ decay ] ) contribute only to electromagnetic energy , as we will discuss below . however , the three - body decays @xmath5 produce hadronic energy when the @xmath6 or @xmath7 decays hadronically . hadronic energy release is very severely constrained by the observed primordial light element abundances @xcite , and so even subdominant hadronic decays could , in principle , provide stringent constraints . these three - body decays may be kinematically suppressed when @xmath8 , but even in this case , four - body decays , such as @xmath9 , contribute to hadronic cascades and may be important . in this paper , we determine the hadronic branching fractions and compare them to bbn constraints on hadronic energy release . although we focus on the supersymmetric case , our results may be extended straightforwardly with minor numerical modifications to the case of graviton superwimps in extra dimensions . in evaluating the constraints , there are two possible approaches . as the superwimp relic abundance is automatically in the right range , a natural assumption is that superwimp gravitinos make up all of the non - baryonic dark matter , with @xmath10 this is the approach taken here . note that @xmath11 need not be identical to @xmath12 not all relic gravitinos need be produced by nlsp decays . our assumption therefore requires that the other sources of gravitinos be insignificant . for the supersymmetric case , this typically implies an upper bound on reheat temperatures of @xmath13 @xcite . higher reheat temperatures may also be allowed @xcite and are desirable , for example , to accommodate leptogenesis . for the universal extra dimension case , the requirement of insignificant kaluza - klein graviton production during reheating implies @xmath14 to @xmath15 , depending , in part , on the number of extra dimensions @xcite . on the other hand , the universe has proven to be remarkably baroque , and there is no guarantee that dark matter is composed of only one component . one might therefore relax the constraint of eq . ( [ omega ] ) on the superwimp energy density and assume , for example , that the nlsp freezes out with its thermal relic density @xmath16 . the superwimp gravitino density is then @xmath17 . in this approach , the gravitino density may be low and even insignficant cosmologically . from a particle physics viewpoint , however , the viability of the gravitino lsp scenario is still worth investigation , as it has strong implications for the superpartner spectrum and collider signatures , independent of its cosmological importance . these approaches differ significantly , not only in the viewpoint taken , but also in their implications . suppose , for example , that the gravitino and nlsp masses are both parametrized by a mass scale @xmath18 . the nlsp number density then scales as @xmath19 if one assumes eq . ( [ omega ] ) , but scales as @xmath18 if one assumes a thermal relic nlsp density , since @xmath20 , where @xmath21 is the thermally - averaged nlsp annihilation cross section . low masses are excluded in the former case , while high masses are disfavored in the latter . we consider both approaches to be worthwhile , but defer discussion of the thermal relic density approach , along with its very different consequences for the superpartner spectrum and implications for collider physics , to a separate study @xcite . assuming a fixed @xmath11 here , we find that gravitino superwimps provide a viable solution to the dark matter problem . we determine the allowed masses for the nlsp and gravitino . at the same time , we find that the hadronic energy constraint is significant and does in fact provide the leading constraint in natural regions of parameter space . we conclude that the analysis below must be done in any scenario with similar late decays . given any standard cosmology , a significant thermal relic abundance of nlsps will be generated , these nlsps will eventually decay to superwimps , and the hadronic constraints we discuss below must be analyzed before the scenario may be considered viable . in passing , we note that another possibility is that the nlsp is the lightest neutralino . for a general neutralino , the decays @xmath22 produce large hadronic energy release . these two - body modes are absent for photino - like neutralinos , where the only two - body decay is @xmath23 . however , even in this case , one can not avoid hadronic activity from decay through a virtual photon , @xmath24 , with a branching fraction of @xmath25 . given the stringency of constraints on hadronic energy release , the neutralino nlsp case is highly constrained . assuming that @xmath11 makes up a significant fraction of the dark matter , natural regions of parameter space are excluded @xcite . in the current paper , we therefore focus on the more promising possibilities of slepton and sneutrino nlsps . the paper is organized as follows . in sec . [ sec : bbn ] , we summarize what is known about constraints on em and hadronic energy release in the early universe . in sec . [ sec : swimp ] , we give a detailed description of the gravitino superwimp scenario with a slepton or sneutrino nlsp . in sec . [ sec : had ] , we determine the hadronic energy release from slepton and sneutrino decays and find the allowed mass regions . we summarize in sec . [ sec : conclusion ] . the relevant gravitino interactions , decay amplitudes , and three - body hadronic decay width formulae are collected in the appendix . the overall success of standard bbn places severe constraints on energy produced by particles decaying after bbn . given the baryon density specified by cmb measurements @xcite , the prediction of standard bbn @xcite for deuterium agrees well with observations @xcite . this concordance is less perfect , but still reasonable , for @xmath26he @xcite . on the other hand , current observations of @xmath4li @xcite are consistently lower than predictions ; this is the leading anomaly in light element abundances at present . the abundances of @xmath27he and @xmath28li also provide constraints , although these are subject to more theoretical uncertainties as discussed below . in this section , we review the constraints from these element abundances on energy release in the early universe . although a unified analysis is necessary for detailed conclusions , we review the constraints on em and hadronic energy release in turn . a schematic summary of the discussion in this section is given in fig . [ fig : bbnconstraints ] . the constraints on em energy release are well understood , and we begin with a brief summary of the results . ( for details , see ref . @xcite . ) high energy photons and leptons produced by late - decaying particles are quickly thermalized through interactions with background ( bg ) particles . high energy photons , for example , initiate em cascades through processes like @xmath29 and @xmath30 , and the initial energy is quickly converted to very soft photons . the same is true for high energy electrons and muons . high energy taus decay to mesons , but these decay and produce em cascades before interacting hadronically . ( see ref . @xcite and the analysis below . ) injected leptons of all three families therefore contribute dominantly to em energy and negligibly to hadronic energy . the energy distribution of the resulting photons is highly suppressed for @xmath31 , where @xmath32\end{aligned}\ ] ] is related to the energy threshold for @xmath33 @xcite . here @xmath34 is the bg photon temperature ; in the radiation - dominated era , it is related to time by @xmath35 ^ 2 \ .\end{aligned}\ ] ] because of this rapid thermalization , constraints on em energy depend on the total amount of energy released and are essentially independent of the shape of the initial energy spectrum . for @xmath36 , the bbn constraints are rather weak since @xmath37 is too small to destroy any light elements . for @xmath38 , the universe is hot and @xmath37 is low , and so only light elements with very low binding energies can be destroyed . the main constraint is from the overdestruction of d , with a binding energy of only 2.2 mev . for @xmath39 , the universe is cooler and @xmath37 becomes high enough that @xmath26he , with binding energy 19.8 mev , is destroyed . even a small amount of @xmath26he photo - dissociation becomes a significant production mechanism for @xmath27he and d , since the observed abundances of @xmath27he and d are much lower than that of @xmath26he . the main constraint is from the overproduction of d. between @xmath40 and @xmath41 there is a region where the overdestruction and overproduction of d cancel . for sufficiently high energy release in this region , the d abundance is that of standard bbn , but @xmath4li , with binding energy 2.5 mev , is destroyed through @xmath42 . the observed @xmath4li abundance @xcite is in fact low compared to the predictions of standard bbn @xcite . we see , then , that the @xmath4li abundance can be brought into agreement with observations without upsetting the d concordance for decay times in this window . for @xmath43 , the best fit values are @xcite @xmath44 where @xmath45 is the emitted em energy density . here @xmath46 is the initial em energy released in nlsp decay , and @xmath47 is the branching fraction of nlsp decay into em components . @xmath48 is the nlsp number density just before nlsp decay , normalized to the bg photon number density @xmath49 . it can be expressed in terms of the superwimp abundance : @xmath50 \left[\frac{{\omega_{\text{swimp}}}}{0.23}\right ] \ . \label{eq : def_y}\ ] ] the parameters of eq . ( [ bestfit ] ) are naturally realized in the superwimp scenario when a weak - scale nlsp decays to a weak - scale superwimp with the observed dark matter density @xcite . the numerical constraints on em energy release as a function of the particle decay lifetime have been determined in ref . @xcite , using the following criteria : @xmath51 we adopt these constraints in our consideration of em energy release bounds . hadronic energy release , and its effects on bbn , has also been considered @xcite . although we are most interested in energy release at times @xmath52 , we summarize here what is known for @xmath53 for completeness . as we will show below , for the decay times of interest , mesons decay quickly and so contribute to em energy . only protons and neutrons are sufficiently long - lived that they might be able to interact hadronically . for times @xmath54 , however , even high energy protons and neutrons ( with energies @xmath55 ) interact through em scattering with bg photons before interacting with bg baryons . injected high energy baryons then quickly become slow baryons , with the initial energy dissipated to em energy . for such early times , this em energy is harmless , as discussed above . constraints are therefore functions of the number density of injected baryons @xmath56 , and independent of the initial baryon energy @xmath57 . slow baryons modify the number of neutrons and protons and therefore bbn predictions . in the very earliest era with @xmath58 , however , the weak interaction @xmath59 is efficient , and so any deviation in the @xmath60 ratio is washed out , and there are no strong bounds . for @xmath53 , the weak interactions decouple . slow baryons produced in late decays cause the @xmath60 ratio to deviate from the standard bbn value . in particular , extra neutrons create d. for @xmath61 , the resulting d is immediately burned to form @xmath26he , and so the dominant constraint is from overproduction of @xmath26he . this constraint is weak since the @xmath26he abundance of standard bbn is already large . for @xmath62 , d can no longer burn into @xmath26he @xcite and the dominant constraint comes from d overproduction . this bound is significantly stronger , since the d / h ratio is @xmath63 . at times @xmath64 , injected baryons interact with bg baryons before thermalizing . these interactions create secondary hadrons that can modify bbn predictions . in this era , constraints depend in principle on both @xmath57 and @xmath65 separately . in principle , the constraints do not simply depend on the product , because the total number of secondary hadrons created by an injected hadron is not always proportional to the injected hadronic energy . the dominant constraint is from secondary baryons destroying @xmath26he to overproduce d. quantitative analyses on the bbn constraints of the hadronic energy injection have been performed in refs . as discussed above , formally , the constraints depend on the lifetime @xmath66 of the decaying particle and on @xmath57 and @xmath56 separately . however , the most recent analysis @xcite finds that for decay times @xmath67 , when the hadronic constraint is most important , the constraints depend on the product @xmath68 to within a factor of 2 for @xmath69 @xcite . we will therefore assume this behavior and apply the hadronic constraint to times @xmath70 , where it is stronger than the em constraint in the superwimp scenario . the authors of ref . @xcite assume the following constraints on the light element abundances ( 95% cl ) : @xmath71 in our analysis of hadronic energy injection , we take the constraints of ref . @xcite derived using these bounds and @xmath72 . note that the constraints of eqs . ( [ dconstraints])-([hadconstraints ] ) adopted for the hadronic energy analysis are less conservative than the constraints of eqs . ( [ d])-([emconstraints ] ) used in the em analysis of ref . @xcite . in particular , a significantly more aggressive d bound is assumed in the hadronic analysis . there is now impressive concordance between the baryon number determinations from d and cmb measurements , which further supports the narrow range of d / h given in eq . ( [ dconstraints ] ) . at the same time , existing anomalies in standard bbn may be indications that systematic errors are underestimated or that there is new physics , which would likely distort the d abundance , since d is very weakly bound . there are several further comments to make . first , we have been treating the em and hadronic effects separately . of course , these effects can add constructively or destructively @xcite . we have explained above that the dominant hadronic constraint for @xmath52 is from d overproduction , whereas for the em constraint it is from d overproduction for @xmath73 and from d overdestruction for @xmath38 . for phenomena contributing to both em and hadronic energy , then , the effects add constructively for @xmath74 , but may cancel for @xmath38 , and there may be fingers of allowed region in the latter time interval . finally , although we consider only the constraints above on d / h , @xmath75 and @xmath76li / h , in ref . @xcite , constraints from @xmath27he / d and @xmath28li / h are also analyzed for late time em and hadronic injections . these two abundances , however , are subject to more uncertainties . @xmath27he / d is considered because the destruction of @xmath26he could produce @xmath27he more easily than d , while the binding energy of @xmath27he is higher than that of d. however , the interpretation of @xmath27he measurements @xcite may be subject to ambiguities arising from stellar production and destruction @xcite . the ratio of @xmath28li/@xmath4li has been observed in low metallicity halo stars and might also constrain late decays . the implications for em cascades have also been explored in refs . again , however , there are arguably significant systematic uncertainties arising from the higher depletion of @xmath28li relative to @xmath4li as a result of its smaller binding energy @xcite . a better understanding of the connection between the primordial and observed abundances of @xmath28li and @xmath4li could provide a very powerful constraint on late decaying particles @xcite . in the superwimp dark matter scenario , superwimps have no effect on the early thermal history of the universe . they are produced only through metastable nlsp decays , nlsp@xmath77 + sm particle , giving a gravitino relic density of @xmath78 @xmath79 is allowed , since @xmath80 . the sm particles produced in the nlsp decay induce em or hadronic cascades , which will affect the primordial abundance of light elements , as discussed in sec . [ sec : bbn ] . the dominant decay mode for slepton nlsps is @xmath81 . most of the energy released from the decay is transferred to em cascades . the total energy released in slepton decay is @xmath82 the total injected em energy density is , for @xmath83 , @xmath84 \left[\frac{\delta m}{300~{\text{gev}}}\right ] \left[\frac{{\text{tev}}}{m_{\tilde{g}}}\right ] \left[\frac{{\omega_{\text{swimp}}}}{0.23}\right ] \ , \end{aligned}\ ] ] where @xmath85 depends on the slepton s flavor , and eq . ( [ eq : def_y ] ) has been used . for selectrons and smuons , energetic electrons and muons quickly thermalize with the background photons . all the visible energy is transferred to the resulting em cascade : @xmath86 , @xmath87 . in the case of staus , the decays of taus could also produce mesons and neutrinos , in addition to electrons and muons . the em activity depends on the lifetime and energy distribution of decay products . as estimated in ref . @xcite , the em energy released in stau decay is in the range @xmath88 to @xmath89 . for the sneutrino nlsp case , the sneutrino mainly decays into neutrino and gravitino . the emitted high energy neutrino interacts with bg neutrinos and loses energy . if the neutrinos annihilate through @xmath90 , an em cascade is initiated . however , the annihilation takes place through weak interactions , and so its cross section may be small . the effect of the expansion of the universe must also be taken into account @xcite . sneutrino decays induce far less em energy release than sleptons , but for decay times @xmath91 , there may still be some impact on bbn . for such decays times , however , much of the parameter space is excluded by considerations of hadronic bbn constraints , which will be discussed below . in what is left of this region , the gravitino mass is @xmath92 . such light gravitinos are less motivated in supergravity and so we have not considered this scenario in great detail here . the decay width of the slepton into lsp gravitino is @xmath93 ^ 4 \ , \label{eq : sleptondecay}\end{aligned}\ ] ] where @xmath94 . for @xmath95 , the slepton decay lifetime is @xmath96 ^ 4 \left[\frac{m_{\tilde{g}}}{{\text{tev}}}\right]\ . \label{eq : decaylifetime}\end{aligned}\ ] ] this expression is valid only when the gravitino and slepton are nearly degenerate , but we present it here as a useful guide for the reader . in the analysis below , we always use eq . ( [ eq : sleptondecay ] ) for the evaluation of the slepton decay lifetime . the slepton lifetime is given in fig . [ fig : lifetime_zetaem]a in the @xmath97 plane , where @xmath98 note that in the definition of @xmath99 , an additional @xmath100 is subtracted from @xmath101 ; three - body decays , to be discussed below , are therefore always kinematically possible for @xmath102 . the typical decay time is @xmath103 . for such decay times , there are strong constraints from the primordial abundance of the light element , as discussed earlier . contours of em energy release @xmath104 are given in fig . [ fig : lifetime_zetaem]b for @xmath105 nlsps , assuming @xmath106 , @xmath87 , and @xmath107 . we see that the em energy release varies from @xmath108 and is largest for large @xmath99 and small @xmath109 , when the decay lifetime is the shortest . the bbn constraints , however , are much weaker for early decays . as discussed in ref . @xcite , considering em energy constraints only , one finds that for a stau nlsp , @xmath110 and @xmath111 are allowed . moreover , late time em energy release from stau decays to gravitinos may resolve the current discrepancy in the @xmath4li for @xmath112 . for selectron or smuon nlsps , the em constraints are slightly stronger since all the released decay energy is transferred to em cascades . the em bbn constraints may be weakened by a number of possibilities . first , for decays in the range @xmath113 , the em constraints may be weakened by cancelation between the effects of em and hadronic energy release , as discussed above . second , if not all of non - baryonic dark matter is superwimps , @xmath114 , the contour lines for @xmath115 will shift toward the left , since @xmath116 , and the allowed parameter space will be enlarged . finally , if the nlsp is not a slepton but a sneutrino , the em bbn constraints are , of course , much weaker , as the two - body decays contribute much less visible energy than in the slepton nlsp case . as discussed in sec . [ sec : bbn ] , the abundances of light elements constrain the amount of hadronic energy injection into the early universe . although , for leptonic nlsps , the hadronic energy release is much smaller than the em energy release , the constraints on the hadronic energy release are much stronger . it is therefore important to evaluate in detail the hadronic activity induced by slepton and sneutrino nlsp decays . let us first study the slepton nlsp possibility . the main decay mode of sleptons is @xmath117 . for selectron and smuon nlsps , such decays induce only em cascades . for stau nlsps , however , the resulting tau lepton may decay into mesons , and these might interact with background hadrons and induce dangerous hadronic cascades . the interaction time of the produced mesons is @xmath118^{-1 } & \simeq & 18~{\text{s}}\left[\frac{100~{\text{mb } } } { \langle \sigma_{\text{had}}v \rangle } \right ] \left[\frac{6\times 10^{-10}}{\eta } \right ] \left[\frac{{\text{kev}}}{t } \right]^3 \nonumber \\ & \simeq & 1.8\times 10^{-8}~{\text{s}}\left[\frac{100~{\text{mb}}}{\langle \sigma_{\text{had } } v \rangle } \right ] \left[\frac{6\times 10^{-10}}{\eta } \right ] \left[\frac{t}{1~{\text{s } } } \right]^{3/2 } \ , \label{eq : interaction}\end{aligned}\ ] ] where @xmath119 is the thermally averaged strong interaction cross section . for the decay times of greatest interest here , @xmath103 , this interaction time is long compared to the typical meson ( @xmath120,@xmath121 ) lifetime , @xmath122 . the produced mesons thus decay before they interact with background hadrons . as a result , the hadronic activity from the dominant two - body decay of slepton is negligibly small . the main contributions to hadronic activity therefore come from three - body decays , @xmath123 with @xmath6 or @xmath7 decaying hadronically , or from four - body decays @xmath124 with @xmath125 . the corresponding feynman diagrams are shown in fig . [ fig : feynmangraphs ] . formulae for the decay amplitudes and decay width calculations are listed in the appendix . the hadronic branching fraction is defined as follow : @xmath126 where @xmath127 are the @xmath6 and @xmath7 hadronic branching fractions . for right - handed sleptons , @xmath128 , and there are additional suppressions because the @xmath6 couplings are smaller than the @xmath7 couplings . thus , the hadronic branching fraction of right - handed sleptons is smaller than that of left - handed ones , especially when @xmath129 . for small @xmath101 , the slepton lifetime increases dramatically , since it grows as @xmath130 by eq . ( [ eq : decaylifetime ] ) . for @xmath131 , three - body decay modes are kinematically suppressed and four - body decay through a virtual photon dominates . the branching fraction of four - body decay is roughly expected to be @xmath132 , which sets a lower limit for the hadronic branching fraction . adjusting this estimation by taking into account details of the kinematics has negligible impact to our results . for the sneutrino nlsp case , the constraints from bounds on em energy release are weak , especially at late times @xmath133 , as discussed in sec . [ sec : swimp ] . however , the hadronic energy release is of the same magnitude for sneutrino nlsps and for left - handed slepton nlsps . the major hadronic activity comes from three - body decays , @xmath134 . the branching fraction for the four - body decay @xmath135 is also similar to the analogous decay in the slepton case , except that diagrams ( a ) and ( b ) in fig . [ fig : feynmangraphs ] are absent since the neutrino is neutral . last , we note that in the special case where the bino and wino masses are nearly equal , diagram ( c ) of fig . [ fig : feynmangraphs ] is highly suppressed , since the gauge boson @xmath136 is almost purely @xmath6 . in fig . [ fig : branch ] we show the hadronic branching fractions for @xmath105 and @xmath137 nlsps for a particular set of values of @xmath138 , @xmath139 , @xmath140 and @xmath141 . varying these parameters typically does not change our results significantly . the reason is that these parameters only affect the neutralino / chargino appearing in diagram ( c ) of fig . [ fig : feynmangraphs ] . the contribution from diagram ( c ) is negligible if the neutralino / chargino is not degenerate to within about 10% with the sleptons . in addition , diagram ( c ) does not interfere with the other diagrams if all couplings are real , which eliminates the possibility of cancelation . hadronic branching fractions for left - handed sleptons are similar to those shown for sneutrinos . the branching fractions are smaller for the right - handed slepton case for the reasons mentioned above . in both cases , the hadronic branching fractions drop quickly for lifetime @xmath142 s , which corresponds to small @xmath99 . for the @xmath105 case , mesons produced in @xmath66 decay may , in principle , induce showers before they decay , as discussed above . this contribution to hadronic energy is also shown in fig . [ fig : branch ] . as can be seen , it is important only at early times @xmath143 and may be safely ignored for the decay times of most interest here . note that the hadronic branching fraction of eq . ( [ hadfraction ] ) is defined at the quark level . hadronization effects are important for example , as we have shown above , mesons produced in @xmath6 and @xmath7 decays do not contribute to hadronic cascades since they decay before they interact . such effects have been included in deriving the hadronic bbn constraints that we use @xcite . [ fig : grid ] shows the hadronic energy injection @xmath144 as a function of decay lifetime @xmath145 for @xmath105 and @xmath137 nlsps . along each curve , @xmath146 is held fixed and @xmath147 varies . the curves are truncated at @xmath148 because for smaller @xmath101 , three - body hadronic decays are highly suppressed , leading to negligible hadronic energy injection . hadronic bbn constraints from ref . @xcite are also shown , with and without @xmath28li ; the regions above the curves are disfavored . it is clear that for larger @xmath109 , larger values of @xmath101 becomes viable . constraints are stronger in the @xmath137 case than in the @xmath105 case because the hadronic branching fraction is larger for left - handed particles than for right - handed ones . notice that we have fixed @xmath149 in these plots . the lines could be simply rescaled for other values of @xmath11 since @xmath150 . we have also adopted the hadronic bbn analysis of ref . @xcite . if updated analyses of the bbn constraints become available , it is straightforward to impose it in these plots to find out the remaining viable parameter regions for @xmath109 and @xmath101 . [ fig : summary ] shows the allowed and disfavored regions of the @xmath151 plane for the two nlsp cases . the cmb @xmath140 distortion constraint , the em bbn constraint of ref . @xcite , and the hadronic bbn constraint of ref . @xcite are included . for the @xmath105 case , the cmb constraint excludes a region with @xmath152 . em bbn constraints from d / h@xmath153 disfavor regions of @xmath99 below 400 gev for @xmath154 and @xmath155 below about 200 gev for larger @xmath109 . hadronic bbn constraints from d / h@xmath153 disfavor the part of the parameter space with large @xmath99 where the slepton decays relatively early . note that the hadronic constraint is extremely important it disfavors a large and natural part of parameter space that would otherwise be allowed . even after all of these constraints , however , the unshaded area in the region @xmath156 and @xmath157 still remains viable . the @xmath105 mass must be above 500 gev , but this is within reach of the lhc . the `` best fit '' point where the @xmath4li discrepancy is fixed by em energy release from late slepton decays is indicated by a circle at @xmath158 . note that it also remains viable , even given the strong hadronic bbn constraints . we also indicate the region that is disfavored by the less solid , but stronger , constraints from @xmath159he / d and @xmath160li / h . the @xmath159he / d constraint excludes parameter space below the solid line . the allowed region would then be pushed up to larger @xmath161 and larger @xmath162 . such a heavy stau nlsp is beyond the reach of the next generation hadron colliders . on the other hand , taking the @xmath160li / h constraint literally would exclude parameter space above the dashed line . taking both constraints literally and combining them would exclude most of the region of parameter space shown . clearly , a more firm understanding of bbn will have important consequences for the stau nlsp superwimp scenario . we have assumed @xmath107 ; of course , the allowed parameter space would be enlarged if @xmath114 . we have also taken @xmath163 in our analysis ; the allowed region could shift if there are significant corrections to this relation . the sneutrino nlsp case is also shown in fig . [ fig : summary ] . hadronic bbn constraints from the d / h and @xmath75 abundances only disfavor @xmath164 100 gev . even including the stronger but more speculative @xmath160li / h constraint , there is still a large region of @xmath165 that is allowed . the origin of this result is that at @xmath166 , the sneutrino nlsp is completely free of hadronic bbn constraints . we have again assumed @xmath167 . to get this relic density from a sneutrino thermal relic density , one would typically require sneutrino masses above 500 gev . however , even assuming a thermal relic abundance , if the superwimp component is only part of the present dark matter , smaller sneutrino masses are allowed . for right - handed selectron and smuon nlsps , the results will be similar to the right - handed stau nlsp case . the em constraints are slightly stronger , since all the released energy induces em cascades . for left - handed slepton nlsps , the region of @xmath168 that is disfavored by hadronic injection is similar to the region disfavored for the sneutrino nlsp . in addition , however , there are em bbn constraints that are similar to those for right - handed slepton nlsps . in this paper , we explored the possibility of superwimps as candidates for dark matter . examples of superwimps are the gravitino in supersymmetric models and the lightest kaluza - klein graviton in universal extra dimension models . superwimps obtain their relic density through late decays : wimp @xmath169 superwimp @xmath170 sm particles . the decay usually occurs between @xmath171 s and @xmath172 s , which has important cosmological implications . such late decays release em and hadronic energy into the universe , which may affect bbn predictions for the primordial abundance of the light elements . the constraints on em injection have been studied in detail in ref . @xcite . in this paper , we have analyzed the hadronic bbn constraints . we have taken the lightest gravitino as a concrete example of a superwimp and have focused on slepton and sneutrino nlsps as the most promising wimp candidates . we have determined the hadronic energy release by calculating three - body decay widths in detail . for the cases of @xmath105 and @xmath137 nlsps , the hadronic decay branching fraction is below the level of @xmath173 when @xmath174 and @xmath175 , respectively , or when the decay lifetime is @xmath176 . we identified the allowed and disfavored regions of the @xmath151 plane , imposing cmb constraints , em bbn constraints , and hadronic bbn constraints . for the sneutrino nlsp case , @xmath177 is allowed for a large range of @xmath109 . for the stau nlsp , the hadronic constraints are weaker : @xmath178 . however , additional constraints from cmb and em energy injection apply . combining all the constraints for the right - handed stau nlsp , the allowed window is @xmath179 , @xmath180 , corresponding to @xmath181 . we have assumed @xmath107 in our analysis . the constraints would be relaxed if superwimps are only part of the present dark matter . in addition , there are still ambiguities in the bbn constraints if the decay time lies in the region where em and hadronic effects are comparable . in particular , as we discussed , their effects on d might cancel , and the allowed parameter space could be enlarged . progress in firming up bbn constraints in this region , and for other elements , such as @xmath27he and @xmath28li may have crucial implications for the superwimp dark matter scenario . although the superwimp itself would escape all direct and indirect dark matter searches , the slepton / sneutrino nlsp will have rich implications for collider phenomenology . such metastable sleptons will not decay inside the detector , resulting in signals of highly ionizing tracks for sleptons and missing energy signals for sneutrinos . discussion of the collider phenomenology , combined with thermal relic density calculations for the nlsp in supergravity models will be presented in ref . @xcite . although we have focused on the susy scenario in this paper , superwimps could in general be any gravitationally interacting particle that obtains its relic density through the late decay of a wimp . the discussion and results for late decay of a leptonic wimp in other models will be qualitatively similar to the discussion and results presented here . this scenario also suggests that a particle that appears to play the role of dark matter at late times , even after bbn , could very well be different from the particle that constitutes dark matter now . the very late decay of `` would be '' dark matter particles may have important cosmological implications , for example , affecting small scale structure . this feature may provide qualitatively new possibilities for explaining puzzling cosmological observations @xcite . we have assumed @xmath182-parity conservation in our discussion . in the case of @xmath182-parity violation ( rpv ) , the gravitino could still constitute dark matter as long as its decay lifetime is comparable to or longer than the age of universe @xcite . gravitino dark matter in an rpv scenario could be distinguished from the superwimp scenario by both cosmological observations and collider experiments . for example , the decay of even a tiny amount of gravitino dark matter into sm particles in an rpv scenario could be seen in the diffuse photon flux . in addition , if the rpv is not extremely small , the collider signatures could be different from those in the superwimp scenario . further work is needed to study how to distinguish these two scenarios . we are grateful to m. fujii , k. kohri , t. moroi , and a. rajaraman for helpful conversations . the work of jlf was supported in part by national science foundation career grant phy0239817 , and in part by the alfred p. sloan foundation . the gravitino - gaugino - gauge boson interaction is @xmath183 \gamma^{\mu}\lambda f_{\nu\rho } \ , \end{aligned}\ ] ] where @xmath184 is the reduced planck mass , and @xmath185 denotes the gravitino field at spacetime point @xmath186 . the gravitino - slepton - lepton interaction is @xmath187 where subscript @xmath188 denotes the lepton chirality . in addition , there are gravitino - slepton - lepton - gauge boson interactions : @xmath189 here @xmath190 is the gauge coupling coefficient , which is given explicitly below . here we present the matrix amplitudes for the three - body decays @xmath194 , where @xmath195 is the sm partner of @xmath196 or its weak isospin partner , and @xmath197 or @xmath7 . the feynman diagrams are shown in fig . [ fig : feynmangraphs ] . we define the following ten operators : @xmath198 where @xmath199 is the polarization of gauge boson @xmath136 . notice that for an on - shell gravitino in the final state , @xmath200 by using the gravitino equation of motion . in terms of these operators , the matrix elements for three - body slepton decay [ diagrams ( a)(d ) in fig . [ fig : feynmangraphs ] ] are @xmath201 \label{eq : melez3}\\ { \cal { m}}_v^d & = & i \frac{\sqrt{2}eg_v^{l_h}}{m_*}{\cal o}_8 \ , \label{eq : melez4}\end{aligned}\ ] ] where @xmath202 denotes a neutralino or chargino , the couplings @xmath190 , @xmath203 , and @xmath204 are given below for each decay process , and @xmath205 notice that @xmath206 . given this relation , the four - momentum scalar products that appear below in the expressions for squared matrix elements can be expressed in terms of @xmath207 , @xmath208 and @xmath209 . the sum of the matrix elements can be written as @xmath222 where the @xmath223 can be read off from eqs . ( [ eq : melez1])([eq : melez4 ] ) above . the squared matrix element is @xmath224 in our calculations , we chose the convention that the diagonalizing matrices @xmath136 and @xmath219 are real . all couplings appearing in eq . ( [ eq : melez1])@xmath225([eq : melez4 ] ) are therefore real . thus , all the @xmath223 are real except for @xmath226 , which has both real and imaginary components . the only non - zero contributions to the last term in eq . ( [ eq : square ] ) come from @xmath227 ( @xmath228 ) and @xmath229 ( @xmath230 ) . expressions for @xmath231 , @xmath232 , @xmath233 ( @xmath228 ) and @xmath234 ( @xmath230 ) are : @xmath235 \left[\mg^2 - \pgpsl + \pgpz\right ] \nonumber\\ & & \!\!\!\!\!\!\!\ ! \times \left[\mg^2m_v^2 + \msl^2m_v^2 - 2m_v^2\pgpsl - \left(\pgpz\right)^2 + 2\pgpz\pslpz - \left(\pslpz\right)^2\right ] \\ \calo_{2,2}&= & -\frac{4}{3\mg^2m_v^2 } \left[\mg^2 - \pgpsl + \pgpz\right ] \left[\msl^2m_v^2 - \left(\pslpz\right)^2\right ] \nonumber \\ & & \!\!\!\!\!\!\!\ ! \times \left[-\left(\pgpsl\right)^2 + 2\pgpsl\pgpz - \left(\pgpz\right)^2 + \mg^2\left(\msl^2 + m_v^2 - 2\pslpz \right ) \right ] \\ \calo_{3,3 } & = & { { 0 } } \\ \calo_{4,4 } & = & \frac{4}{3\mg^2m_v^2 } \left[\mg^2m_v^2 - \left(\pgpz\right)^2\right ] \nonumber \\ & & \times \left[\mg^2m_v^2 - m_v^2\pgpsl + 2\left(\pgpz\right)^2 + \pgpz \left(3m_v^2 - 2\pslpz\right)\right ] \\ \calo_{5,5 } & = & -\frac{4}{3\mg^2 } \left[\mg^2\msl^2 - \left(\pgpsl\right)^2\right ] \nonumber \\ & & \times \left[\mg^2m_v^2 - m_v^2\pgpsl - 4\left(\pgpz\right)^2 + \pgpz \left(-3m_z^2 + 4\pslpz\right)\right ] \\ \calo_{6,6 } & = & -\frac{4}{3\mg^2 } \left[\mg^2m_z^2 - \left(\pgpz\right)^2\right ] \nonumber \\ & & \times \left[\mg^2m_v^2 - m_v^2\pgpsl - 4\left(\pgpz\right)^2 + \pgpz \left(-3m_v^2 + 4\pslpz\right)\right ] \\ \calo_{7,7 } & = & \frac{4}{3\mg^2m_v^2 } \left [ \mg^2m_v^2 - \left(\pgpz\right)^2\right ] \nonumber \\ & & \left\ { \mg^4m_v^2 + 2\mg^2\left(\pgpz\right)^2 + 4\pgpz^3 - \pgpsl \left [ \mg^2m_v^2 + 4\left(\pgpz\right)^2\right ] \right . \nonumber\\ & & \left . - \mg^2\pgpz \left(m_v^2 - 2\pslpz\right)\right\ } \\ \calo_{8,8 } & = & -\frac{4}{3\mg^2m_v^2 } \left[\mg^2 - \pgpsl + \pgpz\right ] \left[2\mg^2m_v^2 + \left(\pgpz\right)^2\right ] \\ \calo_{9,9 } & = & \frac{4}{3\mg^2m_v^2 } \left [ 2\mg^2m_v^2 + \left(\pgpz\right)^2\right ] \nonumber \\ & & \times \left[\mg^2m_v^2 - m_v^2\pgpsl - 2\left(\pgpz\right)^2 - \pgpz \left(m_v^2 - 2\pslpz\right)\right ] \\ \calo_{10,10}&=&\frac{4}{3\mg^2m_v^2 } \left[2\mg^2m_v^2 + \left(\pgpz\right)^2\right ] \nonumber \\ & & \left\ { \mg^4m_v^2 - 2\mg^2\left(\pgpz\right)^2 - 4\left(\pgpz\right)^3 \right . \nonumber\\ & & \left . + \pgpsl\left[-\mg^2m_v^2 + 4\left(\pgpz\right)^2\right ] + \mg^2\pgpz\left(3m_v^2 - 2\pslpz\right)\right\ } \\ \calo_{1,2}^{\text re}&=&\frac{8}{3\mg^2m_v^2 } \left[\mg^2 - \pgpsl + \pgpz \right ] \nonumber \\ & & \times \left [ - \left(\pgpsl\right)^2 + \pgpsl \pgpz + \mg^2\left(\msl^2 - \pslpz\right)\right ] \nonumber \\ & & \times \left [ -\msl^2m_v^2 + m_z^2\pgpsl - \pgpz \pslpz + \left(\pslpz\right)^2\right ] \\ \calo_{1,3}^{\text re } & = & \calo_{1,4}^{\text re } = { { 0 } } \\ \calo_{1,5}^{\text re } & = & \frac{8}{3\mg^2 } \left[\mg^2\msl^2 - \left(\pgpsl\right)^2\right ] \left [ -\left(\pgpz\right)^2 + \mg^2\left(m_v^2 - \pslpz\right ) \right . \nonumber\\ & & \left . + \pgpz\left(-\msl^2 + \pslpz\right ) + \pgpsl \left(-m_v^2 + \pgpz + \pslpz\right)\right ] \\ { \cal{o}}_{1,6}^{\text re } & = & { \cal{o}}_{1,7}^{\text re}={\cal{o}}_{1,8}^{\text re}={\cal{o}}_{1,9}^{\text re}={\cal{o}}_{1,10}^{\text re}=0 \\ { \cal{o}}_{2,3}^{\text re}&=&{\cal{o}}_{2,4}^{\text re}={{0 } } \\ { \cal{o}}_{2,5}^{\text re}&=&\frac{4}{3\mg^2 } \left\ { \left(\pgpsl\right)^2\left[\mg^2m_v^2 + 2\left(\msl^2- m_v^2\right)\pgpz\right ] \right . \nonumber \\ & & + 2\left(\pgpsl\right)^3\left(m_v^2 - \pslpz\right ) - 2\pgpsl \left [ \left(\pgpz\right)^2 \left(\msl^2 - \pslpz\right ) \right . \nonumber \\ & & \left . + \mg^2\left(\msl^2 -\pslpz\right ) \left(m_v^2 - \pslpz\right ) + \mg^2\pgpz\pslpz\right ] \nonumber \\ & & + \mg^2\left[\msl^2\left(\pgpz\right)^2 - 2\pgpz \left(\msl^2 - \pslpz\right)^2 \right . \nonumber \\ & & \left . + \mg^2\left(-\msl^2m_v^2 + \left(\pslpz\right)^2\right)\right]\right\ } \\ { \cal{o}}_{2,6}^{\text re}&=&{\cal{o}}_{2,7}^{\text re}={\cal{o}}_{2,8}^{\text re}={\cal{o}}_{2,9}^{\text re}={\cal{o}}_{2,10}^{\text re}={{0 } } \\ { \cal{o}}_{3,4}^{\text re}&=&{\cal{o}}_{3,5}^{\text re}={\cal{o}}_{3,6}^{\text re}={\cal{o}}_{3,7}^{\text re}={\cal{o}}_{3,8}^{\text re}={\cal{o}}_{3,9}^{\text re}={\cal{o}}_{3,10}^{\text re}={{0 } } \\ { \cal{o}}_{4,5}^{\text re}&=&{{0 } } \\ { \cal{o}}_{4,6}^{\text re}&= & \frac{8}{\mg } \left[\mg^2m_v^2 - \left(\pgpz\right)^2\right ] \left[m_v^2 + \pgpz - \pslpz\right ] \\ \calo_{4,7}^{\text re}&= & \frac{8}{\mg } \left[\mg^2 - \pgpsl + \pgpz\right ] \left[\mg^2m_v^2 - \left(\pgpz\right)^2\right ] \\ \calo_{4,8}^{\text re}&= & \frac{8}{3\mg m_v^2 } \left[-m_v^2\left(\mg^2 - 2\pgpsl\right)\pgpz + \pgpz^ 3 \right . \nonumber\\ & & \left . + \mg^2m_v^2\left(m_v^2 - \pslpz\right ) - \left(\pgpz\right)^2\left(m_v^2 + \pslpz\right)\right ] \\ \calo_{4,9}^{\text re}&= & \frac{8}{3\mg^2 m_v^2 } \left[\mg^4m_v^4 + \mg^2m_v^4\pgpz + \mg^2m_v^2\left(\pgpz\right)^2 - 2\left(\pgpz\right)^4 \right . \nonumber \\ & & \left . - m_v^2\pgpsl\left(\mg^2m_v^2 + \left(\pgpz\right)^2\right ) - \left(\pgpz\right)^3 \left(m_v^2 - 2\pslpz\right)\right ] \\ \calo_{4,10}^{\text re}&= & \frac{8}{3\mg } \left\{\mg^4m_v^2 - \mg^2\left(\pgpz\right)^2 - 3\pgpz^3 + \pgpsl\left[-\mg^2m_v^2 + 3\left(\pgpz\right)^2\right ] \right . \nonumber\\ & & \left . + \mg^2\pgpz\left(3m_v^2 - 2\pslpz\right)\right\ } \\ { \cal{o}}_{5,6}^{\text re}&=&{\cal{o}}_{5,7}^{\text re}={\cal{o}}_{5,8}^{\text re}={\cal{o}}_{5,9}^{\text re}={\cal{o}}_{5,10}^{\text re}={{0 } } \\ \calo_{6,7}^{\text re}&= & \frac{8}{3\mg^2 } \left[\mg^2m_v^2 - \left(\pgpz\right)^2\right ] \nonumber \\ & & \times \left[\left(3\mg^2 - 2\pgpsl\right)\pgpz + 2\left(\pgpz\right)^2 + \mg^2\left(m_v^2 - \pslpz\right)\right ] \\ \calo_{6,8}^{\text re}&=&-\frac{8}{3\mg^2 } \left[\mg^4m_v^2 - \mg^2\left(\pgpz\right)^2 + \left(\pgpz\right)^3 \right . \nonumber\\ & & \left . - \pgpsl \left(\mg^2m_v^2 + \left(\pgpz\right)^2\right ) - \mg^2\pgpz \left(m_v^2 - 2\pslpz\right)\right ] \\ \calo_{6,9}^{\text re}&=&\frac{8}{3\mg } \left[m_v^2\left(3\mg^2 - 2\pgpsl\right ) \pgpz - 3\left(\pgpz\right)^3 \right . \nonumber\\ & & \left . - \left(\pgpz\right)^2 \left(m_v^2 - 3\pslpz\right ) + \mg^2m_v^2\left(m_v^2 - \pslpz\right)\right ] \\ \calo_{6,10}^{\text re}&=&\frac{8}{3\mg^2 } \left[\mg^4m_v^2\pgpz - \left(\mg^2 - 2\pgpsl\right ) \left(\pgpz\right)^3 - 2\left(\pgpz\right)^4 \right . \nonumber \\ & & \left . + \mg^4m_v^2\left(m_v^2 - \pslpz\right ) + \mg^2\left(\pgpz\right)^2 \left(m_v^2 - \pslpz\right)\right ] \\ \calo_{7,8}^{\text re}&=&\frac{8}{3\mg^2m_v^2 } \left[\mg^4m_v^2\pgpz - \left(\mg^2 - 2\pgpsl\right)\left(\pgpz\right)^3 - 2\left(\pgpz\right)^4 \right . \nonumber \\ & & \left . + \mg^4m_v^2\left(m_v^2 - \pslpz\right ) + \mg^2\left(\pgpz\right)^2 \left(m_v^2 - \pslpz\right)\right ] \\ \calo_{7,9}^{\text re}&=&\frac{8}{3\mg } \left[\mg^4m_v^2 - \mg^2\left(\pgpz\right)^2 + \left(\pgpz\right)^3 \right . \nonumber\\ & & \left . - \pgpsl \left(\mg^2m_v^2 + \left(\pgpz\right)^2\right ) - \mg^2\pgpz \left(m_v^2 - 2\pslpz\right)\right ] \\ \calo_{7,10}^{\text re}&= & \frac{8}{3\mg^2m_z^2 } \left\{\mg^6m_v^4 + \mg^4m_v^4\pgpz + \mg^4m_v^2\left(\pgpz\right)^2 \right . \nonumber \\ & & - 2\mg^2\left(\pgpz\right)^4 - 4\left(\pgpz\right)^5 \nonumber \\ & & \left . - \pgpsl\left[\mg^4m_v^4 + \mg^2m_v^2\left(\pgpz\right)^2 - 4\left(\pgpz\right)^4\right ] \right . \nonumber\\ & & \left . + \mg^2\left(\pgpz\right)^3 \left(3m_v^2 - 2\pslpz\right)\right\ } \\ \calo_{8,9}^{\text re}&=&\frac{8}{3\mg m_v^2 } \left [ 2\mg^2m_v^2 + \left(\pgpz\right)^2\right ] \left[m_v^2 + \pgpz -\pslpz\right ] \\ \calo_{8,10}^{\text re}&=&\frac{8}{3\mg^2 m_v^2 } \left[2\mg^2m_v^2 + \left(\pgpz\right)^2\right ] \nonumber \\ & & \times \left[-\left(\mg^2 - 2\pgpsl\right)\pgpz - 2\left(\pgpz\right)^2 + \mg^2\left(m_v^2 - \pslpz\right)\right ] \\ \calo_{9,10}^{\text re}&=&\frac{8}{3\mg } \left[\mg^2 - \pgpsl+ \pgpz\right ] \left[2\mg^2m_v^2 + \left(\pgpz\right)^2\right ] \\ \calo_{1,8}^{\text im}&= & -\frac{8}{3\mg^2m_v^2}(\mg^2-\pgpsl+\pgpz ) \left\{\left[\msl^2m_v^2-(\pslpz)^2+(\pgpz)(\pslpz)\right]\mg^2 \right . \nonumber \\ & & \left . -m_v^2(\pgpsl)^2 + ( \pgpsl)(\pgpz)(\pslpz-\pgpz)\right\ } \\ \calo_{2,8}^{\text im}&= & -\frac{8}{3\mg^2m_v^2}(\mg^2-\pgpsl+\pgpz ) \left[(\msl^2m_v^2-(\pslpz)^2)\mg^2-m_v^2(\pgpsl)^2\right . \nonumber \\ & & \left.-(\pgpz)^2\pslpz + ( \pgpsl)(\pgpz)(m_v^2+\pslpz)\right ] \\ \calo_{3,8}^{\text im}&=&\calo_{4,8}^{\text im}=0\\ \calo_{5,8}^{\text im}&= & \frac{8}{3\mg^2}\left\ { ( \pgpsl-\mg^2)\pslpz \mg^2 -\pgpsl(\pgpz)^2\right . \nonumber \\ & & \left . + \pgpz\left[\pgpsl \mg^2 + ( \pslpz-2 \msl^2)\mg^2+(\pgpsl)^2\right]\right\ } \\ \calo_{6,8}^{\text im}&=&\calo_{7,8}^{\text i m } = \calo_{8,9}^{\text im}=\calo_{8,10}^{\text im}=0 \ .\end{aligned}\ ] ] 99 j. l. feng , a. rajaraman and f. takayama , phys . lett . * 91 * , 011302 ( 2003 ) [ hep - ph/0302215 ] . j. l. feng , a. rajaraman and f. takayama , phys . d * 68 * , 063504 ( 2003 ) [ hep - ph/0306024 ] . j. r. ellis , d. v. nanopoulos and s. sarkar , nucl . b * 259 * , 175 ( 1985 ) . m. bolz , w. buchmuller and m. plumacher , phys . b * 443 * , 209 ( 1998 ) [ hep - ph/9809381 ] . m. bolz , a. brandenburg and w. buchmuller , nucl . b * 606 * , 518 ( 2001 ) [ hep - ph/0012052 ] . j. r. ellis , k. a. olive , y. santoso and v. spanos , hep - ph/0312262 . j. l. feng , a. rajaraman and f. takayama , phys . d * 68 * , 085018 ( 2003 ) [ hep - ph/0307375 ] . superwimp dark matter shares some characteristics with axino dark matter . see l. covi , j. e. kim and l. roszkowski , phys . * 82 * , 4180 ( 1999 ) [ hep - ph/9905212 ] ; l. covi , h. b. kim , j. e. kim and l. roszkowski , jhep * 0105 * , 033 ( 2001 ) [ hep - ph/0101009 ] . h. pagels and j. r. primack , phys . lett . * 48 * , 223 ( 1982 ) . t. moroi , h. murayama and m. yamaguchi , phys . b * 303 * , 289 ( 1993 ) . m. h. reno and d. seckel , phys . d * 37 * , 3441 ( 1988 ) . k. kohri , phys . d * 64 * , 043515 ( 2001 ) [ astro - ph/0103411 ] . s. dimopoulos , r. esmailzadeh , l. j. hall and g. d. starkman , astrophys . j. * 330 * , 545 ( 1988 ) . s. dimopoulos , r. esmailzadeh , l. j. hall and g. d. starkman , nucl . b * 311 * , 699 ( 1989 ) . m. kawasaki , k. kohri and t. moroi , astro - ph/0402490 . w. buchmuller , k. hamaguchi and m. ratz , phys . b * 574 * , 156 ( 2003 ) [ hep - ph/0307181 ] . e. w. kolb , a. notari and a. riotto , phys . d * 68 * , 123505 ( 2003 ) [ hep - ph/0307241 ] . j. l. feng , s. su and f. takayama , in preparation . d. n. spergel _ et al . _ , astrophys . j. suppl . * 148 * , 175 ( 2003 ) [ astro - ph/0302209 ] . s. burles , k. m. nollett and m. s. turner , astrophys . j. * 552 * , l1 ( 2001 ) [ astro - ph/0010171 ] . d. kirkman , d. tytler , n. suzuki , j. m. omeara and d. lubin , astrophys . j. suppl . * 149 * , 1 ( 2003 ) [ astro - ph/0302006 ] . k. a. olive , e. skillman and g. steigman , astrophys . j. * 483 * , 788 ( 1997 ) [ astro - ph/9611166 ] . y. i. izotov and t. x. thuan , astrophys . j. * 500 * , 188 ( 1998 ) . s. g. ryan , t. c. beers , k. a. olive , b. d. fields and j. e. norris , astrophys . j. lett . * 530 * , l57 ( 2000 ) [ astro - ph/9905211 ] . r. h. cyburt , j. r. ellis , b. d. fields and k. a. olive , phys . d * 67 * , 103521 ( 2003 ) [ astro - ph/0211258 ] . m. kawasaki and t. moroi , astrophys . j. * 452 * , 506 ( 1995 ) [ astro - ph/9412055 ] . k. kohri , private communication . e. vangioni - flam , k. a. olive , b. d. fields and m. casse , astrophys . j. * 585 * , 611 ( 2003 ) [ astro - ph/0207583 ] . l. brown and d. n. schramm , astrophys . j. * 329 * , l103 ( 1988 ) . k. jedamzik , phys . lett . * 84 * , 3248 ( 2000 ) [ astro - ph/9909445 ] . p. e. nissen , d. l. lambert , f. primas and v. v. smith , astro - ph/9906306 . k. jedamzik , astro - ph/0402344 . m. kawasaki and t. moroi , phys . b * 346 * , 27 ( 1995 ) [ hep - ph/9408321 ] . j. a. frieman and g. f. giudice , phys . b * 224 * , 125 ( 1989 ) . x. chen and m. kamionkowski , astro - ph/0310473 . k. sigurdson and m. kamionkowski , astro - ph/0311486 . f. takayama and m. yamaguchi , phys . b * 485 * , 388 ( 2000 ) [ hep - ph/0005214 ] . g. moreau and m. chemtob , phys . d * 65 * , 024033 ( 2002 ) [ hep - ph/0107286 ] . h. e. haber and g. l. kane , phys . rept . * 117 * , 75 ( 1985 ) .

the surface states of three - dimensional ( 3d ) topological insulators ( tis ) @xcite possess a helical spin texture in which the spin and momentum are perpendicularly locked to each other . due to this spin - momentum locking , a net spin polarization can be induced by a charge current and _ vice versa_. this peculiar property makes 3d tis a promising platform for spintronic applications @xcite , and as a result , there is growing interest in topological spintronics . already a number of experiments with various approaches have been carried out on 3d tis @xcite , including spin - transfer torque @xcite , spin pumping @xcite , and all - electrical measurement @xcite . among these experiments , the all - electrical measurement of the charge - current - induced spin polarization is of particular importance due to its direct applicability to spintronics . for spintronic applications , it is important to be able to perform spin - transistor " operation , that is , to switch the orientation of the charge - current - induced spin polarization via electrostatic gating . in this regard , the switching of the spin helicity across the dirac point in the topological surface state ( tss ) of a 3d ti may seem promising at first sight , but this expectation is in fact ungrounded , because the fermi velocity of electrons with a given spin orientation does not change across the dirac point ( detailed explanations of this point is given later ) . nevertheless , in the literature two different types of current - induced spin polarizations have been observed . so far both types of results have been claimed to be due to the tss , leaving the situation controversial . if this controversy is sorted out and the origin of the two types of current - induced spin polarizations is understood , such an understanding may make it possible to conceive a spin transistor based on 3d tis . here , we directly address this problem and offer a possible solution . the electrical detection of the current - induced spin polarization has been reported in various 3d ti materials , including bi@xmath0se@xmath1 @xcite , bi@xmath2sb@xmath3te@xmath4se@xmath5 @xcite , ( bi@xmath6sb@xmath7)@xmath0te@xmath1 @xcite , and bi@xmath0te@xmath0se @xcite . most experiments use a ferromagnetic contact as a spin detector and measure spin - dependent voltage between the ferromagnet ( fm ) and ti . however , conflicting interpretations of the spin voltage @xmath8 have been claimed in the literature . specifically , li _ et al . _ @xcite attributed positive @xmath8 to the situation when the magnetization of the fm detector @xmath9 is antiparallel to the induced spin polarization @xmath10 ( i.e. @xmath11 ) , whereas most other papers @xcite attribute positive @xmath8 to the situation @xmath12 . theoretically @xcite , @xmath8 should be positive for @xmath13 . furthermore , as long as the tss is responsible for the induced @xmath10 , the orientation of @xmath10 is expected to be the same for both @xmath14-type and @xmath15-type carriers @xcite . indeed , the expected polarity was observed in a recent experiment in which the carrier type of a ti device was continuously changed from @xmath14-type to @xmath15-type by backgating @xcite . hence , the contribution of the tss to the spin voltage has become largely understood , but the reason of the controversial observation by li _ et al . _ @xcite remains a puzzle . in this respect , it is useful to notice that the possible contribution of the topologically - trivial two - dimensional ( 2d ) states with a large rashba spin splitting , which is caused by surface band bending and often coexist with the tss @xcite , has not been well understood . such rashba states could lead to an opposite sign of the current - induced spin polarization @xcite . one should note that any difference in the work functions between fm and ti would lead to a band bending at the interface and could cause a pronounced effect in spin detection , but this possibility has been neglected in previous works @xcite . in this paper , we report the current - induced spin polarization in devices made from the topological insulator bisbtese@xmath0 . in the as - grown crystals of this material , the chemical potential lies slightly below the dirac point of the surface state , and its position changes easily when the sample is annealed ; this makes it possible for us to investigate spin - voltage devices with different carrier types and densities . in sec . ii , we will present a transparent review of the physics behind the spin - dependent voltage in the ti - based devices and discuss the correct sign of the current - induced spin polarization generated in the tss . we will then present our experiment , in which two different types of spin polarization were detected . the first type was observed in a strongly @xmath14-type doped bisbtese@xmath0 device and its polarity meets the expectation for the tss . the second type was observed in two low - carrier - density devices and their spin polarization can not be ascribed to the tss . we propose that the latter originates from the topologically - trivial 2d states with rashba spin splitting , and discuss why they are expected in low - carrier - density samples . the possibility to utilize the rashba - split 2d states alongside the tss opens new pathways in future topological spintronic devices , in particular the prospect for spin - transistor operations . it has been established that the non - equilibrium spin polarization can be detected by using a fm contact as a spin detector . early theories and experiments of spin detection started from the 1980s @xcite . at that time , ohmic contacts between fm and a normal metal were employed for spin detection . however , it eventually turned out that ohmic contacts are not efficient for detecting the spin polarization occurring in semiconductors due to the so - called conductance mismatch @xcite : when the impedances of the fm and the material beneath it are significantly different , the spin - dependent voltage becomes too small to be measured . in 2000 , it was theoretically proposed that the problem of conductance mismatch can be solved by inserting a tunnel barrier between the fm detector and the semiconductor @xcite . corresponding experimental techniques were developed several years later @xcite and they are recently applied to detect the current - induced spin polarization in tis @xcite . as already mentioned , there is a controversy over the sign of the spin - dependent voltage in the previous studies on tis @xcite . this controversy comes down to the question , what is the proper sign of @xmath8 for @xmath12 ? " here @xmath8 and @xmath9 denote the spin - dependent voltage and the magnetization in the fm spin detector , and @xmath10 is the non - equilibrium spin polarization to be detected . in the first paper by li _ @xcite , it was argued that _ negative _ @xmath16 corresponds to @xmath12 , whereas in the other papers @xcite , the opposite was claimed to be true , that is , _ positive _ @xmath16 corresponds to @xmath17 . in the following , we review the basic principles of spin detection in the case of tis and explain from the ground up why the latter is correct this clarification forms the basis of our interpretation of the two different types of @xmath8 signals observed in bisbtese@xmath0 the principle of spin detection is illustrated in fig . the key idea is that , due to the electron exchange between the fm and the semiconductor , the fm always equilibrate with the electrochemical potential of the subband which has the same magnetization direction . to be concrete , we consider the basic configuration shown in fig . a fm contact with a tunnel barrier , which works as a spin detector , is placed on top of a semiconductor ( sm ) . for simplicity , we assume that the fm is a half - metal , namely , only one subband is at the fermi level . the voltage @xmath8 is measured on the fm with respect to the sm . hereafter , we sometimes use @xmath18 and @xmath19 to denote the @xmath20 and @xmath21 directions , respectively . it is important to notice that , due to the negative charge , the electron magnetic moment @xmath22 is _ antiparallel _ to the spin vector @xmath23 , i.e. , @xmath24 , where @xmath25 is the @xmath25-factor and @xmath26 is the bohr magneton . this means that the @xmath27 subband corresponds to the @xmath28 subband , and the @xmath29 subband corresponds to the @xmath30 subband . note that all the arrows in the band diagrams in fig . 1 denote the direction of the _ magnetization _ , which is opposite to the direction of spin polarization . in the absence of spin polarization in the sm , the electrochemical potentials of @xmath27 and @xmath29 subbands are equal . the chemical potential of the ferromagnet is also at the same level , irrespective of its magnetization direction . in this case , the fm is electrically neutral and hence @xmath31 , as shown in fig . 2(b ) . when a non - equilibrium spin polarization @xmath10 is induced in the sm via methods like spin injection @xcite , a corresponding magnetization @xmath32 is generated in the @xmath33 direction . the non - zero @xmath32 leads to an increase ( decrease ) in the electrochemical potential of the majority ( minority ) subband , with the potential difference @xmath34 ( `` maj '' and `` min '' stand for majority and minority , respectively ) . when @xmath35 , electrons in the majority subband of the sm will move into the fm and raise its potential @xmath36 , until @xmath37 is reached . in this process , the fm become negatively charged , giving rise to a negative voltage @xmath38 . on the other hand , when @xmath39 , electrons in the fm will move into the minority band of the sm and lower @xmath36 , until @xmath40 is reached . in this case , the ferromagnet is positively charged , leading to a positive voltage @xmath41 . all possible combinations of @xmath10 and @xmath9 , as well as the resulting @xmath8 , are illustrated in figs . 1(c ) and 1(d ) . these considerations lead to the following conclusion : @xmath42 this conclusion supports the arguments in refs . it also means that the spin voltage observed in ref . @xcite corresponds to the spin polarization that is opposite to what the authors of ref . @xcite thought to be there . the tss of 3d tis possesses a helical spin texture illustrated in fig . 2(a ) ; here , the spin vector @xmath23 points to @xmath43 with @xmath44 the normal vector of the surface and @xmath45 the group velocity of the bloch electron with momentum @xmath46 . the net spin polarization per unit area , @xmath10 , can be written as @xmath47 where all the occupied states in the surface brillouin zone ( bz ) contribute to the integral . in the absence of charge current , the center of the fermi circle is located at @xmath48 [ solid circle in fig . 2(b ) ] and the integral in eq . ( 2 ) becomes zero due to symmetry . when the tss carries a charge current @xmath49 along the @xmath50 direction , the fermi circle shifts in the @xmath51 direction by @xmath52 [ dashed circle in fig . 2(b ) ] . in this case , the fermi circle is no longer symmetric with respect to @xmath53 and eq . ( 2 ) gives a finite @xmath10 in the @xmath54 direction . based on this picture , we now discuss the current - induced spin polarization in @xmath14-type and @xmath15-type regions of the dirac cone . the longitudinal section cut of the dirac cone along @xmath55 , shown in figs . 2(c ) and 2(d ) , makes it easy to understand the situation . the spin direction of the green branch is @xmath56 and that of the blue branch is @xmath57 . an electric field @xmath58 in the @xmath59 direction drives a charge current @xmath49 along the @xmath59 direction and shifts the fermi surface in the @xmath60 direction , as illustrated in figs . 2(c ) and 2(d ) for the @xmath14- and @xmath15-type regions . notice that the shift of the fermi surface always makes the green branch ( @xmath61 ) more populated , no matter if the fermi level is above or below the dirac point ; this is because the electrons on the green branch are always accelerated by @xmath62 , since their fermi velocity is along @xmath60 . on the contrary , electrons on the blue branch are always decelerated by @xmath62 . the larger population in the @xmath56 branch gives an induced @xmath10 in the @xmath54 direction . therefore , for the @xmath10 originating from the tss , the polarity is always along @xmath63 irrespective of the carrier type . at first glance , this conclusion may look counter - intuitive , because the spin helicity reverses when the fermi level crosses the dirac point . however , for a fixed @xmath46 , the group velocity @xmath45 also changes sign across the dirac point . this means that the sign change in @xmath45 counteracts the spin - helicity reversal and leave the orientation of @xmath10 unchanged . the above discussion implies that the tss of 3d tis can only account for the current - induced @xmath10 directed to @xmath64 . if the detected @xmath10 is along the @xmath65 direction , one must consider a different origin . bisbtese@xmath0 is a bulk - insulating ti with the fermi level located slightly below the dirac point @xcite . we have investigated the current - induced spin polarization in various devices based on exfoliated flakes of bisbtese@xmath0 . the results presented in this paper were obtained in three typical devices , labeled # 1 , # 2 and # 3 . high quality bisbtese@xmath0 single crystals were grown by a modified bridgman method . thin bisbtese@xmath0 flakes were exfoliated from single crystals and transferred onto si / sio@xmath66 substrates . the flakes were examined under a laser confocal microscope . those with uniform thickness and flat surfaces were selected for device fabrications . the thicknesses of the flakes used in devices # 1 , # 2 , and # 3 were 172 , 82 , and 54 nm , respectively . after the exfoliation , an al@xmath66o@xmath67 or mgo tunnel barrier was deposited on top of the flakes . devices # 1 and # 2 had the al@xmath66o@xmath67 barrier , which was fabricated by first evaporating 0.7-nm - thick al in a thermal evaporator and then _ in - situ _ oxidizing the al layer with pure oxygen . the mgo barrier was used in device # 3 , for which about 2-nm - thick mgo was directly deposited from a mgo source by electron - beam ( eb ) evaporation . the tunnel barrier proved to be very important for the spin detection . no spin signal was observed in devices with too thin or too thick a tunnel barrier . the fm and normal - metal contacts were fabricated in two subsequent eb lithography steps . the fm electrodes are made of 30-nm - thick ni@xmath68fe@xmath69 ( py ) capped with about 160-nm - thick au . the au layer prevents py from oxidation and ensures a reliable electrical connection over the edges of the flakes . the normal - metal electrodes are made of ti / au , with the thicknesses of ti and au layers being 3 and 190 nm , respectively . before the deposition of the ti / au electrodes , the tunnel barrier in the contact area defined by eb lithography had been removed by shortly dipping the substrate into a diluted tetramethyl - ammonium hydroxide solution . the device concept is schematically illustrated in fig . 3 , and the optical photographs of the devices are shown in figs . 4(a)-4(c ) . as - grown bisbtese@xmath0 crystals are @xmath15-type , with the bulk carrier density down to @xmath70 @xmath71 @xcite . according to our experience , if high - temperature baking is avoided during the lithography process , the carrier density in bisbtese@xmath0 flakes can be kept low even after they are fabricated into nano - devices ; on the other hand , once the flakes are baked at @xmath72 150@xmath73c for several minutes , they will be strongly @xmath14-type doped and the fermi level will move into the bulk conduction band . based on this knowledge , we have managed to obtain bisbtese@xmath0 devices with various carrier types and concentrations . for example , during the device fabrication process , device # 1 was baked at @xmath74 = 170@xmath73c twice to cure the zep520a resist , each time for 3 minutes ; in contrast , for devices # 2 and # 3 , the curing of resist was done at @xmath74 = 110@xmath73c for 20 minutes , so that the flakes did not experience any high - temperature baking . the difference in baking temperature resulted in different carrier concentrations . to infer the carrier type and density , we measured the hall resistivity @xmath75 of all the devices at @xmath76 k , as plotted in figs . 4(j)-4(l ) . device # 1 showed a linear @xmath77 behavior with a negative slope [ fig . 4(j ) ] ; using the formula @xmath78 , we obtain the two - dimensional ( 2d ) carrier density @xmath79 @xmath80 . here , the negative sign means @xmath14-type carriers . the @xmath77 behavior of device # 2 was also linear , but it has a positive slope much larger in absolute value [ fig . 4(k ) ] , indicating @xmath15-type carriers with much lower carrier density ; the linear fit gives @xmath81 @xmath80 . a highly nonlinear @xmath77 curve was obtained in device # 3 [ fig . 4(l ) ] , indicating the coexistence of @xmath14-type and @xmath15-type channels . to obtain the carrier densities of both channels , we fit the @xmath77 curve to the expression given by the two - band model @xcite , @xmath82 where @xmath83 , @xmath84 , @xmath85 , and @xmath86 are the 2d carrier density and the mobility of the 1st and 2nd channels , respectively . those parameters are constrained by the sheet resistance in zero field which is expressed as @xmath87 the fitting gives @xmath88 @xmath80 , @xmath89 @xmath80 , @xmath84 = 275 @xmath90/vs and @xmath86 = 4068 @xmath90/vs . this result suggests that the top and bottom surfaces have different types of carriers , but they both have chemical potential located close to the dirac point . the detection of current - induced spin polarization was performed in a four - terminal configuration , as schematically depicted in fig . 3 . a @xmath91 current @xmath49 flows along the @xmath92 direction between the outer ti / au contacts . in - plane magnetic field @xmath93 is applied in the @xmath20 direction , which is perpendicular to @xmath49 and parallel to the easy axis of the py spin detector , to control the magnetization @xmath94 in py . upon scanning @xmath95 from negative to positive ( and _ vice versa _ ) , the voltage @xmath96 between a py tunneling contact and a ti / au contact is measured as a function of @xmath95 . to reduce noise , the data for each curve were averaged over tens of independent scans . the measured voltage can be written as @xmath97 , where @xmath8 is the spin - dependent voltage we are interested in , and @xmath98 is a trivial parabolic background mainly contributed by the magnetoresistance of bisbtese@xmath0 . when the py magnetization @xmath94 switches , @xmath8 suddenly changes sign , whereas @xmath98 does not . therefore , it is easy to separate @xmath8 from the measured @xmath96 . the @xmath99 data of all the devices are plotted in figs . 4(d)-4(i ) . curves measured in forward and backward @xmath95-scans form a hysteresis loop . the half width of the loop corresponds to the coercive field of the py spin detector , which is 10 , 20 , and 7 mt for devices # 1 , # 2 , and # 3 , respectively . the coercive field of a nanoscale ferromagnet depends on its shape , and narrower width gives higher coercive field . the width of the py electrodes in devices # 1 , # 2 and # 3 was 1 , 0.5 , and 1.5 @xmath100 , respectively , which is consistent with the observed difference in the coercive field . the amplitude of the spin - dependent voltage can be defined as @xmath101 . with this definition , the sign of @xmath102 is determined by the direction of the jump in the @xmath99 loop , and hence implies the orientation of the detected @xmath10 . for instance , @xmath103 is obtained when @xmath104 shows up for @xmath105 ( i.e. @xmath106 ) , and this is the expected sign of @xmath102 for @xmath107 [ see eq . ( 1 ) ] and thus we can conclude @xmath108 . similarly , @xmath109 implies @xmath110 . hence , the following relation exists between @xmath102 and @xmath10 : @xmath111 in our experiment , @xmath99 loops were measured at both positive and negative currents for each device . the reversal in @xmath49 resulted in the sign change of @xmath102 , indicating that the orientation of @xmath10 is controlled by the direction of @xmath49 . this is consistent with the existence of a helical spin texture in the current - carrying states . however , as one can see in figs . 4(d ) , 4(f ) , and 4(h ) , the orientation of the current - induced @xmath10 is apparently different between device # 1 and the others . to be specific , our data imply @xmath108 for a positive @xmath112 in device # 1 , while it is the opposite in devices # 2 and # 3 . this means that the underlying spin polarization is @xmath113 in the former and @xmath114 in the latter . according to the discussion in sec . ii , only the former can be attributed to the tss of bisbtese@xmath0 . the main result of the present experiment is the reproducible observation of the opposite polarity , which points to the possibility of utilizing the reversed spin helicity in ti - based devices . in passing , a linear dependence between @xmath102 and @xmath112 was observed in the @xmath99 loops measured at various bias currents , as shown in fig . 5 for devices # 1 and # 3 . this linear dependence allows us to define a current - independent spin resistance @xmath115 . as an example , @xmath116 data for various @xmath112 values are shown in fig . 6 for device # 1 . as discussed in sec . iii , the @xmath99 signal in devices # 2 and # 3 can not be interpreted to be due to the tss of a 3d ti . here , we discuss its possible origins . one possibility is that the observed @xmath99 is not stemming from a spin polarization but due to some artifacts like local hall voltages . it was proposed that hall voltages can be locally induced by fringe fields of the fm electrodes and could give rise to a signal similar to the @xmath99 loop @xcite . however , one can rule out this possibility in our experiments for the following reasons . first , if the observed @xmath99 loops are caused by some trivial reasons like the local hall effect , the signal should be insensitive to the details of the tunnel barrier , and thus it should be observed in all devices with similar fm contacts . in this regard , the hysteretic signals in our experiment are very sensitive to the quality and thickness of the tunnel barrier ; we have investigated more than a dozen devices , and the hysteretic voltages were only observed in those devices where the thickness of the tunnel barrier falls into a narrow window and the resulting contact resistance takes a suitable value ( 550 k@xmath117 ) . the voltage loops were never observed in devices with too thin or too thick a tunnel barriers . second , specifically for the local hall effect , since the direction of the unbalanced fringe field depends on both the local geometry of the fm electrode and the morphology of the ti sample @xcite , the polarities of the observed voltage loops should be _ random_. therefore , the voltage loops observed at different fm contacts made on the same device could present opposite polarities . however , in our experiment the voltage loops measured in the same device consistently showed the same polarity at different fm contacts . as an example , the data measured on two different fm contacts in device # 3 are plotted in fig . 5(b ) . to conclusively rule out the possibility that the observed voltage loops were caused by some artifacts ( local hall effect in the ti ake , anomalous nernsteffect in the fm lead , adverse effects of side contacts , etc . ) , we have performed a control experiment on a device which was fabricated in the same way as the other devices except that it does not have a tunnel barrier beneath the fm electrodes . as shown in the supplemental material ( fig . s2 ) @xcite , this control device yielded null result even with a high current of 200 @xmath118a . since all the spurious origins of the voltage loops so far discussed do not require a tunnel contact , the complete absence of a voltage hysteresis in our control device safely rules out the artifacts as the origin of the observed signal . given that the voltage signal is genuinely of spin polarization origin , the most likely origin of the opposite spin voltages observed in devices # 2 and # 3 is the electronic states which have an opposite spin helicity to that of the tss . to the best of our knowledge , the only known states with such a feature are the rashba - split 2d states caused by a surface band bending @xcite . when a downward band - bending occurs at the surface of a 3d ti and confines the bulk states in the potential well , the resulting 2d sub - bands present a rashba spin splitting and form two concentric fermi circles with opposite spin helicities @xcite , as illustrated in fig . it has been elucidated that the helicity of the outer fermi circle is opposite to that of the tss @xcite , so the contribution of the rashba - spit states to the spin signal is also opposite . it is prudent to mention that the opposite spin voltages observed in devices # 2 and # 3 could be due to a _ negative _ spin detection efficiency of the fm contacts ; namely , it is possible that the minority spin has a larger density of states at @xmath119 compared to the majority spin in a ferromagnet , and in such a case the spin voltage is reversed from the case depicted in fig . 1 . also , a reversal in the spin voltage may happen if the tunnel barrier has a spin - selective nature . however , this possibility can be largely ruled out for following reasons : ( i ) for py / al@xmath0o@xmath1 contacts , it is known that the spin polarization of tunneling conductance ( sptc ) is positive ( i.e. the spin polarization of tunneling current is parallel to the majority spin in py ) @xcite . ( ii ) the py / al@xmath0o@xmath1 contacts in devices # 1 and # 2 were made with essentially the same conditions , so it is unlikely that the sptc is reversed between the two . ( iii ) by assuming a positive sptc , our results are consistent with the majority of existing experiments ( ref . 14 - 19 ) , and can be naturally explained by the band - bending theory we discuss in detail later ; if we assume a negative sptc , the experimental results are not understandable . nevertheless , we can not completely exclude the remote possibility that the py / al@xmath0o@xmath1 contact in device # 2 and the py / mgo contact in device # 3 somehow had an unusual negative sptc due to unknown reasons ; if it were the case , the present conclusion would need reconsideration . in passing , very recently li and appelbaum @xcite reported an experiment in which they observed a similar voltage loop at a fm tunnel contact fabricated on a au film , and they claimed that such a result discredits the experiments on ti spin devices . however , since au films are known to exhibit a giant spin hall effect @xcite , when a charge current flows through a au film , an in - plane spin polarization is naturally generated on both top and bottom surfaces and detected by the fm tunnel contact . hence , their experiment seems to be a confirmation that au is not a simple metal but a spin - orbit active material . to judge whether the rashba - split 2d states are indeed relevant to our devices # 2 and # 3 , it is important to understand the band alignment at the py - ti interface and the work functions of py and bisbtese@xmath0 . the work function of ni@xmath68fe@xmath69 permalloy is known to be @xmath120 = 4.83 ev @xcite , which means that the chemical potential of py lies 4.83 ev below the vacuum level . the work function of our as - grown bisbtese@xmath0 crystals was measured by photoemission to be @xmath121 = 5.20 ev @xcite ; since the dirac point of the tss in as - grown bisbtese@xmath0 is located very close to the chemical potential @xcite , one can see that the dirac point lies @xmath1225.20 ev below the vacuum level . therefore , the band alignment of the py - bisbtese@xmath0 interface is such that the dirac point of the ti side always comes @xmath1220.37 ev below the chemical potential of the metallic py side . if the chemical potential of the bisbtese@xmath0 flake used in the device is unchanged from the as - grown state , there occurs a charge accumulation in bisbtese@xmath0 near the interface to compensate for the @xmath1220.37 ev difference in the chemical potential when the two materials are joined , and this is the basic mechanism of the band bending [ compare figs . 8(c ) and 8(d ) ] . however , when the chemical potential in bisbtese@xmath0 is shifted during the device fabrication process , @xmath121 will change and the strength of the band bending will be different , although the band alignment at the interface is always fixed [ compare figs . 8(b ) and 8(d ) ] . note that the existence of a tunnel barrier does _ not _ affect the band bending , as long as the barrier is thin enough to allow electrons to flow until the electrochemical potentials of the two sides ( ti and py ) equilibrate . the above argument for the band bending may not hold when there are a high density of defect states at the interface , because they would pin the chemical potential ; however , our previous top - gating experiments have shown that a thin layer of al@xmath0o@xmath1 deposited on a ti surface does not introduce any noticeable defect states at the interface if the deposition is done at temperatures less than 100@xmath73c @xcite . in fact , since the ( 111 ) surface of tetradymite ti materials does not have dangling bonds , a high density of defect states are not expected at the interface . we emphasize that the fermi level of the ti surface beneath the metallic fm electrode is always fixed and will not change with the chemical potential in the ti sample , as long as the electron exchange is allowed through the tunnel barrier . this means that one can not control the fermi level of the portions of the tss at which the current - induced spin polarization is detected by the fm detector , even if the fermi level of the rest of the tss can be controlled by electrostatic gating . in other words , even when a sign change in the carrier type is observed upon electrostatic gating in a spin - voltage device ( as was the case in ref . @xcite ) , such a change is occurring only in the portion of the sample that is _ not _ in contact with metallic electrodes . therefore , the gate - voltage - dependences of the spin voltages ( such as those reported ref . @xcite ) are not a result of the gating of the fermi level of the tss being probed , but are most likely a reflection of the current redistribution in the ti sample . we have estimated the fermi level @xmath123 in all three devices according to the carrier density obtained from @xmath75 @xcite . taking the dirac point as @xmath124 , the @xmath123 in device # 1 is estimated to be 0.24 ev , corresponding to a work function @xmath125 = 4.96 ev . since the bottom of the conduction band is at @xmath126 0.21 ev @xcite , the fermi level of device # 1 is located inside the conduction band and @xmath127 0.03 ev . in device # 2 , @xmath128 is estimated to be @xmath129 ev , corresponding to the work function @xmath130 = 5.27 ev . the top of the valence band is at @xmath131 ev @xcite , so the fermi level of device # 2 is located inside the bulk band gap and @xmath132 ev . the two surfaces of device # 3 have different types of carriers . the @xmath14-type surface has @xmath128 = 0.16 ev and @xmath133 = 5.04 ev , while the @xmath15-type surface has @xmath134 ev and @xmath135 = 5.21 ev . without a gating experiment , it is difficult to tell which surface is the top surface ; nevertheless , the @xmath128 of both surfaces are located inside the bulk band gap . the band diagrams of bisbtese@xmath0/py contacts are illustrated semi - quantitatively in fig . due to the difference in the work function , when py is in contact with bisbtese@xmath0 through a tunnel barrier , electrons in py will move into bisbtese@xmath0 to bias the electrostatic potential of the latter , until the electrochemical potential matches at the interface . in this process , a downward band bending occurs in bisbtese@xmath0 and moves the surface fermi level to @xmath136 ev , which is @xmath1220.16 ev above the bottom of the conduction band . as already mentioned , @xmath137 in the contact area is determined by the band structures of bisbtese@xmath0 and py , and the carrier density in the bisbtese@xmath0 flakes has no influence on it . the strength of the band bending is expressed by @xmath138 ( # @xmath139 is the device number ) . based on the estimates of @xmath140 presented above , we obtain @xmath141 to be 0.13 ev in device # 1 , 0.44 ev in device # 2 , and either 0.21 or 0.38 ev in device # 3 . the strength of the rashba spin splitting in the confined 2d states is proportional to the electric field in the @xmath142 direction , which depends linearly on @xmath141 . the experiment and model calculation in ref . @xcite showed that a small @xmath141 of 0.13 ev would not cause a measurable rashba spin splitting in bi@xmath0se@xmath1 , and hence one would not expect its contribution to the @xmath8 measurements in device # 1 . according to a photoemission study @xcite , the rashba splitting in bi@xmath0se@xmath1 is barely visible for @xmath143 0.15 ev and becomes as large as @xmath144 for @xmath141 = 0.35 ev , giving the rashba parameter of @xmath145 ev@xmath146 . this suggests that in our devices # 2 and # 3 , the band bending is strong enough to cause significant contribution of the rashba - split 2d states in the @xmath8 measurements . hence , the important question is how the contribution from the rashba - split states can dominate the observed spin signal . when one looks at the situation , the fermi circle of the tss is always larger than that of the rashba - split states , so the carrier density of the tss is also larger ; if the mobility is the same , the tss would carry more current . also , the contribution from the outer fermi circle of the rashba - split states is partially canceled by the inner circle ; therefore , the spin signal from the tss , which experiences no such cancellation , would be stronger . nevertheless , there are some factors which speak for the rashba - split states . first , when @xmath141 is large , there appear more than one pairs of rashba - split states at the fermi level . for instance , a second pair of rashba - split states has been observed in photoemission experiments @xcite . in such a situation , the total spin signal contributed by all pairs could exceed that from the tss . second , as @xmath147 increases , the fermi circles of the tss start to be deformed and show hexagonal warping @xcite . in a deformed fermi circle , spin is no long strictly perpendicular to the wavevector . since the fermi circle of the tss is larger than the others and the hexagonal warping grows as @xmath148 , it is more prone to the deformation and the in - plane spin component gets weaker . therefore , the hexagonal warping helps the rashba - split states to win over the tss in the contribution to the spin signal . third , as shown in fig . 7 , the inner circle of the rashba - split states can shrink to a point when @xmath149 . at this point , theoretical calculation @xcite shows that the spin signal coming from the rashba - split states can be stronger than that from the tss . according to the theory @xcite , for the current - induced @xmath10 generated by a single helical channel , we have @xmath150 where @xmath151 is the fermi wave vector of the helical channel , @xmath152 is the width of the device , @xmath153 is the effective spin polarization of the fm detector , and @xmath154 is the induced spin polarization per unit current . for a py spin detector , we can take @xmath155 @xcite . the value of @xmath154 is determined by the spin texture of the channel and @xmath156 is expected for an ideal tss @xcite . in reality , the spin - detection efficiency @xmath157 is not 100% and eq . ( 3 ) is modified to @xmath158 as discussed in sec . iii - c , for device # 1 , the measured spin signal is mostly due to the tss even though a major fraction of the current is carried by bulk carriers . in this case , @xmath157 can be estimated by using eq . ( 4 ) . specifically , we have @xmath159 m@xmath117 and @xmath160 @xmath118 m for device # 1 , and the band alignment at the interface fixes the fermi level of the tss beneath the fm electrode at @xmath161 ev , which corresponds to @xmath162 @xmath163 @xcite . these values lead to @xmath164 0.54% . such a low detection efficiency is probably due to the coexistence of the bulk states . the spin signals in devices # 2 and # 3 are supposedly dominated by the rashba - split states , but the reason for their dominance is not quantitatively understood . hence , the spin - detection efficiency in devices # 2 and # 3 can not be estimated at this stage . while the exact mechanism of the sign reversal in @xmath8 is to be elucidated in future , the discovery that the current - induced spin polarization can be switched by changing the chemical potential of a ti is of significant practical importance , because it allows us to conceive a spin - transistor device , in which the output spin polarization is switched by electrostatic gating . in this regard , it is conceivable that when the ti layer is thin enough , back gating can change the chemical potential of the ti throughout thickness , leading to a change in the band bending at the top surface beneath the fm electrode . such a tuning of the band bending would allow us to control the dominant spin helicity and the spin - transistor operation can be realized . for the exploitation of this intriguing and useful effect , further studies of its mechanism is strongly called for . we discovered that the current - induced spin polarization in bisbtese@xmath0 flakes can be reversed depending on the fermi level . in particular , samples with a small fermi energy present the spin polarization that is opposite to what is expected for the topological surface states . this is most likely due to the contribution from rashba - split 2d states created by a strong band bending occurring at the interface of the ti and the ferromagnetic spin detector . while its exact mechanism is to be elucidated in future studies , this effect provides an operation principle for a spin - transistor device in which the output spin polarization is controlled by electrostatic gating . we thank m. novak and z. wang for their help in crystal growth , y. maekawa for her help in device fabrications , and s. sasaki for his help in various aspects of the experiments . we also thank t. sato and his group for measuring the work function of bisbtese@xmath0 . this work was supported by jsps ( kakenhi 25220708 ) and dfg ( crc1238 control and dynamics of quantum materials " , projects a04 and c02 ) . 10 m. z. hasan and c. l. kane , _ colloquium : topological insulators _ , rev . phys . * 82 * , 3045 ( 2010 ) . qi and s .- c . zhang , _ topological insulators and superconductors _ , rev . phys . * 83 * , 1057 ( 2011 ) . y. ando , _ topological insulator materials _ , j. phys . . jpn . * 82 * , 102001 ( 2013 ) . a. a. burkov and d. g. hawthorn , _ spin and charge transport on the surface of a topological insulator _ , phys . 105 * , 066802 ( 2010 ) . d. pesin and a. h. macdonald , _ spintronics and pseudospintronics in graphene and topological insulators _ , nat . mater . * 11 * , 409 ( 2012 ) . f. mahfouzi , n. nagaosa , and b. k. nikoli , _ spin - orbit coupling induced spin - transfer torque and current polarization in topological - insulator / ferromagnet vertical heterostructures _ , phys . lett . * 109 * , 166602 ( 2012 ) . a. r. mellnik , j. s. lee , a. richardella , j. l. grab , p. j. mintun , m. h. fischer , a. vaezi , a. manchon , e .- a . kim , n. samarth , and d. c. ralph , _ spin - transfer torque generated by a topological insulator _ , nature ( london ) * 511 * , 449 ( 2014 ) . y. fan , p. upadhyaya , x. kou , m. lang , s. takei , z. wang , j. tang , l. he , l .- t . chang , m. montazeri , g. yu , w. jiang , t. nie , r. n. schwartz , y. tserkovnyak , and k. l. wang , _ magnetization switching through giant spin - orbit torque in a magnetically doped topological insulator heterostructure _ , nat . * 13 * , 699 ( 2014 ) . y. wang , p. deorani , k. banerjee , n. koirala , m. brahlek , s. oh , and h. yang , _ topological surface states originated spin - orbit torques in @xmath165 _ , phys lett . * 114 * , 257202 ( 2015 ) . y. shiomi , k. nomura , y. kajiwara , k. eto , m. novak , k. segawa , y. ando , and e. saitoh , _ spin - electricity conversion induced by spin injection into topological insulators _ , phys . * 113 * , 196601 ( 2014 ) . m. jamali , j. s. lee , j. s. jeong , f. mahfouzi , y. lv , z. zhao , b. k. nikoli , k. a. mkhoyan , n. samarth , and j .- wang , _ giant spin pumping and inverse spin hall effect in the presence of surface and bulk spin - orbit coupling of topological insulator @xmath165 _ , nano lett . * 15 * , 7126 ( 2015 ) . p. deorani , j. son , k. banerjee , n. koirala , m. brahlek , s. oh , and h. yang , _ observation of inverse spin hall effect in bismuth selenide _ , phys . rev . b * 90 * , 094403 ( 2014 ) . c. h. li , o. m. j. vant erve , j. t. robinson , y. liu , l. li , and b. t. jonker , _ electrical detection of charge - current - induced spin polarization due to spin - momentum locking in @xmath165 _ , nat . * 9 * , 218 ( 2014 ) . y. ando , t. hamasaki , t. kurokawa , k. ichiba , f. yang , m. novak , s. sasaki , k. segawa , y. ando , and m. shiraishi , _ electrical detection of the spin polarization due to charge flow in the surface state of the topological insulator @xmath166 _ , nano lett . * 14 * , 6226 ( 2014 ) . a. dankert , j. geurs , m. v. kamalakar , s. charpentier , and s. p. dash , _ room temperature electrical detection of spin polarized currents in topological insulators _ , nano lett . * 15 * , 7976 ( 2015 ) . j. tang , l .- t . chang , x. kou , k. murata , e. s. choi , m. lang , y. fan , y. jiang , m. montazeri , w. jiang , y. wang , l. he , and k. l. wang , _ electrical detection of spin - polarized surface states conduction in @xmath167 topological insulator _ , nano lett . * 14 * , 5423 ( 2014 ) . l. liu , a. richardella , i. garate , y. zhu , n. samarth , and c .- t . chen , _ spin - polarized tunneling study of spin - momentum locking in topological insulators _ , phys . b * 91 * , 235437 ( 2015 ) . j. s. lee , a. richardella , d. r. hickey , k. a. mkhoyan , and n. samarth , _ mapping the chemical potential dependence of current - induced spin polarization in a topological insulator _ , phys . b * 92 * , 155312 ( 2015 ) . j. tian , i. miotkowski , s. hong , and y. p. chen , _ electrical injection and detection of spin - polarized currents in topological insulator @xmath168 _ , sci . * 5 * , 14293 ( 2015 ) . k. vaklinova , a. hoyer , m. burghard and k. kern , _ current - induced spin polarization in topological insulator - graphene heterostructures _ , nano lett . * 16 * , 2595 ( 2016 ) . s. hong , v. diep , s. datta , and y. p. chen , _ modeling potentiometric measurements in topological insulators including parallel channels _ b * 86 * , 085131 ( 2012 ) . m. bianchi , d. guan , s. bao , j. mi , b. b. iversen , p. d. c. king , and p. hofmann , _ coexistence of the topological state and a two - dimensional electron gas on the surface of @xmath169 _ , nat . commun . * 1 * , 128 ( 2010 ) . p. d. c. king , r. c. hatch , m. bianchi , r. ovsyannikov , c. lupulescu , g. landolt , b. slomski , j. h. dil , d. guan , j. l. mi , e. d. l. rienks , j. fink , a. lindblad , s. svensson , s. bao , g. balakrishnan , b. b. iversen , j. osterwalder , w. eberhardt , f. baumberger , and ph . hofmann , _ large tunable rashba spin splitting of a two - dimensional electron gas in @xmath169 _ , phys . rev . lett . * 107 * , 096802 ( 2011 ) . zhu , g. levy , b. ludbrook , c. n. veenstra , j. a. rosen , r. comin , d. wong , p. dosanjh , a. ubaldini , p. syers , n. p. butch , j. paglione , i. s. elfimov , and a. damascelli , _ rashba spin - splitting control at the surface of the topological insulator @xmath169 _ , phys . lett . * 107 * , 186405 ( 2011 ) . m. s. bahramy , p. d. c. king , a. de la torre , j. chang , m. shi , l. patthey , g. balakrishnan , p. hofmann , r. arita , n. nagaosa , and f. baumberger , _ emergent quantum confinement at topological insulator surfaces _ , nat . commun . * 3 * , 1159 ( 2012 ) . m. johnson and r. h. silsbee , _ interfacial charge - spin coupling : injection and detection of spin magnetization in metals _ * 55 * , 1790 ( 1985 ) . m. johnson and r. h. silsbee , _ coupling of electronic charge and spin at a ferromagnetic - paramagnetic metal interface _ , phys . b * 37 * , 5312 ( 1988 ) . g. schmidt , d. ferrand , l. w. molenkamp , a. t. filip , and b. j. van wees , _ fundamental obstacle for electrical spin injection from a ferromagnetic metal into a diffusive semiconductor _ , phys . b * 62 * , r4790 ( 2000 ) . e. i. rashba , _ theory of electrical spin injection : tunnel contacts as a solution of the conductivity mismatch problem _ , phys . b * 62 * , r16267 ( 2000 ) . x. lou , c. adelmann , s. a. crooker , e. s. garlid , j. zhang , k. s. m. reddy , s. d. flexner , c. j. palmstrom , and p. a. crowell , _ electrical detection of spin transport in lateral ferromagnet - semiconductor devices _ , nat . phys . * 3 * , 197 ( 2007 ) . t. arakane , t. sato , s. souma , k. kosaka , k. nakayama , m. komatsu , t. takahashi , z. ren , k. segawa , and y. ando , _ tunable dirac cone in the topological insulator bi@xmath170sb@xmath171te@xmath172se@xmath173 _ , nat . commun . * 3 * , 636 ( 2012 ) . z. ren , a. a. taskin , s. sasaki , k. segawa , and y. ando , _ optimizing bi@xmath170sb@xmath171te@xmath172se@xmath174 solid solutions to approach the intrinsic topological insulator regime _ , phys . b * 84 * , 165311 ( 2011 ) . f. yang , a. a. taskin , s. sasaki , k. segawa , y. ohno , k. matsumoto , and y. ando , _ top gating of epitaxial ( bi@xmath6sb@xmath171)@xmath0te@xmath1 topological insulator thin films _ , 104 * , 161614 ( 2014 ) . e. k. de vries , a. m. kamerbeek , n. koirala , m. brahlek , m. salehi , s. oh , b. j. van wees , and t. banerjee , _ towards the understanding of the origin of charge - current - induced spin voltage signals in the topological insulator bi@xmath0se@xmath1 _ , phys . b * 92 * , 201102 ( 2015 ) . see supplemental material for details regarding the estimation of fermi - level in bisbtese@xmath0 flakes , control experiment on a device without tunnel barrier , surface morphology of exfoliated bisbtese@xmath66 flakes , @xmath175 curves of the exfoliated bisbtese@xmath66 flakes and additional information about the fm tunnel contacts . e. y. tsymbal , o. n. mryasov and p. r. leclair , _ spin - dependent tunnelling in magnetic tunnel junctions _ , j. phys . matter * 15 * , r109 ( 2003 ) . li and i. appelbaum , _ interpreting current - induced spin polarization in topological insulator surface states _ , b * 93 * , 220404 ( 2016 ) . t. seki , y. hasegawa , s. mitani , s. takahashi , h. imamura , s. maekawa , j. nitta and k. takanashi , _ giant spin hall effect in perpendicularly spin - polarized fept / au devices _ , nat . mater . * 7 * , 125 ( 2008 ) . s. saito and t. maeda , _ work function of ferromagnetic metals and alloys _ , vacuum ( japan ) * 24 * , 220 ( 1981 ) . d. takane , s. souma , t. sato , t. takahashi , k. segawa , y. ando , _ work function of bulk - insulating topological insulator bi@xmath170sb@xmath171te@xmath172se@xmath174 _ , arxiv:1606.07933 ( 2016 ) . to estimate the fermi level from the carrier density , one needs to know the dispersion relations of both the surface states and the bulk states of bisbtese@xmath66 . for this purpose , we took the surface - state dispersion of bisbtese@xmath66 obtained by angle - resolved photoemission spectroscopy ( arpes ) @xcite and smoothly extrapolated the curve to a wider energy scale , as plotted in fig . [ fig : figs1 ] . also , according to the same arpes result @xcite , the conduction band bottom and the valence band top are located at @xmath176 ev and @xmath177 ev , respectively , as indicated by green dotted lines in fig . [ fig : figs1](a ) . note that @xmath178 is measured from the dirac point of the surface state . as for the bulk conduction band , since its exact dispersion in bisbtese@xmath66 is not yet known , we used the dispersion in bi@xmath66se@xmath67 @xcite as an approximation to that in bisbtese@xmath66 . the bulk dispersion in fig . [ fig : figs1](a ) is a reproduction of the data for bi@xmath66se@xmath67 reported in ref . @xcite with the location of the band bottom adjusted to @xmath179 ev . at zero temperature , the carrier density contributed by each band is determined by the volume ( or area ) enclosed by the fermi surface ( or fermi circle ) of that band . for a topological insulator , the two - dimensional ( 2d ) carrier density of the topological surface states and bulk states can be respectively written as @xmath180 and @xmath181 where @xmath147 is the fermi wave vector and @xmath182 is the sample thickness . the prefactor 2 in eq . ( 2 ) comes from the spin degeneracy of the bulk states . for a given carrier density , the corresponding energy level @xmath178 of the highest occupied state can be calculated by using eqs . ( 1 ) and ( 2 ) . figure s1(b ) shows such a relation upon considering only the surface sates ; the fermi levels of devices # 2 and # 3 can be read off from this figure , since their carrier densities are low enough for the bulk carriers to be absent . on the other hand , for device # 1 we need to take into account the bulk carriers because of its relatively high carrier density ; by assuming that the total 2d carrier density @xmath183 of device # 1 determined from the hall resistivity ( see sec . iii - b of the main text ) can be approximated by @xmath184 , the @xmath178 vs. @xmath185 relation is calculated as shown in fig . [ fig : figs1](c ) . the fermi level of device # 1 can be estimated from this figure . as a result of the above procedure , the fermi levels in all three devices are estimated to be @xmath186 0.24 ev , @xmath187 ev , @xmath188 0.16 ev and @xmath189 ev for to devices # 1 , # 2 and the two surfaces of devices # 3 , respectively . to rule out the possibility that the observed hysteresis loops in devices # 1 # 3 are due to artifacts like the local hall effect ( within the ti ) or the anomalous nernst effect ( within the fm ) , we fabricated a device ( labeled s1 ) for a control experiment . the thickness of the flake used in device s1 is 298 nm . the ferromagnetic ( fm ) contacts are made of py / au , without a tunnel barrier beneath it . the measurement was performed at @xmath190 k. since the occurrence and detection of both the local hall effect and the anomalous nernst effect do not require a tunnel contact , if the signals observed in devices # 1 # 3 were due to such artifacts , similar signals should also be observed in device s1 . however , as shown in fig . s2(b ) , no hysteresis loop is observed at the excitation current of as large as 200 @xmath118a , the maximum current that is used in this experiment . therefore , possible artifacts due to the local hall effect or the anomalous nernst effect can be safely ruled out . the topological insulator thin films grown by the molecular beam epitaxy ( mbe ) method usually have triangular terraces on the surface . the fringe fields generated at the edge of such triangular features were proposed to be the possible origin of the local hall effect @xcite . here we show that , in contrast to the mbe - grown thin films , the surfaces of exfoliated bisbtese@xmath66 flakes are usually very smooth . one can easily find flakes with atomically flat surfaces . in figs . 3(b)-3(c ) we show the atomic - force - microscope ( afm ) data measured on the top surface of the flake shown in fig . the surface is atomically flat with a height variation of only @xmath191 nm . it is possible that the morphology difference between mbe - grown thin films and exfoliated flakes is the reason why the local hall effect reported in ref . [ 3 ] is not observed in our experiment . the @xmath193 curves of two bisbtese@xmath66 flakes ( labeled f1 and f2 ) are shown in fig . s4(a ) . both flakes f1 and f2 show curved hall resistance , as plotted in fig . the fitting of the f1 data to the two - band model gives @xmath194 @xmath80 and @xmath195 @xmath80 for the two types of carriers ( most likely those on top and bottom surfaces ) , while the same analysis for the f2 data gives @xmath196 @xmath80 and @xmath197 @xmath80 . the areas of the fm tunnel contacts were @xmath198 @xmath199 for device # 2 and @xmath200 @xmath199 for devices # 1 and # 3 . the resistance - area product of the fm contacts is in the range of @xmath201 @xmath202 . the current - voltage ( @xmath112-@xmath203 ) characteristics of the fm contacts are usually non - linear , suggesting tunneling behavior . the @xmath112-@xmath203 curve of a typical al@xmath66o@xmath67/py contact is shown in fig . s4(c ) . t. arakane , t. sato , s. souma , k. kosaka , k. nakayama , m. komatsu , t. takahashi , z. ren , k. segawa , and y. ando , _ tunable dirac cone in the topological insulator bi@xmath170sb@xmath171te@xmath172se@xmath173 _ , nat . commun . * 3 * , 636 ( 2012 ) . y. l. chen , j .- h . chu , j. g. analytis , z. k. liu , k. igarashi , h .- h . kuo , x. l. qi , s .- k . mo , r. g. moore , and d. h. lu , _ massive dirac fermion on the surface of a magnetically doped topological insulator _ , science * 329 * , 659 ( 2010 ) . e. k. de vries , a. m. kamerbeek , n. koirala , m. brahlek , m. salehi , s. oh , b. j. van wees , and t. banerjee , _ towards the understanding of the origin of charge - current - induced spin voltage signals in the topological insulator bi@xmath0se@xmath1 _ , phys . rev . b * 92 * , 201102 ( 2015 ) .

for the last century astronomers have been modelling the structure of ` nebulae ' , and here we focus on those external to the milky way . a key activity performed by many astronomers , past and present , is the catergorisation of these galaxies ( sandage 2005 ) and the quantification of their physical properties . how big are they ? how bright are they ? what characteristics distinguish or unite apparent subpopulations ? answers to such questions , and the establishment of `` scaling relations '' between two or more galactic properties provides valuable insight into the physical mechanisms that have shaped galaxies . understanding how galaxies form , increasingly through the use of simulations and semi - analytic modelling ( e.g. cole 1991 ; white & frenk 1991 ; kauffmann et al . 1993 , 2003 ; avila - reese et al . 1998 ; cole et al . 2000 ; de lucia et al . 2006 ; bower et al . 2006 ; kauffmann et al . 2004 ; di matteo et al . 2005 ; croton et al . 2006 ; naab et al . 2006 ; nipoti et al . 2006 ; covington et al . 2011 ; guo et al . 2011 , and references therein ) , requires an accurate knowledge of galaxy properties and scaling laws , as elucidated by driver ( 2011 ) . not surprising , our knowledge of galaxies is best in the nearby universe out to distances typically measured in megaparsecs rather than by redshift @xmath2 where galaxy structures can be reasonably well resolved . the properties of these galaxies provide the @xmath3 benchmark used in the calibration of galactic evolutionary studies both observed and simulated . popular scaling relations involving global galaxy parameters such as size , surface brightness , luminosity and concentration are reviewed here . as we shall see , many bivariate distributions , which are frequently assumed to be linear , are often only approximately so over a restricted luminosity range . for example , it may come as a surprise for many too learn that the useful kormendy ( 1977b ) relation is only the tangent to the bright arm of a continuous but curved effective radius-(surface brightness ) relation which unifies dwarf and giant elliptical galaxies ( section [ sec_korm ] ) . similarly , the faber - jackson ( 1976 ) relation with a slope of 4 represents the average slope over a restricted luminosity range to what is a curved or broken luminosity-(velocity dispersion ) distribution , in which the slope is 2 rather than 4 at lower luminosities ( section [ sec_faber ] ) . knowing these trends , the bulk of which can not be established when assuming structural homology , i.e. using de vaucouleurs ( 1948 ) @xmath4 model , is vital if one is to measure , model and make sense of galaxies . this article has been structured into four main sections . section 1 provides this general overview plus a further review and introduction to galaxies on the hubble - jeans sequence mag . these galaxies can not be observed ( to date ) at cosmologically interesting distances , and their increased scatter in the colour - magnitude relation may indicate a range of galaxy types ( e.g. penny & conselice 2008 , and references therein ) . ] . included are diagrams showing the location of dynamically hot stellar systems in the mass - size and mass - density plane , revealing that some high-@xmath2 compact galaxies have properties equivalent to the bulges of local disc galaxies . section [ sec_ell ] provides an historical account of how the radial distribution of stars in elliptical galaxies have been modelled , and the iterative steps leading to the development of the modern core - srsic model ( section [ sec_cs ] ) . subsections cover the srsic model ( section [ sec_body ] ) , its relation and applicability to dark matter halos ( section [ sec_dark ] ) , partially - depleted galaxy cores ( section [ sec_def ] ) , excess nuclear light ( section [ sec_enuc ] ) and excess light at large radii in the form of halos or envelopes around giant elliptical galaxies ( section [ sec_cd ] ) . section [ sec_e ] presents and derives a number of elliptical galaxy scaling relations pertaining to the main body of the galaxy . from just two linear relations which unite the faint and bright elliptical galaxy population ( section [ sec_line ] ) , a number of curved relations are derived ( section [ sec_curved ] ) . several broken relations , at @xmath5 mag , are additionally presented in section [ sec_bent ] . for those interested in a broader or different overview of elliptical galaxies , some recent good reviews include renzini ( 2006 ) , cecil & rose ( 2007 ) , ciotti ( 2009 ) and lisker ( 2009 ) . finally , the latter third of this paper is tied up in section [ sec_discs ] which contains a discussion of the light profiles of disc galaxies and their bulge - disc decomposition ( [ sec_bd ] ) . also included are subsections pertaining to dust ( section [ sec_dust ] ) , the difficulties with identifying pseudobulges ( section [ sec_pseudo ] ) , potential bulgeless galaxies ( section [ sec_less ] ) and methods to model bars ( section [ sec_bar ] ) . throughout the article references to often overlooked discovery or pioneer papers are provided . looking out into the milky way arced across our night sky , the notion that we are residents within a pancake - shaped galaxy seems reasonable to embrace . indeed , back in 1750 thomas wright also conjectured that we reside within a flat layer of stars which is gravitationally bound and rotating about some centre of mass . however , analogous to the rings of saturn , he entertained the idea that the milky way is comprised of a large annulus of stars rotating about a distant centre , or that we are located in a large thin spherical shell rotating about some divine centre ( one of the galactic poles ) . while he had the global geometry wrong , he was perhaps the first to speculate that faint , extended nebulae in the heavens are distant galaxies with their own ( divine ) centers . as elucidated by hoskin ( 1970 ) , it was immanuel kant ( 1755 ) , aware of the elliptically - shaped nebulae observed by maupertuis , and working from an incomplete summary of wright ( 1750 ) that had been published in a hamburg journal , who effectively introduced the modern concept of disc - like galactic distributions of stars mistakenly crediting wright for the idea . using his 1.83 m `` leviathan of parsonstown '' metal reflector telescope in ireland , lord william henry parsons , the 3rd earl of rosse , discovered 226 new general catalogue ( ngc : dreyer 1888 ) and 7 index catalogue ( ic : dreyer 1895 , 1908 ) objects ( parsons 1878 ) . important among these was his detection of spiral structure in many galaxies , such as m51 which affectionately became known as the whirlpool galaxy . further divisions into disc ( spiral ) and elliptical galaxy types followed ( e.g. wolf 1908 ; knox shaw 1915 ; curtis 1918 ; reynolds 1920 ; hubble 1926 ) and shapley & swope ( 1924 ) and shapley ( 1928 ) successfully identified our own galaxy s ( gravitational ) center towards the constellation sagittarius ( see also seares 1928 ) . with the discovery that our universe contains doppler shifted nebulae that are expanding away from us ( de sitter 1917 ; slipher 1917 ; see also friedmann 1922 , lundmark 1924 and the reviews by kragh & smith 2003 and shaviv 2011 ) , in accord with a redshift - distance relation ( lemaitre 1927 , robertson 1928 , humasson 1929 , hubble 1929 ) i.e. awareness that some of the `` nebuale '' are external objects to our galaxy came increased efforts to catergorise and organise these different types of `` galaxy '' . as noted by sandage ( 2004 , 2005 ) , sir james jeans ( 1928 ) was the first to present the ( tuning fork)-shaped diagram that encapsulated hubble s ( 1926 ) early - to - late type galaxy sequence , a sequence which had been inspired in part by jeans ( 1919 ) and later popularised by hubble ( 1936a ; see block et al.2004 ) . quantifying the physical properties of galaxies along this sequence , with increasing accuracy and level of detail , has occupied many astronomers since . indeed , this review addresses aspects related to the radial concentration of stars in the elliptical and disc galaxies which effectively define the hubble - jeans sequence . irregular galaxies are not discussed here . for reasons that will become apparent , this review uses the galaxy notation of alan sandage and bruno binggeli , in which dwarf elliptical ( de ) galaxies are the faint extension of ordinary and luminous elliptical ( e ) galaxies , and the dwarf spheroidal ( dsph ) galaxies prevalent in our local group ( grebel 2001 ) are found at magnitudes fainter than @xmath6 to @xmath714 mag ( @xmath8 in stellar mass ; see figure [ fig1]a ) . figure [ fig1]a reveals a second branch of elliptically - shaped object stretching from the bulges of disc galaxies and compact elliptical ( ce ) galaxies to ultra compact dwarf ( ucd ) objects ( hilker et al . 1999 ; drinkwater et al . 2000 ; norris & kannappan 2011 and references therein ) . a _ possible _ connection is based upon the stripping of a disc galaxy s outer disc to form a ce galaxy ( nieto 1990 ; bekki et al . 2001b ; graham 2002 ; chilingarian et al . 2009 ) and through greater stripping of the bulge to form a ucd ( zinnecker et al . 1988 ; freeman 1990 ; bassino et al . 1994 ; bekki 2001a ) . it is thought that nucleated dwarf elliptical galaxies may also experience this stripping process , giving rise to ucds . while the identification of local spiral galaxies is relatively free from debate , the situation is not so clear in regard to elliptically - shaped galaxies . the discovery of ucds , which have sizes and fluxes intermediate between those of galaxies and ( i ) the nuclear star clusters found at the centres of galaxies and ( ii ) globular clusters ( gcs : e.g. haegan et al . 2005 ; brodie & strader 2006 ) , led forbes & kroupa ( 2011 ) to try and provide a modern definition for what is a galaxy ( see also tollerud et al.2011 ) . only a few years ago there was something of a divide between gcs and ucds all of which had sizes less than @xmath930 pc and galaxies with sizes greater than 120 pc ( gilmore et al . 2007 ) . however , as we have steadily increased our celestial inventory , objects of an intermediate nature have been found ( e.g. ma et al . 2007 , their table 3 ) , raising the question asked by forbes & kroupa for which , perhaps not surprisingly , no clear answer has yet emerged . while those authors explored the notion of a division by , among other properties , size and luminosity , they did not discuss how the density varies . as an addendum of sorts to forbes & kroupa ( 2011 ) , the density of elliptically - shaped objects is presented here in figure [ fig1]b . this is also done to allow the author to wave the following flag . apparent in figure [ fig1]b , but apparently not well recognised within the community , is that the bulges of disc galaxies can be much denser than elliptical galaxies . if the common idea of galaxy growth via the accretion of a disc , perhaps from cold - mode accretion streams , around a pre - existing spheroid is correct ( e.g.navarro & benz 1991 ; steinmetz & navarro 2002 ; birnboim & dekel 2003 ; see also conselice et al . 2011 and pichon et al . 2011 ) , then one should expect to find dense spheroids at high-@xmath2 with @xmath10@xmath11 of stellar material , possibly surrounded by a faint ( exponential ) disc which is under development . it is noted here that the dense , compact early - type galaxies recently found at redshifts of 1.4 - 2.5 ( daddi et al . 2005 ; trujillo et al . 2006 ) display substantial overlap with the location of present day bulges in figure [ fig1]a , and that the merger scenarios for converting these compact high-@xmath2 galaxies into today s elliptical galaxies are not without problems ( e.g. nipoti et al . 2009 ; nair et al . 2011 ) . it is also noted that well - developed discs and disc galaxies are rare at the redshifts where these compact objects have been observed alongside normal - sized elliptical galaxies . before trying to understand galaxy structure at high - redshift , and galaxy evolution themes not detailed in this review it is important to first appreciate galaxy structures at @xmath3 where observations are easier and local benchmark scaling relations have been established . over the years a number of mathematical functions have been used to represent the radial distribution of stellar light in elliptical galaxies , i.e. their light profiles . before getting to de vaucouleurs @xmath4 model in the following paragraph , it seems apt to first quickly mention some early competitors . although plummer s ( 1911 ) internal - density model was developed for the nebulae which became know as globular clusters , because of its simplicity it is still used today by some researchers to simulate elliptical galaxies , even though , it should be noted , no modern observers use this model to describe the radial distribution of light in elliptical galaxies . reynold s ( 1913 ) surface - density model , sometimes referred to as hubble s ( 1930 ) model or the reynold - hubble model , was used to describe the nebula which became known as elliptical galaxies . it has an infinite mass and is also no longer used by observers today . the modified hubble model ( rood et al . 1972 ) , which also has an infinite mass , is also still sometimes used by simulators , even though , again , observers do not use this model anymore . oemler s ( 1976 ) exponentially - truncated hubble model , known as the oemler - hubble model , is also not used to represent the observed stellar distribution in elliptical galaxies because it too , like its predecessors , was simply an approximation applicable over a limited radial range , as noted by king ( 1978 ) . it is interesting to note that up until the 1980s , departures at large radii from the reynold s model were attributed to tidal - stripping by external gravitational potentials . that is , for three quarters of a century , reynold s model originally developed from low - quality data for one galaxy was generally thought to describe the original , undisturbed stellar distribution in elliptical galaxies . de vaucouleurs ( 1948 , 1953 ) @xmath4 surface - density model had traction for many years , in part due to de vaucouleurs ( 1959 ) arguing that it fits better than the reynold s model used by hubble , a point re - iterated by kormendy ( 1977a ) and others and the revelation that it fits the radially - extended data for ngc 3379 exceedingly well ( de vaucouleurs & capaccioli 1979 ) . hodge ( 1961a , b ) had however revealed that de vaucouleurs model was inadequate to describe faint elliptical galaxies and hodge ( 1963 , 1964 ) , in addition to king ( 1962 ) , noted that the 3-parameter king model , with its flatter inner profile and steeper decline at large radii , did a better job . for a time , king s ( 1962 , 1966 ) model became popular for describing the light distribution in faint elliptical galaxies , at least until the exponential model also used for the discs of spiral galaxies was noted to provide a good description of some dwarf elliptical galaxies ( hodge 1971 ; faber & lin 1983 ; binggeli et al . 1984 ) and that these galaxies need not have experienced any tidal truncation ( a prescription of the king model with its tidal radius parameter ) . lauer ( 1984 , 1985 ) additionally showed that king s modified isothermal model , with its flat inner core , was inadequate to describe the deconvolved light - profiles of ordinary elliptical galaxies with `` cores '' , i.e. galaxies whose inner light profile displays a nearly flat core . king s model does however remain extremely useful for studies of star clusters , globular clusters , dwarf spheroidal galaxies and galactic satellites which , unlike ordinary elliptical galaxies , can have flat cores in their inner surface brightness profile . today , the model of choice for describing nearby ( and distant ) dwarf and ordinary elliptical galaxies is srsic s ( 1963 ) generalisation of de vaucouleurs @xmath4 model to give the @xmath12 surface - density model ( section [ sec_body ] ) . this model reproduces the exponential model when @xmath13 and de vaucouleurs model when @xmath14 ; it can thus describe the main body of faint and luminous elliptical galaxies . the key advantage that this model has is ( i ) its ability to describe the observed stellar distributions that have a range of central concentrations ( known to exist since at least reaves 1956 ) and ( ii ) it provides a very good description of the data over ( almost ) the entire radial extent . indeed , departures in the light profile from a well - fit srsic s model invariably signal the presence of additional features or components , rather than any failing of the model . expanding upon the srsic model , the core - srsic model ( section [ sec_cs ] ) is nowadays used to quantify those galaxies with `` cores '' . although referring to the king model , the following quote from king ( 1966 ) seems particularly insightful `` ... de vaucouleurs law appears to refer to a particular central concentration and should be appropriate only for galaxy profiles that have that concentration . '' while noted by others , such as oemler ( 1976 ) , capaccioli ( 1985 ) , michard ( 1985 ) and schombert ( 1986 ) , some three decades elapsed before the relevance of king s remark to elliptical galaxies re - surfaced albeit slowly at first in the 1990s . indeed , de vaucouleurs useful , albeit limited , @xmath4 model was referred to as a `` law '' for nearly half a century . however we are now more keenly aware that ( even normal ) elliptical galaxies possess a range of central concentrations : concentrations which are well quantified in terms of the exponent @xmath15 in srsic s @xmath12 model ( see trujillo , graham & caon 2001 ; graham , trujillo & caon 2001 ) . although , it should be confessed that one can still encounter papers which use @xmath4 model parameters alongside some model - independent measure of galaxy concentration , unaware of the inconsistency arising from the fact that every @xmath4 model actually has exactly the same level of concentration . before introducing the equation for srsic s model in the following section , it is pointed out that in addition to modelling what can be considered the main body of the galaxy , one can also find excess stellar light at ( i ) small radii in the form of nuclear ( i.e. centrally located ) discs and dense nuclear star clusters ( section [ sec_enuc ] ) and also at ( ii ) large radii in the form of halos or envelopes in cd and central cluster galaxies ( section [ sec_cd ] ) . as briefly noted above , deficits of stellar flux at the cores of massive galaxies are also observed , and a model to quantify these stellar distribution , relative to the outer non - depleted light profile , is described in section [ sec_cs ] . while non - symmetrical components in elliptical galaxies can also exist , they are not addressed here given the focus on well - structured systems . somewhat random , non - symmetrical components may be a sign of a disturbed morphology ( see e.bartons chapter in this volume ) , of on - going non - uniform star formation ( see s.boissiers chapter in this volume ) or gravitationally - induced tidal features from external forces . jos srsic s ( 1963 , 1968 ) @xmath12 model , which was introduced in spanish , describes how the projected surface - intensity @xmath16 varies with the projected radius @xmath17 , such that @xmath18\right\ } \label{eq_ser}\ ] ] and @xmath19 is the intensity at the ` effective ' radius @xmath20 that encloses half of the total light from the model ( ciotti 1991 ; caon et al . the term @xmath21 ( @xmath22 for @xmath23 , capaccioli 1989 ) is not a parameter but is instead dependent on the third model parameter , @xmath15 , that describes the shape , i.e. the concentration , of the light profile . the exact value of @xmath21 is obtained by solving the equation @xmath24 where @xmath25(2n , x ) is the incomplete gamma function and @xmath26 is the ( complete ) gamma function ( ciotti 1991 ) . useful srsic related expressions have been presented in ciotti ( 1991 ) , simonneau & prada ( 2004 ) and ciotti & bertin ( 1999 ) , while graham & driver ( 2005 ) provide a detailed review of srsic s model plus associated quantities and references to pioneers of this model . the relation between the effective surface brightness ( @xmath27 ) and the central surface brightness ( @xmath28 ) is given by the expression @xmath29 where we have dropped the subscript @xmath15 from the term @xmath21 for simplicity , while @xmath30 \label{eq_mue - mue}\ ] ] gives the difference between the effective surface brightness and the mean effective surface brightness ( @xmath31 ) within @xmath20 . figure [ fig_ser ] shows the behaviour of the srsic model . uniting ccd data with wide and deep photographic images , caon et al . ( 1990 , 1993 , 1994 ) revealed that the srsic @xmath12 model provided a remarkably good description to the stellar distribution over a large radial range , down to surface brightnesses of @xmath928 @xmath32-mag arcsec@xmath33 , for the early - type galaxies brighter than @xmath34 mag in the virgo cluster . this work was in essence an expansion of de vaucouleurs & capaccioli s ( 1979 ) study of ngc 3379 which is very well - fit with @xmath14 . different galaxies were discovered by caon et al . ( 1993 ) to be equally well fit , but required different values of @xmath15 ( see also bertin et al . 2002 ) . importantly , caon et al . ( 1993 ) additionally showed that a correlation existed between stellar concentration , i.e. the srsic index @xmath15 , and ( model - independent ) galaxy size that was not due to parameter coupling in the srsic model ( see also trujillo et al . 2001 , their section 2 ) . one of the commonly overlooked implications of this result is that @xmath4 , and similarly petrosian ( 1976 ) , magnitudes , sizes and surface brightnesses are systematically in error as a function of galaxy concentration ( graham et al . 2005 ; hill et al . that is , application of a model which fails to adequately capture the range of stellar distributions will result in parameters which are systematically biased as a function of galaxy mass . for example , fitting an @xmath4 model to elliptical galaxies which are actually described by an @xmath12 model with @xmath15 less than and greater than 4 will yield sizes and luminosities which are , respectively , greater than and less than the true value ( e.g. binggeli & cameron 1991 ; trujillo et al . 2001 ; brown et al . similarly , fitting an exponential model to bulges that are best described by an @xmath12 model with @xmath15 less than and greater than 1 will yield sizes and luminosities which are , respectively , greater than and less than the true value ( e.g. graham 2001 ) . obviously one does not want to fine tune their galaxy simulations to match scaling relations that contain systematic biases due to poor measurements , and observers are therefore busy fitting @xmath12 models these days . a good approximation to the internal - density profile associated with srsic s model , i.e. with its deprojection , was introduced by prugniel & simien ( 1997 ) . useful expressions for the dynamics , gravitational potential and forces of this model have been developed by trujillo et al . ( 2002 ) , terzi & graham ( 2005 ) and terzi & sprague ( 2007 ) . somewhat more complex than the early light - profile models , such expressions can , importantly , accommodate a range of concentrations , rather than only varying one scale radius and one scale density . such a model is vital if one wishes to properly simulate and understand the mass spectrum of elliptical galaxies , whose srsic index @xmath15 increases with stellar mass . while graham & driver s ( 2005 ) review stated that `` no attempt has been made here to show the numerous scientific advances engendered via application of the @xmath12 model '' , this article reveals how srsic s model , and the core - srsic model 0 ( subsection [ sec_cs ] ) , have become key in unifying and understanding the galaxies around us . like the majority of surface- and internal - density models from the last century , the srsic function is an empirical model created to match data rather than developed from theory , and as such we should be cautious before calling it a law . attempts to find a physical explanation for de vaucouleurs model yielded results which helped to keep it in vogue . dissipational models have long been touted for producing @xmath4 profiles ( e.g. larson 1969 , 1974 ) , and in the 1980s papers based on dissipationless n - body simulations of a cold clumpy collapse or the merger of disc galaxies also claimed to finally produce @xmath4 ( and also reynold ) profiles ( e.g.van albada 1982 ; mcglynn 1984 ; carlberg et al . 1986 ; barnes 1988 ) . however a closer inspection reveals clear departures from the @xmath4 profile , with the simulated profiles better described by an @xmath12 model with @xmath35 . obviously their inability ( or perhaps lack of desire , although see farouki , shapiro & duncan 1983 whose non - homologous merger remnants were initially criticised by @xmath4 aficionados ) to create the range of stellar concentrations now observed in elliptical galaxies highlights a limitation of these early works . nonetheless , these pioneering studies have led to n - body simulations by nipoti et al . ( 2006 ) and aceves et al . ( 2006 ) and farouki et al . ( 1983 ) , whose results with a smaller force softening appeared years ahead of their time have now recovered a range of srsic profile shapes for gravitational collapses in a dark matter halo and for disc galaxy mergers , respectively . given the empirical nature of srsic s @xmath12 model , hjorth & madsen ( 1995 ) revealed how dissipationless merging and violent relaxation provided a physical explanation for the departure from the homologous @xmath4 model . other works have explained how the quasi - constant specific entropy associated with the post violent - relaxation stage of elliptical galaxies results in the observed mass - dependent range of stellar concentrations in elliptical galaxies ( gerbal et al . 1997 ; lima neto 1999 ; mrquez et al . . it does not seem too unreasonable to speculate that elliptical galaxies , whether built by near - monolithic collapse , collisions of disc galaxies , wet or dry mergers , appear to eventually experience the same force(s ) of nature that results in their radial stellar distribution depending on the total stellar mass . that is , it may not matter how the mass was accumulated into an elliptical galaxy , once it becomes a dynamically - heated , bound stellar - system , it appears to eventually obey certain universal scaling relations ( see section [ sec_e ] ) . it is interesting to note that srsic actually introduced his model as a way to parameterise disc galaxies which he thought were comprised of differing ratios of a disc plus an @xmath4-bulge . his model was not initially intended to fit elliptical galaxies , and as such it did not immediately threaten de vaucouleurs model . credit for popularising the use of srsic s @xmath12 model for approximating not only lenticular bulge@xmath36disc galaxies but for describing pure elliptical galaxies resides largely with massimo capaccioli ( e.g. capaccioli 1985 , 1987 , 1989 ; caon et al . 1993 ; donofrio et al.1994)-bulge plus an exponential disc ( see section [ sec_discs ] . ] . however , davies et al . ( 1988 ) had also introduced this model for dwarf elliptical galaxies , while sparks ( 1988 ) developed an early gaussian seeing correction for this model , and ciotti ( 1991 ) developed a number of associated expressions such as the velocity dispersion profile and a distribution function . the important quantification that capaccioli and others provided is how the radial distribution of stars in elliptical galaxies , i.e. their concentration , varies with the size , luminosity and thus the mass of the elliptical galaxy ( see also cellone , forte , & geisler 1994 ; vennik & richter 1994 ; young & currie 1994 , 1995 ; graham et al . 1996 ; karachentseva et al . 1996 , vennik et al . 1996 ) . as we shall see in this article , the implications of this breakthrough have been dramatic , unifying what had previously been considered two distinct species of galaxy , namely dwarf and ordinary elliptical galaxies , previously thought to be described by an exponential and @xmath4 model , respectively . this review would be somewhat incomplete without a few words regarding the connection between srsic s model and ( simulated ) dark matter halos . while modified theories of gravity may yet make dark matter redundant at some level , it is intriguing to note that the prugniel - simien ( 1997 ) internal - density model , developed to approximate the deprojected form of srsic s @xmath12 model , additionally provides a very good representation of the internal - density profiles of simulated dark matter halos . merritt et al.(2006 ) revealed that it actually provides a better description than not only the navarro , frenk & white ( 1997 ) model but even a generalised nfw model with an arbitrary inner profile slope @xmath37 . srsic s former student , navarrro , independently applied srsic s surface - density model to the internal - density profiles of simulated dark matter halos ( navarro et al . 2004 ; merritt et al . jaan einasto ( 1965 ) had previously developed this same function as srsic to describe the internal - density profiles of galaxies . rather than a universal profile shape , as advocated by navarro et al . ( 1997 ) , a range of simulated dark matter density profile shapes is now known to vary with the dark matter halo mass ( avila - reese et al . 1999 ; jing & suto 2000 ; merritt et al . 2005 ; del popolo 2010 and references therein ) . a number of useful expressions related to this `` einasto model '' , which has the same functional form as srsic s model but is applied to the internal rather than projected density profile , can be found in cardone et al . ( 2005 ) , mamon & okas ( 2005 ) and graham et al . ( 2006 ) . an apparent `` bulge - halo conspiracy '' between the radial distribution of stellar mass and dark matter ( after modification by baryons ) has arisen in recent years , such that elliptical galaxies reportedly have _ total _ internal - density profiles @xmath38 described by power - laws ( bertin & stiavelli 1993 ; kochanek 1995 ; gavazzi et al . 2007 ; buote & humphrey 2011 ) . these power - laws were originally claimed to be close to isothermal , such that @xmath39 ( koopmans et al . 2006 , 2009 ; gavazzi et al . recent work now emphasises that only the sample average profile slope is close to @xmath40 , and that a trend in slope with galaxy size exists ( humphrey & boute 2010 ; auger et al . this is a developing field and it is worth noting that the analyses have been confined to massive galaxies with velocity dispersions greater than @xmath9175 km s@xmath41 , and thus with srsic indices @xmath42 ( graham , trujillo & caon 2001 ) . while the light profile shape changes dramatically as the srsic index @xmath15 increases from @xmath91 to @xmath94 , there is not such an obvious change in light profile shape from @xmath94 to higher values of @xmath15 ( see figure [ fig_ser ] ) . the apparent isothermal profiles of elliptical galaxies , and the alleged `` bulge - halo conspiracy '' , may turn out to be a by - product of sample selection ( i.e. choosing galaxies which have approximately the same structure ) . it would be interesting to expand the sloan lens acs ( slacs ) survey ( bolton et al . 2006 ) to a greater range than only bright early - type galaxies that are approximately well fit with an @xmath4 model , and to go beyond the use of simple power - laws to describe the total mass density profile once the data allows this . two component mass models ( prugniel - simien@xmath36einasto ) to a range of galaxy masses await . claims that `` early - type galaxies are structurally close to homologous '' may therefore be premature , as was the case for the distribution of stellar light in elliptical galaxies while the @xmath4 model was thought to be a law . the centres of luminous galaxies have long been known to possess `` cores '' , such that the surface - density profile flattens at the center ( e.g. king & minkowski 1966 ) , and king & minkowski ( 1972 ) remarked on the inability of the reynold s and de vaucouleurs model to match these flattened cores in giant galaxies . although king ( 1978 ) identified a number of galaxies thought to be well described by his model , using seeing - deconvolved ground - based images , lauer ( 1983 , 1984 , 1985 ) analysed 14 galaxies with cores , ranging from 1.55.0 arcseconds in radius , revealing that they , like m87 ( young et al . 1978 ; duncan & wheeler 1980 ; binney & mamon 1982 ) , were not exactly described by the king model which had a completely flat core . similar conclusions , that cores existed but that they do not have flat inner surface brightness profiles , were also reported by kormendy ( 1982 , 1985a ) , creating the need for a new model to describe the stellar distribution in galaxies . nearly a decade later , the hubble space telescope was flying and offered factors of a few improvement over the best image resolution achievable from the ground at that time . not surprisingly , astronomers explored the centres of galaxies . in an effort to quantify these nuclear regions , after the abandonment of the king model and the lack of a flattened core in the @xmath4 model , crane et al . ( 1993 ) , ferrarese et al . ( 1994 ) , forbes et al . ( 1994 ) and jaffe et al . ( 1994 ) used a double power - law model to describe the inner light profiles of large galaxies . grillmair et al . ( 1994 ) , kormendy et al . ( 1994 ) and lauer et al . ( 1995 ) also adopted a double - power - law model but one with an additional , fifth , parameter to control the sharpness of the transition . their model , which they dubbed the `` nuker law '' for describing the nuclear regions of galaxies ( after excluding any apparent excess light ) , has the same functional form as the double power - law model presented by hernquist ( 1990 , his equation 43 ) to describe the internal - density of galaxies ( zhao 1996 ) . however , as noted by the above authors , these double power - law models were never intended to describe the entire radial extent of a galaxy s stellar distribution , and they provided no connection with the outer ( @xmath4-like ) radial profile . this disconnection turned out to be their down fall . due to the curved nature of the outer light profiles beyond the core , which were being fitted by the double power - law model s outer power - law , the five parameters of the nuker model systematically changed as the fitted radial extent changed . this was first illustrated in a number of diagrams by graham et al . ( 2003 ) who revealed that none of the nuker model parameters were robust , and as such they could not provide meaningful physical quantities . for example , trujillo et al . ( 2004 ) reported that the nuker - derived core - radii were typically double , and up to a factor of five times larger , than the radius where the inner power - law core broke away from the outer @xmath4-like profile a result reiterated by dullo & graham ( 2011 , in prep . ) . an additional problem was that these `` break radii '' were being identified in the so - called `` power - law '' galaxies that showed no evidence of a downward departure and flattening from the inward extrapolation of the outer @xmath4-like profile . this situation arose because of the curved nature of what were actually srsic profiles . that is , the so - called `` power - law '' galaxies not only had no distinct `` core '' like the `` core galaxies '' do , but confusingly they do not even have power - law light profiles . given that caon et al . ( 1993 ) and donofrio et al . ( 1994 ) had established that the srsic function fits the brightness profiles of elliptical galaxies remarkably well over a large dynamic range ( see the figures in bertin et al.2002 ) , it is possible to confidently identify departures from these profiles that are diagnostic of galaxy formation . while in this and the following section we deal with partially - depleted cores also referred to as `` missing light '' in luminous galaxies ( thought to be built from dissipationless mergers ) , the ensuing section addresses extra central light above the inward extrapolation of the outer srsic profile ( found in galaxies that have experienced dissipation and star formation ) . building on the work of caon et al . ( 1993 ) , graham et al . ( 2003 ) introduced the core - srsic model , which was applied in trujillo et al . the model represents a modification to srsic s model such that it has an inner power - law core . this represented a dramatic change from what had gone before . while the nuker team ( e.g kormendy et al . 1994 ; lauer et al . 1995 , 2005 ; faber et al . 1997 ) were combining nuker model parameters for the core with @xmath4 model parameters for the main galaxy , graham et al . ( 2003 ) , graham & guzmn ( 2003 ) , balcells et al . ( 2003 ) , graham ( 2004 ) and trujillo et al . ( 2004 ) advocated the measurement of core properties , excesses and deficits of light , measured relative to the outer srsic model . for the first time , the core - srsic model provided an expression capable of unifying the nuclear regions of galaxies with their outer regions and also providing stable physical quantities@xmath43 . ] . the model can be written as @xmath44^{\gamma/\alpha } \nonumber \\ & & \times \exp \left\ { -b[(r^{\alpha}+r_b^{\alpha})/r_{\rm e}^{\alpha}]^{1/(\alpha n)}\right\ } , \label{bomba}\end{aligned}\ ] ] where @xmath45 denotes the break radius separating the inner power law having logarithmic slope @xmath37 from the outer srsic function . the intensity @xmath46 at the break radius is such that @xmath47 . \label{iprime}\ ] ] the sixth and final parameter , @xmath48 , controls the sharpness of the transition between the inner ( power law ) and the outer ( srsic ) regime higher values of @xmath48 indicating a sharper transition . figure [ fig - core ] shows the core - srsic model ( with @xmath49 ) applied to ngc 3348 . terzi & graham ( 2005 ) and terzi & sprague ( 2007 ) provide a number of expressions related to the potential , force and dynamics . the core - srsic model is further discussed and used by ferrarese et al.(2006a , b ) , ct et al . ( 2006 , 2007 ) , kawata , cen & ho ( 2007 ) and ciotti ( 2009 ) . the collisional construction of galaxies from the merger of lesser galaxies is thought to be a common occurrence in the universe . coupled with the presence of a supermassive black hole ( smbh ) at the heart of most galaxies ( wolfe & burbidge 1970 ; magorrian et al . 1998 ) , dissipationless mergers were proposed by begelman , blandford & rees ( 1980 ; see also ebisuzaki , makino , & okumura 1991 ) to explain the depleted nuclei , i.e. the cores , observed in giant elliptical galaxies ( e.g. king 1978 , and references therein ) . it is thought that core - depletion is primarily due to the gravitational slingshot ( saslaw , valtonen , & aarseth 1974 ) effect that the coalescing smbhs from the pre - merged galaxies have on stars while they themselves sink to the bottom of the potential well of the newly we d galaxy . colour gradients in core galaxies such as ngc 1399 suggests that they may not have been built from major , dry merger events ( carter et al . additionally , the globular cluster specific frequency in core galaxies may be at odds with core - galaxy formation through equal - mass merger events because the fainter , intermediate luminosity elliptical galaxies may have lower specific frequencies ( harris & van den bergh 1981 ) . alternative ideas for core - formation in giant galaxies have been proposed : boylan - kolchin et al . 2004 ; nipoti et al . 2006 ; martizzi et al . 2011.[foot_dry ] ] theory predicts that the central mass deficit @xmath50 should scale with @xmath51 , where @xmath52 is the final ( merged ) black hole mass and @xmath0 the number of major `` dry '' ( i.e. gas free , dissipationless ) mergers ( milosavljevi & merritt 2001 ; merritt & milosavljevi 2005 ; merritt 2006a , b ) . graham ( 2004 ) used the core - srsic model to quantify the central deficit of stars relative to the inward extrapolation of the outer srsic profile . figure [ fig_def ] suggests that the luminous elliptical galaxies sampled have experienced an average of 1 or 2 major dry ( i.e. dissipationless ) mergers , a conclusion in agreement with select @xmath53cdm models of galaxy formation ( haehnelt & kauffmann 2002 ; volonteri et al . 2003 ) . quantification of the central stellar deficit relative to the inward extrapolation of the outer srsic profile has also been applied to bright virgo cluster galaxies by ferrarese et al . ( 2006a ) and ct et al.(2007 ) and with the exception of vcc 798 a lenticular ( bulge plus disc ) galaxy provides similar results . of course when an outer disc exists , a core - srsic bulge plus disc fit is required , otherwise the disc will bias the srsic parameters . when an exponential - disc plus ( core - srsic)-bulge fit is performed ( graham , unpublished work ) . ] while galaxies brighter than @xmath54 mag have partially depleted cores ( e.g. faber et al . 1997 ; rest et al . 2001 ; graham & guzmn 2003 ) , fainter galaxies often contain additional stellar components at their centres . such excess nuclear light , above that of the underlying host galaxy , has long been known to exist ( e.g. smith 1935 ; reaves 1956 , 1977 ; romanishin , strom & strom 1977 ) and was systematically studied by binggeli et al . ( 1984 ) , van den bergh ( 1986 ) and binggeli & cameron ( 1991 , 1993 ) in a number of dwarf elliptical galaxies . as far back as larson ( 1975 ) it was known that simulations containing gas can account for these dense nuclear star clusters ; clusters which became easier to detect with the hubble space telescope ( e.g.carollo , stiavelli & mack 1998 ) . for many years it was common practice to simply exclude these additional nuclear components from the analysis of the galaxy light profile ( e.g. lauer et al . 1995 ; byun et al.1996 ; rest et al . 2001 ; ravindranath et al . 2001 ) . using hst data , graham & guzmn ( 2003 ) simultaneously modelled the host galaxy and the additional nuclear component with the combination of a srsic function for the host galaxy plus a gaussian function for the nuclear star cluster . as we will see in section [ sec_discs ] , they also showed that the lenticular galaxies in their early - type galaxy sample could be modelled via a srsic - bulge plus an exponential - disc plus a gaussian-(star cluster ) decomposition of their light profile as done by wadadekar et al . ( 1999 ) with ground - based images . many other studies have since modeled the nuclear star clusters seen in hst images , see figure [ fig_596 ] , with the combination of a nuclear component plus a srsic host galaxy ( e.g. grant et al . 2005 ; ct et al . 2006 ; ferrarese et al . 2006a ; graham & spitler 2009 ) . while graham & guzmn ( 2003 ) and ct et al . ( 2006 ) found that some nuclear star clusters could actually be resolved , it is not yet established what mathematical function best describes their radial distribution of stars . the closest example we have to study is of course the 30 million solar mass nuclear star cluster at the centre of the milky way ( launhardt et al . graham & spitler ( 2009 ) provided what may have been the first ever analysis after allowing for the significant contamination from bulge stars . although they found that the cluster could be well described by a srsic index with @xmath55 ( see figure [ fig_milky ] , it remains to be tested how well a king model can describe the data , or at least the old stellar population known to have a core ( e.g. genzel et al . the excess nuclear light in m32 has also been well fit with an @xmath56 srsic function ( graham & spitler ) but this too may yet be better described by a king model . what is apparent is that neither the underlying bulge nor the nuclear component are described by a power - law . theories which form or assume such power - law cusps appear to be at odds with current observations . -band intensity profile from the centre of the milky way , taken from schdel et al . ( 2009 , their figure 2 ) . the nuclear star cluster is modelled with a srsic function with @xmath55 ( dotted curve ) and the underlying host bulge with an exponential function ( kent et al . 1991 ) that has an effective half - light radius of @xmath94.5 degrees ( e.g. graham & driver 2007 ) and is therefore basically a horizontal line . one parsec equals 25 arcseconds . _ figure from graham & spitler ( 2009 ) . _ , title="fig : " ] [ fig_milky ] this section shall be somewhat cursory given r.bowers detailed chapter on clusters and the intracluster medium in this volume . brightest cluster galaxies ( bcgs ) , residing close to or at the centres of large galaxy clusters , have long been recognised as different from less luminous elliptical galaxies : their light profiles appear to have excess flux at large radii ( e.g. oemler 1976 ; carter 1977 ; van den bergh 1977 ; lugger 1984 ; schombert et al . 1986 ) . however , before exploring this phenomenon , it is important to recall that the light profiles of large galaxies have srsic indices @xmath15 greater than 4 ( e.g. caon et al . 1993 ; graham et al . , the srsic model can be approximated by a power - law ( see graham & driver 2005 ) . ] . subsequently , at large radii in big elliptical galaxies , there will be excess flux above that of an @xmath4 model fit to some limited inner radial range . an initially puzzling result from the sloan digital sky survey ( _ sdss _ ; york et al . 2000 ) survey was the lack of light profiles with srsic @xmath576 ( blanton et al . this was however soon resolved when blanton et al . ( 2005b ) , mandelbaum et al . ( 2005 ) and lisker ( 2005 , 2006b , 2007 ) pointed out a serious sky - subtraction problem with the early sdss photometric pipeline ( photo : ivezi et al . the sky - value to be subtracted from each galaxy had been measured to close to the galaxy in question , and because galaxies with large srsic indices possess rather extended light - profiles , their outer galaxy light was actually measured and then subtracted from these galaxies . as a result , the high srsic index light - profiles were erroneously erased and missed from the sdss . this resulted in the @xmath4 model appearing to provide good fits to bright elliptical galaxies , and consequently a series of papers based on structurally biased radii , magnitudes and surface brightnesses . bearing in mind that large elliptical galaxies have high srsic indices , it is important to distinguish between ( i ) an inadequacy of the @xmath4 model to describe what is actually a single @xmath12 profile , and ( ii ) a distinct physical component such as an envelope of diffuse halo light surrounding a central galaxy . early quantitative photometry of cd galaxies ( supergiant d galaxies : e.g. matthews et al . 1964 ; morgan & lesh 1965 ) suggested the presence of an inner @xmath4 spheroid plus an outer exponential corona ( de vaucouleurs 1969 ; de vaucouleurs & de vaucouleurs 1970 ) . one should however question if this outer corona is a distinct entity or not . to answer this in the affirmative , astronomers can point to how the light profiles can display inflections marking the transition from bcg light to intracluster light . gonzalez et al . ( 2005 ) additionally showed that the inflection in the light profiles of many bcgs was also associated with an a change in the ellipticity profile , signalling the switch from bcg light to intracluster light . gonzalez et al . ( 2005 ) and zibetti et al . ( 2005 ) chose to model their bcg sample using an @xmath58 model to describe the inner galaxy plus the outer halo light . however , given that elliptical galaxies are better described by the @xmath12 model , and the desire to measure the actual concentration of halos rather than assign a fixed @xmath4 profile , seigar et al . ( 2007 ) fitted an @xmath59 model to the light profiles of five bcgs . not surprisingly , they found that the @xmath58 model was not the optimal representation . an @xmath12-galaxy plus an exponential - halo model was found to provide the optimal fit in three instances , with an additional galaxy having no halo detected . the associated galaxy - to - halo luminosity ratios can be found there . this re - revelation of an exponential model , rather than an @xmath4 model , describing the halo has since been confirmed by pierini et al . intriguing is that the halo does not trace the nfw - like dark - matter halo density profiles produced in @xmath60cdm simulations ( section [ sec_dark ] ) . stellar halos around non - bcg galaxies have also now been reported to display an exponential radial distribution ( e.g. gadotti 2011 ; tal & van dokkum 2011 ) . while it is common practice , and somewhat helpful , to call elliptical galaxies fainter than about @xmath61 mag by the term `` dwarf elliptical '' ( sandage & binggeli 1984 ) , it should be noted , and this section will reveal , that on all measures they appear to be the low - mass end of a continuous sequence which unifies dwarf and normal / luminous elliptical galaxies . not only is this true in terms of their structural properties ( e.g. binggeli et al . 1984 ; graham & guzmn 2003 ; gavazzi et al . 2005 ; ferrarese et al . 2006a ; ct et al . 2006 , 2007 , 2008 ; misgeld et al . 2008 , 2009 ; janz & lisker 2009 ; graham 2010 ; chen et al . 2010 ; glass et al . 2011 ) but even the degree of kinematically distinct components is similar ( chilingarian 2009 ) . there are many important relations between stellar luminosity , colour , metallicity , age , and dynamics that reveal a continuous and linear behaviour uniting dwarf and giant elliptical galaxies ( e.g. caldwell 1983 ; davies et al . 1983 ; binggeli et al . 1984 ; bothun et al . 1986 ; geha et al . 2003 ; lisker & han 2008 ) . however , kormendy et al . ( 2009 , their section 8) dismiss all of these apparently unifying relations by claiming that they must not be sensitive to different physical processes which they believe have produced a dichotomy between dwarf and ordinary elliptical galaxies . they claim that it is only relations which show an apparent different behavior at the faint and bright end that are sensitive to the formation physics and the remainder are not relevant . this section explains why such non - linear relations are actually a consequence of the ( dismissed ) linear relations , and as such these non - linear relations actually support a continuum between dwarf and ordinary elliptical galaxies . to begin , it should be reiterated that ( dwarf and ordinary ) elliptical galaxies and the bulges of disc galaxies ( section [ sec_bd ] ) do not have structural homology . instead , they have a continuous range of stellar concentrations quantified by the srsic index @xmath15 ( davies et al . 1988 ; caon et al . 1993 ; donofrio et al . 1994 ; young & currie 1994 , 1995 ; andredakis et al . 1995 ) that varies linearly with both stellar mass and central surface brightness ( after correcting for central deficits or excess light ) . a frequently unappreciated consequence of these two linear relations is that relations involving either the effective half - light radius ( @xmath20 ) or the effective surface brightness ( @xmath62 ) , or the mean surface brightness within @xmath20 ( @xmath63 ) , will be non - linear . such curved relations have often been heralded as evidence that two different physical processes must be operating because the relation is not linear and has a different slope at either end . to further complicate matters , sample selection which includes faint and bright elliptical galaxies , but excludes the intermediate - luminosity population , can effectively break such continuously curved relations into two apparently disconnected relations , as can selective colour - coding . there are three distinct types of ( two - parameter ) relations involving the properties of elliptical galaxies : ( i ) linear relations which are taken to reveal the unified nature of des and es ; ( ii ) curved relations revealing a continuity that had in the past been mis - interpreted to indicate that distinct formation process must be operating ; and ( iii ) broken relations which imply that two physical mechanisms are operating . in the following sections we shall learn how the linear relations result in the existence of curved relations when using effective radii and surface brightnesses . we shall also see that the transition in the broken relations occurs at @xmath64 mag and thus has nothing to do with the previously held belief that des and es are two distinct species separated at @xmath61 mag ( wirth & gallager 1984 ; kormendy 1985b , 2009 ) . this section introduces two key relations , from which a third can be derived , involving structural parameters . they are the luminosity - concentration ( @xmath65@xmath15 ) relation and the luminosity-(central density ) ( @xmath65@xmath66 ) relation . we shall use the srsic shape parameter @xmath15 to quantify the central concentration of the radial light profile , and use the projected central surface brightness @xmath66 as a measure of the central density . it is noted that one can expect a certain level of scatter in the @xmath65@xmath15 and @xmath65@xmath66 diagrams because both the central density and the radial concentration of stars that one observes depends upon the chance orientation of one s triaxial galaxy ( binney 1978 ) . this is of course also true for measurements of effective surface brightness , half - light radii , velocity dispersions , etc . to have the best relations , it is important that we use srsic parameters from elliptical galaxies rather than parameters from single - srsic fits to samples of elliptical and lenticular galaxies given the two - component ( 2d - disc plus 3d - bulge ) nature of the lenticular galaxies . given the offset nature of bulges and elliptical galaxies in the @xmath65@xmath15 diagram ( e.g. graham 2001 ; see also mllenhoff & heidt 2001 ) it is also important that bulges not be combined in this section s analysis of elliptical galaxy scaling relations . caldwell ( 1983 ; his figure 6 ) and bothun et al . ( 1986 , their figure 7 ) revealed that , fainter than @xmath64 mag , there is a continuous linear relation between luminosity and central surface brightness . furthermore , binggeli et al . ( 1984 , their figure 11 ) and binggeli & cameron ( 1991 , their figure 9 and 18 ) revealed that , when using the inward extrapolation of king models , the @xmath65@xmath66 relation is continuous and roughly linear from @xmath67 mag . this same general result was also highlighted by jerjen & binggeli ( 1997 ) and graham & guzmn ( 2003 ) when using the inward extrapolation of the srsic model . the benefit of this approach is that one s central surface brightness is not biased by the presence of a depleted core or any additional nuclear components within the host galaxy . figure [ fig_abc]a displays the elliptical galaxy ( @xmath68 , @xmath66 ) data set from graham & guzmn ( 2003 ) fit by the expression @xmath69 the actual central surface brightness of the luminous `` core galaxies '' is shown in figure [ fig_abc]a , rather than the value obtained from the inward extrapolation of their outer srsic profile . as such these `` core galaxies '' were excluded from the fit , but see the discussion in section [ sec_bent ] . as an aside , if the central supermassive black hole mass @xmath52 in elliptical galaxies is directly related to the central stellar density ( see graham & driver 2007 ) , then the connections between @xmath52 and the global galaxy properties , such as total mass and velocity dispersion , may be secondary . the linear relation between luminosity and srsic index , or strictly speaking the logarithm of these quantities , has been shown many times ( e.g.young & currie 1994 ; jerjen & binggeli 1997 ; graham & guzmn 2003 ; ferrarese et al . this continuous relation between magnitude and concentration for elliptical galaxies had of course been recognised before ( e.g.ichikawa et al . 1986 , their figure 11 ) . the following @xmath68@xmath15 expression is shown in figure [ fig_abc]c , again matched to the sample of elliptical galaxies compiled by graham & guzmn ( 2003 ) . @xmath70 graham & guzmn ( 2003 ) excluded two - component lenticular galaxies fit by others with a single - component srsic model . it may be prudent to continue to exclude these galaxies even after a srsic - bulge plus exponential disc fit because the @xmath68@xmath15 relation defined by bulges , at least in spiral galaxies , is different to that defined by elliptical galaxies ( graham 2001 , his figure 14 ) . combining the above two equations provides an expression between central surface brightness and srsic index such that @xmath71 which is shown in figure [ fig_abc]b , where it can be seen to be roughly applicable for values of @xmath72 . this section explains why , as a direct result of the above linear relations which unite dwarf and giant elliptical galaxies expressions involving either the effective half - light radius @xmath20 , the associated effective surface brightness @xmath62 at this radius , or the mean surface brightness @xmath31 enclosed within this radius , are curved . the following analysis is from graham & guzmn ( 2003 ) . given the empirical @xmath68@xmath15 relation ( equation [ eq_m - n ] ) , one knows what the expected value of @xmath15 is for some value of @xmath68 . one can then convert the empirical @xmath68@xmath66 relation ( equation [ eq_m - mu0 ] ) into an @xmath68@xmath62 relation using the exact relation between @xmath66 and @xmath62 which depends only on the value of @xmath15 ( equation [ eq_mue - mu0 ] ) . doing so , one obtains the expression @xmath73 + 22.8 + 1.086b , \label{eq_m - mue}\end{aligned}\ ] ] where @xmath74 and equation [ eq_m - n ] is used to replace @xmath15 in terms of @xmath68 , such that @xmath75 . one can similarly convert the empirical @xmath68@xmath66 relation ( equation [ eq_m - mu0 ] ) into an @xmath68@xmath31 relation using the exact relation between @xmath66 and @xmath31 which also depends only on the value of @xmath15 ( equation [ eq_mue - mue ] ) . doing this , one obtains the expression @xmath76 , \label{eq_m - muav}\ ] ] where again @xmath74 and equation [ eq_m - n ] is used to replace @xmath15 in terms of @xmath68 . these curves can be seen in figure [ fig_magic ] ( adapted from graham & guzmn 2004 ) . binggeli et al . ( 1984 , their figure 8) and capaccioli & caon ( 1991 ) previously showed with empirical data that the @xmath68@xmath77 relation is curved . what was new in graham & guzmn ( 2003 ) was the explanation . in the past , evidence of non - linear relations involving parameters from dwarf and giant elliptical galaxies were heralded as evidence of a dichotomy . coupled with galaxy sample selection that excluded the intermediate population , and therefore resulted in two apparently disconnected relations , acted to further convince some that they were dealing with two classes of object . now that we know how to play this game , one can additionally make predictions for relations involving the effective radius @xmath20 because we know that the luminosity @xmath78 , with @xmath79 . as explained in graham & worley ( 2008 , their section 5.3.1 ) , one can derive the size - luminosity relation such that @xmath80 = \frac{m_b}{10 } + 1.066 + 0.434n + 0.5\log \left [ \frac{b^{2n } } { { \rm e}^{b}n\gamma ( 2n ) } \right ] \label{eq_m - re}\ ] ] for @xmath23 , with @xmath74 and where equation [ eq_m - n ] is used to replace @xmath15 in terms of @xmath68 . this size - luminosity relation for elliptical galaxies is shown in figure [ fig_magic]c along with real galaxy data . binggeli et al . ( 1984 , their figure 7 ; cf . misgeld & hilker 2011 , their figure 1 ) also demonstrated , with empirical data , that the @xmath81 relation for dwarf and giant elliptical galaxies is curved . their diagram , in addition to figure [ fig_magic]c seen here , reveals why studies which only sample bright elliptical galaxies are often contempt to simply fit a straight line ( e.g. kormendy 1977b ) . the explanation for why this happens is of course akin to approximating the earth as flat when one ( forgivably ) does not sample enough of what is actually a curved profile . as graham et al . ( 2006 , their figure 1b ) re - revealed recently , and as reiterated by bernardi et al . ( 2007 ) , a sample of massive elliptical galaxies will have a steeper size - luminosity relation than a sample of ordinary elliptical galaxies , which will in turn have a steeper size - luminosity relation than a sample of dwarf elliptical galaxies because the size - luminosity relation is curved . graham & worley ( 2008 ) explains why the @xmath81 relation given by equation [ eq_m - re ] , based on two linear relations and the functional form of srsic s model , is curved . interestingly , due to the linear relations between magnitude , central surface brightness and the logarithm of the srsic exponent @xmath15 , the use of faint isophotal radii results in what is roughly a linear size - luminosity relation ( e.g. oemler 1976 ; strom & strom 1978 ; forbes et al . 2008 ; van den bergh 2008 ; nair et al . 2011 ) , with the bright - end slope dependent on the adopted isophotal limit or petrosian radius used . the implications of this important observation shall be detailed elsewhere . helping to propagate the belief that dwarf and ordinary elliptical galaxies are distinct species , dabringhausen et al . ( 2008 ) and lisker ( 2009 ) fit a double power - law to their curved size - luminosity relation for dwarf and ordinary elliptical galaxies , thus yielding distinct slopes at the faint and bright end . in addition , the interesting study by janz & lisker ( 2008 ) reported small deviations from the predicted curved relation . however , galaxies that are well described by srsic s function and which follow linear @xmath82@xmath15 and @xmath82@xmath66 relations _ must _ follow a single curved @xmath82@xmath20 relation . the deviations that they found are therefore mirroring a ) the inadequacy of the fitted linear relations to the @xmath82@xmath15 and @xmath82@xmath66 distribution ( a point noted in the caption of figure [ fig_magic ] ) and/or b ) poor fitting srsic models to their sample of elliptical _ and _ disc galaxies . adding uncertainties to the linear relations in section [ sec_line ] , and propagating those through to the predicted @xmath82@xmath20 relation is required before we can claim evidence of significant deviations . however , searching for such second order effects may indeed be interesting given the different types of dwarf galaxies that are emerging ( lisker et al . 2007 ) . one can additionally derive an expression relating @xmath20 and @xmath15 by substituting the magnitude from the empirical @xmath68@xmath15 relation , expressed in terms of @xmath15 ( equation [ eq_m - n ] ) , into the size - luminosity relation ( equation [ eq_m - re ] ) to give @xmath83 & = & 0.434n - 0.364 -0.94\log(n ) \nonumber \\ & & + 0.5\log \left [ \frac{b^{2n } } { { \rm e}^{b}n\gamma ( 2n ) } \right ] . \label{eq_n - re}\end{aligned}\ ] ] while the @xmath68@xmath15 relation is linear , the @xmath20@xmath15 relation is curved , as can be seen in figure [ fig_n_r ] . is over - plotted empirical data . symbols have the same meaning as in figure [ fig_abc ] . ] in passing it is noted that the form of this relation ( equation [ eq_n - re ] ) matches the bulge data from fisher & drory ( 2010 , their figure 13 ) . they interpret the departure of the low-@xmath15 bulges ( @xmath84 ) from the approximately linear relation defined by the high-@xmath15 bulges ( @xmath85 ) to indicate that a different formation process is operating to produce the less concentrated `` pseudobulges '' . however , based upon linear unifying relations that span the artificial @xmath86 divide , we know that this @xmath20@xmath15 relation must be curved . without an understanding of this relation , and other curved relations ( e.g. greene , ho & barth 2008 ) , they have at times been misinterpretted and used to claim the existence of different physical processes ( see section [ sec_pseudo ] for a discussion of pseudobulges : hohl 1975 and references therein ) . it may be worth better defining the behavior of the @xmath20@xmath15 relation at small sizes in figure [ fig_n_r ] . the data from davies et al . ( 1988 ) suggests than when @xmath87 , values of @xmath20 may range from 1 kpc down to 0.2 kpc ( caon et al . 1993 , their figure 5 ) . such a reduction to the flattening of the @xmath20@xmath15 distribution , below @xmath88 , may in part arise from the inclusion of dwarf spheroidal galaxies ( see misgeld & hilker 2011 , their figure 1 ) . as discussed in graham ( 2010 ) , the first two linear relations in figure [ fig_abc ] naturally explain the curved @xmath31@xmath20 relation in figure [ fig_korm ] . from the empirical @xmath66@xmath15 relation ( equation [ eq_mu0-n ] , figure [ fig_abc]b ) , one can convert @xmath66 into @xmath63 using equations [ eq_mue - mu0 ] and [ eq_mue - mue ] . the effective radius @xmath20 is acquired by matching the empirical @xmath68@xmath66 relation ( equation [ eq_m - mu0 ] , figure [ fig_abc]a ) with the absolute magnitude formula @xmath89 ( see graham & driver 2005 , their equation 12 ) . eliminating the absolute magnitude gives the expression @xmath90 in which we already know the value of @xmath15 associated with each value of @xmath63 . this is achieved by ( again ) using the empirical @xmath66@xmath15 relation ( equation [ eq_mu0-n ] ) with equations [ eq_mue - mu0 ] and [ eq_mue - mue ] , such that @xmath91 \label{eq_mu - n}\ ] ] and @xmath74 . equation [ eq_mur ] , obtained from two linear relations involving srsic parameters , is a curved relation that is shown in figure [ fig_korm ] . overplotted this predicted relation are data points from real galaxies . , the relation between the effective radius ( @xmath20 ) and the mean surface brightness within this radius ( @xmath63 ) is highly curved for elliptical galaxies . the dashed line shows the @xmath92 mag limit for the virgo cluster . ] for those who may have the srsic parameter set ( @xmath93 ) , one can use equation [ eq_mue - mue ] to convert @xmath62 into @xmath94 if one wishes to compare with the relation given by equation [ eq_mur ] . for those who may have the parameter set ( @xmath95 ) , perhaps obtained with no recourse to the srsic model , equation [ eq_mur ] can easily be adjusted using equation [ eq_mue - mue ] to give a relation between @xmath20 and @xmath62 such that @xmath96 where the value of @xmath15 associated with the value of @xmath62 is given by @xmath97 to summarise , due to the linear relations in figure [ fig_abc ] which connect dwarf and ordinary elliptical galaxies across the alleged divide at @xmath61 mag ( kormendy 1985 ) , or at @xmath86 ( kormendy & kennicutt 2004 ) , coupled with the smoothly varying change in light profile shape as a function of absolute magnitude , the @xmath63-@xmath20 and @xmath98-@xmath20 relations are expected to be curved ( figure [ fig_korm ] ) , as previously shown with empirical data by , for example , capaccioli & caon ( 1991 ) . this also explains why the fitting of a linear relation to ( @xmath99 ) data by hoessel et al . ( 1987 ) resulted in slopes that depended on their galaxy sample magnitude . the kormendy relation is a tangent to the bright arm of what is actually a curved distribution defined by the relation given by equation [ eq_mur ] that is taken from graham ( 2010 ) . the apparent deviant nature of the dwarf elliptical galaxies from the approximately linear section of the bright - end of the @xmath63-@xmath20 distribution does not necessitate that a different physical process be operating . moreover , as noted by graham & guzmn ( 2004 ) and graham ( 2005 ) , galaxies which appear to branch off at the faint end of the fundamental plane ( djorgovski & davis 1987 ) the flat portion at the bright end of a curved hypersurface also need not have formed from different physical mechanisms . simulations that assume or reproduce a linear @xmath100-@xmath20 or @xmath62-@xmath20 relation , across too great a magnitude range , have failed to mimic the continuous curved distribution defined by real elliptical galaxies . the same remark is true for simulations of the ` fundamental plane ' . while the relation between a galaxy s absolute magnitude and _ extrapolated _ central surface brightness is remarkably linear ( section [ sec_m - mu0 ] ) , there is a clear break in this relation when using the _ actual _ central surface brightness at the luminous end of the distribution ( figure [ fig_abc]a ) . this departure from the @xmath68-@xmath66 relation by elliptical galaxies brighter than @xmath64 mag ( @xmath101-@xmath102 ) was addressed by graham & guzmn ( 2003 ) in terms of partially depleted cores relative to the outer srsic profile ( see also graham 2004 ; trujillo et al . 2004 ; merritt & milosavljevi 2005 ; ferrarese et al . 2006a ; ct et al . this transition has nothing to do with the alleged divide between dwarf and giant elliptical galaxies at around @xmath103 mag , but is instead directly related with the srsic versus core - srsic transition at around @xmath104 mag . as noted in section [ sec_def ] , such partially depleted cores in luminous core - srsic galaxies are thought to have formed from dry , dissipationless galaxy merger events involving the central coalescence of supermassive black holes ( begelman , blandford , & rees 1980 ; ebisuzaki , makino , & okumura 1991 ; but see footnote [ foot_dry ] ) and resulted in trujillo et al . ( 2004 ) advocating a `` new elliptical galaxy paradigm '' based on the presence of a central stellar deficit versus either none or an excess of light , an approach embraced by ferrarese et al . ( 2006a ) , ct et al . ( 2007 ) and others . further evidence for a division at @xmath104 mag comes from the tendency for the brighter galaxies to be anisotropic , pressure supported elliptical galaxies having boxy isophotes , while the less luminous early - type galaxies may have discy isophotes and often contain a rotating disc ( e.g. carter 1978 , 1987 ; davies et al . 1983 ; bender et al . 1988 ; peletier et al . 1990 ; jaffe et al . 1994 ) . core galaxies also tend to be more radio loud and have a greater soft x - ray flux ( e.g. ellis & osullivan 2006 ; pellegrini 2010 ; richings , uttley & krding 2011 , and references therein ) . it was , in part , from a diagram of central surface brightness versus magnitude that led kormendy ( 1985b , his figure 3 ) to advocate a separation of dwarf and normal elliptical galaxies at @xmath61 mag . however , as noted by graham & guzmn ( 2003 ) , his sample was missing the bridging population near @xmath105 mag . excluding galaxies of this magnitude from figure [ fig_abc]a would also result in two apparently disjoint relations nearly at right angles to each other . it is therefore easy to understand how one may quickly reach the wrong conclusion from an incomplete diagram . although , strom & strom ( 1978 , their figure 8 ; see also binggeli et al.1984 ) had already revealed that a linear relation exists between magnitude and central surface brightness from @xmath106 mag , spanning the magnitude gap in kormendy ( 1985b ) . nonetheless , faber & lin ( 1983 ) had just observed that three of their six dwarf elliptical galaxies had near - exponential light profiles , leading them to speculate that des are more closely related to `` exponential systems '' than ( tidally truncated ) elliptical galaxies , and wirth & gallagher ( 1984 , see also michard 1979 ) had also just advocated a division between exponential - dwarf and @xmath4-giant elliptical galaxies . to further confound matters , kormendy ( 1985b ) had the slope wrong for the distribution of dwarf elliptical galaxies in his @xmath82@xmath66 diagram , which had two consequences . first , the bright end of his dwarf elliptical galaxy distribution did not point towards the faint - end of his luminous elliptical galaxy distribution , and thus there was no suggestion of a connection . this misrepresentation is unfortunately still propagated today ( e.g. tolstoy et al . 2009 , their figure 1 ) , although kormendy et al . ( 2009 ) have now corrected this . second , two points representing flattened disc galaxies were added at the bright end of the mis - aligned dwarf elliptical galaxy sequence by kormendy ( 1985b ) , implying a connection between dwarf elliptical galaxies and disc galaxies rather than ordinary elliptical galaxies . a decade later , the astronomy and astrophysics review paper by ferguson & binggeli ( 1994 ; their figure 3 ) had a big question mark as to `` how '' and indeed `` if '' dwarf and ordinary elliptical galaxies might connect in this diagram . when galaxies spanning the gap in kormendy s ( 1985b ) analysis were included by faber et al . ( 1997 , their figure 4c ) , and shown to follow a trend consistent with the relation from strom & strom ( 1978 ) and binggeli et al.(1984 ) , in which the central surface brightness became fainter with decreasing galaxy luminosity faber et al . ( 1997 ) suggested that this behavior in their data was spurious and due to limited resolution such was the belief in a discontinuity separating dwarf and ordinary elliptical galaxies . at the same time , jerjen & binggeli ( 1997 ) argued exactly the opposite , suggesting that it was instead the `` core '' galaxies which had been displaced in the @xmath82@xmath66 diagram from a linear @xmath82@xmath66 relation . rather than wrong central surface brightness measurements for the faint ( non - dwarf ) elliptical galaxies . as graham & guzmn ( 2003 ) and graham ( 2004 ) later explained , in terms of a @xmath90.1 percent central mass deficit relative to the outer srsic profile in galaxies brighter than @xmath64 mag , jerjen & binggeli ( 1997 ) were right , supporting the views expressed by binggeli et al . ( 1984 ) on a continuity between dwarf elliptical and ordinary elliptical galaxies across the alleged divide at @xmath107 mag . additional support for the dry merging scenario at the high - mass end is the flattening of the colour - magnitude relation above 0.5 - 1@xmath108 . while low luminosity , low srsic index , elliptical galaxies are bluer than bright elliptical galaxies ( e.g. de vaucouleurs 1961 ; webb 1964 ; sandage 1972 ; caldwell & bothun 1987 ) , the brightest galaxies have the same colour as each other . this flattening in the colour - magnitude relation was noted by tremonti et al.(2004 ) and is evident in baldry et al . ( 2004 , their figure 9 ) , ferrarese et al . ( 2006a , their figure 123 ) , boselli et al . ( 2008 , their figure 7 ) and even metcalfe , godwin & peach ( 1994 ) . these observations help alleviate past tension with semi - analytic models that had predicted a relatively flat colour - magnitude relation for bright elliptical galaxies ( e.g. cole et al . previously , based on what was thought to be a linear colour - magnitude relation , bernardi et al . ( 2007 ) had written that `` if bcgs formed from dry mergers , then bcg progenitors must have been red for their magnitudes , suggesting that they hosted older stellar populations than is typical for their luminosities '' . however , the flattening in the colour - magnitude relation has since been recognised in yet more data sets ( e.g. skelton , bell & somerville 2009 ; jimnez et al . 2011 ) although it should perhaps be noted that skelton et al . reported the transition at @xmath109 mag , i.e. @xmath91 mag fainter . in passing it is noted that the relation between luminosity and supermassive black hole mass ( marconi & hunt 2003 ; mclure & dunlop 2004 ) was found to be , after several refinements by graham ( 2007 ) , a linear one - to - one relation for black hole masses predominantly greater than @xmath110 consistent with the concept of dry galaxy merging at this high - mass end . from a sample of 13 early - type galaxies , plus one spiral galaxy , minkowski ( 1962 ) noted that a `` correlation between velocity dispersion and [ luminosity ] exists , but it is poor '' . he wrote that `` it seems important to extend the observations to more objects , especially at low and medium absolute magnitudes '' . this was done by morton & chevalier ( 1973 ) who noted the same `` continuous distribution of dispersions from 60 km / s for m32 to 490 km / s for m87 '' but also did not attempt to quantify this trend . it was faber & jackson ( 1976 ) who , with improved data and a larger sample of 25 galaxies , were the first to quantify minkowski s relation and discovered that @xmath111 for their data set . this result has proved extremely popular and is known as the faber - jackson relation . not long after this , schechter ( 1980 ) and malumuth & kirshner ( 1981 ) revealed that the luminous elliptical galaxies followed a relation with an exponent of @xmath95 rather than 4 . at the same time , tonry ( 1981 ) revealed that expanding the sample to include more faint elliptical galaxies results in an exponent of @xmath93 . this led binney ( 1982 ) to write that `` probably the correlation can not be adequately fitted by a single power law over the full range of absolute magnitudes '' and farouki et al . ( 1983 ) wrote that `` the data suggests the presence of curvature in the @xmath112 relation '' . davies et al . ( 1983 ) , and later held et al . ( 1992 ) , revealed that the dwarf elliptical galaxies followed a relation with an exponent of @xmath92 , which explains why tonry ( 1981 ) had found a slope of @xmath93 when including dwarf and ordinary elliptical galaxies . the relation found by davies et al . , with a slope of @xmath92 , has recently been observed by de rijcke et al . ( 2005 ) and the curved or possibly broken @xmath112 distribution has been interpreted by matkovi & guzmn ( 2005 ) as a change in slope at @xmath113 b - mag ( see also evstigneeva et al.2007 ) in agreement with davies et al . ( 1983 ) . in spite of all the above work , there is a huge body of literature today which appears unaware that the @xmath112 relation is curved or broken . simulations of galaxies which succeed in producing the linear faber - jackson relation , @xmath114 , have actually failed to produce the full distribution of dynamics seen in real elliptical galaxies as a function of magnitude . due to contractual agreements with springer publishing , the full version of this paper can not be posted to astro - ph . all of section [ sec_discs ] , plus the full version of section [ l - mu0 ] , will be available from the springer publication . we have reviewed the progress over the last century in modelling the distribution of stars in elliptical galaxies , plus the bulges of lenticular and spiral galaxies and their surrounding discs . a number of nearly forgotten or poorly recognised references have been identified . the universality , or at least versatility , of srsic s @xmath12 model to describe bulges ( section [ sec_bd ] ) and elliptical galaxies ( section [ sec_body ] ) extends to the stellar halos of cd galaxies ( section [ sec_cd ] ) and simulated dark matter halos ( section [ sec_dark ] ) . dwarf and ordinary elliptical galaxies were shown in section [ sec_line ] to be united by two continuous linear relations between absolute magnitude and a ) the stellar concentration quantified through srsic s ( 1963 ) @xmath12 shape parameter ( section [ sec_m - n ] ) , and b ) the central surface brightness , which is also related to the central density ( section [ sec_m - mu0 ] ) . as discussed in section [ sec_bent ] , a break in the latter relation at @xmath115 mag signals the onset of partially depleted cores relative to the outer srsic profile in luminous elliptical galaxies . additional scaling relations are also noted to show a change in character at this magnitude , which may denote the onset of dry galaxy merging . the identification of depleted galaxy cores and excess nuclear light relative to the outer srsic profile was discussed in sections [ sec_cs ] , [ sec_def ] and [ sec_enuc ] . after accounting for these features , it was revealed how the above two linear relations result in curved scaling relations involving effective half light radii and effective surface brightness ( section [ sec_curved ] ) . specifically , the @xmath82@xmath20 , @xmath82@xmath62 , @xmath82@xmath63 , @xmath62@xmath116 , @xmath63@xmath20 and @xmath15@xmath20 relations are non - linear . these continuous curved relations exist because elliptical galaxies do not have a universal profile shape , such as an @xmath4 profile , but instead a range of profile shapes that vary smoothly with absolute magnitude . without an appreciation of the origin of these curved relations , they had in the past been heralded as evidence for a dichotomy between faint and bright elliptical galaxies . numerical simulations and semi - numerical models which try to reproduce the full elliptical galaxy sequence must be able to reproduce these non - linear relations . this will likely require physical processes which work in tandem , albeit to different degrees over different mass ranges , to produce a continuum of galaxy properties that scale with mass while adhering to the linear @xmath82@xmath15 and @xmath82@xmath66 relations ( subject to core - formation ) . the upcoming 2.6 m vlt survey telescope ( vst , arnaboldi et al . 1998 ; capaccioli et al . 2005 ) , plus the 4@xmath1171.8 m pan - starrs array ( kaiser et al . 2002 ) , the 4 m visible and infrared survey telescope for astronomy ( vista , emerson et al . 2004 ) and the 8.4 m large synoptic survey telescope ( lsst , tyson 2001 ) are expected to deliver sub - arcsecond , deep and wide field - of - view imaging covering thousands of resolvable galaxies . by pushing down the luminosity function into the dwarf galaxy regime , and through the application of improved galaxy parameterisation methods which allow for structural non - homology and the 2- or 3-component nature of disc galaxies , _ both _ statistical and systematic errors will be reduced . this will undoubtedly provide improved constraints on galaxy scaling relations and , in turn , a fuller understanding of galaxy evolution . this research was supported under the australian research council s discovery projects funding scheme ( dp110103509 ) . this research has made use of nasa s astrophysics data system ( ads ) bibliographic services and the nasa / ipac extragalactic database ( ned ) . aceves , h. , velzquez , h. , cruz , f. 2006 , mnras , 373 , 632 ade p.a.r . , et al . 2011 , ( arxiv:1101.2045 ) aguerri , j.a.l . , balcells , m. , peletier , r.f . 2001 , a&a , 367 , 428 aguerri , j.a.l . , beckman , j.e . , prieto , m. 1998 , aj , 116 , 2136 allen p.d . , driver s.p . , graham a. , cameron e. , liske j. , de propris r. 2006 , mnras , 371 , 2 andredakis , y.c . , sanders , r.h . 1994 , mnras , 267 , 283 andredakis , y.c . , peletier , r.f . , balcells , m. 1995 , mnras , 275 , 874 arnaboldi m. , capaccioli m. , mancini d. , rafanelli p. , scaramella r. , sedmak g. , vettolani g.p . 1998 , the messenger , 93 , 30 athanassoula e. , et al . 1990 , mnras , 245 , 130 athanassoula e. 1992 , mnras , 259 , 345 athanassoula e. 2005 , mnras , 358 , 1477 auger , m.w . , treu , t. , bolton , a.s . , gavazzi , r. , koopmans , l.v.e . , marshall , p.j . , moustakas , l.a . , burles , s. 2010 , apj , 724 , 511 avila - reese , v. , firmani , c. , hernndez , x. 1998 , apj , 505 , 37 avila - reese v. , firmani c. , klypin a. , kravtsov a.v . 1999 , mnras , 310 , 527 babcock h.w . 1938 , pasp , 50 , 174 babcock h.w . 1939 , lick observatory bulletin , 19 , 41 babusiaux , c. , et al . 2010 , a&a , 519 , a77 balcells , m. , graham , a.w . , dominguez - palmero , l. , peletier , r.f . 2003 , apj , 582 , l79 balcells , m. , graham , a.w . , peletier , r.f . 2007 , apj , 665 , 1084 baldry i.k . , glazebrook k. , brinkmann j. , ivezi . , lupton r.h . , nichol r.c . , szalay a.s . , 2004 , apj , 600 , 681 barazza , f.d . , jogee , s. , marinova , i. 2008 , apj , 675 , 1194 bardeen , j.m . 1975 , iau symp . , 69 , 297 barnes j.h . 1988 , apj , 331 , 699 barth , a.j . , strigari , l.e . , bentz , m.c . , greene , j.e . , ho , l.c . 2009 , apj , 690 , 1031 barton , i.j . , thompson , l.a . 1997 , aj , 114 , 655 bassino , l.p . , muzzio , j.c . , rabolli , m. 1994 , apj , 431 , 634 begelman , m.c . , blandford , r.d . , rees , m.j . 1980 , nature , 287 , 307 bekki k. , 2010 , mnras , 401 , l58 bekki , k. , couch , w. j. , drinkwater , m. j. 2001a , apj , 552 , l105 bekki , k. , couch , w. j. , drinkwater , m. j. , gregg , m. d. 2001b , apj , 557 , l39 bell , e.f . , de jong , r.s . 2000 , mnras , 312 , 497 bender r. , doebereiner s. , moellenhoff c. 1988 , a&as , 74 , 385 bernardi , m. , et al . 2007 , aj , 133 , 1741 bertin g. , ciotti , l. , del principe , m. 2002 , a&a , 386 , 149 bertin g. , stiavelli m. 1993 , rep . phys . , 56 , 493 bertola , f. , capaccioli , m. 1977 , apj , 211 , 697 bianchi s. , 2008 , a&a , 490 , 461 binggeli , b. , cameron , l.m . 1991 , a&a , 252 , 27 binggeli , b. , cameron , l.m . 1993 , a&as , 98 , 297 binggeli , b. , jerjen , h. 1998 , a&a , 333 , 17 binggeli , b. , sandage , a. , tarenghi , m. 1984 , aj , 89 , 64 binggeli , b. , sandage , a. , tammann g.a . 1985 , aj , 90 , 1681 binney , j. 1978 , mnras , 183 , 501 binney , j. 1982 , ara&a , 20 , 399 binney , j. , mamon , g.a . 1982 , mnras , 200 , 361 birnboim , y. , dekel , a. 2003 , mnras , 345 , 349 bland - hawthorn , j. , vlaji , m. , freeman , k.c . , draine , b.t . 2005 , apj , 629 , 239 blanton , m. , et al . 2005a , aj , 129 , 2562 blanton , m.r . , et al . , 2005b , apj , 631 , 208 block d.l . 2004 , in penetrating bars through masks of cosmic dust . editors : d. block , i. puerari , k.c . freeman , r. groess , e.k . block . kluwer publishers , astrophysics and space science library ( assl ) , 319 , 15 bker , t. , stanek , r. , van der marel , r.p . 2003 , aj , 125 , 1073 bolton , a.s . , burles , s. , koopmans , l.v.e . , treu , t. , moustakas , l.a . 2006 , apj , 638 , 703 bothun , g.d . , gregg , m.d . 1990 , apj , 350 , 73 bothun , g.d . , mould , j.r . , caldwell , n. , macgillivray , h.t . 1986 , aj , 92 , 1007 bower , r.g . , benson , a.j . , malbon , r. , helly , j.c . , frenk , c.s . , baugh , c.m . , cole , s. , lacey , c.g . 2006 , mnras , 370 , 645 boylan - kolchin , m. , ma , c .- p . , quataert , e. 2004 , apj , 613 , l37 bregman , j. , snider , b. , grego , l. , cox , c. 1998 , apj , 499 , 670 brodie , j.p . , strader , j. 2006 , ara&a , 44 , 193 brook c.b . , et al . 2011 , mnras , 415 , 1051 brown , r.j.n . , et al . 2003 , mnras , 341 , 747 buote d.a . , humphrey p.j . 2011 , in hot interstellar matter in elliptical galaxies , eds . kim and s. pellegrini , astrophysics and space science library ( assl ) , springer ( arxiv:1104.0012 ) byun , y .- i . , et al . 1996 , aj , 111 , 1889 caldwell , n. 1983 , aj , 88 , 804 caldwell , n. , bothun , g.d . 1987 , aj , 94 , 1126 cameron , e. , driver , s.p . , graham , a.w . , liske , j. 2009 , apj , 699 , 105 caon , n. , capaccioli , m. , rampazzo , r. 1990 , a&as , 86 , 429 caon , n. , capaccioli , m. , donofrio , m. 1993 , mnras , 265 , 1013 caon , n. , capaccioli , m. , donofrio , m. 1994 , a&as , 106 , 199 capaccioli , m. 1985 , in new aspects of galaxy photometry , ed . nieto , springer - verlag , p.53 capaccioli , m. 1987 , in structure and dynamics of elliptical galaxies , iau symp . 127 , reidel , dordrecht , p.47 capaccioli , m. 1989 , capaccioli , m. 1989 , in the world of galaxies , ed . h. g. corwin , l. bottinelli ( berlin : springer - verlag ) , 208 capaccioli , m. , caon , n. 1991 , mnras , 248 , 523 capaccioli m. , mancini d. , sedmak g. 2005 , the messenger , 120 , 10 cardone , v.f . , piedipalumbo , e. , tortora , c. 2005 , mnras , 358 , 1325 carlberg , r.g . , lake , g. , norman , c.a . 1986 , apj , 300 , l1 carollo , c.m . , scarlata , c. , stiavelli , m. , wyse , r.f.g . , mayer , l. 2007 , apj , 658 , 960 carollo , c.m . , stiavelli , m. , mack , j. 1998 , aj , 116 , 68 carter , d. 1977 , mnras , 178 , 137 carter , d. 1978 , mnras , 182 , 797 carter , d. 1987 , apj , 312 , 514 carter d. , pass s. , kennedy j. , karick a.m. , smith r.j . 2011 , mnras , 414 , 3410 cecil , g. , rose , j.a . 2007 , reports on progress in physics , 70 , 1177 cellone , s.a . , forte , j.c . , geisler , d. 1994 , apjs , 93 , 397 chang c .- k . , ko c .- m . , peng t .- h . 2011 , apj , in press ( arxiv:1107.3884 ) chen , c .- w . , ct , p. , west , a.a . , peng , e.w . , ferrarese , l. 2010 , apjs , 191 , 1 chilingarian , i. 2009 , mnras , 394 , 1229 chilingarian , i. , cayatte , v. , revaz , y. , dodonov , s. , durand , d. , durret , f. , micol , a. , slezak , e. 2009 , science , 326 , 1379 ciotti l. 1991 , a&a , 249 , 99 ciotti l. 2009 , nuovo cimento rivista serie , 32 , 1 ciotti l. , bertin g. 1999 , a&a , 352 , 447 clemens m.s . 2010 , a&a , 518 , l50 cole , s. 1991 , apj , 367 , 45 cole , s. , lacey , c.g . , baugh , c.m . , frenk , c.s . 2000 , mnras , 319 , 168 combes , f. , debbasch , f. , friedli , d. , pfenniger , d. 1990 , a&a , 233 , 82 combes f. , sanders r.h . , 1981 , a&a , 96 , 164 conselice c.j . , bluck a.f.l . , ravindranath s. , mortlock a. , koekemoer a. , buitrago f. , grtbauch r. , penny s. 2011 , mnras , submitted ( arxiv:1105.2522 ) ct , p. , et al . 2006 , apj , 165 , 57 ct , p. , et al . 2007 , apj , 671 , 1456 ct , p. , et al . 2008 , in iau symp . 246 , p.377 crane , p. , et al . 1993 , aj , 106 , 1371 croton d.j . , et al . 2006 , mnras , 365 , 11 curtis , h.d . 1918 , lick obs . pub . 13 dabringhausen , j. , hilker , m. , kroupa , p. 2008 , mnras , 386 , 864 daddi , e. , et al . 2005 , apj , 626 , 680 damjanov , i. , et al . 2009 , apj , 695 , 101 davidge t.j . , beck t.l . , mcgregor p.j . , 2008 , apj , 677 , 238 davies , r.l . , efstathiou , g. , fall , s.m . , illingworth , g. , schechter , p.l . 1983 , apj , 266 , 41 davies , j.i . , phillipps , s. , cawson , m.g.m . , disney , m.j . , kibblewhite , e.j . 1988 , mnras , 232 , 239 de grijs , r. , peletier , r.f . , van der kruit , p.c . 1997 , a&a , 327 , 966 de jong , r.s . , van der kruit , p.c . 1994 , a&as , 106 , 451 de jong r. 1996 , a&as , 118 , 557 de lucia , g. , springel , v. , white , s.d.m . , croton , d. , kauffmann , g. 2006 , mnras , 366 , 499 de rijcke s. , michielsen d. , dejonghe h. , zeilinger w.w . , hau g.k.t . , 2005 , a&a , 438 , 491 del popolo , a. 2010 , mnras , 408 , 1808 de sitter , w. 1917 , mnras , 78 , 3 de vaucouleurs , g. 1948 , annales dastrophysique , 11 , 247 de vaucouleurs , g. 1953 , mnras , 113 , 134 de vaucouleurs , g. 1957 , aj , 62 , 69 de vaucouleurs , g. 1958 , apj , 128 , 465 de vaucouleurs , g. 1959 , in handbuch der physik , ed . s.flgge , springer , berlin , p.311 de vaucouleurs , g. 1961 , apjs , 5 , 233 de vaucouleurs , g. 1963 , apjs , 8 , 31 de vaucouleurs , g. 1969 , apjl , 4 , l17 de vaucouleurs , g. 1974 , iaus , 58 , 1 de vaucouleurs , g. , capaccioli , m. 1979 , apjs , 40 , 669 de vaucouleurs , g. , de vaucouleurs , a. 1970 , apjl , 5 , l219 de vaucouleurs , g. , de vaucouleurs , a. , corwin , h.g . , jr . , buta , r.j . , paturel , g. , fouque , p. 1991 , third reference catalog , springer - verlag berlin heidelberg new york de zeeuw p.t . , franx m. 1991 , ara&a , 29 , 239 devereux , n.a . , young , j.s . 1990 , apj , 359 , 42 di matteo , t. , springel , v. , hernquist , l. 2005 , nature , 433 , 604 djorgovski , s. , davis m. 1987 , apj , 313 , 59 domnguez - tenreiro , r. , tissera , p.b . , siz , a. 1998 , ap&ss , 263 , 35 dong , x.y . , de robertis , m.m . 2006 , aj , 131 , 1236 donghia , e. , burkert , a. , murante , g. , khochfar , s. 2006 , mnras , 372 , 1525 donofrio , m. , capaccioli , m. , caon , n. 1994 , mnras , 271 , 523 donofrio , m. 2001 , mnras , 326 , 1517 draine b.t . 2003 , ara&a , 41 , 241 dreyer , j.l.e . 1888 , memoirs of the royal astronomical society , 49 , 1 dreyer , j.l.e . 1895 , mem . r. astron . soc . , 51 , 185 dreyer , j.l.e . 1908 , mem . r. astron . , 59 , 105 drinkwater m. j. , jones j. b. , gregg m. d. , phillipps s. , 2000 , pub . aust . , 17 , 227 driver s.p . 2010 , amer . 1240 , 17 ( arxiv:1001.4054 ) driver s.p . , allen p.d . , liske j. , graham a. 2007a , mnras , 379 , 1022 driver s.p . , et al . 2007b , apj , 657 , l85 driver , s.p . , popescu , c.c . , tuffs , r.j . , graham , a.w . , liske , j. , baldry , i. 2008 , apj , 678 , l101 duncan , m.j . , wheeler , j.c . 1980 , apj , 237 , l27 ebisuzaki , t. , makino , j. , okumura , s.k . 1991 , nature , 354 , 212 einasto , j. 1965 , trudy inst . astrofiz . alma - ata , 5 , 87 ellis , s.c . , osullivan , e. 2006 , mnras , 367 , 627 elmegreen , b.g . , elmegreen , d.m . , chromey , f.r . , hasselbacher , d.a . , bissell , b.a . 1996 , aj , 111 , 2233 elson r.a.w . , fall s.m . , freeman k.c . 1987 , apj , 323 , 54 emerson , j.p . , sutherland , w.j . , mcpherson , a.m. , craig , s.c . , dalton , g.b . , ward , a.k . 2004 , the messenger , 117 , 27 emsellem e. , et al . 2007 , mnras , 379 , 401 emsellem e. , et al . 2011 , mnras , in press ( arxiv:1102.4444 ) erwin p. , beckman j.e . , pohlen m. 2005 , apj , 626 , l81 erwin , p. , beltrn , j.c.v . , graham , a.w . , beckman , j.e . 2003 , apj , 597 , 929 eskridge p.b . , et al . 2000 , aj , 119 , 536 evstigneeva , e.a . , gregg , m.d . , drinkwater , m.j . , hilker , m. 2007 , aj , 133 , 1722 faber , s.m . , jackson , r.e . 1976 , apj , 204 , 668 faber , s.m . , lin , d.m.c . 1983 , apj , 266 , l17 faber , s.m . , et al . 1997 , aj , 114 , 1771 farouki , r.t . , shapiro , s.l . , duncan , m.j . 1983 , apj , 265 , 597 ferguson , h.c . , binggeli , b. 1994 , a&arv , 6 , 67 ferrarese , l. , et al . 2006a , apjs , 164 , 334 ferrarese , l. , et al . 2006b , ( arxiv : astro - ph/0612139 ) ferrarese , l. , ford , h. 2005 , space science reviews , 116 , 523 ferrarese , l. , van den bosch , f.c . , ford , h.c . , jaffe , w. , oconnell , r.w . 1994 , aj , 108 , 1598 ferrari , f. , dottori , h. , caon , n. , nobrega , a. , pavani , d.b . 2004 , mnras , 347 , 824 ferrers n.m . 1877 , quart . j. pure appl . , 14 , 1 filippenko , a.v . , ho , l.c . 2003 , apj , 588 , l13 fisher , d.b . , drory , n. 2008 , aj , 136 , 773 fisher , d.b . , drory , n. 2010 , apj , 716 , 942 fisher d. , franx m. , illingworth g. 1996 , apj , 459 , 110 fontanot , f. , de lucia , g. , wilman , d. and monaco , p. 2011 , mnras , in press ( arxiv:1102.3188 ) forbes , d.a . , franx , m. , illingworth , g.d . 1994 , apj , 428 , l49 forbes , d.a . , kroupa , p. 2011 , pasa , 28 , 77 forbes , d.a . , lasky , p. , graham a.w . , spitler , l. 2008 , mnras , 389 , 1924 frankston m. , schild r. 1976 , aj , 81 , 500 freeman , k.c . 1966 , mnras , 133 , 47 freeman , k.c . 1970 , apj , 160 , 811 freeman , k.c . 1990 , in dynamics and interactions of galaxies , ed . r. wielen ( berlin : springer ) , 36 freeman , k.c . 2004 , in penetrating bars through masks of cosmic dust - the hubble tuning fork strikes a new note . editors : d. block , i. puerari , k.c . freeman , r. groess , e.k . block . kluwer publishers , astrophysics and space science library ( assl ) , 319 , 639 friedmann , a. 1922 , uner die krmmung des raumes , zeitschrift fr physik , x , 377 - 86 , p.377 gadotti , d.a . , dos anjos , s. 2001 , aj , 122 , 1298 gadotti , d. a. 2008 , mnras , 384 , 420 gadotti d.a . , 2009 , mnras , 393 , 1531 gadotti d.a . 2011 , mnras , submitted ( arxiv:1101.2900 ) gavazzi , g. , donati , a. , cucciati , o. , sabatini , s. , boselli , a. , davies , j. , zibetti , s. 2005 , a&a , 430 , 411 gavazzi , r. , treu , t. , rhodes , j.d . , koopmans , l.v.e . , bolton , a.s . , burles , s. , massey , r.j . , moustakas , l.a . 2007 , apj , 667 , 176 gebhardt , k. , et al . 2001 , aj , 122 , 2469 geha m. , guhathakurta p , van der marel r.p . 2003 , aj , 126 , 1794 genzel r. , thatte n. , krabbe a. , kroker h. , tacconi - garman l.e . 1996 , apj , 472 , 153 gerbal , d. , lima neto , g.b . , mrquez , i. , verhagen , h. 1997 , mnras , 285 , l41 gil de paz , a. , et al . 2007 , apjs , 173 , 185 gilmore , g. , reid , n. 1983 , mnras , 202 , 1025 gilmore , g. , wilkinson , m.i . , wyse , r.f.g . , kleyna , j.t . , koch , a. , evans , n.w . , grebel , e.k . 2007 , apj , 663 , 948 glass , l. , et al . 2011 , apj , 726 , 31 gliozzi , m. , satyapal , s. , eracleous , m. , titarchuk , l. , cheung , c.c . 2009 , apj , 700 , 1759 gonzalez a.h . , zabludoff a.i . , zaritsky d. , 2005 , apj , 618 , 195 gonzlez - garca , a.c . , oorbe , j. , domnguez - tenreiro , r. , gmez - flechoso , m. . 2009 , a&a , 497 , 35 goudfrooij p. , de jong t. , 1995 , a&a , 298 , 784 goudfrooij p. , gorgas j. , jablonka p. , 1999 , ap&ss , 269 , 109 graham a.w . 2001 , aj , 121 , 820 graham a.w . 2002 , apj , 568 , l13 graham a.w . 2004 , apj , 613 , l33 graham a.w . 2005 , in near - field cosmology with dwarf elliptical galaxies , ed . h.jerjen , b.binggeli , iau colloc . 198 , 303 graham a.w . 2007 , mnras , 379 , 711 graham a.w . 2010 , in `` a universe of dwarf galaxies '' , ed . f.prugniel ( arxiv:1009.5002 ) graham a.w . 2011 , arxiv:1103.0525 graham a.w . , colless m.m . , busarello g. , zaggia s. , longo g. , 1998 , a&as , 133 , 325 graham a.w . , erwin , p. , trujillo , i. , asensio ramos , a. 2003 , aj , 125 , 2951 graham a.w . , driver s.p . 2005 , pasa , 22(2 ) , 118 graham a.w . , driver s.p . 2007 , apj , 655 , 77 graham a.w . , driver , s.p . , petrosian , v. , conselice , c.j . , bershady , m.a . , crawford , s.m . , goto , t. 2005 , aj , 130 , 1535 graham a.w . , guzmn r. 2003 , aj , 125 , 2936 graham a.w . , guzmn r. 2004 , in penetrating bars through masks of cosmic dust , ed . d. l. block et al . ( dordrecht : kluwer academic publishers ) , 723 graham a.w . , lauer , t.r . , colless , m.m . , postman , m. 1996 , apj , 465 , 534 graham a.w . , merritt , d. , moore , b. , diemand , j. , terzi , b. 2006 , aj , 132 , 2701 graham a.w . , spitler l.r . 2009 , mnras , 397 , p.2148 graham a.w . , trujillo , i. , caon , n. 2001b , aj , 122 , 1707 graham a.w . , worley , c.c . 2008 , mnras , 388 , 1708 grant , n.i . , kuipers , j.a . , phillipps , s. 2005 , mnras , 363 , 1019 grebel , e.k . 2001 , apsss , 277 , 231 greene , j.e . , ho , l.c . , barth , a.j . 2008 , apj , 688 , 159 grillmair , c.j . , faber , s.m . , lauer , t.r . , baum , w.a . , lynds , r.c . , oneil , e.j . , jr . , shaya , e.j . 1994 , aj , 108 , 102 guo , q. , et al . 2011 , mnras , 413 , 101 haehnelt , m.g . , kauffmann , h. 2002 , mnras , 336 , l61 haegan , m. , et al . 2005 , apj , 627 , 203 harris w.e . , van den bergh s. 1981 , aj , 86 , 1627 hawarden , t.g . , longmore , a.j . , tritton , s.b . , elson , r.a.w . , corwin , h.g . , jr . 1981 , mnras , 196 , 747 held , e.v . , de zeeuw , t. , mould , j. , picard , a. 1992 , aj , 103 , 851 heraudeau , p. , simien , f. 1997 , a&a , 326 , 897 hernquist , l. 1990 , apj , 356 , 359 herschel , j. 1864 , catalogue of nebulae and clusters of stars , royal society of london philosophical transactions series i , 154 , 1 hilker m. , infante l. , vieira g. , kissler - patig m. , richtler t. , 1999 , a&as , 134 , 75 hill d.t . , et al . 2011 , mnras , 412 , 765 hjorth , j. , madsen , j. 1995 , apj , 445 , 55 hodge , p.w . 1961a , aj , 66 , 249 hodge , p.w . 1961b , aj , 66 , 384 hodge , p.w . 1963 , aj , 68 , 691 hodge , p.w . 1964 , aj , 69 , 442 hodge , p.w . 1971 , ara&a , 9 , 35 hoffman , l. , cox , t.j . , dutta , s. , hernquist , l. 2009 , apj , 705 , 920 hohl f. 1975 , iau symp . , 69 , 349 hohl f. , zhang t.a . 1979 , aj , 84 , 585 hoskin , m. 1970 , journal for the history of astronomy , 1 , 44 hoessel , j.g . , oegerle , w.r . , schneider , d.p . 1987 , aj , 94 , 1111 hubble , e. 1926 , apj , 64 , 321 hubble , e. 1929 , proceedings of the national academy of science , 15 , 168 hubble , e. 1930 , apj , 71 , 231 hubble , e. 1934 , the halley lecture , delivered on 8 may 1934 ( oxford : at the clarendon press , 1934 ) , hubble , e.p . 1936a , realm of the nebulae , by e.p . hubble , new haven : yale university press hubble , e.p . 1936b , apj , 84 , 517 hubble , e.p . 1937 , mnras , 97 , 513 hubble , e. , tolman , r. , 1935 , apj , 82 , 302 humasson m. 1929 , proc . , 15 , 167 humphre p.j . , buote d.a . 2010 , mnras 403 , 2143 huxor a.p . , phillipps s. , price j. , harniman r. 2011 , mnras , in press ( arxiv:1103.1257 ) ichikawa , s .- wakamatsu , k .- i . , okumura , s.k . 1986 , apjs , 60 , 475 iodice , e. , donofrio , m. , capaccioli , m. 1997 , asp conf . ser . , 116 , 841 iodice , e. , donofrio , m. , capaccioli , m. 1999 , asp conf . ser . , 176 , 402 ivezi , . , et al . 2004 , astronomische nachrichten , 325 , 583 jablonka p. , gorgas j. , goudfrooij p. , 2007 , a&a , 474 , 763 jaffe , w. , ford , h.c . , oconnell , r.w . , van den bosch , f.c . , ferrarese , l. 1994 , aj , 108 , 1567 janz , j. , lisker , t. 2008 , apj , 689 , l25 janz , j. , lisker , t. 2009 , apj , 696 , 102 jeans j. 1919 . problems of cosmogony and stellar dynamics , cambridge : cambridge univ . press jeans , 1928 , astronomy & cosmogony , ( cambridge : cambridge university press ) , p.332 jerjen , h. , binggeli , b. 1997 , in the nature of elliptical galaxies ; the second stromlo symposium , asp conf . ser . , 116 , 239 jerjen , h. , kalnajs , a. , binggeli , b. 2000b , a&a , 358 , 845 jimnez , n. , cora , s.a . , bassino l.p . , tecce t.e . , smith castelli a.v . 2011 , mnras , in press ( arxiv:1107.0722 ) jing , y.p . , suto , y. 2000 , apj , 529 , l69 jonsson p. 2006 , mnras , 372 , 2 jonsson p. , groves b.a . , cox t.j . 2010 , mnras , 403 , 17 kaiser , n. , et al . 2002 , proc . spie , 4836 , 154 kannappan , s.j . , jansen , r.a . , barton , e.j . 2004 , aj , 127 , 1371 kant , 1755 , allgemeine naturgeschichte und theorie des himmels karachentseva , v.e . , prugniel , p. , vennik , j. , richter , g.m . , thuan , t.x . , martin , j.m . 1996 , a&ass , 117 , 343 kauffmann , g. , white , s.d.m . , guiderdoni , b. 1993 , mnras , 264 , 201 kauffmann , g. , white , s.d.m . , heckman , t.m . , mnard , b. , brinchmann , j. , charlot , s. , tremonti , c. , brinkmann , j. 2004 , mnras , 353 , 713 kautsch , s.j . 2009a , astronomische nachrichten , 330 , 1053 kautsch , s.j . 2009b , pasp , 121 , 1297 kawata , d. , cen , r. , ho , l.c . 2007 , apj , 669 , 232 keel w.c . , white r.e . , iii 2001a , aj , 121 , 1442 keel w.c . , white r.e . , iii 2001b , aj , 122 , 1369 kent , s. 1985 , apjs , 59 , 115 kent , s.m . , dame , t. , fazio , g. , 1991 , apj , 378 , 131 kere , d. , katz , n. , weinberg , d.h . , dav , r. 2005 , mnras , 363 , 2 khosroshahi , h.g . , wadadekar , y. , kembhavi , a. 2000 , apj , 533 , 162 king i.r . 1962 , aj , 67 , 471 king i.r . 1966 , aj , 71 , 64 king i.r . 1978 , apj , 222 , 1 king i.r . , minkowski r. 1966 , apj , 143 , 1002 king i.r . , minkowski r. 1972 , iau symp . , 44 , 87 kochanek c.s . 1995 , apj , 445 , 559 koda j. , milosavljevi m. , shapiro p.r . , 2009 , apj , 696 , 254 kodaira , k. , watanabe , m. , okamura , s. 1986 , apjs , 62 , 703 knapen , j.h . , de jong , r.s . , stedman , s. , bramich , d.m . 2003 , mnras , 344 , 527 knapen , j.h . 2010 , in galaxies and their masks , ed . block d.l . , freeman k.c . , puerari i. , springer , p.201 ( arxiv:1005.0506 ) knox shaw , h. 1915 , helwan obs . 15 , p.129 koopmans , l.v.e . , treu , t. , bolton , a.s . , burles , s. , moustakas , l.a . 2006 , apj , 649 , 599 koopmans , l.v.e . , et al . 2009 , apj , 703 , l51 kormendy , j. 1977a , apj , 217 , 406 kormendy , j. 1977b , apj , 218 , 333 kormendy , j. 1982 , saas - fee advanced course 12 : morphology and dynamics of galaxies , 113 kormendy , j. 1985a , apj , 292 , l9 kormendy , j. 1985b , apj , 295 , 73 kormendy , j. , dressler , a. , byun , y.i . , faber , s.m . , grillmair , c. , lauer , t.r . , richstone , d. , tremaine , s. 1994 , in dwarf galaxies , ed . g. meylan , p. prugniel ( garching : eso ) , eso conf . ser . , 49 , 147 kormendy , j. , fisher , d.b . , cornell , m.e . , bender , r. 2009 , apjs , 182 , 216 kormendy , j. , gebhardt , k. , fisher , d.b . , drory , n. , macchetto , f.d . , sparks , w.b . 2005 , aj , 129 , 2636 kormendy , j. , kennicutt , r.c . , jr . 2004 , ara&a , 42 , 603 khosroshahi , h.g . , wadadekar , y. , kembhavi , a. 2000 , apj , 533 , 162 kragh , h. , smith , r.w . 2003 , history of science , 41 , 141 krajnovi , d. , et al . 2008 , mnras , 390 , 93 larson r.b . 1969 , mnras , 145 , 405 larson r.b . 1974 , mnras , 166 , 585 larson r.b . 1975 , mnras , 173 , 671 lauer t.r . 1983 , in elliptical galaxies , surface photometry , santa cruz : university of california lauer t.r . 1984 , bull . soc . , 16 , 455 ( 08.03 ) lauer t.r . 1985 , apj , 292 , 104 lauer t.r . , et al . 1995 , aj , 110 , 2622 lauer t.r . , et al . 2005 , aj , 129 , 2138 launhardt , r. , zylka , r. , mezger , p.g . 2002 , a&a , 384 , 112 laurikainen , e. , salo , h. 2002 , mnras , 337 , 1118 laurikainen , e. , salo , h. , buta , r. 2005 , mnras , 362 , 1319 laurikainen , e. , salo , h. , buta , r. , knapen , j.h . 2007 , mnras , 381 , 401 lemaitre , g. 1927 , annales de la societe scientique de bruxelles , 47 , 49 liller m.h . 1960 , apj , 132 , 306 liller m.h . 1966 , apj , 146 , 28 lima neto , g.b . , gerbal , d. , mrquez , i. 1999 , mnras , 309 , 481 liske , j. , lemon , d.j . , driver , s.p . , cross , n.j.g . , couch , w.j . 2003 , mnras , 344 , 307 lisker , t. , et al . 2007 , apj , 660 , 1186 lisker , t. 2009 , an , 330 , 1403 lisker , t. , glatt , k. , westera , p. , grebel , e.k . 2006b , aj , 132 , 2432 lisker , t. , grebel , e.k . , binggeli , b. 2005 , iau colloq . 198 : near - fields cosmology with dwarf elliptical galaxies , edited by jerjen , h. , binggeli , b. , cambridge : cambridge university press , 311 lisker , t. , grebel , e.k . , binggeli b. 2006a , aj , 132 , 497 lisker , t. , han , z. 2008 , apj , 680 , 1042 lugger , p.m. 1984 , apj , 286 , 106 lundmark , k. 1924 , mnras , 84 , 747 ma , j. , et al . 2007 , mnras , 376 , 1621 macarthur , l.a . , courteau , s. , holtzman , j.a . 2003 , apj , 582 , 689 macarthur , l.a . , gonzlez , j.j . , courteau , s. 2009 , mnras , 395 , 28 macarthur , l.a . , mcdonald , m. , courteau , s. , jess gonzlez , j. 2010 , apj , 718 , 768 macci , a.v . , murante , g. , bonometto , s.p . 2003 , apj , 588 , 35 magorrian , j. , et al . 1998 , aj , 115 , 2285 maller , a.h . , berlind , a.a . , blanton , m.r . , hogg , d.w . 2009 , apj , 691 , 394 malumuth e.m . , kirshner r.p . 1981 , apj , 251 , 508 mamon , g.a . , okas , e.l . 2005 , mnras , 362 , 95 mandelbaum r. , et al . 2005 , mnras , 361 , 1287 mrquez , i. , lima neto , g.b . , capelato , h. , durret , f. , lanzoni , b. , gerbal , d. 2001 , a&a , 379 , 767 marconi , a. , hunt , l.k . 2003 , apj , 589 , l21 marinova i. , et al . 2011 , apj , submitted martin p. 1995 , aj , 109 , 2428 martizzi d. , teyssier r. , moore b. 2011 , mnras , submitted ( arxiv:1106.5371 ) masters , k.l . , giovanelli , r. , haynes , m.p . 2003 , aj , 126 , 158 matkovi , a. , guzmn , r. 2005 , mnras , 362 , 289 matthews , t.a . , morgan , w.w . , schmidt , m. 1964 , apj , 140 , 35 mcalpine , w. , satyapal , s. , gliozzi , m. , cheung , c.c . , sambruna , r.m . , eracleous , m. 2011 , apj , 728 , 25 mcglynn , t.a . 1984 , apj , 281 , 13 mclure r.j . , dunlop j.s . 2004 , mnras , 352 , 1390 mebold u. , goss w.m . , siegman b. , van woerden h. , hawarden t.g . 1979 , a&a , 74 , 100 merritt d. , 2006a , apj , 648 , 976 merritt d. , 2006b , reports on progress in physics , 69 , 2513 merritt , d. , ferrarese l. , joseph c.l . 2001 , science , 293 , 1116 merritt , d. , graham , a.w . , moore , b. , diemand , j. , terzi , b. 2006 , aj , 132 , 2685 merritt , d. , navarro , j.f . , ludlow , a. , jenkins , a. 2005 , apjl , 624 , l85 merritt , d. , milosavljevi , m. 2005 , living reviews in relativity , 8 , 8 metcalfe , n. , godwin , j.g . , peach , j.v . 1994 , mnras , 267 , 431 michard , r. 1979 , a&a , 74 , 206 michard , r. 1985 , a&as , 59 , 205 mihos j.c . , hernquist l. 1994 . apj , 425 , l13 milosavljevi , m. , merritt , d. 2001 , apj , 563 , 34 minkowski r. 1962 , iau symp . , 15 112 minniti d. , saito r.k . , alonso - garcia j. , lucas p.w . , hempel m. 2011 , ( arxiv:1105.3151 ) misgeld , i. , hilker , m. , mieske , s. 2009 , a&a , 496 , 683 misgeld , i. , mieske , s. , hilker , m. 2008 , a&a , 486 , 697 misgeld i. , hilker m. 2011 , mnras , in press ( arxiv:1103.1628 ) mllenhoff , c. , heidt , j. 2001 , a&a , 368 , 16 morgan , w.w . , lesh j.r . , 1965 , apj , 142 , 1364 morton , d.c . , chevalier , r.a . 1973 , apj , 179 , 55 moorthy , b.k . , holtzman , j.a . 2006 , mnras , 371 , 583 naab t. , burkert a. , hernquist l. , 1999 , apj , 523 , l133 naab , t. , khochfar , s. , burkert , a. 2006 , apj , 636 , l81 nair p. , van den bergh s. , abraham r.g . 2011 , apj , 734 , l31 navarro , j.f . , benz , w. 1991 , apj , 380 , 320 navarro , j.f . , et al . 2004 , mnras , 349 , 1039 navarro , j.f . , frenk , c.s . , white , s.d.m . 1997 , apj , 490 , 493 nieto , j .- l . 1990 , in dynamics and interactions of galaxies , ed . r. wielen ( berlin : springer ) , 258 nipoti , c. , londrillo , p. , ciotti , l. 2006 , mnras , 370 , 681 nipoti , c. , treu , t. , auger , m.w . , bolton , a.s . 2009 , apj , 706 , l86 norman c.a . , sellwood j.a . , hasan h. 1996 , apj , 462 , 114 norris , m.a . , kannappan , s.j . 2011 , mnras , 414 , 739 oemler a.,jr . 1976 , apj , 209 , 693 ohta , k. , hamabe , m. , wakamatsu , k .- 1990 , apj , 357 , 71 ostriker , j.p . 1977 , proceedings of the national academy of science , 74 , 1767 parson l. 1878 , scientific transactions of the royal dublin society vol . ii patterson , f.s . , 1940 , harvard college observatory bulletin , 914 , 9 peletier , r.f . , balcells , m. 1996a , aj , 111 , 2238 peletier r.f . , balcells m. 1996b , iaus , 171 , 29 peletier , r.f . , balcells , m. , davies , r.l . , andredakis , y. , vazdekis , a. , burkert , a. , prada , f. 1999 , mnras , 310 , 703 peletier r.f . et al . 2007 , mnras , 379 , 445 peletier , r.f . , davies , r.l . , illingworth , g.d . , davis , l.e . , cawson , m. 1990 , aj , 100 , 1091 pellegrini , s. 2010 , apj , 717 , 640 pellet a. , 1976 , a&a , 50 , 421 penny , s.j . , conselice , c.j . 2008 , mnras , 383 , 247 peterson c.j , 1978 , apj , 221 , 80 petrosian , v. 1976 , apj , 209 , l1 pfenniger d. 1993 , in galactic bulges , ed . h. dejonghe , h.j . habing ( dordrecht : kluwer ) , 387 pichon c. , pogosyan d. , kimm t. , slyz a. , devriendt j. , dubois y. 2011 ( arxiv:1105.0210 ) pierini , d. , zibetti , s. , braglia , f. , bhringer , h. , finoguenov , a. , lynam , p.d . , zhang , y .- y . 2008 , a&a , 483 , 727 pilbratt , g.l . , et al . 2010 , a&a , 518 , l1 plummer , h.c . 1911 , mnras , 71 , 460 pohlen , m. , beckman , j.e . , httemeister , s.h . , knapen , j.h . , erwin , p. , dettmar , r .- j . 2004 , in penetrating bars through masks of cosmic dust , ed . d. l. block et al . ( dordrecht : kluwer academic publishers ) , 713 popescu , c.c . , misiriotis , a. , kylafis , n.d . , tuffs , r.j . , fischera , j. 2000 , a&a , 362 , 138 popescu , c.c . , tuffs , r.j . , dopita , m.a . , fischera , j. , kylafis , n.d . , madore , b.f . 2011 , a&a , 527 , a109 price j. et al . 2009 , mnras , 397 , 1816 prieto , m. , aguerri , j.a.l . , varela , a.m. , muoz - tun , c. 2001 , a&a , 367 , 405 prieto , m. , gottesman , s.t . , aguerri , j .- a.l . , varela , a .- m . 1997 , aj , 114 , 1413 prugniel , p. , simien , f. 1997 , a&a , 321 , 111 qu , y. , di matteo , p. , lehnert , m.d . , van driel , w. 2011 , a&a , 530 , a10 ravindranath , s. , ho , l.c . , peng , c.y . , filippenko , a.v . , sargent , w.l.w . 2001 , aj , 122 , 653 reaves g. 1956 , aj , 61 , 69 reaves g. 1977 , apjs , 53 , 375 reese , a.s . , williams , t.b . , sellwood , j.a . , barnes , e.i . , powell , b.a . 2007 , aj , 133 , 2846 renzini , a. 1999 , in the formation of galactic bulges , ed . carollo , h.c . ferguson , r.f.g . wyse ( cambridge : cambridge univ . press ) , 1 renzini , a. 2006 , ara&a , 44 , 141 rest , a. , et al . 2001 , aj , 121 , 2431 reynolds , j.h . 1913 , mnras , 74 , 132 reynolds , j.h . 1920 , observatory , 43 , 377 reynolds , j.h . 1927 , observatory , 50 , 185 rich r. m. , origlia l. , 2005 , apj , 634 , 1293 richings a.k . , uttley p. , krding e. 2011 , mnras , in press ( arxiv:1104.1053 ) rix h .- w . , white s.d.m . 1990 , apj , 362 , 52 robertson h.p . 1928 , phil . mag . , 5 , 835 robotham a.s.g . , driver s.p . 2011 , mnras , 413 , 2570 romanishin w. , strom k.m , strom s.e . 1977 , bull . 9 , 347 rood , h.j . , page , t.l . , kintner , e.c . , king , i.r . 1972 , apj , 175 , 627 rozas m. , knapen j.h . , beckman j.e . 1998 , mnras , 301 , 631 rubin v.c . , ford w.k . , krishna kumar c. 1973 , apj , 181 , 61 saha k. , martinez - valpuesta i. , gerhard o. 2011 , mnras , submitted ( arxiv:1105.5797 ) sandage , a. 1972 , apj , 176 , 21 sandage a. , 2004 , in penetrating bars through masks of cosmic dust - the hubble tuning fork strikes a new note . editors : d. block et al .. kluwer publishers , astrophysics and space science library ( assl ) , 319 , 39 sandage , a. 2005 , ara&a , 43 , 581 sandage , a. , binggeli , b. 1984 , aj , 89 , 919 saslaw , w.c . , valtonen , m.j . , aarseth , s.j . 1974 , apj , 190 , 253 scannapieco c. , white s.d.m . , springel v. , tissera p.b . 2011 , mnras , submitted ( arxiv:1105.0680 ) schechter p.l . 1980 aj , 85 , 801 schdel , r. , merritt , d. , eckart , a. 2009 , journal of physics conference series , 131 , 012044 schombert , j.m . 1986 , apjs , 60 , 603 schurer a. , calura f. , silva l. , pipino a. , granato g. l. , matteucci f. , maiolino r. , 2009 , mnras , 394 , 2001 schwarzkopf , u. , dettmar , r .- j . 1997 , agm , 13 , 238 seares , f.h . , 1928 , pasp , 40 , 303 srsic , j .- l . 1963 , boletin de la asociacion argentina de astronomia , vol.6 , p.41 srsic , j.l . 1968 , atlas de galaxias australes seigar m.s . , graham a.w . , jerjen h. 2007 , mnras , 378 , 1575 seigar , m.s . , james , p.a . 1998 , mnras , 299 , 672 sellwood j.a . , wilkinson a. 1983 , rep . 56 , 173 seth , a.c . , blum , r.d . , bastian , n. , caldwell , n. , debattista , v.p . 2008 , apj , 687 , 997 shapley , h. 1928 , nature , 122 , 482 shapley , h. , swope , h.h . 1924 , in studies of the galactic center , astronomical observatory of harvard college , harvard reprint 51 and 52 shaviv g. 2011 , in the quest for chemical element genesis and what the chemical elements tell about the universe , springer pub . heidelberg ( arxiv:1107.0442 ) shaw , m.a . , gilmore , g. 1989 , mnras , 237 , 903 shields , j.c . , walcher , c.j . , bker , t. , ho , l.c . , rix , h .- w . , van der marel , r.p . 2008 , apj , 682 , 104 simien , f. , de vaucouleurs , g. 1986 , apj , 302 , 564 simonneau , e. , prada , f. 2004 , revista mexicana de astronomia y astrofisica 40 , 69 ( astro - ph/9906151 ) skelton r.e . , bell e.f . , somerville r.s . , 2009 , apj , 699 , l9 slipher , v.m . 1917 , proceedings of the american philosophical society , 56 , 403 smith , s. 1935 , apj , 82 , 192 smith castelli a.v . , faifer f.r . , richtler t. , bassino l.p . 2008 , mnras , 391 , 685 soifer b.t . , neugebauer g. 1991 , aj , 101 , 354 somerville r.s . , gilmore r.c . , primack j.r . , dominguez a. 2011 , mnras , submitted ( arxiv:1104.0669 ) sparks w.b . 1988 , aj , 95 , 1569 spergel d.n . 2010 , apjs , 191 , 58 spinrad h. , et al . 1978 , apj , 225 , 56 stebbins j. , whitford a.e . 1934 , pnas , 20 , 93 steinmetz , m. , navarro , j.f . 2002 , new astronomy , 7 , 155 stewart , k.r . , bullock , j.s . , wechsler , r.h . , maller , a.h . , zentner , a.r . 2008 , apj , 683 , 597 stiavelli , m. , miller , b.w . , ferguson , h.c . , mack , j. , whitmore , b.c . , lotz , j.m . 2001 , aj , 121 , 1385 strom , k.m . , strom , s.e . 1978 , aj , 83 , 1293 tal , t. , van dokkum , p.g . 2011 , apj , 731 , 89 temi p. , brighenti f. , mathews w.g . , bregman j.d . 2004 , apjs , 151 , 237 terzi b. , graham a.w . 2005 , mnras , 362 , 197 terzi b. , sprague b.j . 2007 , mnras , 377 , 855 thakur , p. , ann , h.b . , jiang , i .- g . 2009 , apj , 693 , 586 thomas d. , davies r.l . , 2006 , mnras , 366 , 510 tollerud , e.j . , bullock , j.s . , graves , g.j . , wolf , j. 2011 , apj , 726 , 108 tolstoy , e. , hill , v. , tosi , m. 2009 , ara&a , 47 , 371 tonini , c. , maraston , c. , thomas , d. , devriendt , j. , silk , j. 2010 , mnras , 403 , 1749 tonry , j. 1981 , apj , 251 , l1 tran , h.d . , tsvetanov , z. , ford , h.c . , davies , j. , jaffe , w. , van den bosch , f.c . , rest , a. 2001 , aj , 121 , 2928 tremonti , c.a . , et al . 2004 , apj , 613 , 898 trimble , v. 1999 , bull . 31 , 1479 ( # 74.09 ) trujillo , i. , asensio ramos , a. , rubio - martn , j.a . , graham , a.w . , aguerri , j.a.l . , cepa , j. , gutirrez , c.m . 2002 , mnras , 333 , 510 trujillo , i. , et al . 2006 , mnras , 373 , l36 trujillo , i. , erwin , p. , asensio ramos , a. , graham , a.w . 2004 , aj , 127 , 1917 trujillo , i. , graham , a.w . , caon , n. 2001 , mnras , 326 , 869 trumpler r.j . , 1930a , pasp , 42 , 214 trumpler r.j . , 1930b , lick observatory bulletin , 14 , 154 tsikoudi , v. 1979 , apj , 234 , 842 tully , r.b . , fisher , j.r . 1977 , a&a , 54 , 661 tully , r.b . , pierce , m.j . , huang , j .- s . , saunders , w. , verheijen , m.a.w . , witchalls , p.l . 1998 , aj , 115 , 2264 tyson , j.a . 2002 , proc . spie , 4836 , 10 unterborn , c.t . , ryden , b.s . 2008 , apj , 687 , 976 van albada t.s . 1982 , mnras , 201 , 939 van den bergh , s. 1977 , in evolution of galaxies and stellar populations , ed . b.m.tinsley , r.b.larson , p.19 van den bergh , s. 1991 , pasp , 103 , 609 van den bergh , s. 1986 , aj , 91 , 271 van den bergh , s. 2008 , a&a , 490 , 97 van der kruit , p. c. 1979 , a&as , 38 , 15 van der kruit , p. c. 1987 , a&a , 173 , 59 van der kruit , p. c. 1988 , a&a , 192 , 117 van der kruit , p.c . , searle , l. 1981 , a&a , 95 , 105 van houten c.j . 1961 , bull . netherlands , 16 , 1 vennik , j. , richter , g.m . 1994 , astron . nachr . , 315 , h3 , 245 vennik , j. , hopp , u. , kovachev , b. , kuhn , b. , elssser , h. 1996 , a&ass , 117 , 261 volonteri , m. , madau , p. , haardt , f. 2003 , apj , 593 , 661 wadadekar , y. , robbason , b. , kembhavi , a. 1999 , aj , 117 , 1219 wainscoat r.j . , freeman k.c . , hyland a.r . 1989 , apj , 337 , 163 walcher , c.j . , et al . 2005 , apj , 618 , 237 walmsley , c.m . , bertout , c. , combes , f. , ferrara , a. , forveille , t. , guillot , t. , jones , a. , shore , s. 2010 , a&a , 518 , 1 webb , c.j . 1964 , aj , 69 , 442 weiner , b.j . , williams , t.b . , van gorkom , j.h . , sellwood , j.a . 2001 , apj , 546 , 916 weinzirl , t. , jogee , s. , khochfar , s. , burkert , a. , kormendy , j. 2009 , apj , 696 , 411 werner m.w . , et al . 2004 , apjs , 154 , 1 white s.d.m . , frenk c.s . 1991 , apj , 379 , 52 white r.e . , iii , keel , w.c . , conselice , c.j . 2000 , apj , 542 , 761 williams m.j . , et al . 2010 , mnras , 414 , 2163 wilson c.p . 1975 , aj , 80 , 175 wirth , a. , gallagher , j.s . 1984 , apj , 282 , 85 wolf m. 1908 , publ . astrophys . heidelberg , 3 , no . 3 wolf , j. , martinez , g.d . , bullock , j.s . , kaplinghat , m. , geha , m. , muoz , r.r . , simon , j.d . , avedo , f.f . 2010 , mnras , 406 , 1220 wolfe , a.m. , burbidge , g.r . 1970 , apj , 161 , 419 wozniak , h. , friedli , d. , martinet , l. , martin , p. , bratschi , p. 1995 , a&as , 111 , 115 wright , t. , 1750 , in an original theory or new hypothesis of the universe , london wyse , r.f.g . , gilmore , g. 1995 , aj , 110 , 2771 wyse , r.f.g . , gilmore , g. , franx , m. 1997 , ara&a , 35 , 637 yip c .- w . , szalay a.s . , carliles s. , budavari t. 2011 , apj , in press ( arxiv:1101.5651 ) york , d.g . , et al . 2000 , aj , 120 , 1579 yoshizawa , m. , wakamatsu , k. 1975 , a&a , 44 , 363 young , c.k . , currie , m.j . 1994 , mnras , 268 , l11 young , c.k . , currie , m.j . 1995 , mnras , 273 , 1141 young , p.j . , westphal , j.a . , kristian , j. , wilson , c.p . , landauer , f.p . 1978 , apj , 221 , 721 zhao , h. 1996 , mnras , 278 , 488 zibetti s. , white s.d.m . , schneider d.p . , brinkmann j. 2005 , mnras , 358 , 949 zinnecker , h. , keable , c.j . , dunlop , j.s . , cannon , r.d . , griffiths , w.k . 1988 , in iau symp . 126 , globular cluster systems in galaxies , ed . grindlay , a.g.d . philip ( dordrecht : kluwer ) , 603 zoccali m. et al . , 2006 , a&a , 457 , l1

many statistical measures have been developed to distinguish between the various cosmological models , ranging from direct measures of the power spectrum and correlation function from redshift surveys , to measures of velocity dispersion , bulk flows , and mach number from peculiar velocity surveys ( cf . , @xcite for a review ) . each of these measures is designed to be sensitive to different aspects of various models , such as @xmath0 , the initial power spectrum , or the gaussian character of the initial phase distribution . one can get an intuitive understanding of the dependencies of different statistics by considering the following example . compare n - body simulations of four different cosmological models : standard cdm , open cdm , cdm + @xmath1 , and a pure hdm model , the details of which are given in 3 . each simulation is normalized to have the same value of @xmath3 , the standard deviation of the mass fluctuations within an 8 @xmath4 sphere , where @xmath5 is the hubble parameter in units of 100 @xmath6 . the two point correlation function , @xmath7 , for each model is presented in figure 1a . despite their differences in initial power spectra , the final correlation functions are similar in all models . in general , the correlation function does not depend strongly on @xmath0 if one retains the freedom to set the @xmath3 normalization . in contrast , figure 1b shows the pairwise velocity dispersion , @xmath8 , as a function of separation for the same set of models . a strong differentiation between the low and high @xmath0 models is apparent . the velocity dispersion is easy to compute in simulations , but is difficult to measure in the geometric projection of phase space that is measured by redshift surveys . the traditional way to measure the small - scale velocity dispersion of galaxies is via the anisotropy it introduces in the redshift - space two - point correlation function ( @xcite ; @xcite ; @xcite ; @xcite ) . however , as it is a pair - weighted statistic , it is dominated by the densest regions , which are necessarily rare . thus one finds large variance between estimates of the small - scale velocity dispersion between different samples ( @xcite ; @xcite ; @xcite ; @xcite ) , indicating that the small - scale velocity dispersion measured in this way is not very robust . attempts have been made to use direct measures of peculiar velocities of galaxies as a constraint on @xmath9 ( e.g. , @xcite , @xcite ) , but other than the nearby universe , where the velocity field is observed to be very quiet ( @xcite ; @xcite ; @xcite ) , the errors on the individual peculiar velocity measurements swamp the signal from @xmath9 . the primary goal of this paper is to present a statistic the redshift dispersion ( @xmath10)that captures @xmath11 in a way which is naturally applied to volume limited samples drawn from redshift surveys . in the subsequent sections we motivate and present our redshift dispersion statistic . again , because it is a pair - weighted statistic , computing @xmath11 by averaging over all densities results in a value heavily weighted by the densest regions . however , this problem can be alleviated if the dispersion can be calculated _ as a function of density_. the theoretical motivation for our statistic stems from the cosmic virial theorem ( cvt ) , derived in peebles ( 1976a , hereafter p76 ) , which relates @xmath11 to @xmath12 and @xmath13 . in 2 we show that the cvt can be applied to subsets of particles in a system which correspond to surfaces of constant density . many assumptions are necessary to obtain the results shown in 2 , not all of which are obvious . in 3 we explore the relationships derived in 2 with simple n - body simulations , suggesting that the results of 2 hold over a wide range of scales . the redshift dispersion is entirely independent of the results of 2 and 3 , which provide a context for the redshift dispersion that would otherwise be an unmotivated simulation based statistic . 4 describes how to compute the redshift dispersion from a redshift survey . 5 contains our conclusions and remarks on future work . as will be shown in 4 , the redshift dispersion probes the pairwise peculiar velocity dispersion in regions of different density and is designed for comparing simulated and observed redshift surveys . in this section we attempt to provide a theoretical context and some motivation . one of the main challenges of working with the pairwise velocity dispersion arises from its strong density dependence . intuitively , both the number of galaxies and the velocity dispersion should be highest in the densest regions . thus , averaging over all densities will give results dominated by rare , high density peaks . this problem can be eliminated if the dispersion is calculated as a function of density . in the case of clustered , gravitationally interacting particles there exists a scaling relation between the velocity dispersion on small scales @xmath14 averaged over pairs separated by a distance @xmath15 , and the two point correlation function @xmath7 , the excess fractional probability of finding two particles with separation @xmath15 . this result is contained in the cvt derived in p76 ( see also @xcite ; @xcite ; @xcite ) , which can be written ( r ) ( r ) r^2 . several assumptions are used to obtain eq . ( 1 ) ( see p76 ) , including that @xmath13 is given by a power law @xmath16 ; @xmath17 , implying @xmath18 ; the three - point correlation function @xmath19 is given by a symmetrized product of the two - point correlation function ( see eq . [ a23 ] ) ; and the mass of each galaxy is concentrated on scales smaller than their separation . the validity of these assumptions is not obvious ( @xcite ) . for an excellent discussion of the implications of extended dark matter halos on the cvt , see bartlett & blanchard ( 1996 ) . in addition , we are assuming that the _ galaxy _ velocity field is unbiased with respect to that of the dark matter . while theoretical investigations have given strong suggestions that a velocity bias of order 2030% may exist ( @xcite , @xcite , @xcite ) , the difficulty in reliably tracing galaxies in simulations has prevented a good estimate of its magnitude ( @xcite ) . for the motivational purposes of this paper , let us keep these assumptions as we work to derive the density dependence of the cvt . to rewrite eq . ( 1 ) in terms of the density requires some new notation . let @xmath20 denote the cross - correlation of particle sets @xmath21 and @xmath22 . the two - point ( auto ) correlation function of all particles in a system @xmath23 can be written @xmath24 . the correlation function can also be written as the average of the individual cross correlations ( r ) = 1 n _ _ i _ ( r ) , where @xmath25 is the total number of particles . we can write down a similar expression for the velocity dispersion ( r ) = 1 n _ _ i ( r ) , where @xmath26 is the variance of the pairwise velocity between particle @xmath27 and all particles in @xmath23 which lie a distance @xmath15 from @xmath27 . we show in appendix a that these expressions lead to a generalization of the cvt that holds for each particle : ( r ) _ ( r ) r^2 . if the cvt holds for each particle , then it holds for any subset , @xmath28 : ( r ) _ ( r ) r^2 , where ( r ) _ i ( r ) , and _ ( r ) _ i _ ( r ) . let @xmath29 be the number of particles within a radius @xmath15 of the @xmath27th particle , and @xmath30 the set of all particles for which @xmath31 . for this subset , the cvt is ( n , r ) = c_1 ( n , r ) r^2 , where @xmath32 , @xmath33 and @xmath34 is given by eq . the quantity @xmath35 is related to the expected number of particles within a radius @xmath15 around any particle by _ 0^r [ 1 + ( n , r ) ] 4 r^2 dr = n , where @xmath36 is the mean number of particles per unit volume . if @xmath37 and @xmath38 ( as assumed in the original p76 derivation ) , the above expression yields ( n , r ) ( r ) = n , where @xmath39 . inserting the above expression into eq . ( 8) yields ( n , r ) = c_1 c_2 n/(r)^1/3 , where @xmath40 . the above expression shows that the pairwise velocity dispersion is proportional to @xmath12 and the local density ( through @xmath41 ) smoothed on a scale @xmath15 . finally , since we are working with @xmath29 , which is the number of particles _ within _ a volume of radius @xmath15 centered on a particle , it is convenient to work with a similar velocity dispersion . let @xmath42 be the average pairwise velocity dispersion of all the particles _ within _ a volume of radius @xmath15 . @xmath42 can be obtained by integrating @xmath43 under the same assumptions used to derive eq . ( 1 ) ( r ) = c_1 c_3 = c_3 ( r ) , where @xmath44 . substituting eq . ( 12 ) into ( 11 ) results in ( n , r ) n/(r)^1/3 where the constant of proportionality is equal to c_1 c_2 c_3 = , and @xmath45 and @xmath46 are defined in appendix a. the above derivation of eq . ( 13 ) holds whether or not the quantities are averaged over one or many particles with the same density . the above velocity dispersion is computed with respect to the velocity of the central particle , @xmath47 . the velocity dispersion can also be computed with respect to the mean velocity of the particles in a cell , @xmath48 . in a given cell , these two values of the velocity dispersion will differ by @xmath49 . the distribution of these differences for all cells in @xmath30 will peak at zero and have a width on the order of @xmath50 . subsequent averaging over many cells with the same density will result in zero net difference . thus , eq . ( 13 ) also applies for velocity dispersions computed with respect to the mean velocity in the cell if the results are averaged over many cells in @xmath30 . ( 13 ) conveniently relates two readily computable quantities : the velocity dispersion with respect to the mean velocity in a cell of radius @xmath15 to the number of particles in the cell , which is proportional to the density . in the next section , we explore the above form of the cvt with n - body simulations . the previous section presented a derivation for a relationship linking the velocity dispersion to the local density . in this section we explore the range over which eq . ( 13 ) holds using n - body simulations of specific cosmological models with different initial power spectra . we are particularly interested in the dependence on @xmath12 when the number of objects is of the same order as expected from volume limited redshift surveys . the simulations we consider are designed to probe a variety of popular cosmological models . the four models are : standard cdm with @xmath51 , @xmath52 ; hdm with @xmath53 , @xmath54 ; open cdm with @xmath55 , @xmath56 ; and cdm + @xmath1 with @xmath57 , @xmath58 , @xmath56 . the open cdm and cdm + @xmath1 models provide two alternatives to standard cdm that increase the ratio of large scale to small scale power . they differ in evolution in that structure formation `` freezes out '' at an earlier epoch in the open model as the expansion rate exceeds the rate of gravitational collapse . thus , to achieve the same level of structure today , collapse must begin earliest in the open cdm model , later in the cdm + @xmath1 model , and latest in standard cdm model . while hdm is not generally considered a viable theory , it provides a significantly different power spectrum shape with which to explore our ideas . the initial conditions are designed to treat the models , as much as possible , on an even footing . all models assume a harrison - zeldovich primordial power spectrum , and use the same random phases for the fourier modes to generate the initial density field from their respective power spectra . the cdm transfer functions are taken from efstathiou , bond , & white ( 1992 , eq . [ 7 ] ) with the parameter @xmath59 . although this function was not intended for use in open models , it fits more detailed calculations to within 5% ( d. n. spergel 1995 , private communication ) . the hdm transfer function is taken from holtzman ( 1989 , table 2a , line 52 ) . as stated in 1 , each model is normalized to have the same linear value of @xmath60 , so as to provide similar correlation strengths and to isolate out the velocity dependencies . although this normalization does not match that predicted from the observed fluctuations in the cosmic microwave background for some or all of the models ( c.f . , @xcite ; @xcite ) , it roughly equalizes the amount of power on the scales where this paper is focused . each of the simulations follows @xmath61 dark matter particles within a periodic cube of comoving size 100 @xmath4 ( 10,000 @xmath2 ) on a side . the p3mg3a code ( @xcite ) , which implements the p@xmath62 m algorithm ( @xcite ; @xcite ) on the grape-3a hardware board ( @xcite ) , is used to evolve the simulations from redshift @xmath63 to @xmath64 using 1200 time steps . a plummer force law with softening parameter of 156 @xmath65 is used for the gravitational interactions and the mass per particle is @xmath66 . each simulation took approximately 2 hours to run on a sun sparc 10/51 workstation with a 4 chip grape-3a board . the ideal way to construct an artificial galaxy catalog is by identifying concentrations of gaseous and stellar material in high resolution n - body / hydrodynamic simulations . current computer technology and algorithms now permit such simulations on the scale of small groups of galaxies ( @xcite ) . however , at the present time the realm of large volumes remain the domain of strictly gravitational n - body codes . identifying galaxies from dark matter halos is a significant problem that may not be solvable ( @xcite ) ; although a variety of of impressive methods have been developed ( @xcite ; @xcite ) . however , to keep things as simple as possible we choose a model in which mass traces light , and each particle is assumed to be a galaxy . this approach neglects important processes , such as mass and velocity bias , the dependence of bias on cosmological models , and the interaction of dark matter halos . nevertheless this approach should be sufficient for our purely motivational purposes . future simulations which can both cover large volumes and resolve galaxies hydrodynamically will hopefully clarify the nature of the bias . to extract the velocity dispersion of a particular surface of constant density , consider a set of particles where @xmath67 and @xmath68 are the position and velocity of the @xmath27th particle . place spherical cells of radius @xmath15 on a uniform grid over the entire domain . let @xmath69 be the number of particles in cell @xmath70 . the correct way to compute the velocity dispersion in a cell is with respect to the mean motion of the particles . as was argued in 2 , eq . ( 13 ) will apply if the results are averaged over many cells with the same density . denote the mean velocity in the @xmath70th cell by @xmath71 ; the velocity variance , @xmath72 , is then @xmath73 averaged over all particles in the cell . if @xmath30 is the set of all cells having @xmath41 particles ( @xmath74 ) , then the average velocity and variance as a function of @xmath41 is _ v(n , r ) _ j _ n number of particles in the set @xmath30 ( note : @xmath75 ) . these equations provide a specific prescription for computing the density dependence of the mean velocity and the velocity dispersion , which can readily be applied to the simulations . plots of @xmath76 and @xmath77 are shown in figure 2 for the four models . the cell size is @xmath78 , corresponding to @xmath79 . all distances are expressed in units of @xmath2 . figure 2 demonstrates three important points : both @xmath80 and @xmath81 are independent of the shape of the initial power spectrum ; @xmath80 is independent of the local density , while @xmath81 is proportional to the density ; and @xmath80 and @xmath81 have the same strong @xmath12 dependence that @xmath11 demonstrated in figure 1b . formally , the power spectrum independence is explained by the derivation of the cvt ( see appendix a ) . the velocity dispersion depends upon the evolved power spectrum , which is essentially indistinguishable between models ( figure 1 ) . moreover , any remaining difference between models is encoded in the density distribution function , which is not apparent when density is the _ dependent _ variable in figure 2 . to explore the range over which eq . ( 13 ) is valid , @xmath80 and @xmath50 were calculated over the range of cell sizes @xmath82 , corresponding to @xmath83 . we obtained the following empirical fit for the mean velocity magnitude _ v ^1/2 /^ . where @xmath84 . as @xmath80 already contains the desired @xmath12 dependence , it is convenient to represent @xmath81 in terms of the normalized velocity dispersion @xmath85 , with the resulting fit _ v^2/_v^2 n/^1/3 , which agrees with the theory to within the small factor @xmath86 . the quality of the fits can be observed by plotting @xmath80 and @xmath87 against the scaled density ( @xmath88 ) , for each value of @xmath15 . figure 3a shows the ratio of @xmath80 to the fitted value computed from eq . ( 17 ) for the @xmath89 and @xmath90 cdm models . each line of data corresponds to a different value of @xmath15 , and has been offset from the next by one unit . figure 3b shows a similar plot for the normalized velocity dispersion @xmath91 . the solid lines are the best fits given by eq . ( 18 ) ; this has not been divided out . the larger scatter at higher densities and smaller scales is due to the small number of cells contributing to the calculation at these values . the key point to observe from figure 3a is how well the data points follow the horizontal lines corresponding to their fitted values . the fits to the data shown in figure 3 are remarkable , considering that the range includes cells with underdensities ( @xmath92 ) of @xmath93 on scales of @xmath94 and cells with overdensities of nearly 200 on scales of @xmath95 . furthermore , the scaling of @xmath81 is almost exactly that derived from the cvt , indicating that eq . ( 13 ) holds over a large range of scales and densities , even when @xmath96 . the purpose of 2 was to theoretically motivate the local density and @xmath12 dependence of @xmath50 . in 3 the scaling relations in 2 were explored with simple n - body simulations . in addition , 2 and 3 have introduced the ideas which allow us to describe the main point of this paper the redshift dispersion . the formalism we have developed so far can not be directly applied to observations due to the geometric projection of a six dimensional phase space into a three dimensional redshift survey . the redshift of galaxies represents the only probe , albeit indirect , of the peculiar velocity . thus , any statistic that desires to take advantage of the properties of @xmath50 must be defined with the specific geometry of redshift space in mind . in this section we now describe the redshift dispersion , a statistic with the aim of being readily measurable from volume limited samples taken from redshift surveys and which captures the essence of @xmath50 . now let us define quantities analogous to those in 3 , but for redshift space . consider a volume limited sample from a survey out to a maximum redshift of @xmath97 . each data point consists of two angular coordinates on the celestial sphere and a redshift . let us define cells within which to measure density and velocity dispersion in projection on the celestial sphere . the cell @xmath70 consists of a cone emanating from the origin with solid angle @xmath98 around the cell center . the number of points in the cone @xmath70 is @xmath69 and is proportional to the projected density on the sky . the mean and the variance of the redshifts in the cone are denoted @xmath99 and @xmath72 , which are depicted schematically in figure 4 . if @xmath30 is the set of all cones with @xmath74 , then the average and variance of the redshift as a function of @xmath41 is _ z(n , ) _ j _ n u_j . and ^2_z(n , ) _ j _ n ^2_j . in analogy with eqs . ( 15 ) and ( 16 ) . the efficacy with which the @xmath10 statistic might distinguish between models is examined with the simulations discussed in 3 . the simulations were transformed into redshift - space using @xmath100 , which is equal to one half of the simulation box size . since the data are periodic , the origin can be placed at any point , allowing multiple perspectives to be drawn from a single simulation . we computed @xmath101 for @xmath102 and @xmath103 cdm models . the angular cell size was @xmath104 ( @xmath105 ) , and the cell centers were computed by creating a pseudo - uniform grid of points across the celestial sphere ( baumgardner & frederickson 1985 ) . the data have been averaged over 27 different origins regularly distributed throughout the domain , and error bars computed from the standard deviation of this averaging . the @xmath10 curves are plotted in figure 5 . as with the velocity dispersion in figure 1b , the redshift dispersion shows a strong separation between the low and high @xmath12 models . the differences become apparent at moderate angular over - densities ( @xmath106 ) . the shape of the curves in figure 5 is due to the combined spatial and peculiar velocity contributions to @xmath10 , as is illustrated in figure 6 for the @xmath53 cdm simulation . we know the full six - dimensional position of each galaxy in phase space in the n - body simulation , and thus can separate these two contributions . at low overdensities , @xmath10 is dominated by the spatial separation of the particles , as is indicated by the solid points in figure 6 ; the spatial component scales approximately as @xmath107 , due to the more tightly bound nature of denser systems . at higher overdensities the peculiar velocity dominates , scaling approximately as @xmath108 , which is consistent with results of 2 and 3 . for smaller values of @xmath12 the spatial component behaves the same , but the peculiar velocity component is down by a factor of @xmath109 . these results indicate that the greatest separation in the dispersion between models with different values of @xmath12 will occur in the denser regions ; so it is important that we sample many modestly dense regions , which requires a large volume . in addition , to minimize projection effects the sample should not be too deep ( i.e. , @xmath97 not too large ) . therefore , to apply the @xmath10 statistic requires a dense sample over a wide field . if believable simulations of galaxies can be developed it might be possible to constrain models with the same final correlation function by varying @xmath0 in the simulations and finding the best fit to the observations . for any one point , figure 5 indicates an error of about 0.15 in @xmath0 . using many points along the curve , the errors may be small enough to significantly constrain the value of @xmath0 . the current redshift surveys do not have enough data to attempt such a comparison . to get the level of separation shown in figure 5 required @xmath110 particles . we have calculated @xmath101 from the iras 1.2 jansky survey ( @xcite ) , but a volume - limited sample taken from this survey contains only 800 galaxies at best . the error bars from an 800 point sample in the simulations were too large to be able to distinguish between high and low @xmath12 models . fortunately , a substantial increase in the amount of data available will be brought about by the sloan digital sky survey ( sdss ) . the sdss will obtain spectra and measure the redshifts of the @xmath111 galaxies down to @xmath112 ( @xcite ) . we can estimate the size of a volume - limited sample taken from the sdss using the schechter luminosity function @xmath113 fitted to the stromlo - apm survey ( @xcite ) _ s(l ) dl = ^ y^ e^-y dy , y = l / l^ , where @xmath114 , @xmath115 , and @xmath116 are parameters obtained from the fit . the number of galaxies in a given volume , @xmath117 , brighter than @xmath118 , @xmath119 , can be computed by integrating the luminosity function n(v , l_0 ) = v _ l_0^ ( l ) dl = v _ y^ e^-y dy = v ^ ( + 1,y_0 ) , where @xmath120 . for a volume limited survey , @xmath118 is the luminosity an object would have if it had an apparent magnitude @xmath121 and was located at the volume edge @xmath97 ( @xmath122 ) . from this definition it follows that y_0 = z^2_10 ^ 0.4 ( m^ - m_0 + 15 ) , where @xmath121 is the apparent magnitude limit of the survey . the stromlo - apm survey was taken in the @xmath123 band , while the sloan will use the @xmath124 band . the two can be equated by the approximate relation @xmath125 . however , we set @xmath126 which gives a better value for the estimated total number of galaxies in the survey ( @xmath111 ) . inserting @xmath127 into eq . ( 23 ) gives @xmath128 . setting @xmath129 results in @xmath130 . this number might in fact be appreciably larger if there are many more faint galaxies than eq . ( 22 ) implies , and has been suggested by recent surveys for low - surface brightness galaxies ( cf . , @xcite ) . however , these galaxies are unlikely to have redshifts measured as part of the sdss . one could substantially increase the number of galaxies in a volume - limited survey of fixed depth by dropping the requirement that it include the origin . figure 7 plots the expected number of galaxies in a series of volume limited shells for sdss and shows that a redshift shell between @xmath131 and @xmath132 would include almost @xmath133 galaxies ! ultimately , applying the redshift dispersion statistic to the sdss could provide many data points for comparing with simulations and perhaps constraining @xmath0 . the @xmath10 curves can be evaluated for many angular sizes and redshift shells for both the sdss and for several next generation , high resolution , n - body / hydrodynamic simulations with different values of @xmath12 . since the number of estimated objects in the sdss is the same order or more as the simulations shown in figure 5 , one might expect to obtain an equivalent separation between models for each value . it should be interesting to compare @xmath10 measured in different simulations with the sdss as a function of @xmath41 , @xmath134 and shell geometry . our goal has been to present a new statistic the redshift dispersion ( @xmath10)that is sensitive to @xmath12 and is well suited for comparing real and simulated volume limited samples from redshift surveys . given the proper data , @xmath10 is easy to compute and can be applied on many scales without applying arbitrary assumptions . we have used low resolution simulations , which are sufficient for our motivational purposes , to do a simple exploration of @xmath10 , which suggests that applying it to the sdss may be worthwhile . in addition , with the right simulations , it might be possible to constrain @xmath12 in models where the simulations match the observed final correlation function . we have shown that the pairwise velocity dispersion is intrinsically related to the local density and @xmath12 . in 2 and appendix a it is shown that the cvt holds for each particle in a system and subsequently for any subset of particles , if we are careful to define the velocity dispersion and correlation function for these subsets appropriately . the density dependence of the velocity dispersion can be extracted by looking at subsets of particles of a given local density ( see eq . 13 ) . knowing how @xmath50 depends upon the local density gives an indication as to why the standard approach of averaging @xmath11 over all densities is highly sensitive to the presence of rare density peaks . exploring eq . ( 13 ) with n - body simulations of several cosmological models indicates that it holds over a wide range of length scales , @xmath135 . our redshift dispersion statistic is simply the redshift - space analog of the quantity @xmath136 . we have treated our galaxies as equal mass particles containing all the mass , ignoring the hypothesized global stochastic mapping from the dark matter mass and velocity distribution to the distribution of galaxies ( i.e. the mass and velocity bias functions ) , which may differ among the models . a more general treatment would estimate the mass and velocity bias in each model and normalize the models such that the galaxy correlations , not the dark matter correlations , are similar . at best , reliable estimates of these bias functions await the next generation of simulations where galaxies can be resolved within statistically meaningful volumes . however , there are several arguments as to why the differences in the biases between models may not strongly affect our work . first , the initial power spectra of currently favored hierarchical models all have similar slopes on galaxy formation scales . hence , the initial conditions of galaxy formation and the resulting bias may be similar . in addition , the hdm model , which has a very different initial power spectrum , has a similar final correlation function , which suggests that the final power on small scales is dominated by non - linear processes which erase initial differences . any velocity bias which arises through dynamical friction should be similar for models evolved to similar clustering levels . if velocity bias is related to galaxy formation sites preferentially near potential wells , then velocity bias could depend on mass density and the efficacy of this measure could be diminished . however , both types of bias appear to be strong only in the most non - linear regions ( @xmath137 ) ( @xcite ) , while our study focuses on the mildly non - linear regimes ( @xmath138 ) . high resolution , hydrodynamic simulations that include detailed galaxy formation should improve our understanding of bias and how it changes from model to model . intuitively , any bias , by eliminating objects in the underdense regions , should have the overall effect of shifting the data points in figure 5 to the left . higher resolution will also incorporate the interactions of dark matter halos , which may significantly effect the three point correlation function on galaxy scales ( see @xcite ) . all of these effects indicate more detailed simulations , beyond what is currently available , may be necessary for the actual application of the redshift dispersion . the next logical step is to attempt to apply @xmath10 to denser redshift surveys than the iras 1.2 jansky survey . in the mean time , additional studies on larger volume , higher resolution n - body simulations would be useful , but perhaps overkill until a suitable redshift survey becomes available . also , it is not clear that using dark matter halos without the proper means for identifying galaxies would add to these results . further exploration should also be done on a wide variety of survey geometries . here , we only looked at a single small sphere . it is quite possible that a larger sphere or a shell might be the optimal shape to balance the tradeoff between high density and large numbers of clusters that make @xmath10 work best . we would like to thank david spergel and our editor ed bertschinger for their helpful comments , and john baumgardner for the use of his icosahedral mesh code . we gratefully acknowledge the grand challenge cosmology consortium and grants of computer time at the san diego and pittsburgh supercomputer centers . mas acknowledges the support of the wm keck foundation , the alfred p. sloan foundation , and nasa astrophysical theory grant nag5 - 2882 . this work was supported by nsf grants asc 9318185 and ast 9108103 . in 2 we showed that the cvt applied to sets of particles provided it holds for individual particles . to prove the latter result we return to peebles original derivation ( see p76 ) . @xcite later rederived the cosmic virial theorem by placing it in the context of the bbgky hierarchy . we choose to use the earlier approach because of its simpler , more intuitive nature . the essential derivation is still that found in p76 , but a few minor modifications have been added to elaborate key steps in the derivation . the derivation of the cvt proceeds in the following manner . first , an equation linking the velocity , acceleration , and correlation function of a general particle system is obtained from the conservation of phase space density . second , the accelerations are linked through gravity to the two and three point correlation functions . third , the assumptions of a cosmological system are applied to the general particle equation to obtain the scaling between the velocity dispersion and the density . the key modification that we have made in order to show that this result applies for individual particles in the system is to replace the entire phase space with the phase space of a single particle . to start , consider a system of particles with the position , velocity , and acceleration of the @xmath27th particle given by @xmath67 , @xmath68 , and @xmath139 . strictly speaking , these are not phase space variables , but represent a three - dimensional single - particle distribution and its evolution in time . now consider any two particles @xmath27 and @xmath70 having relative position , velocity , and acceleration : @xmath140 , @xmath141 , @xmath142 , as shown in figure 8a . the rate of change in the distance , @xmath15 , between the two particles obeys = v_r = / r , and the acceleration = ( + v^2 - ( /r)^2)/r = ( + v_t^2)/r , where @xmath143 . for a sufficiently large system , the number of particles found in a volume of phase space centered on the @xmath27th particle is @xmath144 , where @xmath145 is the phase space density of all pairs connected to the particle @xmath27 . the function @xmath145 is related to the two point correlation function by dr ddf _ = ( 1 + _ ) 4 r^2 dr , where @xmath36 is the mean number density and @xmath146 is the isotropic two point correlation function of the @xmath27th particle , which is defined with respect to the joint probability @xmath147 . for convenience , we now explicitly drop the subscript @xmath148 notation : @xmath149 , @xmath150 , and @xmath151 . it should be understood that all these quantities are relative to the @xmath27th particle , including the velocities and accelerations . having defined the particle system and the phase space density , @xmath152 , it is now possible to look at the time evolution of the particle within the system . the integral of @xmath152 links it to @xmath13 giving the total number of particles with separation between @xmath15 and @xmath153 . the number of particles with separation less than @xmath15 is n(r , t ) = _ 0^r dr ddf . at some later time @xmath154 there exists a @xmath155 such that n(r , t ) = n(r + , t + ) . without loss of generality , we could have just as well started at the point @xmath156 , and chosen @xmath154 and @xmath155 to satisfy n(r - , t ) = n(r , t + ) , which allows us to expand around @xmath15 and @xmath157 . subtracting @xmath158 from the above expression , we have -[n(r , t ) - n(r - , t ) ] = n(r , t + ) - n(r , t ) . the right side can be readily expanded in powers of @xmath154 and is simply @xmath159 . likewise , from the definition of @xmath158 the left side is given by -[n(r , t ) - n(r - , t ) ] = - dd_r-^r dr f. the taylor expansion of the integral gives _ r^r dr f & & + & & f - ^2 f + & & f + ^2 ( f - ^2 f ) , where the last step made use of the following expansion for @xmath155 = + ^2 + , which is valid for _ fixed _ @xmath160 and @xmath161 . matching powers of @xmath154 gives an expression for @xmath162 & = & - dd(- ^2 ) f + & = & - dd f + dd ^2 f + & = & - 4 , where we have used @xmath163 , but @xmath164 . the time derivative of @xmath165 can be written = 4 _ 0^r dr r^2 = 4 _ 0^r dr r^2 , which is combined with the previous expression to give - r^2 = _ 0^r dr r^2 . substituting in the expressions for @xmath160 and @xmath161 , and expanding the @xmath15 derivative results in the general equation derived in p76 r + = + r^-1 _ 0^r dr r^2 . this equation represents the conservation of phase space for a single particle in a system , and is true for any particle in the system with a well defined phase space density and an isotropic two - point correlation function . for gravitationally interacting particles , the acceleration term is simply @xmath166 , where _ i = g _ j m_j . when averaging over phase space , @xmath167 can be written in terms of the two and three point correlation functions , which requires considering a third particle at @xmath168 ( see figure 8b ) . the average force on @xmath27 is the force from @xmath70 plus the force due to @xmath169 weighted by the conditional probability of the third particle being located at @xmath170 & = & + gm + & = & + d^3 [ 1 + ( r ) + ( s ) + ( q ) + ( r , s , q ) ] where @xmath171 is the average mass per particle and the three point correlation function @xmath19 is defined by the probability dp_3 = dv_i dv_j dv_k ^3 [ 1 + ( r ) + ( s ) + ( q ) + ( r , s , q ) ] . the first , second and third terms in the brackets integrate to zero by isotropy leaving = + d^3 [ ( q ) + ( r , s , q ) ] . likewise , the force on @xmath70 is = - - d^3 [ ( s ) + ( r , q , s ) ] . because @xmath172 , the average of the total force is = - - d^3 [ ( q ) + ( r , s , q ) ] . to obtain the scaling of the velocity dispersion we apply various approximations that are consistent with a cosmological model . first , we consider the situation where the correlation function is assumed to be stationary in comparison to the motions of the particles , the time scale for changes in @xmath13 is much greater than the crossing time @xmath173 . in addition , @xmath47 is randomly oriented so that @xmath174 , and @xmath175 . these assumptions eliminate two terms from eq . ( a14 ) . combining with eq . ( a20 ) gives r = - - 6 g d^3 [ ( q ) + ( r , s , q ) ] . the next step is to evaluate the integrals on the right . at this point we impose a model for @xmath13 and @xmath19 . observations indicate that the correlation functions are well modeled by power laws of the form ( r ) = ( ) ^ , = 1.77 , r_0 = 5.4 and , ( r , s , q ) = q[(r ) ( s ) + ( r ) ( q ) + ( s ) ( q ) ] , q 1.0 . using these relationships , the first integral can be solved with the following identities @xmath176 , and @xmath177 , and by noting that in spherical coordinates @xmath178 d^3 ( q ) & = & s^2 ds dd ( q ) + & = & 2 ds d r ( ( r^2 - 2 s r + s^2)^1/2 ) + & = & 2 r^2 dy d ( ( r^2 - 2 y r^2 + y^2 r^2)^1/2 ) + & = & 2 r^2 _ 0^ dy _ 0^ + & = & . where @xmath179 , @xmath180 , and i_(y ) = ( y+1)^2 - \{1 - ( 2- ) y + y^2 } - similar result d^3 ( r , s , q ) & = & d^3 q [ ( r ) ( s ) + ( r ) ( q ) + ( s ) ( q ) ] + & = & q d^3 ( q ) [ ( r ) + ( s ) ] + & = & , where @xmath181 . for our purposes , it is sufficient to know that @xmath182 and @xmath183 are of order unity . inserting the above results into eq . ( a22 ) gives r = - - - . finally , let @xmath184 . dividing through by @xmath15 , and integrating from @xmath15 to @xmath185 and dividing again by @xmath13 , gives an equation for the velocity dispersion ( r ) = + + for small @xmath15 the third term dominates the second . the first term is negligible in the limit that @xmath171 is very small , i.e. , that the mass is divided up into many tiny individual particles . thus the third term dominates , leaving the classic result ( r ) = . using the definition @xmath186 , we can rewrite the above expression as ( r ) = . which , if we return to the earlier notation ( measuring distances in terms of velocity @xmath187 ) gives the expression quoted in eq . ( 4 ) of 2 ( r ) _ ( r ) r^2 , with a proportionality constant c_1 = . bartlett , j. , & blanchard , a. 1996 , , 307 , 1 baumgardner , j.r . , & frederickson , p.o . 1985 , siam j. numer . , 22 , 1107 bertschinger , e. , & gelb , j. 1991 , computers in physics , 5 , 164 brieu , p.p . , summers , f j , & ostriker , j.p . 1995 , , 453 , 566 brown , m. , & peebles , p.j.e . 1987 , , 317 , 588 burstein , d. 1990 , rep . , 53 , 421 couchman , h. , & carlberg , r. 1992 , , 389 , 453 dalcanton , j. , 1995 , ph.d . thesis , princeton university davis , m. , & peebles , p.j.e . 1977 , , 34 , 425 davis , m. , & peebles , p.j.e . 1983 , , 267 , 465 efstathiou , g. , bond , j. r. , & white , s. d. m. 1992 , , 258 , 1 efstathiou , g. , davis , m. , white , s.d.m . , & frenk , c.s . 1985 , , 57 , 241 efstathiou , g. , frenk , c.s . , white , s.d.m . & davis , m. 1988 , , 235 , 715 evrard , a. e. , summers , f j , & davis , m. 1994 , , 422 , 11 fisher , k.b . , davis , m. , strauss , m.a . , yahil , a. , & huchra , j. 1994 , , 267 , 927 fisher , k.b . , huchra , j. , strauss , m.a . , davis , m. , yahil , a. , & schlegel , d. 1995 , , 100 , 69 gelb , j. , & bertschinger , e. 1994 , , 436 , 491 grski , k. m. , ratra , b. , sugiyama , n. , & banday , a. j. 1995 , , 444 , l65 gunn , j.e . , & weinberg , d.h . 1995 , in s.j . maddox and a. aragn - salamanca eds , wide - field spectroscopy and the distant universe ( singapore : world scientific ) , 3 guzzo , l. , fisher k.b . , strauss , m.a . , giovanelli , r. , & haynes , m.p . 1996 , astroph . lett . & comm . , 33 , 231 hockney , r.w . , & eastwood , j.w . 1981 , computer simulation using particles ( new york : mcgraw hill ) holtzman , j. 1989 , , 71 , 1 loveday , j. , efstathiou , g. , maddox , s. j. , & peterson , b.a . 1996 , , 468 , l1 loveday , j. , peterson , b.a . , efstathiou , g. , & maddox , s. j. 1992 , , 390 , 338 marzke , r.o . , geller , m.j . , da costa , l.n . , & huchra , j.p . 1995 , , 110 , 477 mo , h.j . , jing , y.p . , & brner , g. 1993 , , 264 , 825 okumura , s. k. , et al . 1993 , pasj , 45 , 329 peebles , p.j.e . 1976a , , 45 , 3 ( p76 ) peebles , p.j.e . 1976b , , 205 , l109 peebles , p.j.e . 1980 , the large - scale structure of the universe ( princeton : princeton university press ) sandage , a. 1986 , , 307 , 1 somerville , r.s . , davis , m. , & primack , j.r . 1997 , , 479 , 616 stompor , r. , grski , k. m. , & banday , a. j. 1995 , , 277 , 1225 strauss , m.a . , cen , r.y . , & ostriker , j.p . 1993 , , 408 , 389 strauss , m.a . , & willick , j.a . 1995 , , 261 , 271 summers , f j , davis , m. , & evrard , a. e. 1995 , , 454 , 1 willick , j.a . , strauss , m.a . , dekel , a. , & kolatt , t. 1997 , , in press ( astro - ph/9612240 ) zurek , w.h . , quinn , p.j . , salmon , j.k . , & warren , m.s . , 1994 , , 431 , 559

in the overwhelming majority of the literature on supersymmetric extensions of the standard model , @xmath0-parity is imposed as an ad - hoc symmetry to avoid phenomena such as proton decay which has been very strongly constrained by experiment . however , the proton can remain exactly stable even in the presence of a restricted set of @xmath0-parity violating interactions . in fact , one needs to break both baryon and lepton number for the proton to decay , because the proton is the lightest fermion that carries zero lepton number but nonzero baryon number . various low energy constraints on @xmath0-parity violating couplings have been studied in depth ( see @xcite and references therein ) , however certain aspects of collider phenomenology in the presence of @xmath0-parity violating couplings have not received the same amount of attention compared to the case of exact @xmath0-parity @xcite . in particular , there are two major differences between @xmath0-parity conserving and @xmath0-parity violating supersymmetric models that we wish to focus on in this paper that relate to the signatures at hadron colliders . firstly , @xmath0-parity violation affects production mechanisms by allowing for supersymmetric partner particles to be resonantly produced @xcite . secondly , the lightest supersymmetric particle need no longer be stable , and can decay promptly . since most experimental searches for supersymmetry focus on the presence of missing transverse energy carried off by a neutral lightest superpartner , this requires a major change in search strategies . naturally , in order to account for dark matter , a supersymmetric theory with @xmath0-parity violation would need to have additional matter content with a symmetry that can stabilize the lightest particle in the dark sector . finally , the fact that the lightest supersymmetric particle is no longer stable allows for spectra that are phenomenologically disfavored in models with exact @xmath0-parity and a colored superpartner can be the lightest one in the spectrum@xcite . in the rest of this paper we wish to study some of the phenomenological consequences associated with @xmath0-parity violation with relatively light scalar superpartners , such as the stop . uv motivated models of @xmath0-parity violation have been studied in the context of msugra and grand unified theories @xcite . there also exist motivated models of new physics that are qualitatively different than the mssm , in which @xmath0-parity is violated . an example is provided by the setup of ref . @xcite , where supersymmetry is broken at high energies but reemerges as an accidental symmetry at low energies , and @xmath0-parity violation is unavoidable @xcite . very recently , the cdf collaboration has announced an interesting observation of a resonance in dijets produced in association with a w boson @xcite . the significance of this observation is @xmath4 using data with @xmath5 of integrated luminosity , and the mean value for the resonant dijet mass distribution is around 145 gev . in this paper we will investigate the possibility that the cdf observation arises from the production and decay of new particles , and explore how this final state can arise in models of supersymmetry with @xmath0-parity violation . the topology we consider is characterized by the production of a primary resonance that undergoes a charged - current transition to a secondary resonance by emitting a w - boson , with the secondary resonance subsequently decaying to two jets . for a supersymmetric realization of this topology , we will concentrate on the possibility of squarks or sleptons and sneutrino being the lightest ( s)particles in the spectrum . we give a representative model in which the primary resonance is a bottom squark ( sbottom ) , and the secondary resonance a top squark ( stop ) , the production and decay of which occur through @xmath0-parity violating interactions respectively . we will show that only one or two couplings beyond the mssm are needed for this topology to be consistent with the cdf observation . we will also show that the required couplings can have large enough values without causing conflict with any existing constraints , and we will furthermore argue that the @xmath1 channel is in fact naturally the first place for the new physics to be observed in this scenario . in the next section we present our sbottom - stop model , estimate the size of the necessary @xmath0-parity violating couplings and show that these are not in conflict with any existing bounds . we will then go through various production and decay possibilities and argue why the new physics would appear first in the @xmath1 channel , and evaluate the discovery potential in other final states for the tevatron as well as the lhc . in section [ sec : sleptons ] we give an alternative setup with sleptons instead of squarks that can give rise to similar collider signatures and highlight the differences in phenomenology compared to the stop - sbottom model . it should be emphasized that the most important aspect of the scenarios presented here are the production and decay topologies . while the language of supersymmetry is utilized throughout , the topologies and signatures may be more widely applicable to other frameworks for the underlying physics , including ones with particles of different spin and gauge quantum numbers . the @xmath0-parity violating terms in the superpotential are typically parameterized as @xmath6 where @xmath7 are flavor indices . the first three types of terms violate lepton number while the @xmath8 type terms violate baryon number . among the available @xmath0-parity violating couplings , @xmath8 type terms are generically less constrained than @xmath9 or @xmath10 terms , since searches for new physics involving leptons typically have higher sensitivity . the main constraints on @xmath8 type terms come from neutron - antineutron oscillations and flavor violation in the quark sector@xcite . these constraints can be avoided however , as we will outline below , and for certain choices of @xmath11 , @xmath12 can be of order one . for resonant production we will need at least one coupling to be reasonably large , and this coupling should preferably involve two light quarks in order to give rise to interesting production cross sections . note that the @xmath8 type terms are antisymmetric in the indices @xmath13 and @xmath14 . therefore if two of the indices are to be 1 , the only choices are @xmath15 and @xmath16 , both of which are severely constrained from nuclear decays involving two neutrons@xcite . the next best possibility for resonant production is to make use of a strange quark pdf in the proton , and the there are several couplings of this type that are essentially unconstrained . note however that there are many additional constraints on products of couplings and therefore the safest choice is to turn on as few @xmath0-parity violating terms as possible . below , we will outline a model where the stop and sbottom are the main players for the phenomenology , and where only @xmath17 and @xmath18 is nonzero ( but need not be large ) . at the end of this section we will point to an alternative choice with only @xmath19 turned on and argue that given a similar spectrum , the collider signatures are very similar . as a function of the sbottom mass . for this plot , the mass eigenstate is taken to be pure @xmath20 . ] with @xmath21 turned on , sbottoms can be resonantly produced according to @xmath22 . the cross section for this process is plotted in figure [ fig : rpvmodel ] for a value of @xmath23 . if the sbottom happens to be the lightest state , then it will decay back through the @xmath0-parity violating coupling to a pair of jets . depending on the sbottom mass , this process is constrained by dijet resonance searches from ua2 @xcite ( light sbottoms ) , at the tevatron @xcite ( intermediate mass sbottoms ) and recent constraints from the lhc @xcite for heavy sbottoms . in figure [ fig : lambdabound ] we use the cross section constraints to derive bounds on @xmath21 , or equivalently on @xmath24 , the partial width of the sbottom to @xmath25+@xmath26 . the exact relation between @xmath24 and @xmath21 is given in equation ( [ eqn : partialwidths ] ) . in terms of @xmath21 we find that the constraint from resonant production never goes below @xmath27 . the pair production of sbottoms through qcd has a threshold at twice the sbottom mass , and therefore the cross section for this process is much lower than resonant production . with no other supersymmetric particles lighter than the sbottom , this process would manifest itself in a four jet final state , we will however introduce a different choice of spectrum shortly . it is however interesting to think of the pair production of the lightest colored supersymmetric particle , and we will come back to this point in the next chapter . from the dijet resonance searches at ua2 and cdf . the numbers plotted here are conservative in the sense that we assume perfect acceptance and no suppression from left - right sbottom mixing . in terms of @xmath21 , the bound is always above 0.4 . ] the collider prospects for this very minimal ( s)particle content are bleak for a value of @xmath21 below the bound . the presence of a neutralino below the mass of the sbottom does not lead to much improvement . the sbottom would decay as @xmath28 , the neutralino then decaying through an off - shell sbottom as @xmath29 , thus populating the @xmath30 final state , which has a large background . if there are charginos significantly lighter than the sbottom , this can lead to top quarks in the final state , which is potentially a more optimistic scenario , however the large mass gap necessary for decaying to on - shell tops forces the sbottoms to be heavier , reducing the cross section as a function of the lightest superpartner mass . a possibility that is much more interesting collider - wise , and the one we wish to consider in the rest of this section is the presence of a stop below the sbottom mass . in this case the sbottom can decay as @xmath31 . this mass spectrum is favored by the rg evolution of the squark masses , as the large yukawa couplings tend to drive the stop mass lighter . note that since it is @xmath20 that is resonantly produced , for the w channel to dominate there needs to be some mixing between @xmath20 and @xmath32 . as long as the left handed squarks are not decoupled however , the w decay mode can be dominant . the partial widths of the sbottom into dijets and into @xmath33+@xmath34 are given by @xmath35 where @xmath36 is the triangle function and @xmath37 denote the mixing angles in the squark sector ( for @xmath38 , @xmath39 and likewise for the stop ) . in figure [ fig : sbottombf ] we plot the branching fraction of @xmath31 for @xmath40 as a function of the sbottom mass , taking @xmath41gev . with only @xmath21 nonzero , the stop will have a three - body decay through on off - shell sbottom . note however that the stop can also have two - body decays if there are additional nonzero @xmath0-parity violating couplings , specifically a nonzero value of @xmath42 . note however that neutron - antineutron mixing constrains @xmath43 and @xmath44 to be of order @xmath45 while the constraints are much weaker on @xmath18@xcite . note that the product of @xmath21 and @xmath18 is also not further constrained , and that stop production from a small value of @xmath18 is unobservable due to the pdfs for a @xmath26-@xmath46 initial state . the decay modes of the stop are illustrated in figure [ fig : stopdecays ] . with @xmath18 turned on , the two - body decay of the stop into jets can dominate , and the first signal to be seen would appear in the @xmath47 channel . this is the channel where cdf recently reported observing an interesting signal@xcite . for this specific choice of couplings , one of the jets from the decay of the stop would be a b - jet . currently the cdf analysis states that the heavy flavor content of the excess region is consistent with the sideband . however the analysis does not provide a quantitative measure for whether the existence of a single b - jet is disfavored . as we will describe in section [ sec : scharm ] , the same topology can be realized with a different choice of nonzero @xmath12 such that the final state has no heavy flavor jets . the cdf analysis reports no significant deviation from the standard model expectation in an analogous channel with a dijet resonance produced in association with a z - boson . note that in our minimal setup , this channel is expected to be absent . to @xmath33+@xmath34 . for this plot @xmath41gev , and @xmath48 . ] in order to study the acceptance of the cdf analysis for the @xmath49 signal we have performed a monte carlo study . it is not trivial to generate monte carlo events for this setup due to the completely antisymmetric way that the color indices are contracted in the @xmath0-parity violating vertex . while certain programs such as herwig incorporate aspects of @xmath0-parity violating physics , this is not adequate for our purposes and we have generated @xmath50 events by using cteq parton distribution functions @xcite and integrating over phase space using a custom - made code . we then exported the events into lhe format @xcite by modifying the color flow to remove the antisymmetric color contractions , which can not be faithfully represented in the lhe format . while this may introduce subtle changes at the level of non - perturbative qcd effects , we expect that the impact on the hard scattering observables will be negligible . the events were passed through pythia@xcite for showering and hadronization , and later through pgs@xcite for detector effects . we used the standard cdf parameter set in pgs with a cone jet algorithm and a cone size @xmath51 . we analyzed the signal events using cuts designed to mimic the cdf event selection cuts , namely : * the presence of an electron ( muon ) with @xmath52 ( @xmath53 ) @xmath54gev and @xmath55 and the absence of additional leptons in the event . * missing transverse energy in excess of 25 gev . * the presence of exactly two jets with @xmath56gev and @xmath57 . * a minimum @xmath58 of 0.52 between the lepton and the nearest jet . * a minimum @xmath53 of 40 gev for the dijet system . * a transverse mass in excess of 30 gev for the lepton - neutrino system , where the neutrino is taken to be the only source of missing transverse energy in the event . * an azimuthal angle in excess of 0.4 radians between the missing energy and either jet . * @xmath59 for the two jets . taking @xmath60gev , @xmath41gev and @xmath48 , we find that the number of events observed at cdf in the @xmath61+@xmath62 channels is consistent with @xmath63 , giving a total production cross section of @xmath64pb , a signal acceptance of @xmath65 which includes a 52% branching fraction for @xmath66 ( we take the branching fraction of @xmath67 to be 1 ) . this value of @xmath21 is far below the bound from dijet resonance searches . we plot the dijet invariant mass distribution from the stop decay in our sample in figure [ fig : mjj_300 ] . and @xmath62 channels combined ) using the same cuts as the cdf analysis , with @xmath60gev and @xmath41gev . @xmath63 and @xmath48 are used in order for the number of events after cuts to agree with the cdf observation with 4.3 fb@xmath68 of integrated luminosity . ] another interesting question that is reported to be consistent with the background hypothesis but not ruled out by the cdf analysis is whether a primary resonance can be present in the @xmath1 system . in the topology considered here , this system will reconstruct the sbottom mass . in figure [ fig : mwjj_300 ] we plot the reconstructed sbottom mass in our model for a true value of @xmath60gev using both solutions for the neutrino rapidity in reconstructing the w - mass ( and discarding the event if no real solutions can be found ) . note that while the stop mass is determined by the position of the observed dijet excess , the sbottom mass is an adjustable parameter in this model , and for lighter sbottom masses , the excess in @xmath69 may be difficult to resolve on top of background . a definitive comparison of the primary resonance aspect of this topology with the cdf observation would require a full detector simulation of both signal and backgrounds , as well as the cdf data - driven estimates of background . system from the sbottom decay ( @xmath61 and @xmath62 channels combined ) using the same cuts as the cdf analysis , with @xmath60gev and @xmath41gev . both solutions for the neutrino momentum are included in the distribution . @xmath63 and @xmath48 are used in order for the number of events after cuts to agree with the cdf observation with 4.3 fb@xmath68 of integrated luminosity . ] while in our topology the resonant production of any up - type squark is strongly suppressed due to the smallness of the couplings @xmath70 , the pair production of the stop is determined by qcd and can be an interesting production channel . in figure [ fig : stoppairxsecn ] we show the stop pair production cross section as a function of its mass calculated by using the batch mode of calchep@xcite . this production channel would manifest itself in the four jet channel , with two pairs of jets reconstructing narrow resonances of the same mass . @xcite studied this channel with the conclusion that stops with up to @xmath71 can be discovered . references @xcite also studied the same event topology at the tevatron and the lhc for various scenarios of underlying physics , and showed that there is strong discovery potential . currently no search for new physics has been performed in this channel . note that for a stop mass of 150gev , a search in the four jet channel analysis will be difficult at the lhc because of the larger trigger thresholds which will have low efficiency on the signal . the tevatron has a unique advantage for the case of light stops , and an analysis of existing data could be used to explore this possibility . @xmath0-parity violating interaction at the lhc . ] the initial state for resonant anti - squark production through @xmath0-parity violating interactions is quark quark ( rather than quark anti - quark ) . in the scenario described here , with production through the @xmath21 coupling , one of the initial state quarks is a strange - quark . at the tevatron with @xmath72-@xmath73 collisions , as well as at the lhc with @xmath72-@xmath72 collisions , the parton distribution functions for the leading resonant production are therefore of the valence sea type . aside from the relative center of mass energies , there is no qualitative difference for the overall production rates between the tevatron and lhc for such a resonance . however , the relative rates for sbottom versus anti - sbottom are significantly different . at the tevatron sbottom and anti - sbottom are produced at equal rates because of the charge symmetry of the colliding particles . in contrast , at the lhc , production of anti - sbottom arising from valence sea quark collisions is enhanced compared with sbottom production arising from sea sea anti - quark collisions . this leads to a significant charge asymmetry in the @xmath1 signal at the lhc , as shown in figure [ fig : chargeasymmetry ] . a charge asymmetry is also present in the w+jets background due to the difference between the @xmath25 and @xmath74 valence quark parton distribution functions . however , the w+jets background asymmetry arises predominantly from an order one difference in the relative magnitude of two types of valence sea quark collisions , whereas the signal asymmetry is much larger since it arises from a difference between valence sea versus sea sea collisions . the enhanced charge asymmetry of the @xmath1 signal may provide an additional handle at the lhc to isolate this supersymmetric scenario for the topology of a primary and secondary resonance with a charged - current transition . finally , one can also consider the pair production of the sbottom through qcd . in the scenario with a lighter stop , this channel will go to @xmath75 . while the combinatoric background can be reduced by reconstructing the stops as well as demanding equal sbottom masses , the cross section is small compared to the irreducible standard model backgrounds of @xmath76+jets and @xmath77+jets in the dileptonic decay channel , and the additional background of w+jets for the semileptonic case . as one can see in figure [ fig : stoppairxsecn ] , even for moderately light sbottom masses , the cross section for this process is too low at the tevatron . the leptonic branching fraction of @xmath34 bosons and acceptance effects make it unlikely that a statistically significant excess can be observed at the 7 tev lhc as well ( assuming a total integrated luminosity of @xmath78 ) . this channel may become interesting for the 14 tev lhc with high statistics however . type couplings may be induced at 1-loop from existing ones . the black box denotes a @xmath79 coupling that is nonzero at tree level and the circles represent off diagonal couplings in flavor space . ] we wish to briefly comment on the stability of the choice of nonzero @xmath12 that we use . for instance , the 1-loop diagram in figure [ fig : flav ] shows how an existing @xmath8 coupling may induce additional @xmath0-parity violating couplings in the presence of neutralinos that couple off - diagonally in flavor space . this can be avoided however if all three generations are nearly - degenerate . for the model we outlined in this paper , one can imagine a mass spectrum such that all three generations are present , with @xmath80 . the presence of the additional squarks does not significantly change the phenomenology that we discussed . with only @xmath21 and @xmath18 turned on , the @xmath81 are not resonantly produced at observable rates . they are pair produced , enhancing the cross section by a factor of three compared to that of a single stop in the four jet channel . resonant production of the additional @xmath82 is also suppressed by the pdf s . therefore the phenomenology is essentially identical to the minimal scenario with only a sbottom and a stop . if the left - handed squarks are light enough to be produced , new decay channels with on - shell higgs or z bosons may open up . other operators induced by a diagram similar to figure [ fig : flav ] but with a w in the loop rather than a neutralino are nonrenormalizable , and are small for our choice of nonzero @xmath79 since they are suppressed by quark masses and ckm angles . we do not consider operators induced at higher loops . there is a second choice for which nonzero @xmath79 can give rise to the cdf excess . if @xmath19 is turned on , then both @xmath83 and @xmath84 can be resonantly produced from @xmath74-@xmath85 and @xmath74-@xmath26 initial states , respectively . this choice has the advantage that @xmath19 will also induce decays of both the @xmath83 and the @xmath84 , so no second @xmath12 need be turned on . there are several disadvantages however . since both @xmath83 and @xmath84 are now resonantly produced through the same coupling , the ua2 dijet resonance bound on the lighter state ( assumed to have a mass of 150 gev ) will directly apply to the production cross section of the heavier state , and therefore on the cross section for the @xmath1 decay mode . furthermore , the branching fraction for the w - mode depends on the left - right squark mixing and may be much smaller for the second generation squarks compared to @xmath86 and @xmath33 . finally , the production cross section for a @xmath74-@xmath26 initial state is lower than that of the @xmath25-@xmath26 initial state by a factor of order one . since this affects the bounds from dijet resonances in the same way as the production cross section however , one can simply compensate by using a larger value of @xmath79 as long as one is not already at the upper limit . zero mass mixing is assumed . from bottom to top , the three pairs of curves are for the tevatron , 7 tev lhc and 14 tev lhc respectively . ] for resonant slepton and sneutrino production from ua2 and cdf with only @xmath87 turned on . the numbers plotted here are conservative in the sense that we assume perfect acceptance and we neglect left - right mixing . the bound on @xmath87 never goes below 0.3 for the sneutrino and below 0.2 for the slepton . ] @xmath0-parity violation through @xmath8-type terms is not the only possibility to resonantly produce superpartners at a hadron collider . one can also consider the resonant production of left - handed sleptons through a @xmath9 type coupling . the @xmath88 couplings necessary in this scenario are generically more constrained that the @xmath8 couplings @xcite . while we will not attempt to build a detailed model as in the squark case , we remark that a similar final state can be obtained . resonant production of a left - handed slepton , with charged - current decay through an off - shell w@xmath89 to the sneutrino , which subsequently decays back through the @xmath0-parity violating coupling , gives an @xmath90jj signature . the main difference with a left - handed @xmath91 doublet only , is that the mass splitting between the slepton and sneutrino arises only from electroweak symmetry breaking and is parametrically small . the leptons in this scenario are then significantly softer than those arising from an on - shell w topology . the production cross section and the bound from dijet resonance searches for this scenario are plotted in figures [ fig : sleptonxsec ] and [ fig : sleptonbound ] respectively . rather than go through all the details of the phenomenology again , we highlight here the differences compared to the squark case . since many @xmath88 are constrained through lepton flavor violating processes , there are few choices in which couplings can be turned on to match the cdf @xmath1 observation . one possibility is resonant stau production from a nonzero value of @xmath87 . scenarios of @xmath0-parity violation with @xmath92 and @xmath93 lsp have been considered in msugra scenarios ( see for example @xcite , @xcite and references therein ) . as in the squark case , there is a charge asymmetry in the resonant production of the slepton vs. the anti - slepton . unlike the squark case however , the origin of this asymmetry is the difference between the @xmath25 and @xmath74 valence quark pdf s . therefore the asymmetries of the signal and the w+jets background will be of the same size . additionally , unlike in the case of squarks , sleptons are only pair produced through electroweak interactions , so pair production cross sections will be much smaller , and only @xmath94-@xmath95 initial states can contribute , which is a disadvantage for the lhc compared to the squark case where the @xmath96-@xmath96 initial state contributes . if both right- and left - handed sleptons are light enough to be produced , then the same production and decay topology discussed above can arise with resonant production of the heavier states that charged - current decay through an on - shell w - boson to the lighter states which subsequently decay back to two jets . since in @xmath9-type @xmath0-parity violation both resonant production as well as the charged - current coupling involve the left - handed component of sleptons , significant left - right mixing is required in order for both the cross section as well as the w branching fraction to be sizeable in this case . if kinematically open , decay modes with on - shell higgs and @xmath97-bosons also arise in this scenario . we have explored a production and decay topology for hadron colliders in which a primary resonance undergoes a charged - current transition to a secondary resonance that subsequently decays to dijets . supersymmetric scenarios with @xmath0-parity violation where a light sbottom is resonantly produced and decays to a w - boson and stop , which decays to a pair of jets through the @xmath0-parity violating coupling , give a concrete realization of this topology . other possibilities for underlying theoretical frameworks that could give topologies of this type include technicolor @xcite ( for reviews of phenomenology see @xcite ) , two - higgs - doublet models ( for a detailed overview see @xcite ) , and excited quarks @xcite . we have argued that with such a topology , the new physics may be observed first in the @xmath1 channel , possibly consistent with the recent cdf observation in this channel . we have also pointed to other final states that would be associated with this scenario . in particular , for strongly interacting resonances , other important topologies are pair production of the secondary state giving rise to a 4 jet signature , as well as pair production of the primary state giving rise to a w@xmath2w@xmath3 + 4 jets signature , both with two equal mass dijet resonances within the 4 jets . we argued that due to the smallness of the mass of the secondary state , the former would most effectively be searched for at the tevatron , while the latter has better prospects at the upgraded lhc . we have outlined how a less minimal supersymmetric model can be set up with all three generations of squarks present , without significantly changing the collider phenomenology , but forbidding additional @xmath0-parity violating couplings to be induced at loop - level from existing ones . we have also shown that there is an alternative choice of couplings for the squark scenario , where the phenomenology is focused on the second generation squarks rather than the third . this alternative model needs only one @xmath0-parity violating coupling to be nonzero , however the branching fraction for the w mode may be small unless there is significant left - right mixing in the squark sector . finally we have outlined an alternative supersymmetric scenario with the same production and decay topology , but where the new particles are sleptons rather than squarks . many additional production channels such as pair production through qcd are absent in this case . to conclude , we believe that even if the cdf @xmath1 observation proves not to be due to new physics , that collider signatures of @xmath0-parity violating supersymmetric models should be further explored by both theorists and experimental searches . * note added : * after this paper was submitted , several developments took place . cdf updated their analysis with a @xmath98 data set in which the significance of the excess was reported as @xmath99 after accounting for systematic uncertainties @xcite . in this analysis , various other consistency checks were performed on the analysis , such as using different monte carlo generators , making the jet cuts inclusive ( rather than demanding exactly two jets ) , varying the jet energy scale and some of the analysis cuts . the excess was reported as being robust under these variations . this is true of the signal in our scenario as well , the effect of these variations to the acceptance of our monte carlo sample is mild . there is however a result in the updated cdf analysis that has a larger impact on our scenario , namely the heavy flavor content in the jets accompanying the @xmath34-boson has been reported quantitatively . this result is difficult to reconcile with one of the two jets being a b - jet in each event , and therefore disfavors our sbottom - stop benchmark . note however that both the second generation squark benchmark as well as the slepton benchmark which we proposed produce light jets in the final state and are therefore still consistent with the cdf update . another dramatic development was the release of a d search for new physics in the same final state @xcite . this analysis reports that d does not observe an excess consistent with the cdf result , and for a gaussian centered at the same mass they obtain a best fit for the cross section of @xmath100 . considering that the d analysis uses cuts very similar to the cdf analysis , it is not at the moment clear why the two experiments obtain such drastically different results . while this development casts doubt on the existence of new physics in this final state , it is also possible that there is a signal with a cross section smaller than the one reported by cdf but still consistent with the d result , for example around @xmath101 . if that were to be the case , the signal in our scenario could easily have a smaller cross section by using a higher mass for the first resonant squark or slepton , or by having smaller @xmath0-parity violating couplings . as we remarked at the end of our conclusions , as theorists we believe this topology for the resonant production of supersymmetric particles to be an interesting one that should be studied further even if the eventual resolution of the discrepancy between cdf and d results disfavors a new physics explanation . we also wish to mention a variety of other new physics explanations that were proposed to explain the cdf excess after the submission of this paper . these include technicolor models @xcite , new gauge bosons @xcite , extended higgs sectors @xcite and other possibilities @xcite . since technicolor is arguably the most commonly associated type of new physics associated with the @xmath1 final state , we want to point to a few phenomenological differences between our scenario and a technicolor model , which could be used in order to distinguish the two signals . for one , since a techni - pion is expected to couple to sm fermions proportionally to their mass , it is expected to decay preferentially to @xmath102 pairs . in our scenario , the alternative squark and slepton benchmarks would only give rise to light jets in the final state whereas the sbottom - stop benchmark would produce a single @xmath46-jet . as mentioned earlier in this note , the cdf update appears to disfavor the existence of @xmath46-tagged jets in the excess . another possibility that can distinguish between our scenario and a technicolor model is to look for pair production of the primary resonance . in both our sbottom - stop as well as our alternative squark benchmarks , both the heavier as well as the lighter squarks are expected to be pair produced from qcd processes , giving rise to a @xmath103 final state for the lighter squark and a @xmath104 final state for the heavier squark , as we mentioned at the end of section [ sec : squarks ] , whereas techni - rhos and techni - pions in minimal technicolor models do not carry color charge and will not be pair produced in the same way . note however that pair production is not a good discriminant if the resonantly produced superpartners are sleptons which also do not carry color . r. barbier _ et al . _ , phys . rept . * 420 * , 1 ( 2005 ) [ arxiv : hep - ph/0406039 ] . s. dawson , e. eichten , c. quigg , phys . rev . * d31 * , 1581 ( 1985 ) ; h. baer , x. tata , cambridge , uk : univ . pr . ( 2006 ) 537 p ; k. nakamura _ et al . _ [ particle data group collaboration ] , j. phys . g * g37 * , 075021 ( 2010 ) . s. dimopoulos , r. esmailzadeh , l. j. hall and g. d. starkman , phys . d * 41 * , 2099 ( 1990 ) . h. k. dreiner and g. g. ross , nucl . b * 365 * , 597 ( 1991 ) . u. sarid and s. d. thomas , phys . lett . * 85 * , 1178 ( 2000 ) [ arxiv : hep - ph/9909349 ] . n. sakai and t. yanagida , nucl . b * 197 * , 533 ( 1982 ) . l. j. hall and m. suzuki , nucl . b * 231 * , 419 ( 1984 ) . d. e. brahm and l. j. hall , phys . d * 40 * , 2449 ( 1989 ) . r. hempfling , nucl . b * 478 * , 3 ( 1996 ) [ arxiv : hep - ph/9511288 ] . a. y. smirnov and f. vissani , nucl . b * 460 * , 37 ( 1996 ) [ arxiv : hep - ph/9506416 ] . k. tamvakis , phys . b * 382 * , 251 ( 1996 ) [ arxiv : hep - ph/9604343 ] . k. tamvakis , phys . b * 383 * , 307 ( 1996 ) [ arxiv : hep - ph/9602389 ] . r. barbieri , a. strumia and z. berezhiani , phys . b * 407 * , 250 ( 1997 ) [ arxiv : hep - ph/9704275 ] . g. f. giudice and r. rattazzi , phys . b * 406 * , 321 ( 1997 ) [ arxiv : hep - ph/9704339 ] . h. k. dreiner , arxiv : hep - ph/9707435 . m. a. diaz , j. c. romao and j. w. f. valle , nucl . phys . b * 524 * , 23 ( 1998 ) [ arxiv : hep - ph/9706315 ] . b. c. allanach , a. dedes and h. k. dreiner , phys . d * 69 * , 115002 ( 2004 ) [ erratum - ibid . d * 72 * , 079902 ( 2005 ) ] [ arxiv : hep - ph/0309196 ] . r. sundrum , jhep * 1101 * , 062 ( 2011 ) [ arxiv:0909.5430 [ hep - th ] ] . r. sundrum , private communication . t. aaltonen _ et al . _ [ cdf collaboration ] , phys . d * 79 * , 112002 ( 2009 ) [ arxiv:0812.4036 [ hep - ex ] ] . g. aad _ et al . _ [ atlas collaboration ] , arxiv:1103.3864 [ hep - ex ] . v. khachatryan _ et al . _ [ cms collaboration ] , phys . lett . * 105 * , 211801 ( 2010 ) [ arxiv:1010.0203 [ hep - ex ] ] . j. pumplin , d. r. stump , j. huston , h. l. lai , p. m. nadolsky and w. k. tung , jhep * 0207 * ( 2002 ) 012 [ arxiv : hep - ph/0201195 ] . e. boos _ et al . _ , arxiv : hep - ph/0109068 . t. sjostrand , s. mrenna and p. z. skands , `` pythia 6.4 physics and manual , '' jhep * 0605 * , 026 ( 2006 ) [ arxiv : hep - ph/0603175 ] . j. conway _ et al . _ , `` pgs 4 : pretty good simulation of high energy collisions , '' 2006 , www.physics.ucdavis.edu/@xmath105conway/research/software/pgs/pgs4-general.htm a. pukhov _ et al . _ , arxiv : hep - ph/9908288 ; a. pukhov , arxiv : hep - ph/0412191 . d. choudhury , m. datta and m. maity , phys . d * 73 * , 055013 ( 2006 ) [ arxiv : hep - ph/0508009 ] . r. s. chivukula , m. golden and e. h. simmons , nucl . b * 363 * , 83 ( 1991 ) . b. a. dobrescu , k. kong and r. mahbubani , phys . b * 670 * , 119 ( 2008 ) [ arxiv:0709.2378 [ hep - ph ] ] . c. kilic , t. okui and r. sundrum , jhep * 0807 * , 038 ( 2008 ) [ arxiv:0802.2568 [ hep - ph ] ] . c. kilic , s. schumann and m. son , jhep * 0904 * , 128 ( 2009 ) [ arxiv:0810.5542 [ hep - ph ] ] . y. bai and b. a. dobrescu , arxiv:1012.5814 [ hep - ph ] . k. desch , s. fleischmann , p. wienemann , h. k. dreiner and s. grab , phys . d * 83 * , 015013 ( 2011 ) [ arxiv:1008.1580 [ hep - ph ] ] . h. k. dreiner and s. grab , phys . b * 679 * , 45 ( 2009 ) [ arxiv:0811.0200 [ hep - ph ] ] . l. susskind , `` dynamics of spontaneous symmetry breaking in the weinberg - salam theory , '' phys . rev . d * 20 * , 2619 ( 1979 ) ; s. weinberg , `` implications of dynamical symmetry breaking , '' phys . d * 13 * , 974 ( 1976 ) ; s. weinberg , `` implications of dynamical symmetry breaking : an addendum , '' phys . d * 19 * , 1277 ( 1979 ) . e. farhi and l. susskind , `` technicolor , '' phys . rept . * 74 * , 277 ( 1981 ) ; c. t. hill and e. h. simmons , `` strong dynamics and electroweak symmetry breaking , '' phys . rept . * 381 * , 235 ( 2003 ) [ erratum - ibid . * 390 * , 553 ( 2004 ) ] [ arxiv : hep - ph/0203079 ] . k. lane and s. mrenna , `` the collider phenomenology of technihadrons in the technicolor straw man model , '' phys . d * 67 * , 115011 ( 2003 ) [ arxiv : hep - ph/0210299 ] . g. h. brooijmans _ et al . _ , `` new physics at the lhc : a les houches report . physics at tev colliders 2007 new physics working group , '' arxiv:0802.3715 [ hep - ph ] . j. f. gunion , h. e. haber , g. l. kane and s. dawson , front . phys . * 80 * , 1 ( 2000 ) . h. terazawa , m. yasue , k. akama and m. hayashi , phys . b * 112 * , 387 ( 1982 ) . f. m. renard , nuovo cim . a * 77 * , 1 ( 1983 ) . a. de rujula , l. maiani and r. petronzio , phys . b * 140 * , 253 ( 1984 ) . m. r. buckley , d. hooper , j. kopp , e. neil , phys . rev . * d83 * , 115013 ( 2011 ) . [ arxiv:1103.6035 [ hep - ph ] ] ; f. yu , phys . rev . * d83 * , 094028 ( 2011 ) . [ arxiv:1104.0243 [ hep - ph ] ] ; x. -p . wang , y. -k . wang , b. xiao , j. xu , s. -h . zhu , phys . * d83 * , 117701 ( 2011 ) . [ arxiv:1104.1161 [ hep - ph ] ] ; k. cheung , j. song , phys . lett . * 106 * , 211803 ( 2011 ) . [ arxiv:1104.1375 [ hep - ph ] ] ; a. e. nelson , t. okui , t. s. roy , [ arxiv:1104.2030 [ hep - ph ] ] ; s. jung , a. pierce , j. d. wells , [ arxiv:1104.3139 [ hep - ph ] ] ; p. ko , y. omura , c. yu , [ arxiv:1104.4066 [ hep - ph ] ] ; p. j. fox , j. liu , d. tucker - smith , n. weiner , [ arxiv:1104.4127 [ hep - ph ] ] ; d. -w . jung , p. ko , j. s. lee , [ arxiv:1104.4443 [ hep - ph ] ] ; s. chang , k. y. lee , j. song , [ arxiv:1104.4560 [ hep - ph ] ] ; k. s. babu , m. frank , s. k. rai , [ arxiv:1104.4782 [ hep - ph ] ] ; z. liu , p. nath , g. peim , phys . lett . * b701 * , 601 - 604 ( 2011 ) . [ arxiv:1105.4371 [ hep - ph ] ] ; j. l. hewett , t. g. rizzo , [ arxiv:1106.0294 [ hep - ph ] ] ; a. e. faraggi , v. m. mehta , [ arxiv:1106.5422 [ hep - ph ] ] . q. -h . cao , m. carena , s. gori , a. menon , p. schwaller , c. e. m. wagner , l. -t . wang , [ arxiv:1104.4776 [ hep - ph ] ] ; b. dutta , s. khalil , y. mimura , q. shafi , [ arxiv:1104.5209 [ hep - ph ] ] ; c. -h . chen , c. -w . chiang , t. nomura , y. fusheng , [ arxiv:1105.2870 [ hep - ph ] ] . x. -p . wang , y. -k . wang , b. xiao , j. xu , s. -h . zhu , phys . rev . * d83 * , 115010 ( 2011 ) . [ arxiv:1104.1917 [ hep - ph ] ] ; r. sato , s. shirai , k. yonekura , phys . * b700 * , 122 - 125 ( 2011 ) . [ arxiv:1104.2014 [ hep - ph ] ] ; l. a. anchordoqui , h. goldberg , x. huang , d. lust , t. r. taylor , phys . * b701 * , 224 - 228 ( 2011 ) . [ arxiv:1104.2302 [ hep - ph ] ] ; b. a. dobrescu , g. z. krnjaic , [ arxiv:1104.2893 [ hep - ph ] ] ; h. b. nielsen , [ arxiv:1104.4642 [ hep - ph ] ] ; b. bhattacherjee , s. raychaudhuri , [ arxiv:1104.4749 [ hep - ph ] ] ; g. segre , b. kayser , [ arxiv:1105.1808 [ hep - ph ] ] ; t. enkhbat , x. -g . he , y. mimura , h. yokoya , [ arxiv:1105.2699 [ hep - ph ] ] .

the @xmath0-calculus is a powerful model for representing functions . in its un - typed version , every recursive function can be represented . but , in this model , a term can be applied to itself and a computation may not terminate . to avoid this problem , types are used . in the simplest case , they are built from atomic types with the arrow and the typing rules say that a function of type @xmath2 may only be applied to an argument of type @xmath3 . this discipline ensures that every typed term is strongly normalizing , i.e. a computation always terminate . in this system ( the simply typed @xmath0-calculus ) , church numerals , i.e. the terms of the form @xmath4 , are codes for the integers . they are the only terms ( in normal form ) of type @xmath5 . thus , functions on the integers can be represented but schwichtenberg @xcite has shown that very few functions are so . he showed that the extended polynomials ( i.e. polynomials with positive coefficients together with a conditional operator ) are the only functions that can be represented there . other type systems were then designed to allow the representation of more functions . they are built in different ways . the first one consists in extending the set of terms . for example , in gdel system @xmath6 , the terms use the usual constructions of the @xmath0-calculus , the constant @xmath7 , the constructor @xmath8 and an operator for recursion . the types are built from the atomic type @xmath9 with the arrow . this system represents exactly the functions whose totality can be shown in peano first order arithmetic . the second one consists in keeping the same terms but extending the type system . this is , for example , the case of girard system @xmath10 where the types can use a second order universal quantifier . there , the type of the integers is given by @xmath11 . this system represents exactly the functions whose totality can be shown in peano second order arithmetic . a third way consists in extending the _ logic_. in the curry - howard correspondence , the previous systems correspond to _ intuitionistic _ logic . other systems correspond to _ classical _ logic . there , again , new constructors for terms are introduced . this is , for example , the case of parigot s @xmath1-calculus @xcite . since the introduction of girard system @xmath10 for intuitionistic logic and parigot s @xmath1-calculus for classical logic , many others , more and more powerful , type systems were introduced . for example , the calculus of constructions ( coquand & huet @xcite ) and , more generally , the pure type systems . it is also worth here to mention the system @xmath12 of parigot @xcite where some types are defined as the least fixed point of an operator . this system was introduced , not to represent more functions , but to represent more _ algorithms_. for example , to be able to represent the integers in such a way that the predecessor can be computed in constant time , which is not the case for the previous systems . these systems all satisfy the subject reduction ( i.e. the fact that the type is preserved by reduction ) , the strong normalization ( i.e. every computation terminates ) and , for the systems based on simple types , the decidability of type assignment . we study here other kinds of extension of the simply typed @xmath0-calculus , i.e. systems where _ equations _ on types are allowed . these types are usually called _ recursive types_. for more details see , for example , @xcite . they are present in many languages and are intended to be able to be _ unfolded _ recursively to match other types . the subject reduction and the decidability of type assignment are preserved but the strong normalization may be lost . for example , with the equation @xmath13 , the term @xmath14 where @xmath15 is typable but is not strongly normalizing . with the equation @xmath16 , every term can be typed . by making some natural assumptions on the recursive equations the strong normalization can be preserved . the simplest condition is to accept the equation @xmath17 ( where @xmath10 is a type containing the variable @xmath18 ) only when the variable @xmath18 is positive in @xmath10 . for a set @xmath19 of mutually recursive equations , mendler @xcite has given a very simple and natural condition that ensures the strong normalization of the system . he also showed that the given condition is necessary to have the strong normalization . his proof is based on the reducibility method . the condition ensures enough monotonicity to have fixed point on the candidates . but this proof ( using candidates of reducibility ) can not be formalized in peano arithmetic and the strength of the system needed for a proof of the strong normalization of such systems was not known . in this paper , we give an _ arithmetical _ proof of the strong normalization of the simply typed @xmath0-calculus ( and also of the @xmath20-calculus ) with recursive equations on types satisfying mendler s condition . this proof is an extension of the one given by the first author for the simply typed @xmath0-calculus . it can be found either in @xcite ( where it appears among many other things ) or as a simple unpublished note on the web page of the first author @xcite . apparently , proof methods similar to that used here were independently invented by several authors ( levy , van daalen , valentini and others ) . the proof for the @xmath1-calculus is an extension of the ones given in @xcite or @xcite . the paper is organized as follows . in section [ s - lambda ] we define the simply typed @xmath0-calculus with recursive equations on types . to help the reader and show the main ideas , we first give , in section [ proof ] , the proof of strong normalization for the @xmath0-calculus . we generalize this proof to the @xmath20-calculus in section [ s - mu ] . in section [ applic ] , we give two examples of applications of systems with recursive types . we conclude in section [ conc ] with some open questions . let @xmath21 be an infinite set of variables . 1 . the set @xmath22 of @xmath0-terms is defined by the following grammar + @xmath23 2 . the relation @xmath24 on @xmath22 is defined as the least relation ( compatible with the context ) containing the rule @xmath25 $ ] . as usual , @xmath26 ( resp . @xmath27 ) denotes the reflexive and transitive ( resp . transitive ) closure of @xmath24 . let @xmath28 be a set of atomic constants and @xmath29 be a set of type variables . 1 . the set @xmath30 of types is defined by the following grammar + @xmath31 2 . when @xmath32 is a set of types , the congruence @xmath33 generated by @xmath34 is the least congruence on @xmath30 such that @xmath35 for each @xmath36 . let @xmath33 be a congruence on @xmath30 . the typing rules of the typed system are given below where @xmath37 is a context , i.e. a set of declarations of the form @xmath38 where @xmath39 and @xmath40 . @xmath41 @xmath42 @xmath43 @xmath44 [ congru ] let @xmath33 be a congruence generated by a set of types . 1 . if @xmath45 , then @xmath46 or @xmath47 . 2 . if @xmath48 , then @xmath49 and @xmath50 . 3 . if @xmath51 , then @xmath38 occurs in @xmath37 for some @xmath52 . 4 . if @xmath53 , then @xmath54 for some @xmath55 such that @xmath56 . 5 . if @xmath57 , then @xmath58 , @xmath59 for some @xmath60 and @xmath3 . 6 . if @xmath61 and @xmath62 , then @xmath63 . 7 . if @xmath64 and @xmath65 , then @xmath66 : t$ ] . easy . [ sr ] if @xmath67 and @xmath68 , then @xmath69 . it is enough to show that if @xmath70 , then @xmath71:t$ ] . assume @xmath70 . by lemma [ congru ] , @xmath72 , @xmath73 and @xmath74 . thus , @xmath75 and @xmath76 . by lemma [ congru ] , we have @xmath77 and @xmath78 . thus , @xmath79 . since @xmath73 and @xmath74 , the result follows immediately . let @xmath80 . we define the subsets @xmath81 and @xmath82 of @xmath30 as follows . * @xmath83 * if @xmath84 , then @xmath85 . * if @xmath86 and @xmath87 , then @xmath88 and @xmath89 . [ cong ] we say that a congruence @xmath33 is good if the following property holds : for each @xmath80 , if @xmath90 , then @xmath91 . * examples * in each of the following cases , the congruence generated by the given equations is good . 1 . @xmath92 and @xmath93 . 2 . @xmath94 and @xmath95 . 3 . the same equations as in case 2 and @xmath96 where @xmath10 is any type using only the variables @xmath97 . 4 . the same equations as in case 3 and @xmath98 , @xmath99 where @xmath100 are any types using only the variables @xmath101 . _ in the rest of the paper , we fix a finite set @xmath32 of types and we denote by @xmath33 the congruence generated by @xmath34 . we assume that @xmath33 is good . _ * notations and remarks * * we have assumed that the set of equations that we consider is finite . this is to ensure that the order on @xmath102 given by definition [ ordre ] below is well founded . it should be clear that this is not a real constraint . since to type a term , only a finite number of equations is used , we may consider that the other variables are constant and thus the general result follows immediately from the finite case . * if @xmath103 is a term , @xmath104 will denote the structural complexity of @xmath103 . * we denote by @xmath105 the set of strongly normalizing terms . if @xmath106 , we denote by @xmath107 the length of the longest reduction of @xmath103 and by @xmath108 the pair @xmath109 . * we denote by @xmath110 the fact that @xmath103 is a sub - term of a reduct of @xmath9 . * as usual , some parentheses are omitted and , for example , we write @xmath111 instead of @xmath112 . more generally , if @xmath113 is a finite sequence @xmath114 of terms , we denote by @xmath115 the term @xmath116 and by @xmath117 the fact that @xmath118 . * if @xmath119 is the substitution @xmath120 $ ] , we denote by @xmath121 the set @xmath122 , by @xmath123 the set @xmath124 and by @xmath125 the fact that @xmath126 . * if @xmath119 is a substitution , @xmath127 and @xmath103 is a term , we denote by @xmath128 $ ] the substitution @xmath129 defined by @xmath130 for @xmath131 and @xmath132 . * in a proof by induction , will denote the induction hypothesis . when the induction is done on a tuple of integers , the order always is the lexicographic order . we give the idea for one equation @xmath133 . the extension for the general case is given at the beginning of section [ 3.4 ] . it is enough to show that , if @xmath134 are in @xmath105 , then @xmath135 \in sn$ ] . assuming it is not the case , the interesting case is @xmath136 with @xmath137 where @xmath138 \in sn$ ] . this implies that @xmath139 and @xmath140 \not \in sn$ ] . if we know that the type of @xmath9 is an arrow type , we get a similar situation to the one we started with , but where the type of the substituted variable has decreased . repeating the same argument , we get the desired result , at least for @xmath9 whose type does not contain @xmath18 . if it is not the case , since , by repeating the same argument , we can not come to a constant type ( because such a term can not be applied to something ) , we come to @xmath18 . thus , it remains to show that , if @xmath134 are in @xmath105 and the type of @xmath141 is @xmath18 , then @xmath135 \in sn$ ] . to prove this , we prove something a bit more general . we prove that , if @xmath142 where @xmath119 is a substitution such that the types of its image are in @xmath81 , then @xmath143 \in sn$ ] . the proof is done , by induction on @xmath144 as follows . as before , the interesting case is @xmath145 , @xmath146 \in sn$ ] and @xmath140 \not\in sn$ ] . thus , there is a sub - term of a reduct of @xmath147 of the form @xmath148 such that @xmath149 ) \not \in sn$ ] but @xmath150 \in sn$ ] . thus @xmath151 must reduce to a @xmath0 . this @xmath0 can not come from some @xmath152 , i.e. @xmath153 . otherwise , the type of @xmath154 would be both positive ( since @xmath155 and the type of @xmath156 is positive ) and negative ( since , in @xmath103 , @xmath154 is an argument of @xmath141 whose type also is positive ) . thus the type of @xmath151 ( the same as the one of @xmath154 ) does not contain @xmath18 . but since @xmath157 are in @xmath105 , we already know that @xmath140 $ ] must be in @xmath105 . a contradiction . thus , @xmath158 and we get a contradiction from the induction hypothesis since we have @xmath159 \not \in sn$ ] for @xmath160 strictly less than @xmath103 . the case when @xmath161 has more than one argument is intuitively treated by `` repeat the same argument '' or , more formally , by lemma [ prepa ] below . as a final remark , note that many lemmas are stated in a negative style and thus may seem to hold only classically . this has been done in this way because we believe that this presentation is closer to the intuition . however , it is not difficult to check that the whole proof can be presented and done in a constructive way . [ untyped ] assume @xmath162 and @xmath163 . then , for some term @xmath164 , @xmath165 and @xmath166\;\overrightarrow{o } ) \not \in sn$ ] . since @xmath162 , an infinite reduction of @xmath167 looks like @xmath168\ ; \overrightarrow{o ' } ) \rhd \ldots $ ] and the result immediately follows from the fact that @xmath166\;\overrightarrow{o } ) \rhd^ * ( m'[x:=n']\;\overrightarrow{o'})$ ] . [ subterm ] let @xmath103 be a term and @xmath119 be a substitution . assume @xmath169 and @xmath143 \not\in sn$ ] . then @xmath170 } ) \not \in sn$ ] for some @xmath171 such that @xmath172 } \in sn$ ] . a sub - term @xmath164 of a reduct of @xmath103 such that @xmath173 is minimum and @xmath174 \not\in sn$ ] has the desired form . [ lambda ] let @xmath103 be a term and @xmath119 be a substitution such that @xmath143 \rhd^ * \lambda z m_1 $ ] . then + - either @xmath175 and @xmath176 \rhd^ * m_1 $ ] + - or @xmath177 for some @xmath178 and @xmath179 } ) \rhd^ * \lambda z m_1 $ ] . this is a classical ( though not completely trivial ) result in @xmath0-calculus . note that , in case @xmath106 ( and we will only use the lemma in this case ) , it becomes easier . the proof can be done by induction on @xmath108 by considering the possibility for @xmath103 : either @xmath180 or @xmath181 or @xmath182 ( for @xmath141 in @xmath183 or not ) . [ ordre ] we define on @xmath102 the following relations * @xmath184 iff @xmath185 for some @xmath6 such that @xmath186 . * @xmath187 iff @xmath184 and @xmath188 . * @xmath189 iff @xmath184 and @xmath190 it is clear that @xmath191 is an equivalence on @xmath102 . 1 . let @xmath192 and @xmath193 . 2 . for @xmath194 , let @xmath195 where @xmath196 is the set of type variables occurring in @xmath6 . 3 . for @xmath197 , we will abbreviate by @xmath198 the set @xmath199 and by @xmath200 the set @xmath201 . if @xmath202 , @xmath203 will denote the opposite of @xmath204 . the opposite of + is - and conversely . let @xmath36 . the class of @xmath205 can be partitioned into two disjoint sets @xmath206 and @xmath207 satisfying the following properties . 1 . if @xmath202 , @xmath208 and @xmath209 , then for each @xmath210 , @xmath211 and for each @xmath212 , @xmath213 . 2 . let @xmath214 . then , if @xmath215 , @xmath216 and @xmath217 and if @xmath218 , @xmath219 and @xmath220 . this follows immediately from the following observation . let @xmath221 and @xmath222 . choose an occurrence of @xmath223 in @xmath6 and in @xmath3 . then , these occurrences have the same polarity . this is because , otherwise , since @xmath184 , there is a @xmath224 such that @xmath225 and @xmath226 occurs in @xmath224 . but then , replacing the mentioned occurrences of @xmath223 by @xmath224 in @xmath6 and @xmath3 will contradict the fact that @xmath33 is good . [ class ] let @xmath36 and @xmath202 . we denote @xmath227 / for each @xmath228 , @xmath229 and for each @xmath230 , @xmath231 . [ propr ] let @xmath36 and @xmath202 . 1 . @xmath232 . 2 . if @xmath233 and @xmath234 , then @xmath235 . 3 . if @xmath233 and @xmath236 , then @xmath237 and @xmath238 . immediate . * notations , remarks and examples * * if the equations are those of the case 4 of the examples given above , we have @xmath239 and , for example , @xmath240 and @xmath241 , @xmath242 , @xmath243 , @xmath244 and @xmath245 . * if @xmath6 is a type , we denote by @xmath246 the size of @xmath6 . note that the size of a type is , of course , not preserved by the congruence . the size of a type will only be used in lemma [ sansx ] and the only property that we will use is that @xmath247 and @xmath248 are less than @xmath249 . * by the typing rules , the type of a term can be freely replaced by an equivalent one . however , for @xmath36 and @xmath250 , the fact that @xmath233 does not change when @xmath3 is replaced by @xmath224 for some @xmath251 . this will be used extensively in the proofs of the next sections . to give the idea of the proof , we first need a definition . let @xmath252 be a set of types . denote by @xmath253 $ ] the following property : let @xmath254 . assume @xmath255 and @xmath65 for some @xmath256 such that @xmath257 . then @xmath135 \in sn$ ] . to get the result , it is enough to show @xmath258 $ ] . the proof that any typed term is in @xmath105 is then done by induction on @xmath104 . the only non trivial case is @xmath259 . but @xmath260 $ ] and the result follows from @xmath258 $ ] and the _ ih_. we first show the following ( see lemma [ sansx ] ) . let @xmath261 . to prove @xmath262 $ ] , it is enough to prove @xmath263 $ ] for each @xmath264 . it is thus enough to prove of @xmath265 $ ] for each @xmath36 . this is done by induction on @xmath205 . assume @xmath266 $ ] for each @xmath267 . thus , by the previous property , we know @xmath268 $ ] . we show @xmath265 $ ] essentially as we said in section [ 3.1 ] . the only difference is that , what was called there `` @xmath18 is both positive and negative in @xmath6 '' here means @xmath6 is both in @xmath269 and @xmath270 . there we deduced that @xmath18 does not occur in @xmath6 . here we deduce @xmath271 and we are done since we know the result for this set . [ sansx ] let @xmath194 be such that @xmath263 $ ] holds for each @xmath264 . then @xmath262 $ ] holds . let @xmath134 be terms in @xmath105 . assume @xmath255 and @xmath272 and @xmath273 . we have to show @xmath274 \in sn$ ] . this is done by induction on @xmath275 . assume @xmath135 \not\in sn$ ] . by lemma [ subterm ] , let @xmath276 be such that @xmath277 and @xmath278 where @xmath279 $ ] and @xmath280}$ ] . by lemma [ untyped ] , @xmath281 and @xmath282 \;\ ; \overrightarrow{q_1 } ) \not \in sn$ ] . if @xmath3 is a variable ( which is in @xmath283 since @xmath284 ) , we get a contradiction since we have assumed that @xmath263 $ ] holds for each @xmath264 . the type @xmath3 can not be a constant since , otherwise @xmath141 could not be applied to some arguments . thus @xmath285 . in the typing of @xmath286 , the congruence may have been used and thus , by lemma [ congru ] , there are @xmath287 , @xmath288 , @xmath289 and @xmath290 and @xmath291 . but then , we also have @xmath292 and @xmath293 . now , by the _ ih _ , we have @xmath294 \in sn$ ] since @xmath295 . since @xmath296 for some @xmath297 and @xmath298 : u_2 $ ] , by the _ ih _ since @xmath299 , we have @xmath282 \ ; \overrightarrow{q_1})=(z \ ; \overrightarrow{q_1})[z = n_1[x_1:=p_1 ] ] \in sn$ ] . contradiction . _ for now on , we fix some @xmath205 and we assume @xmath266 $ ] for each @xmath300 . thus , by lemma [ sansx ] , we know that @xmath268 $ ] holds . it remains to prove @xmath265 $ ] i.e. proposition [ avecx ] . _ let @xmath103 be a term , @xmath119 be a substitution , @xmath37 be a context and @xmath3 be a type . say that @xmath301 is adequate if the following holds . * @xmath302 : u$ ] and @xmath303 . * for each @xmath131 , @xmath304 and @xmath305 . [ prepa ] let @xmath306 be integers , @xmath307 be a sequence of terms and @xmath308 be adequate . assume that 1 . @xmath309 and @xmath310 \ , \overrightarrow{s } ) : w$ ] for some @xmath311 . @xmath312 , @xmath313 and @xmath314 . 3 . @xmath143 \in sn$ ] for every adequate @xmath315 such that @xmath316 . then @xmath317 \ ; \overrightarrow{s } ) \in sn$ ] . by induction on the length of @xmath307 . if @xmath307 is empty , the result follows from ( 3 ) since @xmath314 . otherwise , let @xmath318 and assume that @xmath319 \rhd^ * \l z \ ; r$ ] . by lemma [ lambda ] , there are two cases to consider : * @xmath320 . we have to show that @xmath321\ \overrightarrow{s_2 } ) \in sn$ ] . since @xmath322 , by lemmas [ congru ] and [ propr ] , there are types @xmath323 such that @xmath324 and @xmath325 and @xmath326 and @xmath327 and @xmath328 . since @xmath329 and @xmath330 , \delta \cup \ { z : b_1\ } , r ' , b_2)$ ] is adequate , it follows from ( 3 ) that @xmath331 \in sn$ ] . + - assume first @xmath332 . since @xmath333 and @xmath334 $ ] , the result follows from @xmath268 $ ] . + - otherwise , the result follows from the _ ih _ since @xmath335 , \delta \cup \ { z : b_1\ } , r ' , b_2)$ ] is adequate and the length of @xmath336 is less than the one of @xmath307 . * if @xmath337 for some @xmath338 . then @xmath339 } ) : b $ ] . by the definition of adequacy , the type of @xmath161 is in @xmath340 and @xmath341 contradiction . [ x>0 ] assume @xmath342 is adequate . then @xmath343 \in sn$ ] . by induction on @xmath108 . the only non trivial case is @xmath344 for some @xmath345 . let @xmath346 . by the _ , @xmath347 , \overrightarrow{o[\sigma ] } \in sn$ ] . by lemma [ congru ] , we have @xmath348 , @xmath349 : w_1 $ ] and @xmath350 ) : w_2 $ ] . moreover , by lemma [ propr ] , @xmath351 and @xmath352 . since @xmath353 = ( z \ ; \overrightarrow{o})[\sigma + z : = ( n \ , q[\sigma])]$ ] , @xmath354 , @xmath355 and @xmath356 , it is enough , by the _ ih _ , to show that @xmath357 ) \in sn$ ] . assume that @xmath358 . we have to show that @xmath359\in sn$ ] . - assume first @xmath360 . the result follows from @xmath361 $ ] . - otherwise , assume @xmath359 \not \in sn$ ] . since @xmath362 \in sn$ ] , by lemma [ subterm ] , @xmath363 for some @xmath364 such that @xmath365 } \in sn$ ] and @xmath366 \ ; \overrightarrow{l[y:=q [ \sigma ] ] } ) \not \in sn$ ] . but this contradicts lemma [ prepa ] . note that , by the , condition ( 3 ) of this lemma is satisfied . [ avecx ] assume @xmath367 and @xmath368 and @xmath369 . then @xmath274 \in sn$ ] . this follows from lemma [ x>0 ] since @xmath370 , \g , m , u ) $ ] is adequate . let @xmath372 be an infinite set of variables such that @xmath373 . an element of @xmath21 ( resp . @xmath372 ) is said to be a @xmath0-variable ( resp . a @xmath374-variable ) . we extend the set of terms by the following rules + @xmath375 2 . we add to the set @xmath28 the constant symbol @xmath376 and we denote by @xmath377 the type @xmath378 . 3 . we extend the typing rules by + @xmath379 @xmath380 + where @xmath37 is now a set of declarations of the form @xmath381 and @xmath382 where @xmath141 is a @xmath0-variable and @xmath383 is a @xmath384-variable . we add to @xmath385 the following reduction rule @xmath386 $ ] where @xmath387 $ ] is obtained by replacing each sub - term of @xmath103 of the form @xmath388 by @xmath389 . this substitution will be called a @xmath384-substitution whereas the ( usual ) substitution @xmath135 $ ] will be called a @xmath0-substitution . * remarks * * note that we adopt here a more liberal syntax ( also called de groote s calculus @xcite ) than in the original calculus since we do not ask that a @xmath390 is immediately followed by a @xmath391 ( denoted @xmath392 m$ ] in parigot s notation ) . * we also have changed parigot s typing notations . instead of writing @xmath393 , @xmath394 we have written @xmath395 but , since the first introduction of the @xmath1-calculus , this is now quite common . * unlike for a @xmath0-substitution where , in @xmath135 $ ] , the variable @xmath141 has disappeared it is important to note that , in a @xmath384-substitution , the variable @xmath383 has not disappeared . moreover its type has changed . if the type of @xmath9 is @xmath3 and , in @xmath396 , the type of @xmath383 is @xmath397 it becomes @xmath398 in @xmath399 $ ] . * the definition of good congruence is the same as before . as a consequence , we now have the following facts . if @xmath400 , then @xmath401 and , if @xmath402 , then @xmath403 . * we also extend all the notations given in section [ s - lambda ] . finally note that lemma [ congru ] remains valid . moreover , they are easily extended by lemma [ typed1 ] below . [ typed1 ] 1 . if @xmath404 , then @xmath405 for some @xmath224 such that @xmath62 . 2 . if @xmath406 , then @xmath407 and @xmath408 . 3 . if @xmath409 and @xmath65 , then @xmath410 : t$ ] . [ sr ] if @xmath67 and @xmath68 , then @xmath69 . it is enough to show that , if @xmath411 , then @xmath412:t$ ] . assume @xmath413 . by lemma [ congru ] , @xmath414 , @xmath415 and @xmath416 . thus , @xmath417 and @xmath418 . by lemma [ congru ] , we have @xmath419 . since @xmath73 and @xmath416 , @xmath420 : \bot$ ] . then @xmath421 : v$ ] and @xmath422 : t$ ] . [ mu.lambda ] let @xmath103 be a term and @xmath423 where @xmath424 ( resp . @xmath425 ) is @xmath0 ( resp . @xmath384 ) substitution . assume @xmath143 \rhd^ * \m\al m_1 $ ] ( resp . @xmath180 ) . then - either @xmath426 ( resp . @xmath427 ) and @xmath176 \rhd^ * m_1 $ ] - or ( @xmath177 for some @xmath428 and @xmath179 } ) \rhd^ * \m \al m_1 $ ] ( resp . @xmath180 ) . a @xmath384-substitution can not create a @xmath0 or a @xmath384 ( see , for example , @xcite ) and thus , the proof is as in lemma [ lambda ] . [ mu.untyped ] assume @xmath429 and @xmath430 . then either ( @xmath431 and @xmath432 \;\overrightarrow{q } ) \not \in sn$ ] ) or ( @xmath433 and @xmath434 \;\overrightarrow{q } ) \not \in sn$ ] ) . as in lemma [ untyped ] . [ mu.subterm ] let @xmath103 be a term and @xmath119 be a @xmath0-substitution . assume @xmath169 and @xmath143 \not\in sn$ ] . then @xmath170 } ) \not \in sn$ ] for some @xmath171 such that @xmath172}\in sn$ ] . as in lemma [ subterm ] . a @xmath374-substitution @xmath119 is said to be fair if , for each @xmath435 , @xmath436 where @xmath437 ( resp . @xmath438 ) means that @xmath439 ( resp . @xmath440 ) for some @xmath441 . [ goodsub ] let @xmath119 be is a fair @xmath374-substitution , @xmath442 and @xmath443 ( resp . @xmath444 ) , then @xmath143[x:=\sigma(\alpha)]= m[x:=\sigma(\alpha)][\sigma]$ ] ( resp . @xmath143[\beta = \sigma(\alpha)]= m[\beta = \sigma(\alpha)][\sigma]$ ] ) . immediate . [ mu.nonsn ] let @xmath134 be terms and @xmath119 be a fair @xmath384-substitution . assume @xmath143 , n \in sn$ ] but @xmath445 \ ; n ) \not \in sn$ ] . assume moreover that @xmath143 \ras \m\al m_1 $ ] . then , for some @xmath446 , we have @xmath447 \ ; n ) \not\in sn$ ] and @xmath448 \in sn$ ] where @xmath449 $ ] . by lemma [ mu.lambda ] , we know that @xmath450 for some @xmath451 such that @xmath452 \ras m_1 $ ] . let @xmath164 be a sub - term of a reduct of @xmath103 such that @xmath453 ) , cxty(m')\rangle$ ] is minimum and @xmath454 \not\in sn$ ] . we show that @xmath455 and has the desired properties . by minimality , @xmath164 can not be of the form @xmath456 , @xmath457 nor @xmath458 for @xmath459 or @xmath460 . if @xmath461 . by the minimality of @xmath164 , @xmath462 , p_2[\sigma ' ] \in sn$ ] . thus , by lemma [ mu.lambda ] and [ mu.untyped ] , @xmath463 ( resp . @xmath464 ) such that @xmath465[x:=p_2[\sigma']]= q[x:=p_2][\sigma ' ] \not\in sn$ ] ( resp . @xmath465[\be = p_2[\sigma']]=q[\be = p_2][\sigma ' ] \not\in sn$ ] ) and this contradicts the minimality of @xmath164 . if @xmath466 for some @xmath467 . then @xmath468 \ ; \sigma(\be ) ) \not\in sn$ ] and , by the minimality of @xmath164 , @xmath469 \in sn$ ] . thus , by lemmas [ mu.lambda ] , [ mu.untyped ] and [ goodsub ] , @xmath470 ( resp . @xmath471 ) such that @xmath465[x:=\sigma(\be)]= q[x:=\sigma(\be)][\sigma ' ] \not\in sn$ ] ( resp . @xmath465[\gamma=\sigma(\be)]=q[\gamma=\sigma(\be)][\sigma ' ] \not\in sn$ ] ) and this contradicts the minimality of @xmath164 . thus @xmath455 and its minimality implies @xmath448 \in sn$ ] . we use the same notations as in section [ proof ] . [ mu.sansx ] let @xmath194 be such that @xmath263 $ ] holds for each @xmath264 . then @xmath262 $ ] holds . assume that @xmath263 $ ] holds for each @xmath264 . the result is a special case of the following claim . _ claim : _ let @xmath103 be a term , @xmath55 be types such that @xmath472 and @xmath119 be a @xmath0-substitution such that , for each @xmath141 , @xmath473 $ ] where @xmath474 is a fair @xmath374-substitution such that @xmath475 ) = \emptyset$ ] . assume @xmath476 and for each @xmath131 , @xmath477 . assume finally that @xmath103 and the @xmath478 $ ] are in @xmath105 . then , @xmath143 \in sn$ ] . _ by induction on @xmath479 where @xmath480 @xmath481 and @xmath482 and , in the sums , each occurrence of a variable counts for one . for example , if there are two occurrences of @xmath483 and three occurrences of @xmath484 , @xmath485 . note that we really mean @xmath486 and not @xmath487)$ ] and similarly for @xmath488 . the only non trivial case is when @xmath489 for @xmath131 . by the _ ih _ , @xmath347 , \overrightarrow{o[\sigma]}\in sn$ ] . it is enough to show that @xmath490 \ ; q[\sigma ] ) \in sn$ ] since @xmath143 $ ] can be written as @xmath454 $ ] where @xmath491})$ ] and @xmath492 \ ; q[\sigma])$ ] and ( since the size of the type of @xmath493 is less than the one of @xmath3 ) the gives the result . by lemma [ mu.untyped ] , we have two cases to consider . * @xmath478 \ras \l y n_1 $ ] . by lemma [ mu.lambda ] , @xmath494 and the proof is exactly the same as in lemma [ sansx ] . * @xmath478 \ras \m \al n_1 $ ] . by lemma [ mu.nonsn ] , let @xmath495 be such that @xmath496 \in sn$ ] and @xmath497 \ ; q[\sigma ] ) \not \in sn$ ] where @xmath498 $ ] . but @xmath499 can be written as @xmath500 $ ] where @xmath129 is the same as @xmath119 except that @xmath501 $ ] . note that @xmath502 is the same as ( or less than ) @xmath103 but one occurrence of @xmath141 has been replaced by the fresh variable @xmath161 . the substitution @xmath503 is fair and @xmath504 . the gives a contradiction since @xmath505 . note that the type condition on @xmath129 is satisfied since @xmath506 has type @xmath3 , thus @xmath383 has type @xmath377 and thus @xmath507 also has type @xmath3 . _ for now on , we fix some @xmath205 and we assume @xmath266 $ ] for each @xmath300 . thus , by lemma [ mu.sansx ] , we know that @xmath268 $ ] holds . it remains to prove @xmath265 $ ] i.e. proposition [ mu.avecx ] . _ let @xmath103 be a term , @xmath508 where @xmath424 ( resp . @xmath425 ) is a @xmath0 ( resp . @xmath384 ) substitution , @xmath37 be a context and @xmath3 be a type . say that @xmath301 is adequate if the following holds : * @xmath509 : u$ ] and @xmath303 . * for each @xmath510 , @xmath511 and @xmath512 . note that nothing is asked on the types of the @xmath384-variables . [ mu.prepa ] let @xmath306 be integers , @xmath307 be a sequence of terms and @xmath308 be adequate . assume that 1 . @xmath513 and @xmath514 \ , \overrightarrow{s } ) : w$ ] for some @xmath311 . @xmath312 , @xmath313 and @xmath314 . 3 . @xmath143 \in sn$ ] for every adequate @xmath315 such that @xmath515 . then @xmath317 \ , \overrightarrow{s } ) \in sn$ ] . by induction on the length of @xmath307 . the proof is as in lemma [ prepa ] . the new case is @xmath516 \rhd^ * \m \al r$ ] ( when @xmath517 ) . by lemma [ mu.lambda ] , we have two cases to consider . * @xmath518 . we have to show that @xmath519\ \overrightarrow{s_2 } ) \in sn$ ] . by lemma [ typed1 ] , the properties of @xmath33 and since @xmath520 , there are types @xmath323 such that @xmath521 and @xmath522 : b_2 $ ] and @xmath523 . since @xmath524 and @xmath525 , \delta \cup \ { \al : \neg b_2\ } , \mu \al r ' , b_2)$ ] is adequate , it follows from ( 3 ) that @xmath526 \in sn$ ] . + - assume first @xmath332 . since @xmath333 and @xmath527 $ ] , the result follows from @xmath268 $ ] . + - otherwise , the result follows from the _ ih _ since @xmath528 , \delta \cup \ { \al : \neg b_2\ } , \mu \al r ' , b_2)$ ] is adequate and the length of @xmath336 is less than the one of @xmath307 . * @xmath337 for some @xmath0-variable @xmath529 . as in lemma [ prepa ] . [ mu.2 ] assume @xmath530 is adequate . then @xmath343 \in sn$ ] . as in the proof of the lemma [ mu.sansx ] , we prove a more general result . assume that , for each @xmath510 , @xmath531 $ ] where @xmath474 is a fair @xmath374-substitution such that @xmath475 ) = \emptyset$ ] . we prove that @xmath143 \in sn$ ] . by induction on @xmath532 and , by secondary induction , on @xmath533 where @xmath534 and @xmath535 are defined as in lemma [ mu.sansx ] . the proof is as in lemma [ mu.sansx ] . the interesting case is @xmath489 for some @xmath510 . the case when @xmath536 \ras \l y n'$ ] is as in lemma [ x>0 ] . the new case is when @xmath478 \rhd^ * \m \al n'$ ] . this is done as in lemma [ mu.sansx ] . note that , for this point , the type was not used . [ mu.avecx ] assume @xmath367 and @xmath368 and @xmath537 . then @xmath274 \in sn$ ] . this follows from lemma [ mu.2 ] since @xmath370 , \g , m , u ) $ ] is adequate . by using recursive types , some terms that can not be typed in the simply typed @xmath0-calculus become typable . for example , by using the equation @xmath538 , it is possible to type terms containing both @xmath539 and @xmath540 as sub - terms . just take @xmath541 and @xmath542 . by using the equation @xmath543 , it is possible to apply an unbounded number of arguments to a term . it is thus natural to try to extend schwichtenberg s result and to determine the class of functions that are represented in such systems and , in particular , to see whether or not they allow to represent more functions . note that doyen @xcite and fortune & all @xcite have given extensions of schwichtenberg s result . here is an example of function that can not be typed ( of the good type ) in the simply typed @xmath0-calculus . let @xmath544 and @xmath545 where @xmath546 are type variables . let @xmath547 be the church numeral representing @xmath548 and @xmath549 , @xmath550 be the terms representing _ false _ and _ true_. note that @xmath551 has type _ nat _ and @xmath552 , @xmath553 have type @xmath554 . the term @xmath555 where @xmath556 has been introduced by b.maurey . it is easy to see that , for every @xmath557 , the term @xmath558 reduces to * 1 * if @xmath559 and to * 0 * otherwise . krivine has shown in @xcite that the type @xmath560 can not be given to @xmath561 in system @xmath10 but , by adding the equation @xmath562 , it becomes typable . our example uses the same ideas . let @xmath33 be the congruence generated by @xmath563 . for each @xmath564 , let @xmath565 where @xmath566 . [ p1 ] for each @xmath564 we have @xmath567 . we have @xmath568 , then @xmath569 , thus @xmath570 . but @xmath571 , therefore @xmath572 . we have @xmath573 , then @xmath574 , thus @xmath575 . but @xmath576 , therefore @xmath577 . we deduce that @xmath578 , then @xmath579 and thus @xmath580 . [ p2 ] for each @xmath564 and @xmath581 , @xmath582 reduces to * 1 * if @xmath583 and to * 0 * otherwise . @xmath584 + @xmath585 + @xmath586 * remarks * note that for the ( usual ) simply typed @xmath0-calculus we could have taken for @xmath18 and @xmath587 the same variable but , for propositions [ p1 ] and [ p2 ] , we can not assume that @xmath588 because then the condition of positivity would not be satisfied . this example is thus not completely satisfactory and it actually shows that the precise meaning of the question `` which functions can be represented in such systems '' is not so clear . the strong normalization of a typed @xmath1-calculus can be deduced from the one of the corresponding typed @xmath0-calculus by using cps translations . see , for example , @xcite for such a translation . there is another , somehow simpler , way of doing such a translation . add , for each atomic type @xmath18 , a constant @xmath589 of type @xmath590 . using these constants , it is not difficult to get , for each type @xmath6 , a @xmath0-term @xmath591 ( depending on @xmath6 ) such that @xmath592 has type @xmath593 . this gives a translation of the @xmath594-calculus into the @xmath0-calculus from which the strong normalization of the @xmath1-calculus can be deduced from the one of the @xmath0-calculus . this translation , quite different from the cps translations , has been used by krivine @xcite to code the @xmath1-calculus with second order types in the @xmath595-calculus . with recursive equations , we do not have to add the constant @xmath589 since we can use the equation @xmath596 . we give here , without proof , the translation . we denote by @xmath597 the simply typed @xmath0-calculus where @xmath33 is the congruence on @xmath598 ( where @xmath599 ) generated by @xmath600 for each @xmath18 and by @xmath601 the usual ( i.e. without recursive types ) @xmath1-calculus . 1 . we define , for each type @xmath6 , a closed @xmath0-term @xmath591 such that @xmath602 as follows . this is done by induction on @xmath6 . * @xmath603 where @xmath604 . * if @xmath80 , @xmath605 . * @xmath606 2 . we define a translation from @xmath601 to @xmath597 as follows . * @xmath607 . * @xmath608 . * @xmath609 . * @xmath610 if @xmath383 has the type @xmath377 . * @xmath611 . for a better understanding , in the translation of @xmath612 and @xmath613 , we have kept the same name to the variable @xmath383 but it should be clear that the translated terms are @xmath0-terms with only on kind of variables . [ trtype ] if @xmath614 then @xmath615 . let @xmath134 be typed @xmath1-terms . if @xmath616 , then @xmath617 . it is enough to check that @xmath618)^*$ ] . the strong normalization of @xmath597 implies the one of @xmath601 . by lemmas [ trtype ] and [ trred ] . * remark * note that the previous translation can not be used to show that the @xmath1-calculus with recursive types is strongly normalizing since having two equations ( for example @xmath596 and @xmath619 ) is problematic . the proof of the strong normalization of the system d of intersection types @xcite is exactly the same as the one for simple types . is it possible to extend our proof to such systems with equations ? note that the sort of constraints that must be given on the equations is not so clear . for example , what does that mean to be positive in @xmath620 ? to be positive both in @xmath621 and @xmath622 ? in one of them ? it will be interesting to check precisely because , for example , it is known that the system given by system d and the equations @xmath623 and @xmath624 is strongly normalizing ( but the proof again is not formalized in peano arithmetic ) though the positivity condition is violated . 2 . we could add other typing rules and constructors to ensure that , intuitively , @xmath18 represents the _ least fixed point _ of the equation @xmath133 . this kind of thing is done , for example , in @xmath12 . what can be said for such systems ? there are many translations from , for example , the @xmath1-calculus into the @xmath0-calculus that allows to deduce the strong normalization of the former by the one of the latter . these cps transformations differ from the one given in section [ transfo ] by the fact that the translation of a term does not depend on its type . what is the behavior of such translations with recursive equations ? we would like to thank p urzyczyn who has mentioned to us the question solved here and has also indicated some errors appearing in previous versions of our proofs . thanks also to the referees and their valuable remarks .

the impressive development of low - dimensional electronic and optoelectronic devices , brought a new urgency to the essential measurement and modeling of charge carriers transmission time and particularly of the tunneling time through specific potential regions . when electrons and holes are involved , the low - dimensional device response depends on the slower charge - carrier s traveling time @xcite . undoubtedly , for such systems it is crucial to study the tunneling time of holes . an equally important reason for address investigations on heavy holes @xmath2 and light holes @xmath3 , is their intrinsic band mixing effects . since the hole s transport is essentially an interconnected - multichannel process , the theoretical calculation of transmission quantities is in some sense more cumbersome than that for electrons . currently , the research of under - synchronic anomalous scattering events , is one of the major areas in the quantum transport theory@xcite . regarding the long - standing controversy around superluminal velocities and total transmission inconsistency , together with their wide - known anomalous character @xcite , they are welcome due to all possible benefits they open up for design and development of nano - optoelectronic devices@xcite . there is a large accumulated knowledge about these topics -mostly dedicated to electrons , optical pulses and electromagnetic waves @xcite- , accurately enough to unambiguously provide solid grounds for superluminal propagation , the paradoxical hartman premonition and legitimacy of the phase time approach . worthwhile noticing however that , hole transport analogous phenomena has been less studied yet . resonant mechanisms between propagating modes oscillations and the self oscillations of the system are the fundamental cause , which gives rise to an isolated maximums of transmission , often referred to as giant conductance ( gg ) phenomena . events of gg have been observed for electrons and optical pulses in many finite periodic systems , which possess time reversal symmetry , namely : ballistic mesoscopic conductors , luminal pulses and superlattices @xcite . recent theoretical studies in molecular devices suggested that could rise up particularly large values for dynamic conductance @xcite and novel experimental achievements pointed out to a giant enhancement of electronic tunneling in a two - dimensional lattice of graphene@xcite . lately , was found a strong dependence of the conductance fluctuations with the density of states in graphene nanoribbons , which represent an alternative insight to study the band structure of these physical systems@xcite . owing to its relevance for quantum transport , the tunneling time has played a key role for a correct evaluation of electronic devices @xcite . a large number of proposals in reply the challenge to settle how long does it takes to a particle to tunnel a single potential barrier @xcite , have been put forward @xcite . the lively and long standing debate concerning this matter @xcite -presumed not to finish completely- , it seems to have finally found its cornerstone in precise experimental measurements for photons and optical pulses @xcite . these genuine measurements put to the test , before all else , the very question loosely formulated by mccoll and directly bear on truly temporal features of tunneling phenomenon . besides , they are consistent with the phase time conception derived from group delay within the stationary - phase method @xcite . an additional motivation was the accurate description of above mentioned experiments by careful phase - time calculations @xcite . in this paper we will present numerical calculations of the conductance and the phase time for an arbitrary number of alternate layers of @xmath4 materials , in the framework of the multicomponent scattering approach ( msa)@xcite . the main idea is to search evidences of anomalous events in @xmath5 and @xmath6 transmission . in sec.[msa ] some theoretical remarks within the msa are exposed . brief complementary insights to the uncoupled regime formula of scattering matrix unitarity relations and phase time will also be presented . analytical expressions for the phase time and the conductance are applied in sec.[numericalresults ] to study the anomalous events associated with these physical quantities in the double barrier resonant tunneling ( dbrt ) and the superlattice . phase time behavior for holes shows a very good agreement with the previous results for electron - tunneling through superlattices@xcite . evidences of the giant conductance phenomena for uncoupled hole transport are shown . finally , in sec.[conclusions ] we summarize and conclude . the msa is based on standard @xmath7 effective hamiltonians and the multichannel transfer matrix ( tm ) method . in this approach the scattering amplitudes can be obtained straightforwardly @xcite . the main purpose here , is to evaluate quantities of synchronic multi - mode transport for massive charge carriers moving through multi - layered structures at coupled and decoupled channel regimes . owing to consistency , we found useful to recall few basic outlines of the kohn - lttinger ( kl ) two - band model @xcite , due to its widely accepted accuracy for describing dynamics of elementary excitations as well as electronic properties in the valence band , within the framework of the multiband effective mass approach . this model , is highly common regarding to the field of hole quantum transport through periodic heterostructures . it considers in the @xmath8 approximation the highest two valence bands degenerated in the @xmath9-point of the brillouin zone . the usual @xmath10 kl hamiltonian @xcite in the broido - sham representation @xcite , has the form @xmath11 with @xmath12 being @xmath13 in atomic units . the matrix elements @xmath14 show a direct dependency of the semi - empirical parameters of lttinger @xmath15 , that typify each slab of the heterostructure . while @xmath16 and @xmath17 , stand for the periodic potential , and the two - components in - plane _ quasi_-momentum , respectively @xcite . to study the quantum transport properties of holes through semiconductor heterostructure , we consider the system shown in fig.([fig : heterostructure ] ) , described by the kohn - lttinger hamiltonian in the expression ( [ 4x4hkl ] ) . the four accessible channels of the system , denoted as : @xmath18 , and @xmath19 , are considered simultaneously , following one of the bare bones of the msa . we emphasize this upmost advantage , as it allows clearly identify inter- and intra - subband hole transitions together with transitions from kramer - up toward kramer - down states to time reversal , and the other way around . the latter is not the usual case . in this approach , the segment between @xmath20 and @xmath21 in fig . ( [ fig : heterostructure ] ) , is defined as a _ cell _ and represents the period of the heterostructure superlattice @xcite . schematic illustration of multichannel @xmath5 and @xmath6 quantum transport processes through a @xmath22 superlattice.,width=302 ] combining the @xmath23 kohn - lttinger hamiltonian with the tm formalism , a direct connection with the scattering theory is established @xcite , then we have @xmath24 . here , @xmath25 defines a relation between the traveling wave vectors of the tunneling problem and their derivatives , with the @xmath26dimensional state vectors that describe a propagating or evanescent mode . this crucial transformation of the msa , connects two different kinds of tm . the first class of tm , @xmath27 , relates the linearly independent state functions and their derivatives ( @xmath28 ) between two points of the heterostructure in the form @xmath29 , where the matrix @xmath30 can be constructed once the linearly independent solutions are obtained @xcite . the second kind of tm , @xmath31 , whose elements describe a propagating or evanescent mode depending on the energy , connects the state vectors ( @xmath32 ) in two different points of the system . the mathematical structure of these two matrices is conditioned by requirements of time - reversal invariance and flux conservation @xcite . a very strong criterium to verify the calculations is based in the determinant of these transfer matrices which must be constant and equal to the unity , mostly . then applying the well known relations between the @xmath33 , and the scattering matrix + @xmath34 blocks @xcite , one can directly evaluate the relevant physical quantities of the scattered transport theory . initially we have to obtain the complex transmission amplitudes @xcite @xmath35 for each transition from the @xmath36-_th _ incoming channel ( incident propagating mode ) to the @xmath37-_th _ outgoing channel ( transmitted or reflected propagating modes ) . it is worth noticing that within the msa , none of the incoming amplitudes is required to cancel , as it is commonly done . hence , an authentic multichannel and synchronic description of the hole transmission processes is naturally achieved . this is to say that , we are able to distinguish accurately the contribution from each input channel to the scattering process , modulated by the simultaneous presence of the rest of the channels , which are , in principle , equally accessible to the incident hole stream . none of the usual simplifications of the original hamiltonians are invoked , thus the time reversal symmetry is assured . given the transmission amplitudes @xmath38 , it is possible to evaluate the total transmission probability to channel @xmath37 @xmath39 the total two - probe landauer conductance defined as @xcite @xmath40 and the transmission amplitude phase for each of the @xmath41 available paths @xmath42 @xcite , @xmath43 . the transmission phase time is obtained from @xcite @xmath44 whose advantages and applicability criteria has been widely discussed in the literature @xcite . the sums in equations ( [ gi ] ) and ( [ gt ] ) run over all the open channels and the expressions ( [ gi])-([phase_time ] ) and ( [ phase_time_prop ] ) are defined for the @xmath45 cells periodic structure showed in fig.([fig : heterostructure ] ) . hereafter @xmath46 stands for the incoming particle s energy . tunable parameter @xmath17 , is essential to modify the in - plane dynamics of holes , _ i.e. _ the degree of freedom transverse to the main transport direction , also refereed as the valence - band mixing for holes . this later is crucially relevant in order to understand the main problem envisioned here , and how we face it within the msa . whenever the _ quasi_-momentum parallel to the interfaces , @xmath47 , the crossed transitions are not allowed and the system is said to yield an uncoupled hole regime . under this particular condition , only direct incoming - outgoing channels are connected , therefore @xmath48 blocks @xmath49 , @xmath50 , @xmath51 and @xmath52 of the tm @xmath53 are diagonal matrices . the later , straightforwardly leads us to new unitarity relationships , that satisfy the blocks of the multiband scattering matrix under an uncoupled hole regime , acquiring a more simple form : @xmath54 where @xmath55 . consequently , transmission and reflection amplitudes have reduced expressions respect to ( [ transmission_amplitudes ] ) : @xmath56 rewriting the phase time ( [ phase_time ] ) for this specific case , we obtain : @xmath57 expressions like ( [ transmission_reflection_amp ] ) were previously reported for the resonant tunneling and band mixing in multichannel superlattices of propagating electron systems @xcite . for the numerical calculations that will be presented here we shall consider our layered media with barrier thickness @xmath58 , and well width @xmath59 . for the barrier height we will have either @xmath60 or @xmath61 . the cases where the in - plane wavenumber ( parallel to the interfaces ) @xmath62 , with vanishing band mixing effects , and the cases where @xmath63 , with coupling phenomena effects on the quantum transport@xcite will be separately study . focusing to study the behavior of the phase time @xmath64 and the free motion time @xmath65 ( being @xmath66 the cell longitude , @xmath67 the effective mass of holes and @xmath68 the eigenvalues of the scattering problem solved with the msa@xcite ) as a function of the incoming energy @xmath69 , the barriers height and thickness were fixed at @xmath70 and @xmath71 , respectively . both , a double barrier ( @xmath72 cells ) and a superlattice of @xmath73 cells were considered . for consistency with the reported results for electrons@xcite , a @xmath74 concentration of @xmath75 was fixed inside the barrier : @xmath76 . phase time @xmath64 for a tunneling of holes through a single barrier of @xmath77 @xmath78 as a function of the hole incident energy at @xmath79 with @xmath80 . the solid ( dashed ) line stands for @xmath81 direct path . in the broken interval a negative dip extents up to @xmath82 ps.,width=283 ] in figure [ fig2 ] , we plot the phase time for @xmath6 and @xmath5 through a single potential barrier . it is clear from this figure that the @xmath6 modes ( dashed line ) are faster than @xmath5 modes ( solid line ) , although the difference is not large . in an effort to validate our numerical simulation with experimental data , we compare the results of figure [ fig2 ] with those of dragoman _ et . _ @xcite the tunneling times measured for holes , for null or small magnetic field in a ballistic - regimen nano - device , are of the order of @xmath83 . to reproduce that situation , we canceled the interaction between holes by assuming @xmath47 ( uncoupled regimen ) , in the absence of magnetic field . in our calculation the phase time for @xmath5 and @xmath6 is also of the order of @xmath83 , which is in an acceptable qualitative agreement with experiment and it is just one order less than the traveling time for electrons through a single barrier @xcite , as was detected experimentally for a dbrt @xcite . we have observed broaden peaks of the phase time for @xmath84 direct transition at @xmath85 and for @xmath86 direct transition at @xmath87 . we have observed , although not plotted , that the transmission coefficient maxima are located at comparable energies . it was shown in ref.@xcite that , the tunneling time follows the band structure . this feature of hole tunneling agrees also with calculations of ref.@xcite , where a particular case of the envelope function approximation was used . evaluating the tunneling times for energies in the range @xmath88 we have found , although not shown , a reduction of @xmath64 in @xmath89 orders of magnitude with respect to that shown in figure [ fig2 ] when the in - plane momentum changes as @xmath90 . this behavior of @xmath64 , was earlier detected in experiments for a hole passage through a slightly different system @xcite -a double asymmetric quantum well ( qw)- , and is considered to arise from valence subbands mixing @xcite . direct phase time @xmath64 for hole tunneling through a single barrier as a function of the hole incident energy for several values of the barrier thickness at @xmath91 , with @xmath80 and @xmath92.,width=283 ] figure [ fig3 ] displays a raising of the phase time while @xmath93 increases in the case of uncoupled regime ( @xmath47 ) of holes incidence . the same is achieved for light holes , although not shown in this figure . this behavior agrees qualitatively with the results of an experiment on time - resolved luminescence spectroscopy @xcite . for @xmath94 , @xmath64 is sensitive to the raising of @xmath93 and it is shown an oscillating behavior of the transmission phase time as function of the incident energy . the resonances shift to smaller energies , increasing their intensities with @xmath93 . it is assumed understood that , when the de broglie wavelength of the hole propagating mode is a multiple of @xmath93 , there is a resonance in the transmission and correspondingly the phase time delay raises at these energies . as @xmath93 increases , new resonances appear , thus the oscillating behavior is achieved . these oscillations correspond to the quasi - stationary levels of the virtual qw with increasing width . in that sense , from @xmath64 properties we are able to forecast a longer lifetime for the lower unbound hole levels . on general grounds , we found this in correspondence with the idea of the quasi - stationary states lowering , whenever the virtual qw becomes wider . phase time _ versus _ barrier thickness for the direct transitions @xmath95 given as solid diamonds / circles ( @xmath96 ) , in the uncoupled case ( @xmath97 ) . a fragmented exponential fitting is shown in dashed line ( with full circles ) at @xmath98 . the barrier height @xmath60.,width=264 ] figure [ figtauvslbn1 ] shows phase transmission time _ versus _ barrier thickness for the direct transitions @xmath95 given as solid diamonds / circles ( @xmath99 ) , in the uncoupled case . we prescribe appealing tunneling events of holes for @xmath100 under uncoupled propagating regime . this figure displays @xmath64 for resonant direct @xmath95 transitions ( diamonds ) , which rapidly becomes autonomous from barrier broadening as hartman predicted for electrons @xcite . these results are in accordance with the transit time for optical pulses through opaque barriers ( which transmit about @xmath101 of the incident radiation ) @xcite and also with theoretical predictions for the delay time in one and two dimensions @xcite . the circles in figure [ figtauvslbn1 ] correspond to numerical results within the msa , for the tunneling time of direct paths ( indicated in the legend ) for @xmath102 . here we must stress , that the resonant oscillating behavior ( on top of an average slope of smooth growth @xcite ) , is an expected phenomena due to the barrier s interference effects with the continuum states , wide accepted as ramsauer - townsend oscillations @xcite . for @xmath102 we briefly investigated segments in figure [ figtauvslbn1 ] , pursuing a locally exponential - like behavior of @xmath64 , as was suggested in a reported experiment with holes @xcite , within a rank comparable with their measurements . the interval of barrier thickness length variance in their measurements , ranges from @xmath103 @xmath104 up to @xmath105 . notice in figure [ figtauvslbn1 ] a sectional fitting , using an exponential growth function @xcite ( dashed line with full circles ) . we do believe that oscillations are already intrinsically present in heberle s , _ et al . _ accurate investigation @xcite and could be shown by including more experimental points or enlarging in rank the search . direct phase time @xmath64 for hole tunneling through a dbrt structure as a function of the energy of an impinging hole stream at @xmath62 , with @xmath106 , @xmath107 and @xmath108 , for @xmath109 ( solid line ) , and @xmath110 ( dashed line ) . the inset displays transmission coefficients for the same direct transitions , through an identical dbrt heterostructure , as a function of the incident energy . the labelling for the hole levels follows that of the reference @xcite.,width=283 ] figure [ fig5 ] is devoted to analyze briefly , the spectrum and the quasi - stationary bound states lifetime in a dbrt . the curves of figure [ fig5 ] ( inset ) , show several transmission phase times @xmath64 resonances for the uncoupled hole regime . in particular we have analyzed a semiconductor heterostructure of the form @xmath111 . it was shown in ref . @xcite that the tunneling time follows the band structure profile . figure [ fig5 ] ( inset ) displays clearly , at @xmath100 , the familiar discrete quasi - stationary hole - level spectrum of the embedded qw , widely accepted from hole quantum transport calculations @xcite . we have found at low energies ( @xmath100 ) , that the @xmath64 resonances are allocated at energies which correctly reproduce the hole spectrum previously reported for the dbrt experiment @xcite , as expected ( see table [ tab1 ] ) . notice the good quantitative agreement of our approach in table [ tab1 ] , with similar values reported elsewhere @xcite . a previous comparison @xcite with a robust and widely accepted theoretical model @xcite , exhibits the advantages of the present approach . lcc + levels @xcite & + & experiment@xcite . & theory ( msa ) . + @xmath112 & 0,028 & 0,027 + @xmath113 & 0,073 & 0,071 + @xmath114 & 0,111 & 0,110 + @xmath115 & 0,245 & 0,247 + @xmath116 & 0,299 & 0,307 + @xmath117 & 0,428 & 0,434 + + tunneling time can not be quantitatively determined from the resonant tunneling line width , @xcite nevertheless some qualitative prognosis is allowed . notice the general trend of larger values for @xmath5 states phase time in comparison to that of @xmath6 states . this is in correspondence with the properties we have found for the transmission - coefficient resonances depicted in the inset of figure [ fig5 ] , which shows , as tendency , that @xmath5 resonances are thinner than the @xmath6 ones . this observation foretells a stronger confinement for heavy holes , and we guess it could be used in resonant - tunneling diode design and nano - modelling @xcite . among the different formulations of tunneling time with recurrent experimental agreement @xcite , the phase time have clearly advantages . in reference @xcite , was demonstrated that the phase time associated with the passage ( tunneling or not ) of particles is a relevant quantity of the quantum transport , which was calculated and successfully compared with precise experimental measurements @xcite . moreover in reference @xcite , some appealing effects for electron phase time were predicted , namely : superluminal phase propagation and resonant - band structure . in an attempt to distinguish similar behaviors for holes , we perform some numerical simulations , exhibited in the next two figures . in figure [ figtauvsen2 ] the phase transmission time @xmath118 , through a single cell ( @xmath119 ) and a dbrt ( @xmath120 ) respectively , together with @xmath121 are shown in connection with the free motion time @xmath122 ( dotted line ) . the last , is the propagation time of the quasi - particle in absence of scattering potentials . the direct paths under analysis , are indicated in the legends of each panel . a resonant behavior for @xmath123 is very clear for both @xmath5 states ( top panel ) and @xmath6 states ( bottom panel ) . in the forbidden energy region , @xmath123 approaches @xmath124 ( dashed red line ) , which represents its lower bound as had been foretold for electrons @xcite . the relationship between @xmath123 and @xmath121 ( blue line ) resonances - as one can see from ( [ phase_time_prop])- , justify @xmath123 to accurately reproduce , the quasi - stationary states of the embedded @xmath0 qw . another appealing feature arises for forbidden energy regions ( @xmath121 vanishes ) , where @xmath125 . the opposite is predicted for allowed energies . for @xmath5 states ( top panel ) of incoming energies around @xmath126 , @xmath127 , while for @xmath6 states ( bottom panel ) @xmath128 . this earlier arrival time for @xmath5 and @xmath6 , was predicted for electron tunneling through a same heterostructure @xcite , and suggest a more speedily passage of holes through a dbrt like @xmath129 . top(bottom ) panel shows the evolution of the phase time , transmission probability , free motion time through @xmath120 cells ( dbrt ) and the phase time through one single cell as function of the incoming energy for direct transitions of @xmath130 . with @xmath131 , @xmath132 and @xmath133 a @xmath74 concentration of @xmath75 was fixed inside the barrier.,width=302 ] it remains to notice that @xmath134 for the dbrt , having a bigger difference at low energies , with respect to that of high energies . another remarkable distinction is the width of @xmath121 resonances , which are noticeable widest for @xmath6 . this difference yields to a larger confinement for @xmath5 inside the embedded qw , with respect to that for @xmath6 . thus greater lifetime for @xmath5 in the scattering region is expected . top(bottom ) panel shows the evolution of the phase time , transmission probability , free motion time through @xmath73 cells ( dbrt ) and the phase time through one single cell as function of the incoming energy for direct transitions of @xmath130 . with @xmath131 , @xmath132 and @xmath133 . a @xmath74 concentration of @xmath75 was fixed inside the barrier.,width=302 ] in figure [ figtauvsen8 ] the phase time @xmath135 and transmission coefficient @xmath136 , through a superlattice ( @xmath73 ) undoubtedly follow a resonant - mini - band structure . direct hole transitions were considered and appear indicated in the legends of each panel . for finite periodic systems -which is the case envisioned here- , a finite number of intra - band energy levels @xcite are typically observed , at variance with the continuous band structures predicted for infinite periodic systems . in the mini - gaps , @xmath136 vanishes and @xmath135 tends to @xmath124 , which is similar to the dbrt case discussed before . on the other hand , in the mini - bands region @xmath135 exhibits a resonant behavior approaching @xmath122 , which behaves as a lower bound @xcite . in the scattering regions , @xmath135 increases as the incoming energy approximates to any allowed energy level . while , in the forbidden regions the holes propagation is away faster , reflecting the evanescent behavior and the lack of hospitality of the barrier . these evidences for @xmath5 and @xmath6 phase time , resemble those for electrons as was foretold for the tunneling time in a similar superlattice @xcite . phase transmission time as a function of the barrier thickness for heavy hole transition @xmath137 within the uncoupled regime , @xmath138 . both incident energies @xmath46 : less ( full triangles ) and greater ( empty triangles ) than the barrier height @xmath60 were fixed at @xmath139 ev and @xmath140 ev , respectively . the two bottom and top panels represents the numerical values of this time for the dbrt and the superlattice of @xmath141 cells respectively.,width=302 ] figure [ figtauvslbn24 ] exhibits two dissimilar behaviors of @xmath64 for direct transitions @xmath95 as function of the barrier thickness . for energies under the potential barrier ( full triangles ) @xmath64 becomes independent of the barrier thickness , being consistent with hartman s classical premonition for electrons @xcite . in the dbrt , @xmath64 becomes autonomous starting from @xmath142 , meanwhile , in a superlattice of @xmath141 cells , this quantity saturates at about @xmath143 . we can also note a decrease of the phase time in the region @xmath144 for the dbrt and around @xmath145 for the superlattice , showing the attractive effect of the barrier , as a reminiscence of the similar behavior widely quoted in the literature for electrons through a simple barrier @xcite . for energies higher than the potential barrier ( empty triangles ) @xmath64 , in the dbrt and in the superlattice , exhibits a resonant oscillating behavior similar to those of the single barrier ( see figure [ figtauvslbn1 ] ) . this is a consequence of an interference effect between the reflected wave and the continuum states , known as ramsauer - townsend oscillations as commented above . by comparing figure [ figtauvslbn1 ] and figure [ figtauvslbn24 ] , it is clear that oscillations @xcite remain robust for the dbrt . the robustness of this feature vanishes for the superlattice , provides that well - resolved oscillating curves are no longer observed for @xmath141 cells . the bottom panel plots the @xmath146 outgoing channel conductances @xmath147 through a single barrier heterostructure with @xmath107 , where @xmath148 ( solid line ) represents the @xmath149 channels , respectively , and @xmath150 ( dashed line ) represents the @xmath151 channels , respectively . the top panel displays the cosine of phase shift for the same transitions.,width=264 ] it is quite well established in the literature the strong relation between the transmission magnitudes and the wave function s phase @xcite . specifically , the giant conductance events are nearly related with phase shifts and phase interference phenomena @xcite . it has been shown for various mesoscopic systems , the correlation between the variations of the conductance and an even function of the phase shift , due to an alternatively appearance of maximum - minimum valued sharp peaks or dips , whenever an even functional of the phase shift , changes by almost @xmath152 radians @xcite . here , the cosine of the phase shift was chosen to verify , whether strike maximum - transmission peaks occur or not , in a negligible conductance background . commonly , the @xmath153 is plotted , as it is a suitable even function to show evidences of giant conductance events @xcite . figure [ fig6 ] ( bottom panel ) , presents the conductance @xmath154 to direct channels @xmath95 ( solid line ) , @xmath155 ( dashed line ) , and the cosine of phase shift for the same transitions ( top panel ) for the uncoupled hole regime as function of the incident energy . it is clear that the maximum values in the conductance for both output channels are always accompanied by an extreme change in the cosine of the phase difference . one can see , that the curve for @xmath5 output channel ( solid line ) changes two times in @xmath152 radians and in the bottom panel appear two @xmath147 maximum correspondingly ( at @xmath156 and at @xmath157 ) . meanwhile for the @xmath6 output channel ( dashed line ) , there is only one change ( at @xmath158 ev ) . direct paths contributions are the largest ones , as it follows from vanishing band mixing effects as @xmath97 . the maximum values shown in the bottom panel of figure [ fig6 ] , for an even function of @xmath159 ( @xmath160 ) , at even multiples of @xmath152 @xcite , are not enough to confirm an abrupt increment of transmission . indeed , there is nothing particularly large about hole transmission through a single cell , as that reported for electrons in finite periodic systems with time reversal symmetry @xcite . besides , in the single - cell case , does not exist a resonant mechanism that allows hole conductance to become giant through the barrier . however , it is tempting to have a fairly reference frame to compare to what we will see next for a dbrt . worthwhile noticing -although not shown here- , the largest contribution of the imaginary scattering amplitude component , that give rise @xmath147 to maximize in the same energy intervals , as we have shown in the figure [ fig6 ] bottom panel . dependency of the @xmath146 outgoing channels conductance and the cosine of phase shift with hole incident energy , through a dbrt and a superlattice with @xmath141 cells of width @xmath107 @xmath78 and height @xmath60.,width=302 ] well - established time reversal symmetry of the @xmath10 kohn - lttinger model , is not fulfilled by the @xmath161 subspaces as was recently demonstrated @xcite . events of gg have been observed in many finite periodic systems , which possess time reversal invariance but as yet not observed in time asymmetric ones . probably , this explains why other accurate calculations made in the @xmath161 kohn - luttinger subspaces , did not report gg events for holes @xcite . figure [ figgg ] displays the conductance and phase shift behavior as a function of the incident energy , for direct transitions to @xmath149 ( solid line ) and @xmath151 ( dashed line ) output channels . the @xmath162 curves exhibit evidences pointing out the existence of well - defined resonances for @xmath163 . it is relevant in this case , the presence of striking conductance events , as can be clearly noticed in the panels for @xmath120 of figure [ figgg ] . indeed , peaks @xmath116 and @xmath117 , have found to have a clear signature of gg phenomenon since : ( i ) they are sharp resonances , ( ii ) they keep a simple relation with an even function @xmath164 of the transmission phase @xmath165 , thereby peaks locations at energy , are coincident with those of @xmath165 whenever the last changes by almost @xmath152 radians in a relative short energy interval @xcite , and finally ( iii ) there is a strong phase interference mechanism between propagating or evanescent modes , giving rise to resonant transmission @xcite . as follows from the scattering theory , the gg mechanism in a dbrt , derives from the interference between states , whenever the incoming modes energy perfectly matches with one of the quasi - stationary states within the embedded qw of the scatterer system . this agreement produces the essential conditions for energy to maximize , _ i.e_. a resonant transmission occurs . sharp peaks become visible at energies @xmath166 ( dashed line ) for direct transitions of @xmath6 , and for @xmath167 ( solid line ) for direct transitions of @xmath5 . we have disregarded the @xmath113 sharp resonances and others in the low energies region , because they do not reach maximum and also the corresponding phase shifts differ significantly from @xmath152 radians . in addition , it is necessary to mention the loss of the gg phenomenon in the superlattice . as one can see in the top panels of figure [ figgg ] ( for @xmath141 and @xmath100 ) , the @xmath168 changes are not about @xmath152 radians . in this case the leak of maximum transmission for the superlattice becomes visible . we conjecture the existence of a competitive mechanism when the number of cells grow , and consequently a phase coherence effect rises . for higher energies , the direct @xmath162 might give rise to a _ plateau _ , and we foretell the probable existence of another related striking feature : the giant resistance . despite a more ambiguous interference process in this case , a sharp dips are expected as a consequence of a transmission rate decrease , due to a resonant transfer of holes flux toward evanescent modes . it is likely to underline that the energies of the resonances -whether they are giant or not- , accurately reproduce the quasi - bounded hole spectrum as was published for the dbrt @xcite , were another approaching frame was applied . this is to say that , none gg phenomenological analysis can possibly be complete if a resonant means is missing . in this sense , to endorse the gg evidences showed in figure [ figgg ] , the most immediate tool within the msa , is the tm technique to estimate the hole states energies at the qw of the dbrt . the basic idea here , is to search several states of the embedded finite quantum well ( fqw ) , by cautious tuning those of the external infinite quantum well ( iqw ) . the last we had achieved , by imposing that @xmath169 slowly tends to @xmath170 , respectively ( see figure [ figdbrt ] ) . this method approaches the iqw boundary conditions to those of the fqw , and is a reasonably good sketch for an estimation of some fqw levels . schematic representation of the quasi - bound - state levels ( horizontal lines ) and valence band lineup , within an infinite walls quantum well . the ground state wave function ( dotted line ) is represented qualitatively . the energy is taken as positive for convenience.,width=302 ] figure [ figdbrt ] shows schematically , holes quasi - bound - state levels ( horizontal lines ) and valence band lineup , within the iqw . for a numerical estimation of the quasi - stationary levels within the fqw , we consider a cell enclosed between the @xmath20 and @xmath171 points . the barriers and well have the same dimensions of the figure [ figgg ] . thus , it is undemanding and straightforward to reproduce a dbrt or a superlattice , by a periodic concatenations of cells from @xmath171 @xcite . the latter considers facts of equivalence for points @xmath172 and @xmath173 , respectively , as they have the same materials to the left and to the right . as boundary condition for the iqw , is standard to impose the wave function goes to zero at the points @xmath174 and @xmath175 @xmath176 being @xmath177 the splitting distance between these points and the infinite potential . in the framework of the tm approach , the latter requirement can be written as : @xmath178 therefore , the non trivial solutions of the expression ( [ annulm_cond ] ) can be numerically determined , by solving the equation @xmath179 being @xmath180 the _ amplitude - derivative _ matrix block of the tm @xmath181 . up to this end , it is turn clear from table [ tab2 ] , a reasonably good agreement between our estimation for the energies of several quasi - steady levels of the fqw and the resonant - band structure displayed in figure [ figgg ] . now it is undoubted , the very existence of these fqw levels , and therefore their role as a fair resonant mechanism of the gg phenomena in the dbrt we claim for . the @xmath116 and @xmath117 energy values showed here , are a truthful prove of the gg phenomena , predicted for the uncoupled hole transport in the dbrt for sharp resonances allocated at similar energies ( see figure [ figgg ] ) . two - probe conductance through the four direct accessible channels as a function of the incoming energy on the @xmath138 limit , for a superlattices of @xmath182 cells ( from bottom to top ) in the range of energies between @xmath183\ ev$].,width=302 ] ll + levels ( labeling from ref.@xcite ) & resonances @xmath184 $ ] + @xmath114 & 0.1362 + @xmath116 & 0,2978 + @xmath117 & 0,4112 + + phase transmission time through one single cell _ versus _ barrier thickness for the direct and crossed passages of holes in a coupled case ( @xmath185 @xmath186 ) for @xmath187.,width=302 ] figure [ figgn789 ] presents the dependency of the conductance @xmath188 of @xmath5 and @xmath6 through a superlattices of @xmath182 cells with the energy of the incident flux . the holes travel along all four accessible direct channels @xmath189 ( solid line ) and @xmath190 ( dotted - dashed line ) . it is wide open visible the thinner peaks of the @xmath5 resonances , respect to those of the @xmath6 ones , since we are restricted on the uncoupled regime and thereby @xmath5-spectrum states are the closest due to their larger effective mass . further outstanding behavior , is the filter - like effect that can be observed on both flavors of holes in the selected region . indeed , the barriers become opaque for the @xmath191 states at low / high energies , leading to a stronger confinement for the @xmath191 at low / high energies . this fact yields a larger / shorter lifetime for @xmath191 inside the allowed region of the scatterer system for low energy , at variance with the opposite for higher energies . dealing with a superlattice of @xmath192 cells ( topmost panel ) , we found an @xmath193 overlap region at the vicinity of @xmath194 ev . the phenomenology described in figure [ figgn789 ] , allows us hypothesize it as concerns applications in hole - based electronics , pursuing maximum - minimum conductivity response as function of the charge carrier effective mass . the bottom panel shows the crossed - path contributions to the conductance through the first outgoing channel , @xmath195 , for a dbrt with barriers of width @xmath107 , and height @xmath60 . the top panel displays the cosine of two phase shifts that match with the same outgoing channel , @xmath196 and @xmath197 , both in the same range of energy.,width=302 ] the bottom panel shows the the crossed - path contributions to the conductance through the second outgoing channel , @xmath198 , for a dbrt with barriers of width @xmath107 , and height @xmath60 . the top panel displays the cosine of two phase shifts that match with the same outgoing channel , @xmath199 and @xmath200 , both in the same range of energy.,width=302 ] let us considers first the hartman effect under this regime . figure [ figtauvslb8akt510 - 4 ] depicts the phase transmission time as a function of the barrier thickness for several crossed transitions paths , namely : @xmath201 and @xmath202 , of the sixteen possible ones , through a single - cell scatterer system . this plot shows the phase transmission time curves as a function of the barrier thickness @xmath93 , given fixed @xmath100 . notice the changes of @xmath203 through each path while @xmath93 increases , until @xmath203 reaches saturation values . this autonomous behavior of the phase time with @xmath93 , unambiguously gives rise to the paradoxical hartman effect @xcite , which remains robust , even for a relative strong coupling of holes . owing to simplicity we dropped the remain transitions , provided they follow a similar tendency . worthwhile underlying some characteristics of @xmath203 transient interval for several paths . crossed transitions @xmath197 ( + ) and @xmath199 ( up empty triangles ) , show negative slope before reach saturation . meanwhile the crossed path @xmath196 ( down full triangles ) , exhibits a double change of slope . these behaviors clearly departs ( in the transient interval ) from the classical hartman premonition for electrons and deserve more investigation to shed light on this topic . in the presence of interference effects ( @xmath204 ) at @xmath205 , we have found abrupt tunneling phase changes when @xmath206 radians , which seems to be an anomalous behavior of the phase time . in figure [ figtauvslb8akt510 - 4 ] , we have found an overlap of @xmath203 for several crossed transitions . for example : with @xmath207 at @xmath208 , also with @xmath209 at @xmath210 $ ] , and with @xmath211 at @xmath212 $ ] . in some cases the effective mass changes , and the incoming channels belong to a so called kramer-_up _ degenerated hole states @xcite , while the outgoing channels are kramer-_low _ ones . it can be readily observed , similar @xmath203 matches , for other crossed transitions at different values of @xmath93 . these coincidences relate to the strong @xmath193 propagating modes mixing , to the simultaneous treatment of the @xmath146-coupled channels of the present approach , and to @xmath193 interaction with an effective potential @xmath213 due the barrier broadening . accordingly , whenever a match in @xmath203 takes place , the time spent by the @xmath5 and @xmath6 in the classically forbidden region , shall be the same . full contributions to the conductance through the second outgoing channel , @xmath198 , for a dbrt with barriers of width @xmath107 , and height @xmath60.,width=302 ] we turn now to discuss how robust are the hole giant conductance features , in the presence of valence - band mixing . figure [ figcros1 ] and figure [ figcros2 ] present the conductance and the @xmath164 through the outgoing channels @xmath84 and @xmath86 , respectively , as a function of the incoming energy under interference between channels . we had taken @xmath214 @xmath186 . despite the presence of sharp peaks at @xmath100 ( qualitatively similar to those showed in figure [ figgg ] ) , as one can observe at @xmath215 , @xmath216 and @xmath217 in fig.[figcros1 ] , it is not seen a neat concordance with the widely accepted gg formulation posted above @xcite , and illustrated in fig.[figgg ] . firstly there is not a nice correspondence between the highest peak of @xmath195 at the vicinity of @xmath218 ev ( see fig.[figcros1 ] bottom panel ) and the fluctuation of @xmath165 , which for this peak is sensitive less than @xmath152 radians ( see fig.[figcros1 ] top panel ) . on the other hand , the values of @xmath219 exhibited in fig.[figcros1 ] and fig.[figcros2 ] , respectively , are in average almost one order smaller than those shown by fig.[figgg ] , being the comparative behavior observed in fig.[figcros2 ] even more restrictive in this sense , with respect to fig.[figgg ] . following these evidences , we conjecture so far , that the gg events are not straightforward signatures for hole tunneling traversing a dbrt , under a selected range of parameters in the coupled - particle regime . additional remarkable differences become clear from a comparison of this trend in the coupled hole regime and figure [ figgg ] . for instance , there is a small shift between the position in energies of the peaks and the changes in the @xmath164 . we guess that the last reflects the loss of maximum transmission when the valence - band mixing , _ i.e. _ the interference between paths is incremented . besides , the @xmath220 oscillations spectra in the low - energy interval shown in fig.[figcros2 ] , seem to be of the fano - type leading to assume the existence of interference of embedded - qw discrete levels with a continuum . the later gives rise to characteristically asymmetric peaks in transmission spectra @xcite . the bottom panel plots tunneling phase time for the crossed - transition @xmath221 through a single barrier heterostructure at @xmath222 @xmath186 , with @xmath107 @xmath223 and @xmath224 ev . top panel shows the cosine of phase shift for the same transition @xmath221 ( dashed line ) where the first number stands for the input channel @xmath36 and the other stands for the output one @xmath37 . in dotted line the same magnitude is shown for the direct path @xmath225 with respect to the crossed path @xmath221.,width=302 ] the porpoise of figure [ figcros3 ] , is to briefly analyze further anomalous features in the transmission phase time spectra . fig.[figcros3 ] displays in panel ( a ) the @xmath226 for the crossed @xmath227 transition @xmath228 ( dashed line , @xmath229 ) . in panel ( a ) with dotted line , we plot the difference of the the same magnitude for the path @xmath230 with respect to the path @xmath231 . the bottom panel ( b ) plots tunneling phase time for the crossed - transition @xmath232 through a single barrier heterostructure . in the presence of interference effects ( @xmath233 ) at @xmath234 ev , the phase changes abruptly when @xmath235 radians , as can be seen from the dashed line in the top panel of figure [ figcros3 ] . both curve in panel ( a ) show a clear correspondence of abrupt changes in @xmath236 with observed peaks(dips ) of @xmath237 , which seem to be an anomalous behavior of the phase time , as they resemble fano - like resonances due theirs asymmetry hallmark @xcite . in the neighborhood of @xmath238 ev [ see panel ( b ) ] we have found fano - like positive ( negative ) peaks ( dips ) of @xmath237 with @xmath239 , displayed in the bottom panel of figure [ figcros3 ] . they are associated with changes of the phase in almost @xmath240@xmath241@xmath242 radians , and are plotted in the top panel with dashed line . we also have observed , although will not be shown here , local increments of transmission coefficients in the energy vicinity , where negative dips appeared , though regretfully they need more investigation to derive proper conclusions . most interestingly , though , we found indications that might be connected to possible giant conductance events in the single - cell case , which we omit graphically here because of brevity . there were found two prompt shifts in transmission phase amplitude , equivalent to an even multiple of @xmath243 ( @xmath244 ) at @xmath245 @xmath186 , similar to those plotted in figure [ figgg ] . one of them , appear in @xmath246 intra - band transition ( @xmath247 ) when @xmath93 varies within @xmath248 $ ] @xmath223 range . the second one , was observed through @xmath249 inter - band transition ( @xmath250 ) when @xmath93 varies within @xmath251 $ ] @xmath223 range . then , maxima of @xmath147 could be expected for those crossed paths , not only varying the incident energy , but the barrier thickness as well . however , a clear resonant mechanism in this case remains a puzzle and further investigations need to be addressed to clarify wether the effect of giant conductance for holes , is expected to arise under valence - band coupling phenomena during transport trough a single - cell system . relevant and striking quantum transport features of holes , directly connected to the phase , the phase shift and the phase resonant interference phenomena are : the giant conductance - resistance events , the hartman effect , and the earlier phase propagation in the barrier . several results nicely resemble those previously reported for electrons @xcite , while some predictions of the msa model are in acceptable qualitative compliance with the experiments @xcite and with theoretical predictions @xcite . theoretical evidences of the giant conductance phenomena for decoupled hole transmission through dbrt were exhibited . in this case a trustworthy resonant mechanism supporting this anomaly , is the alignment of the incident propagation energy with one of the quasi - bond hole states of the embedded quantum well in the dbrt . this mechanism was confirmed with an independent numerical calculations of several quasi - bound levels . the giant conductance effect vanishes in the presence of valence - band particles coupling and by tuning the number of superlattice layers as well . the hole phase transmission time for a @xmath252 superlattice , shows a resonant mini - bands structure in the scattering regions , having the free motion as an inferior limit time . moreover , in the forbidden regions the holes propagation became away faster and was shown @xmath64 to be inferiorly bounded by the single - cell phase time @xmath124 . in the dbrt mini - gaps regions , @xmath64 moves forward from the free motion time , suggesting a more speedily phase propagation within the barrier . the hartman s classic prediction is observed robust for null and finite hole band mixing . was presented an alternate - selective confinement strength independently for both flavors of holes , implying filter - like effects on the effective mass _ via _ the manipulation of the incident energy in the uncoupled hole regime . we thanks dr . h. rodrguez - coppola for clue suggestions in some numerical treatments . one of the authors ( l.d - c ) gratefully acknowledges the visiting academic program of the uia / mxico and the facilities of the ifsc - usp / so carlos , brasil . 00 h. schneider , h. t. grahn , k. klitzing , and k. ploog , phys . b * 40 * , 10040 ( 1989 ) . d. dragoman and m. dragoman , j. appl . phys . * 93 * , 6133 ( 2003 ) . l. diago - cisneros , h. rodrguez - coppola , r. prez - lvarez , and p. pereyra , phys . b * 74 * , 045308 ( 2006 ) . l. diago - cisneros , h. rodrguez - coppola , r. prez - lvarez and p. pereyra , rev . fs * 53 * , 7 ( 2007 ) . p. pereyra and h. simanjutak , phys . e * 75 * , 1 ( 2007 ) . h. g. winful , phys . lett * 91 * , 26 ( 2003 ) . h. g. winful , phys . lett * 90 * , 2 ( 2003 ) . e. h. hauge and j. a. stvneng , rev . phys * 61 * , 917 ( 1989 ) . r. landauer and th . martin , rev . phys * 66 * , 217 ( 1994 ) . p. pereyra , v. g. ibarra - sierra and j. l. cardoso , microelectronics journal * 40 * , 779 ( 2009 ) . p. pereyra , phys . rev . lett . * 84 * , 1772 ( 2000 ) . a. m. steinberg , p. g. kwiat and r. y. chiao , phys . lett * 71 * , 708 ( 1993 ) . spielmann , r. szipcs , a. stingl , and f. krausz , phys . lett . * 73 * , 2308 ( 1994 ) . s. chu and s. wong , phys . lett * 48 * , 738 ( 1982 ) . s. bosanac , phys . a * 28 * , 557 ( 1983 ) . a. b. shvartsburg , m. marklund , g. brodin and l. stenflo , phys . e * 78 * , 016601 ( 2008 ) . a. kadigrobov , a. lagoskin , r. i. shekter and m. jonson , phys . b * 52 * , r8662 ( 1995 ) . a. kadigrobov and m. jonson , applied physics report * 94 * , 6 ( 1994 ) . p. pereyra , physica e * 17 * , 209 ( 2003 ) . n. k. allsopp , j. snchez - caizares , r. raimondi and c. g. lambert , j. phys . condens . matter * 8 * , l377 ( 1996 ) . roberto romo , phys . b * 66 * , 245311 ( 2002 ) . jorge villavicencio and roberto romo , phys . b * 68 * , 153311 ( 2003 ) . junling wu , baigeng wang , jian wang and hong guo , phys . b * 72 * , 195324 ( 2005 ) . yuanbo zhang , victor w. brar , feng wang , caglar girit , yossi yayon , melissa panlasigui , alex zettl and michael f. crommie , nature physics * 4 * , 627 ( 2008 ) . a. k. geim and k. s. novoselov , nature materials * 6 * , 183 ( 2007 ) . guangyu xu , carlos m. torres jr . , emil b. song , jianshi tang , jingwei bai , xiangfeng duan , yuegang zhang and kang l. wang , arxiv.org ( to appear in nano letters ) ( 2010 ) . r. wessel and m. altarelli , phys . b * 39 * , 12802 ( 1989 ) . e. p. wigner , phys . rev . * 98 * , 145 ( 1955 ) . t. e. hartman , j. appl . phys * 33 * , 12 ( 1962 ) . l. a. maccoll , phys . rev * 40 * , 621 ( 1932 ) . d. bohm , _ quantum theory _ , ( prentice - hall , new york , 1951 ) . f. t. smith , phys . rev . * 118 * , 349 ( 1960 ) . . j. nucl.phys . * 4 * , 182 ( 1967 ) . m. bttiker and r. landauer , phys . 49 * , 1739 ( 1982 ) . m. bttiker phys . b * 27 * , 6178 ( 1983 ) . g. nimtz , a. haibel , and r .- m . vetter , phys . e * 66 * , 037602 ( 2002 ) . a. ranfagni , d. mugnai , p. fabeni and g. p. pazzi , appl . . lett . * 58 * , 774 ( 1991 ) . a. enders and g. nimtz , j. phys . i2 * , 1693 ( 1992 ) . j. m. lttinger and w. kohn , phys . rev . * 97 * , 869 ( 1955 ) . l. diago - cisneros , _ tunelaje multicanal y simetras de los huecos en heteroestructuras semiconductoras _ , universidad de la habana , facultad de fsica , la habana ( 2005 ) . p. a. mello , p. pereyra and n. kumar , ann . ( n.y ) * 181 * , 290 ( 1988 ) . j. imry and r. landauer , rev . * 71 * , 306 ( 1999 ) . a. p. heberle , x. q. zhou , a. tackeuchi , w. w. rhle , and k. khler , semicond . * 9 * , 519 ( 1994 ) . t. kumar , m.cahay , and k. roenker , phys . b * * 5**6 , 4836(1997 ) . g. klimeck , r. chris bowen , and t. b. boykin , supperlattices and microstructures * 29 * , 188 ( 2001 ) . r. landauer , nature ( london ) * 341 * , 567 ( 1989 ) . m. f. krol , s. ten , b. p. mcginnis , m. j. hauduk , g. khitrova , and n. peyghambarian , phys . b * 52 * , r14344 ( 1995 ) . s. ten , m. f. krol , p. t. guerreiro , and n. peyghambarian , appl . . lett . * 69 * , 3387 ( 1996 ) . m. morifuji and c. hamaguchi , phys . b * 52 * , 14131 ( 1995 ) . m. u. erdogan , k. w. kim , and m. a. stroscio , app . lett . * 62 * , 1423 ( 1993 ) . a. d. snchez and c. r. proetto , j. phys . condens . matter * 7 * , 2059 ( 1995 ) . k. v. rosseau , k. l. wang , and j. n. schulman , superlattices and microstructures * 6 * , 67 ( 1989 ) . e. e. mendez , w. i. wang , b. rico , and l. esaki , appl . lett * 47 * , 415 ( 1985 ) . p. g. n. de vegvar , t. a. fulton , w. h. mallison , and r. e. miller , phys . lett . * 73 * , 1416 ( 1994 ) . h. pothier , s. guron , d. esteve , and m. h. devoret , phys . * 73 * , 2488 ( 1994 ) . l. diago - cisneros , h. rodrguez - coppola , r. prez - alvarez , and p. pereyra , phys . scripta * 71 * , 1 ( 2005 ) . g. nimtz , _ proceedings of the adriatico research conference on tunneling and its implications _ , pp.@xmath253 ( ictp , trieste , italy , 1996 ) . a. m. steinberg and r. y. chiao , phys . a * 49 * , 3283 ( 1994 ) . we have obtained the amplitude of the exponential fitting function , for the phase transmission time with different sections of barrier thickness , to be constant starting from @xmath254 . luis de la pea , _ introduccin a la mecnica cuntica _ , universidad nacional autnoma de mxico , fondo de cultura econmica , 2006 , cap.6 . in additions of an exponential growing functions , adjustments _ via _ @xmath45-order polynomial ( with @xmath255 ) and cycloid functions are suitable as well . p. pereyra and e. castillo , phys . b * 65 * , 205120 ( 2002 ) . v. t. petrashov , v. n. antonov , p. delsing and r. claeson , phys . rev . lett * 70 * , 3 ( 1993 ) . u. fano , phys . rev . * 124 * , 1866 ( 1961 ) .

after almost two decade of intense research about the cuprates , there is still plenty of open questions in this problem . however , it is recognized that the electrons which move on the @xmath2 planes are the most relevant to describe their physical properties @xcite . in the undoped regime , these compounds are insulators and exhibit antiferromagnetic order at sufficient low temperatures @xcite . the physical properties of the insulating phase can be well described by the heisenberg model @xcite . upon doping , these systems suppress the antiferromagnetic order and become superconductors . in this scenario there is no doubt that the @xmath3-@xmath3 electron correlations play a fundamental role . the study of the electronic structure near the fermi level @xmath4 in such strongly correlated systems is very important to understand their physical properties @xcite . earlier angle - resolved photoemission experiments ( arpes ) have showed the presence of flat bands close to @xmath4 in a region centered around the point @xmath5 in the @xmath6-type cuprates like @xmath7@xmath8@xmath9 and @xmath10@xmath11@xmath12 @xcite . due to the presence of strong correlations , to study some physical properties of these cuprate compounds , the one - band hubbard model @xcite can be used . _ @xcite have done monte carlo calculations in the one - band hubbard model . their results show bands with a flat region near @xmath5 point for a given doping which agreed with the previously mentioned arpes results @xcite . beenen and edwards @xcite , using the roth s two - pole approximation @xcite in the one - band hubbard model , have studied the normal state of the model obtaining flat quasi - particle bands , which agree well with those found with monte carlo simulations @xcite . the roth s two - pole approximation has been proposed to improve the hubbard - i approximation @xcite by considering a decoupling scheme which produces an additional energy shift ( the roth s band shift ) in the peaks of the spectral function . that result is in agreement with those obtained by harris and lange @xcite . they looked at the moments of individual peaks in the spectral function . the presence of the exchange term @xmath13 in the roth s band shift exhibits in it a spin dependence . as consequence , the roth s method raises the possibility of magnetic solution in the hubbard model while this feature is not present in the hubbard - i approximation . recently , due to the good agreement between the roth s and the monte carlo data , beenen and edwards have extended the roth s two - pole approximation in order to investigate the superconducting properties of the one - band hubbard model . their main achievement has been to show the emergence of the pairing with @xmath14 symmetry in a given amount of doping . in that approach , the gap equation for @xmath3-wave symmetry depends on a particular four operator correlation function which , in principle , can be found extending the roth s formalism to obtain two particle green s functions . however , the authors have introduced two decoupling schemes to calculate the gap . the first one ( the factorization procedure ) has been formulated to treat the problem for intermediated values of @xmath15 ( the coulomb interaction ) and it provides an upper estimate for the gap and @xmath16 ( the critical temperature ) . the second one is adapted for very large @xmath15 scenario which preserves the proper limit for @xmath17 where the gap function vanishes . nevertheless , the one - band models neglect the presence of the oxygen sites . due to the strong correlations at the @xmath18-sites , the oxygen sites may be occupied by holes when the system is doped @xcite . for instance , the hubbard one - band model suffers some limitations to describe the low - energy physical properties of the cuprate superconductors @xcite . in the doped regime , the one - band hubbard model gives a wrong description of various properties , like , for example , the asymmetric magnetic doping - temperature phase diagram @xcite . therefore , a model which take into account also the oxygen can be more adequate to treat the cuprate systems in the doped regime @xcite . this raises the question whether it is possible to extend the beenen and edwards analysis to investigate the @xmath3-wave symmetry superconductivity when the hybridization is present . recently , the present authors have used the extended hubbard model @xcite with the roth s method to study the role of hybridization in the superconductivity following closely the approach introduced by beenen and edwards @xcite . as discussed in the references @xcite and @xcite , the flattening of the bands is directly related to the band shift . the presence of flat bands at fermi level @xmath4 in the @xmath6-type cuprates @xcite suggests a high density of states at the fermi level , which can favor pair formation . therefore , considering that the main responsible elements for the density of states at the fermi level are the @xmath3-electrons , it has been assumed in ref.@xcite that the @xmath19 pairs are the most relevant ones for superconductivity @xcite . the band shift plays an important role in the study of the superconducting properties of the model using the roth s or some similar procedures . in reference @xcite , the factorization procedure @xcite has been used to investigate the effects of the hybridization on the superconductivity . it has been shown @xcite that the hybridization has strong effects in the shift and , therefore , in some superconducting physical properties such as the critical temperature @xmath16 . however , as a first approach , in reference @xcite , the band shift has been evaluated taking into account the hybridization effects , but disregarding temperature effects , superconducting properties and , most important , it has been considered in the limit @xmath20 . as a consequence of this limit , many correlation functions , which appear in the shift , are vanished . the important point is that these correlation functions are very relevant in the sense of to include correctly the hybridization effects . therefore , it would be necessary to calculate the shift with the @xmath15 finite in order to include the hybridization effects in a more complete way . in this work , the superconductivity problem has been studied using the roth s method , following closely reference @xcite , but adapted to the @xmath3-@xmath6 extended hubbard model . here , special attention is devoted to the effects of the hybridization and superconductivity in the band shift . in order to have the effects of the hybridization included properly in the superconductivity , the gap function is obtained using the factorization procedure @xcite and the shift is evaluated with finite @xmath15 . this procedure is justified because it preserves some correlation functions present in the band shift , which are non vanishing for finite @xmath15 . as consequence , it captures the effects of the hybridization properly . some preliminary results of this approach have been given in ref . @xcite . there are some shortcomings in the beenen and edwards approach @xcite . for instance , the @xmath14 pairing is quite dependent on the choice of the decoupling scheme for the correlation functions related to the gap . however , in the present work , the main goal is to study the effects of hybridization . therefore , as discussed in the previous paragraph , the natural choice is the decoupling scheme for intermediated @xmath15 , which is also the simplest one . one is allowed to find in that procedure , at least , a better estimate for the gap ( and therefore for @xmath16 ) as a function of hybridization within the same decoupling procedure . the paper presents the following organization . in section [ sec:2 ] it is introduced the model and given a short introduction of the roth s method @xcite . also , some analytic expressions for quasi - particle bands and the green s functions are derived . in section [ sec:3 ] , the factorization procedure proposed by beenen and edwards @xcite is applied for the present case . in section [ sec:4 ] , the band shift is discussed in detail . the numerical results are showed and discussed in section [ sec:5 ] . finally , in section [ sec:6 ] , a short summary and some concluding remarks are given . the model considered here assumes overlapping bands . it is characterized by a narrow @xmath3-like band with a large density of states and a wide @xmath6-like band with low density of states . the extended hubbard model is defined as : @xmath21 where @xmath22 is the chemical potential . the @xmath23 , @xmath24 and @xmath25 , @xmath26 are the creation and annihilation operators of the @xmath3- and @xmath6-electrons , respectively , with spin @xmath27 at a lattice site @xmath28 . the @xmath29 and @xmath30 are the centers of the on site energies of the occupied orbitals of the copper and oxygen respectively . the second term of the hamiltonian given in eq . ( [ eq2.0 ] ) describes a narrow @xmath3-band with a hopping amplitude @xmath31 . the hamiltonian ( [ eq2.0 ] ) considers also a @xmath6-band which is wider than the @xmath3-band . the following relation between @xmath31 and @xmath32 can be established @xmath33 with @xmath34 . also , @xmath35 to coincide the bottom of the @xmath3- and @xmath6-bands with the @xmath36 point @xmath37 as suggested by experimental results @xcite . the third term corresponds to the coulomb interaction @xmath15 that represents the repulsion between two holes in the same @xmath3-orbital . the last term of the hamiltonian ( [ eq2.0 ] ) is the @xmath0 hybridization and describes the nearest neighbor hopping process between the @xmath3-orbital of the cu - atom and the @xmath6-orbital of the o - atom . considering a rectangular two dimensional lattice , the unperturbed @xmath3- and @xmath6-energy bands are given by @xmath38 and @xmath39 where @xmath40 is the lattice constant . in this work , the hamiltonian given in eq . ( [ eq2.0 ] ) has been investigated using the roth s two - pole approximation @xcite to obtain the green s function in the zubarev s formalism . in the roth s procedure , a set of operators @xmath41 is introduced in order to describe the relevant one particle excitations of the system . these operators satisfy in some approximation the following relation : @xmath42 _ { \left ( -\right ) } = \sum_{m}k_{nm}a_{m } \label{eq2.1}.\ ] ] anticommuting both sides of eq . ( [ eq2.1 ] ) with each operator of the set @xmath41 and taking the thermal average , the equation ( [ eq2.1 ] ) becomes : @xmath43 where @xmath44 and @xmath45 are the energy and normalization matrices , given by @xmath46 _ { \left ( -\right ) } , a_{m}^{\dag}\right ] _ { \left ( + \right ) } \right\rangle \label{eq2.3}\ ] ] and @xmath47_{\left ( + \right ) } \rangle \label{eq2.4}.\ ] ] in matrix notation , eq . ( [ eq2.2 ] ) is written as @xmath48 , where , if @xmath49 is nonsingular , then the @xmath50 matrix can be obtained . with the equation of motion ( in the zubarev s formalism ) of the green s function @xmath51 and the eqs . ( [ eq2.1])-([eq2.4 ] ) , it is possible to obtain the following general green s functions @xmath52_{\left ( + \right ) } \rangle \label{eq2.6}.\ ] ] in the particular case , where @xmath53 , the elements of the green s function matrix @xmath54 are given by eq . ( [ eq2.5 ] ) . thus , using the matrices @xmath55 and @xmath49 , the matrix @xmath54 is given by @xmath56 where @xmath57 considering the fact that the operators of the set @xmath58 describe the particle excitations of the system , the choice of these operators is very relevant to study the physical properties of the system . in order to discuss superconductivity , beenen and edwards , in their approach with the one - band hubbard model , mixed electron and hole operators and evaluated anomalous correlation functions @xcite . therefore , using a set of four operators @xmath59 @xmath60 , it has been obtained a four - pole approximation to the green s functions . however , in order to discuss the role of the hybridization , it is necessary to adapt the formalism to include a @xmath6-operator in the original set of operators used by beenen and edwards . thus , the new set of operators is given by @xmath61 in the present work , only the singlet pairing is considered , and particularly the _ d_-wave symmetry . in this particular case , @xmath62 and @xmath63 , where @xmath64 are the nearest neighbors of @xmath28 . using the set of operators given in eq . ( [ eq2.9 ] ) , and introducing the symmetries discussed above , the elements of the energy matrix defined in eq . ( [ eq2.3 ] ) can be obtained as : @xmath65\nonumber \\ \label{eq2.16}\end{aligned}\ ] ] where @xmath66 and @xmath67 are the fourier transform of @xmath68 and @xmath69 respectively . the matrix @xmath70 present in the energy matrix @xmath71 is given by : @xmath72 % \nonumber \\ \label{eq2.10}\end{aligned}\ ] ] where @xmath73 . it is assumed that the system considered here is translationally invariant , then @xmath74 . the quantity @xmath75 is the fourier transform of @xmath76 in eq . ( [ eq2.11 ] ) , the band shift @xmath77 is defined as : @xmath78 where @xmath79 can be separated into two explicit contributions @xmath80 the terms @xmath81 and @xmath82 are associated with the hopping @xmath83 and the hybridization @xmath69 , respectively . the hybridized term of @xmath79 may be written as @xmath84\delta _ { ij } \label{eq2.14},\ ] ] and the part associated to the hopping @xmath83 is given by @xmath85 the calculation of the correlation functions presented in eqs . ( [ eq2.14 ] ) and ( [ eq2.15 ] ) will be discussed in detail in section [ sec:4 ] . one of the most important elements of the matrix @xmath71 is @xmath86 , where @xmath87 and @xmath88 the correlation function @xmath89 gives the gap of the superconductor state in the @xmath90wave case . the elements of the normalization matrix @xmath91 are given from eq . ( [ eq2.4 ] ) as : @xmath92 and @xmath93 the remaining elements of the normalization matrix @xmath91 , due to the _ d_-wave symmetry and the anticommution rules , have been found to be zero . using the energy and the normalization matrices @xmath71 and @xmath91 , respectively , the matrix green s function @xmath94 defined in eq . ( [ eq2.7 ] ) can be obtained . for simplicity , only the _ most relevant _ elements ( for the purposes of this work ) of this @xmath95 @xmath96 matrix are shown . following the roth s notation @xcite , the correlation function @xmath97 is related to the green s function @xmath98 as : @xmath99 where @xmath100 is the fermi function . the chemical potential @xmath22 is obtained in the standard way , using the element @xmath101 of the matrix @xmath96 and the relation given in eq . ( [ eq2.21 ] ) . the matrix element @xmath101 is given by @xmath102 } { \overline{d}\left ( \omega \right ) } , \label{eq2.22}\ ] ] where @xmath103 and @xmath104 are elements of the energy matrix @xmath105 , defined in eq . ( [ eq2.16 ] ) . in eq . ( [ eq2.22 ] ) , it is also necessary to introduce the following definitions : @xmath106 with @xmath107 @xmath108 and @xmath109 the quantities @xmath110 and @xmath111 are defined as @xmath112 and , @xmath113 in the particular case , when @xmath114 , @xmath115 and @xmath69 are zero , @xmath110 and @xmath111 represent the quasi - particle bands in the paramagnetic normal state of the one - band hubbard model . the term @xmath116 is given by : @xmath117 where @xmath118 is the fourier transform of @xmath77 given in eq . ( [ eq2.12 ] ) . the denominator of the green s function @xmath119 given in eq . ( [ eq2.22 ] ) is defined as : @xmath120\nonumber \\ \label{eq2.30}\end{aligned}\ ] ] where @xmath121 with @xmath122\nonumber\\ & & \times[(\omega+e_{11})(\omega n_{-\sigma}^d+e_{22})-(\omega n_{-\sigma}^d+e_{12})^2].\nonumber\\ \label{eq2.32}\end{aligned}\ ] ] in eq . ( [ eq2.32 ] ) , @xmath123 and @xmath124 are elements of the energy matrix @xmath125 given in eq . ( [ eq2.10 ] ) . the use of a set of five operators @xmath126 results in a five - pole approximation to the green s functions . then , the @xmath127 defined in eq . ( [ eq2.30 ] ) may be also written as : @xmath128 where the quasi - particle bands @xmath129 ( with @xmath130 ) satisfy @xmath131 . therefore , the resulting green s function can be written as a sum of five terms : @xmath132 where @xmath133 express the spectral weights which satisfy @xmath134 in the case of @xmath3-wave symmetry , the traditional correlation function @xmath135 is always zero . therefore , this correlation function can not be used to determinate the pairing gap in the @xmath3-wave channel @xcite . in the factorization procedure proposed by beenen and edwards in ref . @xcite , the correlation function given by eq . ( [ eq2.18 ] ) is rewritten as @xmath136\langle d_{i-\sigma}d_{l\sigma}\rangle \label{eq3.0}\ ] ] where the symmetry @xmath137 is conserved , and the products @xmath138 and @xmath139 are split up . it is also introduced @xmath140 which allows to rewrite eq . ( [ eq3.0 ] ) as @xmath141 where @xmath142 can be calculated from @xmath143 . considering the _ d_-wave symmetry , the fourier transform of @xmath144 given by eq . ( [ eq2.17 ] ) becomes @xmath145 \label{eq3.3}\ ] ] where @xmath146 is the gap - function amplitude . due to the @xmath3-wave symmetry , @xmath147 for @xmath148 in the @xmath149 direction and @xmath150 when @xmath148 is in the @xmath151 direction . the fourier transform of the correlation function @xmath152 is given by @xmath153 where @xmath154 is the number of sites in the system . the correlation function @xmath155 can be evaluated using the green s function @xmath156 and the relation given by eq . ( [ eq2.21 ] ) . the green s function @xmath156 can be rewritten as : @xmath157 where @xmath158 and @xmath127 is defined in eq . ( [ eq2.30 ] ) . combining the equation ( [ eq2.17 ] ) with the eqs . ( [ eq3.2 ] ) to ( [ eq3.6 ] ) , the gap equation can be written as : @xmath159 where @xmath160 with @xmath161 ^ 2 f_{q\sigma } ^{13}(\omega ) \label{eq3.9}.\ ] ] using the definition ( [ eq2.13 ] ) in eq . ( [ eq2.12 ] ) , the band shift @xmath162 can be written as : @xmath163 where @xmath164 + \lambda _ { ij\sigma } ^d}{n_{-\sigma } ^{d } ( 1-n_{-\sigma } ^{d } ) } \label{eq4.1}\ ] ] and @xmath165 the quantity @xmath166 is given by eq . ( [ eq2.14 ] ) . the correlation function @xmath167 present in @xmath166 can be obtained from de green s function @xmath168v_k^{pd } } { \overline{d}\left ( \omega \right ) } \label{eq4.3}.\ ] ] the remaining correlation function @xmath169 present in @xmath166 is calculated from the green s function @xmath170v_k^{dp } } { \overline{d}\left ( \omega \right ) } \label{eq4.4},\ ] ] where @xmath171 with @xmath172 defined in eq . ( [ eq2.23 ] ) . the quantity @xmath173 , in terms of the elements of the energy matrix ( [ eq2.16 ] ) , is given by : @xmath174 the green s function @xmath175 tends to zero as @xmath20 , consequently , the correlation function @xmath176 also vanishes recovering the result of ref . @xcite for @xmath177 . the quantity @xmath178 present in eq . ( [ eq4.1 ] ) is given by eq . ( [ eq2.15 ] ) . the fourier transform of @xmath179 is given by : @xmath180 substituting eq . ( [ eq2.15 ] ) into eq . ( [ eq4.1 ] ) and then putting the result into eq . ( [ eq4.7 ] ) , the fourier transform of @xmath181 can be written as : @xmath182 the correlation functions present in @xmath183 are evaluated following the original roth s procedure @xcite . introducing extra operators @xmath184 , the correlation functions of the form @xmath185 can be calculated by using eqs . ( [ eq2.6 ] ) and ( [ eq2.21 ] ) . in refs . @xcite , the sum present in eq . ( [ eq2.6 ] ) has been considered only over the operators which describe the normal state of the system . in the present work , the sum includes also the hole operators which describe the superconducting properties of the system . thus , @xmath183 is given by : @xmath186 where the term @xmath187 is directly related to the gap function @xmath188 through the green s functions @xmath189 and @xmath190 ( see appendix a ) . the quantities @xmath191 , @xmath192 and @xmath187 are given in appendix a. in this section , the numerical results obtained in this work are presented . one of the most important parameters of the model given in eq . ( [ eq2.0 ] ) is the @xmath0 hybridization @xcite , which is defined as @xmath193 \label{eq5.0}.\ ] ] in this work , as in ref . @xcite , the hybridization has been assumed @xmath194-independent @xmath195 , where @xmath196 is the average over the brillouin zone . in ref . @xcite sengupta and ghatak have also used a @xmath194-independent hybridization due to the fact that the pairs occur within a small energy interval around the fermi level , therefore the dispersion of the hybridization can be neglected . the total occupation number is given by @xmath197 , where @xmath198 is obtained combining @xmath143 ( eq . ( [ eq2.22 ] ) ) and the relation given in eq . ( [ eq2.21 ] ) . the charge transfer energy @xmath199 is positive . this means that the first hole added to the system will energetically prefer to occupy the @xmath3-orbital of the copper ions @xcite . all results presented in this section are obtained with @xmath200 and @xmath201 . consequently , @xmath202 , as estimated in ref . @xcite . llllll & @xmath15 & @xmath188 & @xmath203 & @xmath204 & @xmath187 + beenen and + edwards @xcite & @xmath205 & 0 & 0 & 0 & 0 + + ref.@xcite & @xmath206 & 0 & 0 & @xmath205 & 0 + + present work & @xmath205 & @xmath205 & @xmath205 & @xmath205 & @xmath205 + as discussed in refs . @xcite , the band shift @xmath207 ( see eq . ( [ eq4.0 ] ) ) can be evaluated considering different approximations . in the limit @xmath20 , some terms of the band shift vanish ( see ref . @xcite ) . in ref . @xcite , the present authors estimate @xmath207 in the limit of @xmath20 , but with finite @xmath15 in other parts of the problem . in ref . @xcite , beenen and edwards evaluated @xmath207 in the normal state ( where @xmath208 ) and considering @xmath203 equal to zero and finite @xmath15 using the one - band hubbard model ( hybridization null ) . in the present work , the correlation functions present in @xmath207 given in eq . ( [ eq4.8 ] ) are evaluated following closely the procedure used by roth in ref . @xcite . nevertheless , here , also the hole operators given in the set of eq . ( [ eq2.9 ] ) are used to evaluate the correlation functions . as consequence , a new term ( @xmath187 ) appears in @xmath207 ( [ eq4.9 ] ) ) . the approximations used to evaluate @xmath207 are shown in table [ tab:1 ] . in figure [ fig1](a ) , the quasi - particle bands @xmath129 , with @xmath209 ( see eq . ( [ eq2.33 ] ) ) , are plotted along the symmetry lines @xmath210 , in the two - dimensional brillouin zone . the quasi - particle energies @xmath129 , in the superconducting state , are relative to the chemical potential @xmath22 . the circles bf show the @xmath211 band , where the center of @xmath211 is shifted by @xmath212 relative to the zero of energy . all results shown in this paper are obtained with @xmath213 . the dashed line corresponds to the noninteracting @xmath214 band @xmath215 relative to the noninteracting chemical potential . the figure [ fig1](b ) shows the superconducting gap between the electron and hole bands in the neighborhood of the @xmath5 point , while on the @xmath216 diagonal ( fig . [ fig1](a ) ) the gap is zero . this fact reflects the @xmath3-wave symmetry proposed in this work . the dashed lines show the absence of the gap in the normal state . in figure [ fig1](c ) , the region near to the @xmath217 point shows the gaps produced by the @xmath0 hybridization @xmath204 . the dashed lines show the result for @xmath218 . in figure [ fig1.1 ] , the electron and hole quasi - particle bands are shown for two different values of hybridization . as can be observed , the hybridization shifts the quasi - particle bands to lower energy by breaking the symmetry in relation to the @xmath194 axis . the figure [ fig2 ] shows the spectral weights @xmath133 for two different hybridizations . the dashed line corresponds to the sum of the five spectral weights which is equal to one ( see eq . ( [ eq10.4 ] ) ) . in figure [ fig2](b ) , the effects of the hybridization on the spectral weights are shown . such effects cause a small change in the chemical potential and consequently in the superconductivity . the figure [ fig3 ] shows the behavior of gap function amplitude @xmath219 as a function of the hybridization @xmath204 . it is clear that there is a decreasing of @xmath219 with increasing @xmath204 . the analysis of the function @xmath220 introduced in eq . ( [ eq3.81 ] ) and defined in eq . ( [ eq3.9 ] ) is important to understand the behavior of the gap function amplitude showed in figure [ fig3 ] . the figure [ fig4 ] shows the function @xmath220 for @xmath221 and two different values of hybridization . as can be seen in the dashed line , the magnitude of the function @xmath220 decreases when the hybridization is enhanced . moreover , the function is shifted to lower energy , breaking the symmetry respect to @xmath222 . the symmetry break has been also observed in figure [ fig1.1 ] for the electron and hole quasi - particle bands . for @xmath221 , the product @xmath223 given in eq . ( [ eq3.81 ] ) vanishes when @xmath224 . that is because the fermi function @xmath100 is zero for that range of @xmath225 . as consequence of the shift and the suppression of @xmath220 , the value of @xmath226 , which is given by the integral in eq . ( [ eq3.81 ] ) , decreases when the hybridization increase . however , from eq . ( [ eq3.8 ] ) , it is necessary a minimum value for @xmath226 to obtain a nonzero solution for @xmath227 . but , for very strong values of hybridization , the minimum value for @xmath226 is not reached and only the zero solution exists . according to this analysis , there is a critical value of hybridization ( @xmath228 ) , above which , the superconductivity is suppressed . similar results , which show a critical value for the hybridization , were also obtained in ref . @xcite , for a @xmath194-dependent hybridization and using the hartree - fock approximation for the electron - electron interaction . in refs . @xcite , although the high @xmath229 was not considered , the hybridization effects play an important role for resonant states . the discussion above is also valid if the values of the temperature @xmath203 are raised with @xmath230 constant . the only difference is that in this case the fermi function becomes sloping smoothly , changing the product @xmath223 . the effect of the temperature in the fermi function causes a decreasing of @xmath226 and consequently of @xmath229 . since the hybridization is directly related to the applied pressure @xcite , the transition temperature @xmath229 may have a dependence on pressure through the hybridization however , the pressure dependence of @xmath229 is very complicated in high temperature superconductors . as discussed in ref . @xcite , at least , in conventional superconductivity where the electron pairing is mediated by phonons , two effects are responsible for the pressure dependence of @xmath229 . the first one is related to the lattice vibrations , while the second one comes from the electronic contribution . as long the pressure is increased , the lattice vibrations tend to increase @xmath229 , whereas the effects of the electronic contribution associated with the hybridization cause a decreases of @xmath229 . the figure [ fig5 ] shows the function amplitude @xmath219 , as a function of temperature @xmath203 for several values of hybridization . if the hybridization is enhanced , the critical temperature @xmath229 decreases . therefore , the present results agree with the discussion above in the scenario where the electronic effects dominate . this behavior for @xmath229 is also shown in figure [ fig6 ] , in phase diagrams displaying @xmath229 versus the total occupation number @xmath231 . the numerical results obtained show that there is a critical value of hybridization where the superconducting phase vanishes . these results agree with the ones obtained by the authors of ref . @xcite for heavy - fermion superconductivity with an @xmath232-boson treatment . the solid lines in figures [ fig6](a)-(b ) show the present result for @xmath229 , where the effects of the temperature , the superconductivity and the coulomb interaction have been included in the calculation of the band shift @xmath207 . the dashed lines correspond to the results obtained in ref . @xcite , where the band shift has been evaluated considering @xmath221 , @xmath208 and @xmath20 ( see table [ tab:1 ] ) . the difference between the results can be explained by the analysis of the eq . ( [ eq2.14 ] ) where some of the correlation functions present in the band shift vanish in the @xmath17 limit . in eq . ( [ eq2.14 ] ) , the correlation function @xmath233 , which is directly related to the hybridization effects in the band shift , vanishes for @xmath17 . it is important to highlight that the correlation functions in eq . ( [ eq2.14 ] ) are both negative . therefore , for large @xmath15 , the correlation function @xmath234 decreases and the hybridized shift @xmath235 is enhanced . however , for intermediate values of @xmath15 , both correlation functions remain finite . as consequence , the hybridization effects in the band shift and , therefore , in the superconductivity , are weakened . the figures [ fig6](c)-(d ) show the present results when the value of @xmath15 is increased . the main consequence is , within the factorization procedure , to shift the window of doping where superconductivity is found , as in ref @xcite . in figure [ fig7](a ) , the chemical potential is show as a function of the total occupation number @xmath231 for @xmath236 and two different hybridizations . in ref . @xcite , the authors criticized the roth s method because the compressibility @xmath237 is negative in the vicinity of half - filling in the beenen and edwards result . in ref . @xcite , by using a composite operator approach and imposing the pauli principle , the authors have showed that the compressibility remains negative . however , they also showed that the pairing decreases the strength of the negative compressibility . in the present work , a careful study about the nature of the negative compressibility and the effect of the hybridization near half - filling in roth s approximation has been carried out . it has been verified that the most important contribution to provide negative compressibility comes from the spin - term @xmath238 present in the @xmath3-part @xmath239 of the roth s band shift @xmath207 ( see eqs . [ eq4.8 ] and [ apx26 ] ) . in reference @xcite , it has been showed that the correlation function @xmath238 plays an important role on the flattening of the quasi - particle bands . the correlation function @xmath238 increases with occupation and its effect is pronounced near half - filling . nevertheless , when the hybridization is present , the numerical results show that it acts in the sense of suppressing the negative compressibility near half - filling . because the hybridization considered here is @xmath194-independent @xcite , the hybridization term @xmath235 of the band shift is constant within the brillouin zone . its main effect is to shift the poles of the green s functions and consequently to change the value of the chemical potential suppressing the negative compressibility . in figure [ fig7](a ) , it is clear that the effect of the hybridization in the chemical potential decreases the negative compressibility . the figure [ fig7](b ) shows the gap function amplitude @xmath219 as a function of the total occupation number . this result agrees with those obtained in figure [ fig3 ] , where @xmath219 decreases with increasing of @xmath230 . in this work , the roth s two - pole approximation is extended to study the superconducting properties of the extended hubbard model given in eq . ( [ eq2.0 ] ) . the quality of the roth s two - pole approximation had been investigated in a previous work by beenen and edwards @xcite . in their work , they showed the remarkable agreement between the roth s and the monte carlo results @xcite for the one - band hubbard model in the paramagnetic normal state . moreover , the flat bands obtained with roth s procedure show a qualitative agreement with the arpes experiment data @xcite in cuprates . it is important to point out that the flattening observed in the quasi - particle bands which produces a peak in the density of states , can be connected with the van hove scenario . in cuprate systems the van hove singularity is present in the vicinity of the fermi energy . therefore , it is believed that the van hove scenario play a fundamental role in order to clarify the mechanism which drives the transition to superconductivity in these interesting materials @xcite . the accuracy of the roth s results is very related to the adequate evaluation of the band shift . therefore , the focus of the present work has been to evaluate the roth s band shift taking into account relevant effects as coulomb interaction , temperature , superconductivity and hybridization . also , the effect of the hybridization in the superconducting of the model has been studied . this work has been carried out following the factorization procedure proposed by beenen and edwards @xcite . in order to study superconductivity , beenen and edwards proposed to include hole operators in the original set of operators that describes the normal state of the system . these operators can introduce the pairing formation in the @xmath3-band . the factorization procedure proposed by beenen and edwards @xcite and the @xmath3-wave symmetry are considered to obtain the gap function amplitude . the hybridization effects are considered by also including a @xmath6-operator . thus , the set of operators is enlarged to five , which results in a five - pole approximation to the green s functions . the hybridization effects present in the band shift come from some correlation functions . the important point is that part of them vanish when @xmath17 , as it have been done in ref . @xcite . in order to consider properly the hybridization effects , the band shift should be obtained for finite @xmath15 . in fact , the obtained phase diagrams show that the presence of superconducting order exists in a larger range of doping when compared with the @xmath17 limit @xcite , for the same hybridization . therefore , this result suggests that , in the @xmath17 limit , the hybridization effects are overestimated . that is the ultimate justification for the use of the factorization procedure @xcite , which is valid for intermediated values of @xmath15 for the gap function . the beenen and edwards s @xcite results are recovered taking @xmath218 in the present work . the hybridization @xmath204 breaks the symmetry between the electron and hole quasi - particle bands , respect to @xmath194 axis . also , the gap amplitude function @xmath219 and the critical temperature @xmath229 are suppressed with increasing the hybridization @xmath204 . the results show that the chemical potential does not change significantly away the half - filling . however , near half - filling , it is showed that the negative compressibility decreases with increasing @xmath204 . the correlation functions present in the @xmath3-part of the band shift @xmath207 were discussed in detail . when the hole operators are also considered to obtain this correlation functions , a new term appears in the @xmath3-part of the band shift @xmath207 . the new term is directly associated with the superconducting properties of the system . nevertheless , this term is quite small and therefore may be disregarding in the calculation of the band shift . the authors are grateful to the grupo de fsica estatstica - ifm , universidade federal de pelotas where part of numerical calculations were performed . this work was partially supported by the brazilian agencies capes ( coordenao de aperfeioamento de pessoal de nvel superior ) and fapergs ( fundao de amparo pesquisa do rio grande do sul ) . the correlation functions present in the band shift @xmath239 can be evaluated by introducing extra @xmath240 operators , as in the original roth s procedure . combining the eq . ( [ eq2.6 ] ) and the relation given in eq . ( [ eq2.21 ] ) , it is possible to write @xmath241_{(+)}\rangle \label{apx1},\ ] ] where @xmath126 and @xmath242 are members of the set of operators given in eq . ( [ eq2.9 ] ) . for evaluate @xmath243 , it has been necessary to introduce the following @xmath240 operators : @xmath244 and @xmath245 by considering the operator given in the eq . ( [ apx2 ] ) , the correlation function @xmath246 can be written as : @xmath247 where the right side of eq . ( [ apx4 ] ) may be obtained using the relation given by eq . ( [ apx1 ] ) . therefore , it is necessary to evaluate the anticommutators @xmath248_{(+)}$ ] for the set of operators @xmath249 given in eq . ( [ eq2.9 ] ) . for @xmath250 , the @xmath251 operators are given by : @xmath252 @xmath253 @xmath254 @xmath255 and @xmath256 thus , the following results have been obtained @xmath257_{(+)}\rangle = n_{0\sigma}^{d } -e^{i\vec{k}\cdot \vec{r}_j}\langle d_{0\sigma}^{\dagger}d_{j\sigma}\rangle \label{apx9.1},\ ] ] @xmath258_{(+ ) } \rangle=\langle n_{0-\sigma } ^{d}n_{j\sigma}^{d}\rangle -e^{i\vec{k}\cdot \vec{r}_j}\langle d_{0\sigma}^{\dagger}n_{j-\sigma}^{d}d_{j\sigma}\rangle \label{apx10},\ ] ] @xmath259_{(+)}\rangle=0 \label{apx11},\ ] ] @xmath260_{(+)}\rangle= -\langle n_{j\sigma } ^{d}d_{0\sigma}^{\dagger } d_{0-\sigma}^{\dagger}\rangle \label{apx12},\ ] ] @xmath261_{(+)}\rangle=0 \label{apx13}\ ] ] where , it has been assumed that the brackets are real and unchanged when the indices 0 and @xmath262 are interchanged . also , due to translational invariance of the system , @xmath263 . considering the relations given by eqs . ( [ apx1 ] ) and ( [ apx4 ] ) with the results from eq . ( [ apx9.1 ] ) to eq . ( [ apx13 ] ) , the correlation function @xmath246 can be written as : @xmath264 where @xmath265 . in eq . ( [ apx14 ] ) , it has been introduced the following definitions : @xmath266 @xmath267 @xmath268 @xmath269 and @xmath270 where @xmath119 is given in eq . ( [ eq2.22 ] ) . the remaining green s functions @xmath271 and @xmath190 are given respectively by @xmath272 } { \overline{d}\left ( \omega \right ) } \label{apx19.1}\ ] ] and @xmath273 where @xmath274 is defined in eq . ( [ eq4.5 ] ) and @xmath275 in eq . ( [ eq2.33 ] ) . it is also necessary to define @xmath276 @xmath277 and @xmath278 where @xmath189 is given in eq . ( [ eq3.6 ] ) . the correlation function @xmath279 present in eq . ( [ apx14 ] ) , can be obtained by repeating the procedure above using the operator @xmath280 ( given by eq . ( [ apx3 ] ) ) . thus , @xmath281 where @xmath282 @xmath283 and @xmath284 with @xmath285 reversing the spin labels i.e. , @xmath286 in eq . ( [ apx19.6 ] ) and substituting the result into eq . ( [ apx14 ] ) , then @xmath287 \label{apx23}.\end{aligned}\ ] ] for evaluate the two last correlation functions present in eq . ( [ eq4.8 ] ) , the following operators have been introduced @xmath288 and @xmath289 using @xmath290 , and following the procedure outlined above , the correlation function @xmath291 is given by @xmath292 . \nonumber \\ \label{apx26}\end{aligned}\ ] ] similarly , using @xmath293 @xmath294 where the @xmath3-wave symmetry has been considered , therefore , @xmath295 . the four @xmath296 operators introduced up to now are exactly the same operators used by roth in ref . @xcite to obtain the band shift @xmath297 in the normal state and without hybridization . however , in the present work , due to the presence of the hole operators ( see eq . ( [ eq2.9 ] ) ) , a new @xmath240 operator , which is given by @xmath298 has been introduced . with this operator , the correlation function @xmath299 present in eq . ( [ apx27 ] ) may be evaluated . thus , the result given in eq . ( [ apx29 ] ) can be used in eq . ( [ apx23 ] ) to obtain @xmath303 \label{apx31}.\end{aligned}\ ] ] finally , with the results ( [ apx29 ] ) , ( [ apx30 ] ) and ( [ apx31 ] ) into eq . ( [ eq4.9 ] ) , the following result has been obtained @xmath304 where @xmath305 @xmath306 @xmath307}{1-\beta_{\sigma}\beta_{-\sigma}}\nonumber\\ & & -\frac{\alpha_{j\sigma}^{(1)}n_{0j-\sigma}^{(1 ) } + \beta_{j\sigma}^{(1 ) } m_{j-\sigma}^{(1)}}{1+\beta_{\sigma } } \label{apx35}\end{aligned}\ ] ] with @xmath308

in 1969 , garside solved the conjugacy problem in the braid group by investigating the monoid of positive braids @xcite . at the end of the 90 s , dehornoy and paris introduced in @xcite the notion of a garside group , that captures most of the properties used in garside approach . after this seminal paper , the notion has been extended and numerous articles have dealt with garside monoid / group structures ( @xcite,@xcite,@xcite,@xcite,@xcite for instance ) . actually various groups can be equipped with a garside group structure . recently , krammer @xcite , bessis @xcite and digne - michel @xcite have extended the notion of a garside group into the notion of a garside groupoid , which turned out to be crucial in the proof of the long - standing question of the @xmath0 property for complex reflection arrangements @xcite . the standard examples of garside groups are braid groups and their generalization , the so - called artin - tits groups of spherical type . one of the main properties of artin - tits groups is the existence of a family of distinguished subgroups , namely the _ standard parabolic subgroups _ , which are the subgroups generated by a subset of the standard generating set . in order to describe normalizers of such subgroups , we were naturally led in @xcite ( see also @xcite ) to introduce a groupoid called the _ ribbon groupoid_. in an unpublished paper @xcite digne and michel prove that the ribbon groupoid of an artin - tits group of spherical type is a garside groupoid . actually , this result implicit in @xcite , even though the notion of a garside groupoid had not been introduced yet . in @xcite , we extended the notion of a parabolic subgroup into the framework of garside groups . our objective here is to go further and extend the notion of ribbon groupoid . we prove that in the context of garside groups , ribbon groupoids remain a garside groupoids . indeed we prove ( see next sections for definition ) that [ thintror(a)]for every garside group @xmath1 with a @xmath2-structure , the positive ribbon category @xmath3 is a garside category and the ribbon groupoid @xmath4 is the associated garside groupoid . every artin - tits group of spherical type has a @xmath2-structure . then , theorem [ thintror(a ) ] applied in this special case . in order to prove theorem [ thintror(a ) ] , we consider the category @xmath5 of subcategories of a category @xmath6 , whose morphisms are natural transformations . it turns out that for a garside category @xmath6 , the category @xmath5 is not a garside category in general . however , this category shares nice properties with garside monoids ( see section [ section2section ] for precise definition and results ) . we also provide a groupoid presentation of the ribbon groupoid @xmath4 for every garside group @xmath1 with a @xmath2-structure . to obtain this presentation we investigate the quasi - centralizer @xmath7 of @xmath1 , that is the subgroup of @xmath1 made of those elements that permute the atoms of the monoid @xmath8 . in particular we generalize a result of picantin @xcite : [ thintror(a)2 ] for every garside group @xmath1 , the group @xmath7 is a finitely generated free commutative group . furthermore , we describe the normalizer of the standard parabolic subgroups by proving ( see theorem [ theodescnormal ] ) : [ thintroprodsemdir]let @xmath1 be a garside group . let @xmath9 be the groupoid whose objects are the parabolic subgroups @xmath10 of @xmath1 , whose hom - set from @xmath10 to @xmath11 is @xmath12 , and whose composition is the product in @xmath1 . let @xmath13 be the totally disconnected groupoid whose objects are the parabolic subgroups of @xmath1 and whose vertex group at @xmath10 is @xmath10 . then , there exists a standard parabolic subgroupoid @xmath14 of @xmath4 such that @xmath9 is equal to the semi - direct product of groupoids @xmath15 . this theorem generalizes ( * ? ? ? finally , the subgroups of the vertex groups of the ribbon groupoid @xmath4 that are generated by the atoms contained in the vertex group are garside groups . this result is similar to ( * ? ? ? * theorem b ) , which deals with coxeter groups . the reader may wonder what is the interest of proving that ribbon groupoids are garside groupoids . the main point is that it provides a complete description of the connection between standard parabolic subgroups . the notion of a parabolic subgroup is crucial to build important tools like the so - called _ deligne complexes _ of artin - tits groups . therefore to understand their normalizers and their mutually conjugating elements is certainly crucial . furthermore the fact that ribbon groupoids are garside groupoids shows that , even in order to study classical groups such as braid groups , the notion of a garside groupoid is natural and useful . finally , our groupoid presentations of ribbon groupoids of spherical type artin - tits groups provide garside groupoids with _ braid - like _ defining relations . these groupoids are then special among garside groupoids and therefore should provide seminal examples of garside groupoids that are not groups . the paper is organized as follows . in section 2 , we introduce the notions of a garside groupoid and of a parabolic subgroupoid ; we investigate the category @xmath5 . in section 3 , we introduce the positive ribbon category associated with a garside category and prove theorem [ thintror(a ) ] . finally , in section 4 we provide a groupoid presentation of the ribbon groupoid ; along the way we prove theorem [ thintror(a)2 ] . we conclude with the proof of theorem [ thintroprodsemdir ] . here we set the necessary terminology on _ small categories_. in particular we recall the notion of a _ natural transformation _ between two _ functors_. we introduce the central notion of a _ garside groupoid _ and provide examples . the present paper deals with categories , but the reader should not be scared since we only consider _ small categories_. let us introduce the notion of a small category and some basic notions in a naive way . [ definsmcat ] a _ quiver _ @xmath16 is a pair @xmath17 of sets equipped with two maps @xmath18 and @xmath19 such that @xmath20 for every element @xmath21 of @xmath22 . the elements of @xmath22 are called the _ vertices _ and the elements of @xmath23 are called the edges . an edge @xmath24 is said to be _ oriented _ from its _ source _ @xmath25 to its _ target _ @xmath26 . the _ path semigroup @xmath27 associated with the quiver _ @xmath16 is the semigroup with @xmath28 defined by the presentation whose generating set is @xmath29 and whose defining relations are @xmath30 for every @xmath24 in @xmath29 and @xmath31 for every two _ distinct _ elements of @xmath22 . clearly , we can identify the elements of @xmath32 with their images in @xmath27 . note that a product @xmath33 of edges is not equal to @xmath28 in @xmath27 if and only if for every index @xmath34 , one has @xmath35 . if @xmath24 lies in @xmath27 and is not @xmath28 , then one can define its source and its target to be @xmath36 and @xmath37 for any of its decomposition @xmath33 as a product of elements in @xmath29 . the reader may note that the semigroup algebra of @xmath27 is equal to the path algebra of the quiver @xmath16 as defined in @xcite . an element of @xmath27 distinct from @xmath28 is called a _ path _ in the sequel . now , let @xmath16 be a quiver and let @xmath38 be a congruence on @xmath27 such that @xmath28 is alone in its equivalence class @xmath39 and such that two distinct @xmath38-equivalent paths in @xmath27 have the same source and the same target . we denote by @xmath40 the quotient semigroup of @xmath27 by the congruence @xmath38 . by assumption on @xmath38 , to each equivalence classes of paths @xmath41 one can associate a source and a target by setting @xmath42 and @xmath43 , respectively . the ( _ small _ ) _ category @xmath44 associated with the quiver @xmath16 and the congruence @xmath38 _ is the set @xmath45 of non - zero equivalence classes of paths equipped with the source and target maps , and the partial product induced by the product in @xmath40 . in that context , the vertices of @xmath16 are called the objects of the category , and the elements of the category are called the _ morphisms _ of the category . the _ free category _ @xmath46 on @xmath16 is obtained by considering the trivial congruence . in the sequel , we denote by @xmath47 the set of objects of a category @xmath6 . for every @xmath48 in @xmath47 , we denote by @xmath49 the equivalence class of the trivial path @xmath48 . for @xmath48 and @xmath50 in @xmath47 , we denote by @xmath51 the set of morphisms from @xmath48 to @xmath50 . in order to be consistent with @xcite and with the above naive definition of a small category , if @xmath21 lies in @xmath52 and @xmath53 lies in @xmath54 , then we denote by @xmath55 the morphism of @xmath56 obtained by composition . we denote by @xmath57 and by @xmath58 the set of morphisms of the category @xmath6 whose source and target , respectively , are @xmath48 . the seminal example to keep in mind is the case of a category with one object @xmath48 : this is just a monoid whose unity is @xmath49 . here we introduce the notion of a _ natural transformation _ which is crucial in our article . a map @xmath59 from a category @xmath60 to a category @xmath61 is a _ functor _ if , firstly , there exists a map @xmath62 such that @xmath63 for every vertex @xmath48 in @xmath47 and , secondly , one extends @xmath64 to a morphism of semigroups from @xmath40 to @xmath65 by setting @xmath66 . note that the second property is equivalent to say that @xmath67 is included in @xmath68 for every two objects @xmath69 of @xmath6 , and @xmath70 for every two morphisms @xmath53 , @xmath21 of @xmath6 such that @xmath71 . for instance , for every category @xmath44 , there is a surjective functor from the free category @xmath46 to @xmath44 . isomorphism of categories _ is a functor that is bijective . in that case , the inverse is also an isomorphism of categories and the set @xmath67 is equal to @xmath72 . an automorphism of @xmath6 is an isomorphism from @xmath6 to itself . let @xmath60 and @xmath61 be two categories with @xmath73 . let @xmath64 and @xmath74 be two functors from @xmath6 to @xmath75 . then , a _ natural transformation _ @xmath76 from @xmath64 to @xmath74 is a map @xmath77 that sends every vertex @xmath48 in @xmath22 to a morphism @xmath78 from @xmath79 to @xmath80 such that for every two objects @xmath69 in @xmath16 and every morphism @xmath21 in @xmath51 , one has @xmath81 ( see figure [ diagrammetransfnaturelle ] ) . ( 100,70 ) ( 18,0 ) ( 8,0)@xmath82(8,46)@xmath79(78,0)@xmath83 ( 78,46)@xmath80 ( 0,25)@xmath84(38,54)@xmath78(38,6)@xmath85 ( 88,25)@xmath86 for instance , if @xmath6 and @xmath87 are groups and @xmath64 , @xmath74 are two morphisms of groups from @xmath6 to @xmath87 , then a natural transformation from @xmath64 to @xmath74 is an element @xmath88 of the group @xmath87 such that @xmath89 for every @xmath90 in @xmath6 . to introduce the central notion of a garside groupoid , we need first to fix some definitions . the _ atoms _ of a category @xmath44 are the non - identity morphisms whose representing paths are edges only . the _ atom graph _ @xmath91 of the category is the graph that has the same set of vertices than @xmath16 and whose edges are the atoms of the category . we recall that a binary relation ( defined on a set ) is noetherian when every decreasing infinite sequence stabilizes . we say that a morphism @xmath53 of @xmath44 is a factor of the morphism @xmath92 when there exist two morphisms @xmath93 and @xmath94 such that @xmath95 . we say that the category @xmath44 is _ noetherian _ if the set of morphisms is noetherian for factors , in other words , if for every infinite sequence @xmath96 of morphisms such that @xmath97 is a factor of @xmath98 , there exists an index @xmath99 such that @xmath100 for every @xmath101 . in this case , the category is _ atomic _ : each morphism can be written as a finite product of atoms ; furthermore , the category @xmath44 is isomorphic to @xmath102 , where @xmath103 is the congruence on @xmath104 induced by the congruence @xmath38 . this property holds in particular when for every morphism there is a bound on the lengths of its representing paths in the quiver @xmath16 ( where each edge has length @xmath105 ) . if in the category @xmath44 there is no loop for the factor relation , then for every object @xmath48 the left - divisibility and the right - divisibility induce partial orders on the sets @xmath57 and @xmath58 , respectively . this is the case when the category is noetherian . we say that the category @xmath44 is cancellative when for every two objects @xmath48 , @xmath50 and every morphism @xmath21 in @xmath51 the maps @xmath106 and @xmath107 are injective on the sets @xmath108 and @xmath58 , respectively . finally , a _ groupoid _ is a category @xmath6 such that every morphism has an inverse , in other words , such that for every morphism @xmath21 in @xmath51 there exists a morphism @xmath109 in @xmath110 with @xmath111 and @xmath112 . to each category @xmath60 one can associate a groupoid @xmath113 and a functor @xmath114 in the following way . the quiver @xmath115 is obtained by adding to @xmath16 an edge @xmath109 from @xmath116 to @xmath117 for each edge @xmath21 of @xmath16 . the congruence @xmath118 is the congruence generated by @xmath38 and the relations @xmath111 and @xmath112 for each edge @xmath21 from @xmath48 to @xmath50 in @xmath16 . we call the groupoid @xmath119 the _ groupoid of formal inverses _ of @xmath6 . for instance , the _ free groupoid _ on @xmath16 is the groupoid @xmath120 . it should be noted that every functor @xmath121 between two categories induces a functor @xmath122 such that the diagram @xmath123 is commutative . we are now ready to introduce the definition of the special kind of groupoids we consider in this paper . we refer to @xcite for the general theory on lattices . in @xcite , bessis defines the notion of a garside category as follows . a _ categorical garside structure _ is a pair @xmath124 such that : * @xmath6 is a small category equipped with an automorphism @xmath64 such that @xmath76 is a natural transformation from the identity functor @xmath125 to @xmath64 ; * @xmath6 is noetherian and cancellative ; * for every object @xmath48 of @xmath6 , the sets @xmath126 and @xmath58 are lattices for the left - divisibility order and the right - divisibility order , respectively ; * for every two objects @xmath69 of @xmath6 and every atom @xmath21 in @xmath51 , there exists a morphism @xmath127 such that the following diagram is commutative : + ( 100,70 ) ( 8,0 ) ( 8,0)@xmath50(8,46)@xmath48(68,0)@xmath82 ( 68,46)@xmath79 ( 2,25)@xmath21(32,54)@xmath78(32,6)@xmath85 ( 78,25)@xmath84 ( 37,28)@xmath127 in this case , we say that the category @xmath6 is a _ garside category_. + we say that @xmath124 is a _ monoidal garside structure _ when furthermore the category @xmath6 has a unique object . in this case , we say that the monoid @xmath6 is a _ garside monoid_. the reader should note that in the above definition , the automorphism @xmath64 is uniquely defined by @xmath76 . a _ garside groupoid _ @xmath119 is the groupoid of formal inverses of a garside category @xmath6 . when furthermore @xmath6 is a monoid , then we say that @xmath119 is a garside group . if @xmath6 and @xmath87 are two categories and @xmath64 is a functor , then the subset @xmath128 of @xmath6 has a natural structure of category . we say that such a subset is a _ subcategory _ of @xmath6 . in the sequel , we denote by @xmath129 the _ canonical functor _ from the subcategory @xmath130 to the category . the reader should note that , by hypothesis , a garside category @xmath6 verifies the so - called _ ore relations _ and , therefore , the functor @xmath114 is injective . in the sequel we consider @xmath6 as a subcategory of @xmath119 . clearly , the automorphism @xmath64 extends to an automorphism of @xmath119 , still denoted by @xmath64 , and that @xmath76 extends to a natural transformation from the identity functor @xmath131 to @xmath64 . [ exagpat]the seminal case of a garside groupoid @xmath119 occurs when the category has one object @xmath48 . in this case , the category @xmath6 is a garside monoid and @xmath119 is its group of fractions , that is a garside group ( @xcite ) . the element @xmath78 is called the _ garside element _ of the group @xmath119 , or of the monoid @xmath6 . in such a garside group , a natural transformation from the identity morphism to an automorphism corresponds to an element that realizes this automorphism as an inner automorphism . _ braid groups _ and , more generally , _ artin - tits groups of spherical type _ are garside groups which are defined by a group presentation of the type @xmath132 where @xmath133 are positive integers greater than 2 with @xmath134 . in this case , the associated _ coxeter group _ , which obtained by adding the relations @xmath135 for @xmath136 in @xmath137 to the presentation , is finite . if we consider @xmath138 with @xmath139 for @xmath140 and @xmath141 otherwise , we obtain the classical presentation of the braid group @xmath142 on @xmath143 strings . in that case , the associated coxeter group is the permutation group on @xmath143 elements . if @xmath1 is such an artin - tits group , and @xmath144 is its coxeter group , then the canonical surjection @xmath145 has a canonical section ( * ? ? ? * chap . 4 ) . if we denote by @xmath76 the image by that section of the greatest element of @xmath144 , then @xmath76 belongs to the submonoid @xmath8 of @xmath1 generated by @xmath137 , and the associated inner group automorphism induces an automorphism of @xmath8 . it turns out that the pair @xmath146 is a monoidal garside structure with @xmath1 for associated garside group . a group can be a garside group for several monoidal garside structures . for instance , the braid group @xmath142 has the alternative group presentation : @xmath147 where @xmath148 for the generators @xmath149 , and if @xmath150 denotes the submonoid of @xmath142 generated by the @xmath149 , and @xmath151 , then the inner group automorphism of @xmath142 induced by @xmath88 restricts to an automorphism of @xmath150 . again , the pair @xmath152 is a monoidal garside structure with @xmath142 for associated garside group . this alternative garside structure is called the _ dual garside structure _ of the braid group . this point of view was generalized by bessis in @xcite to every artin - tits groups of spherical type . @xcite the free group @xmath153 on two letters @xmath154 has a presentation @xmath155 where @xmath156 and @xmath157 . let @xmath158 . if we denote by @xmath159 the submonoid of @xmath153 generated by the elements @xmath160 , and by @xmath64 the automorphism of @xmath153 that sends @xmath160 onto @xmath161 , then @xmath162 is a monoidal garside structure with group @xmath153 as the associated garside group . the automorphism induced by @xmath76 is @xmath64 .[exef2 ] we end this subsection with a definition and two properties of garside groupoids that we need in the sequel . they are immediate generalizations of well - known properties of artin - tits groups . 2.2 ) [ defisimplelem ] let @xmath124 be a categorical garside structure and denote by @xmath64 the automorphism of @xmath6 induced by @xmath76 . a morphism @xmath21 in @xmath51 is _ simple _ if there exists @xmath163 in @xmath164 such that @xmath165 . note that in the above definition the morphism @xmath163 is also simple . by definition of a garside category , the atoms are simple . [ uniquedecomposprop ] let @xmath124 be a categorical garside structure . + ( i ) @xcite every morphism @xmath21 that belongs to @xmath51 can be decomposed as a product @xmath166 of simple elements . + ( ii ) let @xmath21 lie in @xmath167 . there exist an object @xmath168 , a morphism @xmath169 in @xmath170 and a morphism @xmath171 in @xmath54 such that @xmath169 and @xmath171 are prime to each other in @xmath172 and @xmath21 is equal to @xmath173 . the triple @xmath174 is unique . point @xmath175 of the above proposition does not appear in @xcite , but the proof is similar to the case of garside groups . we call the triple @xmath174 the _ right greedy normal form _ of @xmath21 there is a similar unique _ left greedy normal form _ @xmath176 . [ defcatsbcat]let @xmath6 be a ( small ) category . we define _ the category of subcategories _ @xmath5 of @xmath6 as follows : * the objects are the non - empty subcategories of @xmath6 ; * if @xmath177 and @xmath178 are two subcategories of @xmath6 , then @xmath88 lies in @xmath179 if there exists an isomorphism @xmath180 from @xmath177 to @xmath178 such that @xmath88 is a natural transformation from @xmath181 to @xmath182 . * if @xmath183 and @xmath184 belong to @xmath179 and @xmath185 , respectively , then the composition morphism @xmath186 is defined by @xmath187 , where @xmath188 is the isomorphism associated with @xmath183 . note that in the second item of definition [ defcatsbcat ] the isomorphism @xmath180 is uniquely defined by @xmath88 because @xmath6 is cancellative ; in the third item , @xmath186 is a natural transformation from @xmath181 to @xmath189 , where @xmath190 is such that @xmath184 belongs to @xmath191 . ( 150,70 ) ( 8,0 ) ( 8,0)@xmath50(8,46)@xmath48(68,0)@xmath192 ( 68,46)@xmath193 ( 142,0)@xmath194 ( 142,46)@xmath195 ( 2,25)@xmath21(32,54)@xmath196(32,6)@xmath197 ( 80,25)@xmath198 ( 98,54)@xmath199(98,6)@xmath200 ( 155,25)@xmath201 our objective below is to investigate the properties of @xmath5 and , in particular , to decide whether or not @xmath5 is a garside category . [ lempropcccat]let @xmath124 be a categorical garside structure and denote by @xmath64 the automorphism of @xmath6 induced by @xmath76 . + ( i ) there exist an automorphism @xmath202 of @xmath5 and a natural transformation @xmath203 from the identity functor of @xmath5 to @xmath202 that are induced by @xmath64 and @xmath76 . + ( ii ) the category @xmath5 is cancellative . + ( iii ) for every object @xmath177 of @xmath5 , the left - divisibility and the right - divisibility define partial orders on @xmath204 and @xmath205 , respectively . _ ( i ) _ let @xmath177 , @xmath178 be two objects of @xmath5 and assume @xmath88 lies in @xmath179 ; denote by @xmath206 the isomorphism defined by @xmath88 . then , @xmath207 is an isomorphism from @xmath208 to @xmath209 , and the map @xmath210 is a natural transformation from the canonical embedding @xmath211 to @xmath212 . therefore , we define an isomorphism @xmath202 of @xmath5 by setting @xmath213 and @xmath214 . now , for every object @xmath177 of @xmath5 we denote by @xmath215 the restriction of the map @xmath76 to the subcategory @xmath177 . then , @xmath215 lies in @xmath204 . indeed , the map @xmath215 is a natural transformation from @xmath129 to @xmath216 , where @xmath217 is the restriction of @xmath64 to @xmath177 . clearly , the map @xmath203 is a natural transformation from the identity functor of @xmath5 to @xmath202 . + _ ( ii ) _ the cancellativity of the category @xmath5 follows directly from the cancellativity of the category @xmath6 . + _ ( iii ) _ there is no loop for the factor relation in @xmath5 because there is no loop for the factor relation in @xmath6 . therefore , @xmath204 and @xmath205 are partially ordered sets for the left - divisibility and the right - divisibility , respectively . in the sequel , an object @xmath48 of a category @xmath6 such that @xmath218 is called a trivial object _ of the category . let @xmath124 be a categorical garside structure such that @xmath6 has no trivial object.[proppropribbcat ] the category @xmath5 is noetherian if and only if its object set is finite . assume @xmath6 has a finite number of objects . let @xmath219 be a sequence of morphisms of @xmath5 such that for every @xmath220 , the morphism @xmath221 is a factor of the morphism @xmath222 . we denote by @xmath223 the isomorphism defined by @xmath222 . let @xmath224 and @xmath225 be the sequences of morphisms of @xmath5 such that @xmath226 , and denote by @xmath227 and @xmath228 the isomorphisms associated with @xmath224 and @xmath225 , respectively . then , for every positive integer @xmath220 , every object @xmath48 of @xmath229 and every morphism @xmath21 of @xmath229 , we have @xmath230 and @xmath231 . since @xmath6 is noetherian , for every object @xmath48 of @xmath232 , there exists some integer @xmath233 such that for every @xmath234 , we have @xmath235 . in particular , for every @xmath236 we have @xmath237 , @xmath238 , @xmath239 and @xmath240 . let @xmath241 . from the above discussion , it follows that for every @xmath236 one has @xmath242 , @xmath243 and @xmath244 . + assume finally that @xmath6 has an infinite number of objects ; denote by @xmath245 a sequence of distinct objects of @xmath6 such that @xmath246 is not empty for every @xmath34 . we fix a morphism @xmath247 in @xmath246 for each @xmath34 . let @xmath248 be the subcategory of @xmath6 whose set of objects is @xmath249 such that @xmath248 has trivial objects only . let @xmath250 be the subcategory whose object set is @xmath251 such that @xmath250 has trivial objects only . consider @xmath252 the map from the set of objects of @xmath248 to the morphisms @xmath5 that sends @xmath253 on @xmath254 if @xmath255 and @xmath256 on @xmath257 otherwise . let @xmath258 be the map from the objects of @xmath248 to the morphisms of @xmath5 that sends @xmath259 on @xmath247 and @xmath260 on @xmath261 otherwise . then , @xmath252 clearly belongs to @xmath262 and @xmath258 belongs to @xmath263 . by construction , @xmath264 for every @xmath34 . hense , the category @xmath5 is not noetherian . consider a categorical garside structure @xmath124 and denote by @xmath64 the automorphism of @xmath6 induced by @xmath76 . assume @xmath177 is an object of @xmath6 such that @xmath204 is a lattice for the left - divisibility . assume @xmath88 is an atom of the category @xmath5 that belongs to @xmath179 . then , there exists a morphism @xmath265 in @xmath266 such that the following diagram is commutative : ( 100,70 ) ( 8,0 ) ( 6,0)@xmath178(6,48)@xmath177(70,0)@xmath267(\mathcal{q})$ ] ( 70,48)@xmath267(\mathcal{p})$ ] ( 2,25)@xmath88(30,55)@xmath268(\mathcal{p})$](30,6)@xmath268(\mathcal{q})$ ] ( 82,25)@xmath267(\delta)$ ] ( 40,27)@xmath265 indeed , since @xmath88 is an atom , the greatest common divisor of @xmath88 and @xmath215 is either trivial or equal to @xmath88 . but , @xmath88 is a non - identity morphism of @xmath5 . therefore , there exists an object @xmath48 of @xmath177 such that @xmath269 is not ( the identity morphism ) @xmath48 . this implies that @xmath270 is not trivial and the claim follows . gathering the above results we get : [ thgsfvs2 ] consider a categorical garside structure @xmath271 whose object set is finite . assume that for every object @xmath177 of @xmath5 the sets @xmath204 and @xmath205 are lattices for the left - divisibility and the right - divisibility , respectively . then , the pair @xmath272 is a categorical garside structure . the category @xmath5 is cancellative and noetherian , and for every non empty subcategory @xmath177 of @xmath6 the sets @xmath204 and @xmath205 are lattices for the left - divisibility and the right - divisibility , respectively . furthermore , @xmath203 is a natural transformation from the identity functor of @xmath5 to the automorphism @xmath202 , where @xmath64 is the automorphism of @xmath6 induced by @xmath76 . there is no clear argument to expect that the left / right - divisibilities induce lattice orders in general . however there is one particular case where this property always hold : [ thgsfvs3 ] consider a monoidal garside structure @xmath124 . then , for every non - empty subcategory @xmath177 of @xmath6 the sets @xmath204 and @xmath205 are lattices for the left - divisibility and the right - divisibility . consider a categorical garside structure @xmath273 and a subcategory @xmath177 of @xmath6 . let @xmath183 , @xmath184 belong to @xmath204 . for every object @xmath48 of @xmath177 , we set @xmath274 and @xmath275 ( we recall that @xmath57 is a lattice ) . denote by @xmath188 and @xmath276 the isomorphisms associated with @xmath183 and @xmath184 , respectively . we denote by @xmath277 and @xmath278 the objects of @xmath6 such that @xmath279 and @xmath280 lie in @xmath281 and @xmath282 , respectively . let @xmath50 be another object of @xmath177 , and assume there exists @xmath21 in @xmath283 . write @xmath284 and @xmath285 . we have the sequence of equalities @xmath286 . since @xmath57 is a lattice , there exists a unique morphism @xmath287 in @xmath288 which verifies the equalities @xmath289 and @xmath290 . consider the map @xmath291 that sends an object @xmath48 of @xmath177 to @xmath277 and an edge @xmath21 of @xmath177 to @xmath287 . then , by the cancellativity property , for every two objects @xmath69 of @xmath177 , the restriction of this map to @xmath283 is injective . in particular , when @xmath177 has one object , @xmath291 is an isomorphism from the subcategory @xmath177 to the subcategory @xmath292 . assume that @xmath6 is a monoid , that is a category with one object . set @xmath293 for @xmath294 . clearly , the maps @xmath295 , @xmath296 and @xmath297 are natural transformations from @xmath298 , @xmath299 and @xmath129 to @xmath300 , @xmath301 and @xmath302 , respectively . thus , @xmath303 is a common multiple of @xmath183 and @xmath184 in @xmath5 . now , assume @xmath304 is a common multiple of @xmath183 and @xmath184 in @xmath5 with @xmath305 . let @xmath306 the isomorphism associated with @xmath304 . then , for every object @xmath48 of @xmath177 , there exists @xmath307 in @xmath308 such that @xmath309 and @xmath310 . therefore , the map @xmath311 lies in @xmath312 , and @xmath303 is the least common multiple of @xmath183 and @xmath184 in @xmath204 . similarly , for every morphism @xmath21 in @xmath283 there exists a unique morphism @xmath313 in @xmath314 such that @xmath315 . again , if @xmath6 is a monoid the map @xmath316 is injective , and @xmath317 is the greatest common divisor of @xmath183 and @xmath184 in @xmath204 . thus , the set @xmath204 is a lattice for the left - divisibility . by similar arguments , the set @xmath205 is a lattice for the right - divisibility . as a direct consequence of corollary [ thgsfvs2 ] and proposition [ thgsfvs3 ] , we get let @xmath124 be a monoidal garside structure . then , @xmath272 is a categorical garside structure.[thgsfvs4 ] a subcategory that is a groupoid will be called a _ subgroupoid _ in the sequel . when @xmath124 is a categorical garside structure such that @xmath5 is a garside category , then @xmath318 is a garside groupoid . there is another way to associate a groupoid with @xmath6 . indeed , let @xmath319 be a small groupoid . the _ groupoid of subgroupoids _ of @xmath319 is the full subcategory @xmath320 of the category @xmath321 whose objects are the subgroupoids of @xmath319 . [ defngc ] clearly , the category @xmath320 is a groupoid . then , @xmath322 is a groupoid for every small category @xmath6 . if @xmath323 is a group and @xmath324 is a subgroup , then @xmath325 is the normalizer of @xmath324 in @xmath323 . we recall that we denote by @xmath326 the functor induced by a functor @xmath121 . let @xmath6 be a category . the canonical functor @xmath327 induces a functor @xmath328 . for every object @xmath177 of the category @xmath5 we set @xmath329 . in other words , @xmath330 is the subgroupoid of @xmath119 generated by the subcategory @xmath331 . now , consider @xmath88 in @xmath179 , and denote by @xmath180 its associated isomorphism . let @xmath332 be the map that sends an object @xmath48 of @xmath330 to the morphism @xmath333 . then @xmath332 belongs to @xmath334 . indeed , consider the map @xmath335 which sends an object @xmath48 of @xmath330 to @xmath336 and a morphism @xmath21 in @xmath337 to @xmath338 . then @xmath339 is an isomorphism from @xmath330 to @xmath340 and @xmath332 is a natural transformation from @xmath341 to @xmath342 . then we obtain a map @xmath328 by setting @xmath343 and @xmath344 . clearly , this map is a functor . we remark that @xmath345 is not injective in general : for two distinct subcategories @xmath177 and @xmath178 we can have @xmath346 . for instance , consider the braid monoid @xmath347 which is defined by the monoid presentation @xmath348 ; its group of fractions is the braid group @xmath349 . the sets @xmath350 and @xmath351 generate distinct submonoids of @xmath352 , but both generate the braid group @xmath349 , as a group . moreover , the functor @xmath345 is neither faithful in general : the restriction of @xmath345 to a set @xmath179 is not always injective ; if we consider the monoid @xmath353 defined by the monoid presentation @xmath354 , and we denote by @xmath355 , @xmath356 the submonoids generated by @xmath357 and @xmath358 , respectively , then @xmath359 is a non - trivial morphism in @xmath360 , whereas @xmath361 is trivial in @xmath362 . in the sequel , we write @xmath363 for @xmath318 . [ propfaithfulpropert ] let @xmath124 be a monoidal garside structure . denote by @xmath364 the canonical embedding functor of @xmath6 into @xmath119 . the functor @xmath365 is faithful . let @xmath177 and @xmath178 be two objects of @xmath363 and consider @xmath88 , @xmath304 in @xmath366 such that @xmath367 . then , @xmath368 . then , it is enough to assume for the rest of the proof that @xmath369 and @xmath370 . since the category @xmath5 is a garside category , we can identify @xmath5 with its image in its group of fractions @xmath363 , and decompose the morphism @xmath88 as a product @xmath371 with @xmath183 and @xmath184 in @xmath372 for some object @xmath373 of @xmath5 . since @xmath374 , we have @xmath375 , that is , @xmath376 . by definition , this means that for every object @xmath48 of @xmath177 , we have @xmath377 . but @xmath364 is an injective functor . then @xmath378 and @xmath379 . as a final comment to this section , we remark that the morphism @xmath345 is not injective on the objects . therefore , the groupoid @xmath363 can not be seen as a subgroupoid of @xmath322 . in this section , we introduce and investigate the _ positive ribbon category _ and the _ ribbon groupoid_. this ribbon category is defined as a full subcategory of the category @xmath5 . the ribbon groupoid is defined as a full subgroupoid of @xmath322 , which turns out to be , under some restriction , the group of fractions of the positive ribbon category . we first define the notion of a parabolic subgroupoid . here we extend the notion of a parabolic subgroup of a garside group @xcite into the framework of garside groupoids . if @xmath6 is a category and @xmath88 lies in @xmath57 , then we denote by @xmath380_{\mathcal{c}}$ ] the set of elements between @xmath49 and @xmath88 for the left - divisibility in @xmath57 . [ defparasucat ] let @xmath124 be a categorical garside structure . a subcategory @xmath177 of @xmath6 is a _ parabolic subcategory _ of @xmath6 if for every object @xmath48 of @xmath177 the sets @xmath381 and @xmath382 are sublattices of @xmath57 and @xmath58 , respectively , and , there exists @xmath88 in @xmath383 such that for every object @xmath48 of @xmath177 , one has @xmath384_\mathcal{c } = [ 1_x,\delta(x)]_\mathcal{c}\cap \mathcal{p}_{x\to\cdot } = \delta(x)\wedge\mathcal{p}_{x\to\cdot}\ ] ] in the lattice @xmath57 . with the notation of the above definition , when @xmath177 is a parabolic subcategory then one has @xmath385_\mathcal{c}\cap \mathcal{p}_{x\to\cdot } = [ 1_x,\delta(x)]_{\mathcal{p}}$ ] , and the atoms of @xmath177 are the atoms of @xmath6 that lie in @xmath177 . it is immediate that the pair @xmath386 is a categorical garside structure . furthermore , the canonical embedding functor from @xmath177 to @xmath6 induces an injective functor from @xmath387 to @xmath119 such that the following diagram is commutative : @xmath388 therefore , we can identify @xmath387 with a subgroupoid of @xmath119 . in that context , we have @xmath389 ( where the notion of an intersection of two subcategories is defined in an obvious way ) . let @xmath124 be a categorical garside structure . + a subgroupoid of the garside groupoid @xmath119 is a _ standard parabolic subgroupoid _ when it is the groupoid of formal inverses @xmath387 of a parabolic subcategory @xmath177 of @xmath6 . a non empty subgroupoid @xmath319 of @xmath119 is a _ parabolic subgroupoid _ if there exists a standard parabolic subgroupoid @xmath387 such that @xmath390 is not empty . [ exematdeux ] in the context of garside groups , we recover the notion of a standard parabolic subgroup introduced in @xcite , which extends the notion of a standard parabolic subgroup of an artin - tits group . standard parabolic subgroups of an artin - tits group @xmath1 are the subgroups generated by any set of atoms of the associated artin - tits monoid @xmath8 ( _ cf . _ example [ exagpat ] ) ) . in the sequel , we denote by @xmath391 the standard parabolic subgroup of a garside group @xmath1 generated by a set @xmath392 of atoms of @xmath8 . we only consider parabolic subgroupoids that are standard in the sequel . therefore , we write _ parabolic _ for _ standard parabolic _ in the sequel . [ exef22 ] consider the free group @xmath153 on two letters @xmath154 and the garside structure of example [ exef2 ] . in this context , except the whole group @xmath153 and the trivial subgroup , a parabolic subgroup of @xmath153 is any subgroup generated by one generator @xmath160 . then , parabolic proper subgroups are all isomorphic to @xmath393 . [ exef23 ] direct product of garside groups are garside groups . for instance , the group @xmath1 defined by the group presentation @xmath394 is a garside group , with @xmath395 as garside element . its parabolic subgroups are the four subgroups @xmath396 , @xmath397 , @xmath398 and @xmath1 . [ exef24 ] the group @xmath1 defined by the presentation @xmath399 is a garside group , with @xmath400 as garside element . indeed , amalgam products of garside groups above their garside elements are garside groups . the parabolic subgroups of @xmath1 are the four subgroups @xmath401 , @xmath402 , @xmath403 and @xmath1 . it is quite easy to see that most of the properties proved in @xcite can be extended into the context of garside groupoid , with similar proof . we are now ready to define the ribbon groupoid , whose objects are the parabolic subgroupoids . let @xmath271 be a categorical garside structure . the _ positive ribbon category _ @xmath404 associated with @xmath6 is the full subcategory of @xmath5 whose objects are the parabolic subcategories of @xmath6 . we recall that for every parabolic subcategory @xmath177 of @xmath6 , we have identified @xmath387 with a subgroupoid of @xmath119 . we also recall that @xmath405 and that the categories @xmath177 and @xmath387 have the same objects . in particular every element @xmath88 of @xmath406 belongs to @xmath407 , where @xmath180 is the isomorphism associated with @xmath88 . let @xmath124 be a categorical garside structure . + ( i ) if @xmath387 and @xmath408 are two parabolic subgroupoids , then a ribbon from the former to the latter is a morphism @xmath88 in @xmath406 that belongs to @xmath409 . + ( ii ) the _ ribbon groupoid _ @xmath410 of @xmath119 is the subcategory of @xmath411 whose objects are the standard parabolic subgroupoids of @xmath119 and whose morphisms are the ribbons . consider the free group @xmath153 on two letters @xmath154 , and the garside structure of examples [ exef2 ] and [ exef22 ] . then , each object of the groupoid @xmath412 distinct from the trivial subgroup and the whole group can be identified with one of the @xmath160 . clearly , there are morphisms between @xmath160 and @xmath413 in @xmath412 if and only if @xmath414 is even . the category @xmath410 is clearly a groupoid and the image by @xmath345 of a morphism of @xmath404 is a ribbon . in the sequel such a ribbon is called a _ positive _ ribbon . indeed , let @xmath124 be a categorical garside structure such that the vertex set of @xmath6 is finite . denote by @xmath64 the automorphism of @xmath6 induced by @xmath76 . + ( i ) the category @xmath404 is cancellative and noetherian . + ( ii ) the functor @xmath202 restricts to an automorphism @xmath202 of @xmath404 . the restriction of @xmath203 to the vertex set of @xmath404 is a natural transformation from the identity functor of @xmath404 to this restricted automorphism . + ( iii ) when @xmath6 is a monoid , the functor @xmath328 is injective on to the category @xmath404 and the image @xmath415 is a subcategory of @xmath410 . moreover the morphism @xmath416 induces an isomorphism of groupoids from @xmath417 to @xmath410 . @xmath418 @xmath404 is cancellative and noetherian since it is a subcategory of @xmath5 . + @xmath175 from the equality @xmath419 and the definition of a parabolic subcategory , it follows that @xmath208 is a parabolic subcategory if and only if @xmath177 is a parabolic subcategory . moreover , in this case the map @xmath215 lies in @xmath420 . + @xmath421 by construction , the image of @xmath404 by @xmath345 is a subcategory of @xmath410 . furthermore , we already know that @xmath345 is faithful by proposition [ propfaithfulpropert ] . now , if @xmath422 , @xmath423 are two objects of @xmath404 , then one has @xmath424 hence , @xmath345 lies injective on @xmath404 . using that @xmath5 is a garside category , we get that @xmath416 induces an isomorphism of groupoids from the subgroupoid @xmath417 to its image , which lies in @xmath410 . let @xmath88 lie in @xmath425 and @xmath426 be its associated isomorphism . we can decompose @xmath88 as a product @xmath427 for some non negative integer @xmath220 and some @xmath183 in the monoid @xmath6 . now , the element @xmath183 belongs to @xmath428 since @xmath429 . therefore , the image of @xmath417 by @xmath345 is @xmath410 . as a consequence of the previous result , we get [ trtrtrrt]let @xmath124 be a categorical garside structure such that @xmath6 is a monoid . assume the sets @xmath430 and @xmath431 are sublattices of @xmath204 and @xmath205 , respectively , for every parabolic submonoid @xmath177 of @xmath6 . then @xmath432 is a categorical garside structure with @xmath410 as associated garside groupoid . generalizing the terminology of @xcite to garside groupoids , we could say under the assumption of proposition [ trtrtrrt ] that @xmath404 is a _ garside subcategory _ of @xmath5 . in the previous sections , we prove , under some technical assumptions , that the category of subcategories of a garside @xmath5 category @xmath6 is itself a garside category . in particular this is always the case when @xmath6 is a garside monoid . we prove that , moreover , the positive ribbon category @xmath404 is also a garside category , with the ribbon groupoid @xmath410 as associated garside groupoid , under the assumption that this positive ribbon category @xmath404 respects the lattice structure of the category of subcategories @xmath5 . in this section , we consider a simple criteria which insures that @xmath404 respects the lattice structure of @xmath5 . the criteria is verified by artin - tits groups of spherical type . we recall that a monoidal garside structure @xmath146 is a categorical garside structure such that @xmath8 is a monoid ; we denote by @xmath1 the associated garside group . we denote by @xmath433 a parabolic submonoid of @xmath8 whose atom set is @xmath434 . we denote by @xmath10 the associated standard parabolic subgroup . we recall that , in this context , an element @xmath435 of @xmath436 is an element of @xmath8 such that @xmath437 ( which is equivalent to @xmath438 ) . from now on , for @xmath439 in @xmath440 , we denote by @xmath441 and @xmath442 their lcm and their gcd for the left - divisibility , respectively , and we denote by @xmath443 and @xmath444 their lcm and their gcd for the right - divisibility , respectively . [ defnustructure]let @xmath146 be a monoidal garside structure . let @xmath137 be the atom set of @xmath8 . + ( i ) we say that @xmath445 _ has a @xmath2-function _ if for each parabolic submonoid @xmath433 there exists a map @xmath446 such that 1 . for every @xmath136 in @xmath137 , the morphism @xmath447 is an atom of @xmath3 ; 2 . for every @xmath136 in @xmath137 and every @xmath435 in @xmath448 , if @xmath136 left - divides @xmath435 then @xmath447 left - divides @xmath435 ; 3 . for every @xmath449 in @xmath137 , the lcm of @xmath447 and @xmath450 for left - divisibility lies in @xmath448 . \(ii ) we say that @xmath445 _ has a @xmath451-function _ if for each parabolic submonoid @xmath433 there exists a map @xmath452 such that 1 . for every @xmath136 in @xmath137 , the morphism @xmath453 is an atom of @xmath3 ; 2 . for every @xmath136 in @xmath137 and every @xmath435 in @xmath454 , if @xmath136 right - divides @xmath435 then @xmath453 right - divides @xmath435 ; 3 . for every @xmath449 in @xmath137 , the lcm of @xmath453 and @xmath455 for right - divisibility lies in @xmath454 . \(iii ) we say that @xmath445 has a _ @xmath2-structure _ if it has both a @xmath2-function and a @xmath451-function . \(i ) we do not require that @xmath136 left - divides @xmath447 . + ( ii ) when @xmath1 has a @xmath2-function ( _ resp . _ a @xmath451-function ) , then each atom of @xmath456 that belongs to @xmath457 ( _ resp . _ to @xmath458 ) is equal to some @xmath447 ( _ resp . _ some @xmath453 ) . furthermore , if @xmath435 is an atom of @xmath456 that belongs to @xmath457 ( _ resp . _ to @xmath458 ) and @xmath136 in @xmath137 left - divides ( _ resp . _ right - divides ) @xmath435 , then @xmath459 ( _ resp . _ @xmath460 ) . [ exemplnufunctspheratgp]every artin - tits monoid of spherical type ( _ cf . _ example [ exef23 ] ) has a @xmath2-structure . consider the notation of example [ exef23 ] . let @xmath434 be a subset of @xmath137 and @xmath136 lie in @xmath137 . if @xmath136 does not belong to @xmath434 , we set @xmath461 and @xmath462 ; if @xmath136 lies in @xmath434 , we set @xmath463 where @xmath464 is the indecomposable component of @xmath434 that contains @xmath136 . we recall that an artin - tits group of spherical type is said to be _ indecomposable _ when it is not the direct product of two of its proper standard parabolic subgroups . it follows from @xcite that these functions provide a @xmath2-structure to the artin - tits monoid . [ rtgvfresc ] let @xmath146 be a monoidal garside structure . + ( i ) if @xmath1 has a @xmath2-function , then the set @xmath457 is a sublattice of @xmath465 for every parabolic submonoid @xmath433 . + ( ii ) if @xmath1 has a @xmath451-function then the set @xmath458 is a sublattice of @xmath466 for every parabolic submonoid @xmath433 . ( 150,110 ) ( 10,15 ) ( 4,6)@xmath467(61,6)@xmath468(126,6)@xmath469 ( 4,106)@xmath434(61,106)@xmath470(126,106)@xmath471 ( 48,60)@xmath472 ( 4,60)@xmath435(20,107)@xmath473 ( 80,106)@xmath474(36,81)@xmath475(52,36)@xmath476(94,73)@xmath477(128,60)@xmath478 @xmath418 let @xmath137 be the atom set of @xmath8 . let @xmath479 be a parabolic submonoid of @xmath8 , and consider @xmath439 that lie in @xmath448 . since @xmath3 is a full subcategory of @xmath480 , it is enough to prove that the lcm and the gcd for the left - divisibility in @xmath8 of @xmath435 and @xmath90 belong to @xmath448 . let @xmath136 lie in @xmath137 . + gcd s : by the second defining property of a @xmath2-function , if @xmath136 left - divides both @xmath435 and @xmath90 then @xmath447 left - divides both @xmath435 and @xmath90 . since @xmath447 lies in @xmath3 , by the noetherianity property , it follows that the gcd of @xmath435 and @xmath90 for the left - divisibility lies in @xmath448 . + lcm s : let @xmath435 be in @xmath481 and @xmath90 be in @xmath436 . we have to prove that the lcm @xmath441 of @xmath435 and @xmath90 belongs to @xmath448 . clearly , we only need to consider the case when @xmath459 for some @xmath136 in @xmath137 . furthermore we can assume without restriction that @xmath435 does not left - divide @xmath90 . write @xmath482 . it is enough to prove that there exist @xmath483 in @xmath137 and @xmath484 in @xmath8 such that @xmath485 . indeed , in this case one has @xmath486 , and if @xmath487 we can write @xmath488 ; in particular , @xmath489 . we can repeat the process as long as @xmath490 . by the noetherianity property of @xmath8 the process has to stop . so , consider a decomposition @xmath491 where @xmath492 and @xmath473 belongs to @xmath493 . write @xmath494 and @xmath495 ( see figure [ diagramme ] ) . since @xmath8 has a @xmath2-function and @xmath435 is equal to @xmath447 , there exist @xmath468 and @xmath472 such that @xmath496 , @xmath476 and @xmath477 belong to @xmath497 , @xmath498 and @xmath499 , respectively . by assumption , @xmath496 does not left - divide @xmath474 . in particular @xmath500 , and there exists @xmath501 in @xmath137 such that @xmath502 left - divides @xmath476 ; furthermore , @xmath476 left - divides @xmath503 . therefore , if @xmath504 we are done . otherwise , replacing @xmath435 and @xmath90 by @xmath505 and @xmath477 , respectively , we can find @xmath506 in @xmath137 such that @xmath507 . furthermore , we can write @xmath508 and @xmath509 with @xmath510 , @xmath511 and @xmath512 in @xmath513 , @xmath514 and @xmath515 , respectively , for some subsets @xmath516 , @xmath517 of @xmath137 . again , there exists @xmath518 in @xmath137 such that @xmath519 left - divides @xmath511 , and @xmath511 left - divides @xmath520 . we can repeat the process as long as @xmath521 . but , @xmath522 . by the noetherianity property , there exists @xmath34 such that @xmath523 . in this case , @xmath524 left - divides @xmath478 , and we are done . + @xmath175 the proof is similar to the proof of @xmath418 . ( 190,60 ) ( 10,6 ) ( 50,0)@xmath160(90,0)@xmath161(133,0)@xmath525 ( 50,36)@xmath160(92,36)@xmath161(135,36)@xmath525 ( 25,11)@xmath526(67,11)@xmath527(115,11)@xmath528(158,11)@xmath529 from propositions [ trtrtrrt ] and [ rtgvfresc ] we obtain theorem [ thintror(a ) ] , which applies in the case of artin - tits groups of spherical type . the latter groups are not the only garside groups which have a @xmath2-structure : \(i ) consider the notation of example [ exef2 ] . the category @xmath530 is a garside category with @xmath531 as associated garside groupoid . + ( ii ) the category @xmath3 of the garside group @xmath1 of example [ exef23 ] is a garside category with @xmath4 as associated garside groupoid . + ( iii ) the category @xmath3 of the garside group @xmath1 of example [ exef24 ] is a garside category with @xmath4 as associated garside groupoid . we only need to explicit a @xmath2-structure in each case : + @xmath418 for @xmath532 and @xmath533 , one has @xmath534 ; for @xmath535 and @xmath533 , one has @xmath536 ; for @xmath537 and @xmath533 with @xmath538 , one has @xmath539 , @xmath540 , and @xmath541 . + @xmath175 for @xmath532 or @xmath542 and @xmath543 , one has @xmath544 ; for @xmath545 and @xmath543 , one has @xmath546 ; for @xmath547 , one has @xmath548 , and @xmath549 . + @xmath421 for @xmath532 and @xmath543 , one has @xmath544 ; for @xmath545 and @xmath543 , one has @xmath550 ; for @xmath551 , one has @xmath552 , @xmath553 and @xmath554 ; for @xmath555 , one has @xmath556 , @xmath557 and @xmath558 . given a group , it is a classical question to obtain a presentation of this group . here , we consider the similar problem in the context of groupoids . for all this section we fix a monoidal garside structure ( @xmath559 , such that the garside monoid @xmath8 has a @xmath2-structure . we denote by @xmath137 the atom set of @xmath8 and by @xmath1 its group of fractions . in the previous section , we have seen that the ribbon groupoid @xmath4 is a garside groupoid . let us first precise the notion of a groupoid presentation . we recall that the free category @xmath46 associated with a graph @xmath16 has been defined in section [ definsmcat ] . let @xmath6 be a small category . let @xmath16 be a quiver whose vertex set is the object set of @xmath6 and whose edges are morphisms of @xmath6 . + ( i ) we say that @xmath16 _ generates @xmath6 as a category _ if the canonical map from @xmath16 to @xmath6 extends to a surjective functor from the free category @xmath46 to the category @xmath6 . + ( ii ) if @xmath6 is a groupoid , we say that @xmath16 _ generates @xmath6 as a groupoid _ if the canonical map from @xmath16 to @xmath6 extends to a surjective functor from the free groupoid @xmath120 to the groupoid @xmath6 . let @xmath6 be a small category . consider a quiver @xmath16 and a set @xmath560 of morphisms of @xmath46 with the same source and the same target . denote by @xmath561 is the congruence on @xmath27 generated by @xmath562 . + ( i ) we say that @xmath563 is a _ presentation of @xmath6 as a category _ if @xmath16 generates @xmath6 as a category and the surjective functor from @xmath46 to @xmath6 induces an isomorphism of categories from @xmath564 to @xmath6 . + ( ii ) when @xmath6 is a groupoid , we say that @xmath563 is a _ presentation of @xmath6 as a groupoid _ if @xmath16 generates @xmath6 as a groupoid and the surjective functor from @xmath120 to @xmath6 induces an isomorphism of groupoid from @xmath565 to @xmath6 . [ generquivera]let @xmath568 be the graph whose vertices are the standard parabolic subgroups of @xmath1 , and whose edge set is the image of the @xmath2-functions : to each element @xmath447 that lies in @xmath569 corresponds an edge from @xmath10 to @xmath11 in @xmath568 . the quiver @xmath568 generates @xmath3 as a category and @xmath4 as a groupoid . the atom graph of @xmath3 is a generating quiver and the image of the @xmath2-functions is precisely the atom set of the category @xmath3 ( see section [ sectiongardgrd ] ) . furthermore , @xmath4 is the group of fractions of the garside category @xmath3 . in the case of an artin - tits group of spherical type , we have seen in example [ exemplnufunctspheratgp ] that there is two types of atoms in @xmath4 , namely the @xmath574 and the @xmath575 . we are going to see that this is a general fact ; we need first to study the quasi - centralizer of a garside group . let @xmath576 be a garside monoid with @xmath323 as garside group . the _ quasi - centralizer _ of the monoid @xmath576 is the submonoid @xmath577 defined by @xmath578 the _ quasi - centralizer _ of the group @xmath323 is the subgroup @xmath579 defined by @xmath580 in other words , @xmath577 and @xmath581 are equal to @xmath582 and @xmath583 , respectively . by definition , the garside element @xmath76 belongs to @xmath577 . in this section , we extend the main result of @xcite into every garside group . the proof is in the spirit of the latter reference . however , some definitions and technical arguments of the proof of @xcite can not be simply extended to every garside group . this is because in general the set of divisors of @xmath76 is not finite . in our proof , this argument is replaced by the noetherianity property of @xmath576 . we need some preliminary results . [ defindesdeltag ] let @xmath584 be a monoidal garside structure . + ( i ) for every @xmath435 in @xmath576 the set @xmath585 has a lcm @xmath586 for the left - divisibility . furthermore , @xmath435 left - divides @xmath586 and @xmath587 . + ( ii ) for every @xmath435 in @xmath576 the set @xmath588 has a @xmath589 for the right - divisibility . furthermore , @xmath435 right - divides @xmath589 and @xmath590 . by symmetry , it is enough to prove @xmath418 . from the noetherianity property of @xmath576 it follows that every none - empty subset of @xmath576 that has a common multiple for the left - divisibility has a lcm for the left - divisibility . but @xmath76 is quasi - central and there exists a positive integer @xmath220 such that @xmath435 left - divides @xmath591 . therefore , @xmath591 is a common multiple of @xmath585 , and therefore @xmath586 exists . by definition we have @xmath587 . finally , considering @xmath592 , we obtain that @xmath435 left - divides @xmath586 . [ lemsuitetaun ] let @xmath584 be a monoidal garside structure and @xmath435 lie in @xmath576 . + define @xmath593 by @xmath594 , @xmath595 and @xmath596 . the sequence @xmath593 stabilizes to an element @xmath597 which belongs to @xmath598 . furthermore , for every element @xmath90 in @xmath598 , @xmath435 left - divides @xmath90 if and only if @xmath597 left - divides @xmath90.[defindestaug ] in particular , for every quasi - central element @xmath90 , we have @xmath599 . by the previous lemma , @xmath601 left - divides @xmath602 , and @xmath603 right - divides @xmath601 . furthermore , with the notation of lemma [ defindesdeltag ] s proof , all the @xmath604 divides @xmath591 ( for both the left - divisibility and the right - divisibility since @xmath591 is a quasi - central element ) . by the noetherianity property , there exist @xmath586 in @xmath576 and @xmath605 such that @xmath606 for every @xmath236 . by lemma [ defindesdeltag ] , we get the sequence of inclusions @xmath607 . thus , @xmath597 is a quasi - central element . assume @xmath90 lies in @xmath598 such that @xmath435 left - divides @xmath90 , and write @xmath608 . by assumption , we have @xmath609 . therefore , @xmath90 is a common multiple of the set @xmath610 and @xmath586 left - divides @xmath90 . but @xmath90 is quasi - central , then @xmath586 , that is @xmath611 , also right - divides @xmath90 . inductively we get that @xmath612 left - divides and right - divides @xmath90 for every @xmath220 . finally , by lemma [ defindesdeltag ] , @xmath602 left - divides @xmath601 and @xmath601 right - divides @xmath603 for every @xmath220 ; then we can write @xmath613 for some @xmath614 in @xmath576 . since @xmath597 is a quasi - central , it follows that @xmath435 left - divides @xmath597 . [ remlemsuitetaun]with the notation of lemma [ lemsuitetaun ] and a similar proof of this lemma , we can obtain an element @xmath615 by considering the sequence @xmath616 defined by @xmath617 , @xmath618 and @xmath619 . using that an element of @xmath577 has the same set of left - divisors and of right - divisors , we deduce that @xmath620 for every @xmath435 in @xmath576 . [ centrcommprop]let @xmath584 be a monoidal garside structure , and denote by @xmath137 the atom set of @xmath576 . for every @xmath449 in @xmath137 , either @xmath621 or @xmath622 in @xmath576 . furthermore , in any case , @xmath623 . the main argument of the first part s proof is as in @xcite . since @xmath624 is a quasi - central element , it has the same left - divisors and right - divisors . assume that @xmath136 divides @xmath624 and let us prove that @xmath621 . by lemma [ defindestaug ] , we can write @xmath625 . still by lemma [ defindestaug ] , if @xmath483 divides @xmath626 , then @xmath624 divides @xmath626 and we are done : @xmath621 . assume this is not the case . since @xmath483 is an atom of @xmath576 , the elements @xmath435 and @xmath626 are prime to each other for the right - divisibility in @xmath576 . this implies that @xmath624 is the lcm of @xmath435 and @xmath90 for the left - divisibility . in particular , @xmath627 . then , by definition , @xmath483 left - divides @xmath628 and , therefore , left - divides @xmath629 . in particular , we can write @xmath630 for some @xmath478 in @xmath631 . but @xmath626 and @xmath624 are quasi - central elements , then @xmath90 is also a quasi - central element and @xmath632 . a contradiction : we get @xmath633 with @xmath634 . now , if @xmath635 then there exists @xmath636 in @xmath137 that divides both @xmath626 and @xmath624 . by the above argument , we have @xmath637 . assume finally that @xmath638 . we can write @xmath639 and @xmath640 where @xmath641 lie in @xmath576 . clearly @xmath641 have to be quasi - central elements . as @xmath642 , by the first part of the proof , the decomposition @xmath643 is a left greedy normal form ( _ cf . [ uniquedecomposprop]@xmath175 ) . but we have @xmath644 . since @xmath643 is not in @xmath576 , @xmath624 and @xmath626 do not right - divide @xmath505 and @xmath477 , respectively . this implies that both @xmath645 and @xmath646 are right greedy normal forms . by unicity of this normal form , we get @xmath647 and @xmath648 . thus , @xmath623 . [ theoremcentralisateurgeneralisationpicantin]let @xmath584 be a monoidal garside structure , and denote by @xmath137 the atom set of @xmath576 . + ( i ) the monoid @xmath598 is a free commutative monoid . furthermore there exists a projection @xmath649 such that 1 . the set @xmath650 is a free base of @xmath577 ( where some @xmath626 are possibly equal ) ; 2 . the map @xmath304 respects the left - divisibility and the right - divisibility , in particular if @xmath651 left / right - divides @xmath652 then @xmath626 left / right - divides @xmath435 ; 3 . the map @xmath304 is a semilattice homomorphism for @xmath653 and for @xmath654 ; in particular , @xmath655 for every @xmath439 in @xmath576 . this theorem should be compared with the definition of @xmath2-functions ( _ cf . _ definition [ defnustructure ] ) . note that the map @xmath304 is not a semilattice homomorphism for @xmath656 ( nor @xmath657 ) in general @xcite . @xmath418 we already know by lemma [ defindestaug ] that @xmath658 , that @xmath659 and that @xmath304 respects the left - divisibility and the right - divisibility . furthermore , using the noetherianity property and the same lemma , we get that the set @xmath650 is a generating set for @xmath577 . by proposition [ centrcommprop ] , it follows that the monoid @xmath577 is commutative , and that for distinct elements @xmath626 and @xmath624 , the element @xmath660 is a common multiple of @xmath626 and @xmath624 for both the left - divisibility and the right - divisibility . using lemma [ defindestaug ] , we get that @xmath660 is the lcm of @xmath626 and @xmath624 ( for both the left - divisibility and the right - divisibility ) : otherwise , there would be some @xmath636 in @xmath137 that divides both @xmath626 and @xmath624 and then we should have @xmath637 . therefore , @xmath598 is a free commutative monoid with @xmath650 as a free base . now , let @xmath435 and @xmath90 belong to @xmath576 . they left - divides @xmath441 . therefore @xmath597 , @xmath629 and @xmath661 left - divides @xmath662 . but @xmath597 and @xmath629 have a common multiple in @xmath577 for the left - divisibility ( that is some power of @xmath76 ) . by the noetherianity property , there exists a minimal common multiple @xmath663 of @xmath626 and @xmath624 for the left - divisibility in @xmath598 . assume there exists @xmath636 in @xmath137 that right - divides both @xmath664 and @xmath665 ; then by remark [ remlemsuitetaun ] , @xmath666 also right - divides @xmath664 and @xmath665 , which is impossible by minimality . therefore @xmath661 belongs to @xmath577 . since @xmath441 left - divides @xmath661 , the element @xmath662 left - divides @xmath661 . as a conclusion , @xmath662 is equal to @xmath661 . + @xmath175 let @xmath435 lie in @xmath579 and @xmath667 be its left greedy normal form ( _ cf . _ proposition [ uniquedecomposprop ] ) . we claim that @xmath359 and @xmath357 belong to @xmath577 . this proves that @xmath579 is ( isomorphic to ) the group of fractions of @xmath577 and , therefore , the second point of the proposition . actually , every element @xmath435 of @xmath323 can be written as @xmath668 with @xmath669 and @xmath670 in @xmath576 . but @xmath76 is a quasi - central element . therefore , when @xmath435 lies in @xmath579 , then @xmath670 and @xmath671 belong to @xmath577 . now assume that @xmath672 where both @xmath358 and @xmath673 lie in @xmath577 . if some @xmath636 in @xmath137 left - divides both @xmath358 and @xmath673 , then @xmath666 also left - divides @xmath358 and @xmath673 . we conclude as in the proof of @xmath418 . in the next section , we are going to apply the above results to parabolic submonoids . we recall that we have fixed a garside monoid @xmath8 with a @xmath2-structure in the introduction of section [ sectionpresgrp ] . [ cornuspeciaux ] let @xmath674 be a parabolic submonoid of @xmath8 . let @xmath136 belong to @xmath137 . + ( i ) if @xmath136 lies in @xmath434 , then @xmath675 ; + ( ii ) if @xmath447 belongs to @xmath479 , then there exists @xmath483 in @xmath434 such that @xmath676 . @xmath418 assume @xmath136 lies in @xmath434 . since @xmath136 left - divides @xmath626 , by the defining properties of a @xmath2-structure , @xmath447 left - divides @xmath626 . but @xmath626 belongs to @xmath479 which is a parabolic submonoid . therefore , @xmath447 belongs to @xmath433 . consequently , @xmath447 lies in @xmath677 , and has to be equal to @xmath626 . + @xmath175 assume @xmath447 belongs to @xmath479 . this implies that @xmath447 lies in @xmath677 , and therefore is equal to a product @xmath678 with @xmath679 in @xmath137 . but @xmath447 is an atom of @xmath3 , hence @xmath680 . we still consider the general notations of section [ sectionpresgrp ] . in lemma [ generquivera ] , we have seen that the quiver @xmath568 whose edges are the @xmath447 is a generating quiver for the groupoid @xmath4 . in corollary [ cornuspeciaux ] , we have seen that among the @xmath447 there is two kinds of atoms of @xmath3 , like in the case of artin - tits groups of spherical type ( _ cf . _ example [ exemplnufunctspheratgp ] ) . in the sequel , for every parabolic submonoid @xmath681 of @xmath8 , and every element @xmath136 that belongs to @xmath137 , we denote by @xmath682 the element @xmath447 when it belongs to @xmath479 , and by @xmath683 otherwise . similarly , we write @xmath684 or @xmath685 for @xmath453 , depending whether or not this atom lies in @xmath479 . for instance , for every @xmath136 in @xmath434 , one has @xmath686 . [ lemfghbvcasc ] let @xmath479 and @xmath687 be parabolic submonoids of @xmath8 . assume @xmath682 lies in @xmath677 and @xmath688 belongs to @xmath436 . then , there exists @xmath689 in @xmath690 such that @xmath691 since @xmath682 lies in @xmath677 and @xmath688 belongs to @xmath436 , the element @xmath692 belongs to @xmath690 . but @xmath682 is an atom of the category @xmath3 . therefore @xmath692 is also an atom of @xmath3 . by corollary [ cornuspeciaux ] , we are done . [ mpolkiujyh]let @xmath479 and @xmath687 be parabolic submonoids of @xmath8 . assume that @xmath435 is equal to @xmath693 and lies in @xmath436 , where @xmath694 ; assume furthermore that @xmath477 belongs to @xmath479 and @xmath512 belongs to @xmath695 . then , + ( i ) the lcm and the gcd in @xmath8 of @xmath435 and @xmath477 for the left - divisibility is @xmath696 and @xmath105 , respectively . + ( ii ) the lcm and the gcd in @xmath8 of @xmath435 and @xmath512 for the right - divisibility is @xmath697 and @xmath105 , respectively . assume @xmath680 . if @xmath636 in @xmath137 left - divides both @xmath435 and @xmath477 , then @xmath636 lies in @xmath434 and @xmath698 left - divides @xmath683 . this is impossible because in this case we would have @xmath699 . therefore , the gcd in @xmath8of @xmath435 and @xmath477 for the left - divisibility is @xmath105 . similarly , the gcd of @xmath435 in @xmath8 and @xmath512 for the right - divisibility is @xmath105 ( we recall that @xmath688 is equal to some @xmath700 ) . let @xmath701 in @xmath433 be such that @xmath702 . the gcd of @xmath701 and @xmath435 for the left - divisibility is @xmath105 by the first part of the proof . thus , @xmath697 is the lcm of @xmath435 and @xmath512 for the right - divisibility . similarly , @xmath696 is the lcm in @xmath8 of @xmath435 and @xmath477 for the left - divisibility . assume @xmath703 . by an easy induction on @xmath478 we deduce from the case @xmath680 that the element @xmath696 is the lcm of @xmath435 and @xmath477 for the left - divisibility . similarly @xmath697 is the lcm of @xmath435 and @xmath512 for the right - divisibility . finally , we deduce from the latter result the last part of the lemma : let @xmath704 be the gcd of @xmath435 and @xmath477 for the left - divisibility . since @xmath479 is a parabolic submonoid , @xmath704 belongs to @xmath479 . in particular the lcm of @xmath704 and @xmath435 for the left - divisibility , which is @xmath435 , has to be equal to @xmath705 . then @xmath706 . similarly , the gcd of @xmath435 and @xmath512 for the right - divisibility is @xmath105 . [ lemsecdecpodeslcmdesnu ] let @xmath479 be a parabolic submonoid of @xmath8 and consider @xmath687 , @xmath707 two parabolic submonoids of @xmath8 which are distinct from @xmath433 . assume @xmath683 and @xmath688 belong to @xmath436 and @xmath481 , respectively . then , there exist a parabolic submonoid @xmath708 , a path @xmath709 in @xmath710 and a path @xmath711 in @xmath712 such that @xmath713 we can assume @xmath714 . the category @xmath3 is a garside category with @xmath568 for atom graph . then we can decompose @xmath715 in the category @xmath3 : @xmath716 by lemma [ lemfghbvcasc ] , we can assume without restriction that there exists @xmath484 and @xmath717 such that @xmath718 is equal to @xmath719 if and only if @xmath720 , and that @xmath721 is equal to @xmath722 if and only if @xmath723 . let us set @xmath724 and @xmath725 , and consider @xmath726 such that @xmath727 . since @xmath435 and @xmath90 belong to @xmath687 , the elements @xmath505 and @xmath477 lie in @xmath687 . but @xmath505 has to right - divide @xmath728 . then , by lemma [ mpolkiujyh ] , we have @xmath729 . similarly , @xmath730 that is @xmath731 . finally we get @xmath732 since the lcm of @xmath435 and @xmath90 for the right - divisibility has to be @xmath105 . [ theorpreseconjplusa ] let @xmath146 be a monoidal garside structure with a @xmath2-structure . denote by @xmath137 the atom set of @xmath440 , and by @xmath1 its group of fractions . consider the set @xmath562 of equalities of paths in @xmath733 : that hold in @xmath3 and such that in equalities of type ( [ deregalite ] ) the two paths correspond to two decompositions of @xmath715 as in equation ( [ lemsecdecpodeslcmdesnu2 ] ) of lemma [ lemsecdecpodeslcmdesnu ] . + then , @xmath737 is a presentation of @xmath4 as a groupoid . the garside group @xmath1 has a @xmath2-function , then @xmath4 is a garside groupoid , with @xmath3 as garside category . therefore , it is enough to prove that @xmath737 is a presentation of @xmath3 as a category . we already know that @xmath568 is a generating quiver of @xmath3 . now , let @xmath38 be the congruence on @xmath738 defined by the relations considered in the theorem . let @xmath739 and @xmath740 be two paths in @xmath741 . clearly , if the two paths are @xmath38-equivalent , they represent the same element in @xmath742 . conversely , assume they represent the same element in @xmath742 and let us prove that they are @xmath38-equivalent . using defining relations ( [ reldeuxdelaprese ] ) of @xmath38 , we obtain a path @xmath743 which is @xmath38-equivalent to the path @xmath739 . similarly , we also obtain a path @xmath744 which is @xmath38-equivalent to the path @xmath740 . by lemma [ mpolkiujyh ] the equalities @xmath745 and @xmath746 hold in @xmath8 ( see the proof of lemma [ lemsecdecpodeslcmdesnu ] ) . from theorem [ theoremcentralisateurgeneralisationpicantin ] , it follows that the two paths @xmath747 and @xmath748 are @xmath38-equivalent , using the relations of type ( [ relundelaprese ] ) . finally , @xmath749 and @xmath750 are @xmath38-equivalent by lemma [ lemtechnordset ] . define @xmath434 to be the set made of the paths in @xmath751 whose images in @xmath752 right - divides @xmath749 . we define a noetherian partial order on @xmath434 by saying that for @xmath753 in @xmath434 , one has @xmath754 if @xmath755 in @xmath751 , or the image of @xmath756 in @xmath3 strictly right - divides the image of @xmath757 . we define @xmath23 as the set of pair of elements of @xmath434 that have the same image in @xmath752 . we fix for every element @xmath53 in @xmath752 a representing path @xmath758 in @xmath751 . we define @xmath759 in the following way . let @xmath760 belong to @xmath23 . write @xmath761 and @xmath762 and assume they are distinct . let @xmath34 be minimal such that @xmath763 . choose an arbitrary defining relation @xmath764 of type ( [ deregalite ] ) . there exists @xmath53 in @xmath752 such that @xmath765 is equal to the image of @xmath756 in @xmath752 . we set @xmath766 and @xmath767 . by lemma [ lemtechnordset ] , and with its notations , there exists @xmath220 such that @xmath768 . the results follows easily . given a garside monoid with a right partial action on a set , dehornoy shows in @xcite that under a technical hypothesis , a natural garside category can be associated with this action . we note that the positive ribbon category @xmath3 does not arised trivially in this way . actually , we can not defined a right partial action of @xmath8 on the set of its parabolic submonoids by setting for an atom @xmath136 that @xmath769 where @xmath471 is such that @xmath447 belongs to @xmath770 . for instance , consider the braid group @xmath771 with the notations of example [ exagpat ] ; then @xmath772 whereas @xmath773 . indeed , we do not have the equality @xmath774 but @xmath775 which is a relation of type ( [ deregalite ] ) in theorem [ theorpreseconjplusa ] . [ lemtechnordset ] let @xmath776 be a noetherian partially ordered set with an infimum element denoted by @xmath105 . assume @xmath23 is a subset of @xmath777 that contains @xmath778 and which is stabilized by the maps @xmath779 . assume @xmath780 is such that @xmath781 for @xmath48 in @xmath23 , and @xmath782 and @xmath783 for @xmath784 . consider @xmath785 defined inductively by @xmath786 and @xmath787 for @xmath788 . for every @xmath789 in @xmath23 , there exists @xmath220 such that @xmath768 . \(i ) for every @xmath435 which lies in @xmath448 , there exists a unique pair @xmath791 such that @xmath792 with @xmath505 in @xmath677 and @xmath793 in @xmath794 . + ( ii ) let @xmath439 belong to @xmath795 . assume @xmath792 and @xmath796 with @xmath726 in @xmath677 and @xmath797 in @xmath794 . then @xmath441 is equal to @xmath798 , and @xmath442is equal to @xmath799 . [ lemstructconjnu ] @xmath418 it has been proved when proving theorem [ theorpreseconjplusa ] . + @xmath175 it is clear that @xmath798 is a common multiple of @xmath435 and @xmath90 . conversely , any common multiple of @xmath435 and @xmath90 is a common multiple of @xmath800 and @xmath801 for the left - divisibility . but the lcm of the latter is @xmath798 by lemma [ mpolkiujyh ] ( the lcm of @xmath793 and @xmath512 lies in @xmath795 by lemma [ lemsecdecpodeslcmdesnu ] ) . similarly , @xmath799 is a common left - divisor of @xmath435 and @xmath90 . now , let @xmath802 belong to @xmath448 with @xmath484 in @xmath677 and @xmath803 in @xmath794 , and assume @xmath478 left - divides both @xmath435 and @xmath90 . using that @xmath435 is the lcm of @xmath435 and @xmath478 for the left - divisibility , we get that @xmath484 and @xmath803 left - divide @xmath505 and @xmath793 , respectively . by the same argument , @xmath484 and @xmath803 also left - divide @xmath477 and @xmath512 , respectively . we deduce that @xmath484 and @xmath803 left - divide @xmath804 and @xmath805 , and finally that @xmath478 left - divides @xmath799 . let @xmath146 be a monoidal garside structure such that @xmath8 has a @xmath2-structure . denote by @xmath137 its atom set , and by @xmath1 its group of fractions . the subgroupoid @xmath14 of @xmath4 generated by the @xmath683 is a standard parabolic subgroupoid of @xmath4 . furthermore , let @xmath806 denote the subgraph of @xmath568 with the same vertex set and whose edges are the @xmath683 . let @xmath807 be the set of relations of type ( 3 ) that appear in theorem [ theorpreseconjplusa ] . then @xmath808 is a presentation of @xmath14 as a groupoid . we recall that the notion of a parabolic subcategory has been introduced in definition [ defparasucat ] . we only need to prove that @xmath790 is a parabolic subcategory of @xmath404 and that @xmath808 is a presentation of @xmath790 as a category . let @xmath479 be a parabolic subcategory of @xmath8 with @xmath809 as garside element . the lattice @xmath794 is a sublattice of @xmath448 by lemma [ lemstructconjnu ] . for similar reason , @xmath810 is a sublattice of @xmath454 . by definition @xmath268(a_x^{\scriptscriptstyle + } ) $ ] is equal to @xmath76 and we can decompose @xmath76 as a product @xmath811 . clearly , @xmath812 lies in @xmath448 . as @xmath479 is a parabolic submonoid , no element of @xmath434 can left - divides @xmath812 . this implies that @xmath812 lies in @xmath794 . we claim that the equalities @xmath813_{\conjplusa } = [ 1,\delta]_{\conjplusa}\bigcap\ { \conjplusanu}_{a_x^{\scriptscriptstyle + } \to\cdot } = \delta\land{\conjplusanu}_{a_x^{\scriptscriptstyle + } \to\cdot}\ ] ] hold . this proves that @xmath790 is a parabolic subcategory of @xmath404 . indeed , this equalities are direct consequences of lemma [ lemstructconjnu ] : assume @xmath435 lie in @xmath814 and assume it belongs to @xmath815_{\conjplusa}$ ] . write @xmath792 with @xmath505 in @xmath677 and @xmath793 in @xmath794 . by lemma [ lemstructconjnu ] , @xmath505 and @xmath793 left - divide @xmath809 and @xmath812 , respectively , and @xmath729 if and only @xmath435 left - divides @xmath812 . we recall that in the case of an artin - tits group @xmath1 with @xmath137 as a generating set , every subset @xmath816 of @xmath137 generates a standard parabolic subgroup with @xmath816 as atom set . in this particular case , we can precise the statement of the theorem [ theorpreseconjplusa ] . [ antelastprop ] let @xmath146 be a monoidal garside structure such that @xmath8 is an artin - tits monoid of spherical type . denote by @xmath137 its atom set and by @xmath1 its associated artin - tits group . assume that @xmath434 is a non empty subset of @xmath137 and let @xmath449 lie in @xmath817 and be distinct . then @xmath818 has exactly two representing path in the free category @xmath733 . one begins with @xmath683 and the other with @xmath688 . one has @xmath819 and @xmath820 ( _ cf . _ example [ exemplnufunctspheratgp ] ) . therefore , @xmath818 is equal to @xmath821 and , in particular , @xmath822 is equal to @xmath823 , which is a simple element of @xmath8 ( _ c.f . _ definition [ defisimplelem ] ) . consider in @xmath733 a representing path @xmath824 of @xmath818 such that @xmath825 and @xmath826 . given @xmath827 and @xmath828 , the set @xmath829 is uniquely defined by the equality @xmath830 . we can write @xmath831 now it is well - known that the simple elements of an artin - tits monoid are _ square free _ : no representing word of a simple element has a square of an element of @xmath137 as a subfactor . this implies that for every @xmath832 , the element @xmath828 is uniquely defined by the two conditions @xmath833 and @xmath834 . thus , there is a unique representing path of @xmath818 in @xmath733 that starts with @xmath683 . similarly , there is a unique representing path of @xmath818 in @xmath733 that starts with @xmath688 . in @xcite , bessis defined a _ weak garside group _ to be a group that is isomorphic to some vertex group @xmath835 of a garside groupoid @xmath6 . clearly a garside group is a weak garside group , but there is no obvious general reason for a weak garside group to be a garside group @xcite . in this section , we consider a garside group @xmath1 with a @xmath2-function and the subgroup @xmath323 of some vertex group @xmath836 that is generated by the atoms of @xmath4 that belong to @xmath836 . under some assumptions that hold in the case of an artin - tits group of spherical type , we prove ( _ cf . _ theorem [ thgarssubg ] ) that @xmath323 is a garside group . this section is inspired by ( * ? ? ? * theorem b ) . we still assume that @xmath146 is monoidal garside structure such that @xmath8 has a @xmath2-structure . we still denote by @xmath137 and @xmath1 the atom set and the group of fractions of @xmath8 , respectively . for all the section , we fix a standard parabolic subgroup @xmath10 of @xmath1 . we set @xmath837 following @xcite , we call the elements of @xmath838 the _ shakers _ of @xmath10 . it is immediate that the set of shakers @xmath838 is equal to @xmath839 . there is two kinds of shakers , namely the @xmath682 and the @xmath683 ; in the sequel , we denote by @xmath840 the set @xmath841 , which is equal to the set @xmath842 . furthermore , we denote by @xmath843 and by @xmath844 the sets @xmath845 and @xmath846 respectively . there is no reason to expect that @xmath847 in general . similarly , we set @xmath848 and @xmath849 . finally , in the sequel we extend the notation previously used for parabolic submonoids and parabolic subgroups : for every subset @xmath434 of @xmath8 , we denote by @xmath850 and @xmath851 the submonoid of @xmath8 and the subgroup of @xmath1 , respectively , generated by @xmath434 . below , we consider @xmath852 and @xmath853 , which are submonoids of @xmath854 . [ lemthgarsubg ] assume @xmath855 is a parabolic submonoid of @xmath8 , and denote by @xmath856 its garside element . then , + ( i ) the elements of @xmath838 belong to @xmath855 . + ( ii ) one has @xmath857 , and every atom of @xmath3 that appears in any decomposition of @xmath856 as a product of atoms in @xmath448 belongs to @xmath838 . + ( iii ) one has @xmath858 . write @xmath859 . then @xmath812 belongs to @xmath860 , and every atom of @xmath790 that appears in any decomposition of @xmath861 in @xmath794 as a product of atoms belongs to @xmath840 . \(i ) all the elements of @xmath843 left - divide @xmath856 , which lies in @xmath448 . then , by the defining properties of the @xmath2-structure , it follows that the elements of @xmath862 left divides @xmath856 . but @xmath855 is a parabolic submonoid , therefore all the elements of @xmath862 belong to @xmath855 . + ( ii ) the set @xmath852 is included in @xmath863 by @xmath418 . conversely , consider an element @xmath435 of @xmath8 which lies in @xmath863 and choose an arbitrary decomposition @xmath864 of @xmath435 in @xmath448 ( where @xmath825 ) . by definition of a @xmath2-structure , we can assume without restriction that each @xmath828 left - divides @xmath718 ; otherwise , if we consider @xmath865 that left - divides @xmath718 , then we have @xmath866 . the assumption that @xmath855 is a parabolic submonoid implies that all the @xmath828 belong to @xmath855 . therefore , we get step - by - step that all the @xmath827 are equal to @xmath434 . then @xmath435 belongs to @xmath852 . if we applied the above argument to @xmath867 we get the second part of @xmath175 . + ( iii ) the element @xmath809 belongs to @xmath677 and no element of @xmath434 can left - divides @xmath812 . indeed , if an @xmath136 in @xmath137 left - divides @xmath812 , then @xmath868 is a simple element of @xmath8 and therefore can not lie in @xmath479 . then , @xmath812 lies in @xmath860 ( _ cf . _ lemma [ lemstructconjnu]@xmath418 ) and the atoms of @xmath790 that appears in any decomposition of @xmath861 belongs to @xmath840 . [ thgarssubg ] let @xmath146 be a monoidal garside structure such that @xmath8 has a @xmath2-structure . denote by @xmath137 its atom set and by @xmath1 its associated garside group . let @xmath433 be a parabolic submonoid of @xmath8 such that @xmath869 . if the monoid @xmath855 is a parabolic submonoid of @xmath8 whose garside element is denoted by @xmath856 , then + ( i ) the subgroup @xmath870 is a standard parabolic subgroup of the groupoid @xmath4 with @xmath10 as unique object . in particular , @xmath870 is a garside group . the garside element of @xmath871 is @xmath856 and its atom set is @xmath838 . + ( ii ) write @xmath859 . the subgroup @xmath872 of @xmath1 is a ( standard ) parabolic subgroup of @xmath870 with @xmath812 for garside element , and @xmath873 @xmath418 the element @xmath856 belongs to @xmath874 and we have @xmath875_{\conjplusa } \subseteq [ 1_{a_x^{\scriptscriptstyle + } } , \delta]_{\conjplusa } \cap a^{\scriptscriptstyle + } _ { { \sh}(x ) } \subseteq \left(\delta\wedge a^{\scriptscriptstyle + } _ { \lsh(x)}\right)\cap { \conjplusa}_{a_x^{\scriptscriptstyle + } \to\cdot } \hfill\cr\hfill \subseteq \left(\delta_{{\lsh}(x)}\wedge a^{\scriptscriptstyle + } _ { \lsh(x)}\right)\cap { \conjplusa}_{a_x^{\scriptscriptstyle + } \to\cdot } \subseteq [ 1_{a_x^{\scriptscriptstyle + } } , \delta_{\lsh(x)}]_{\conjplusa}.}\ ] ] the first inclusion follows from lemma [ lemthgarsubg]@xmath175 , and the third one from the fact that @xmath433 be a parabolic submonoid of @xmath8 . by lemma [ lemthgarsubg]@xmath175 , we also get that @xmath874 is a sublattice of @xmath448 for the left - divisibility : let @xmath435 and @xmath90 belong to @xmath874 , then their gcd and their lcm for the left - divisibility lie in @xmath876 because the latter is a parabolic submonoid of @xmath8 . but , they also lie in @xmath448 . then , by lemma [ lemthgarsubg]@xmath175 , they lie in @xmath874 , and they are common divisor and common multiple , respectively , of @xmath435 and @xmath90 in @xmath876 . since we assume @xmath869 , by symmetry , @xmath874 is a sublattice of @xmath454 for the right - divisibility . + @xmath877 we have to prove that @xmath853 is a parabolic submonoid of @xmath874 , with @xmath812 as garside element . this is a consequence of lemma [ lemthgarsubg]@xmath421 and [ lemstructconjnu ] with a proof similar to @xmath175 . it is clear that @xmath853 normalizes @xmath878 and both subgroups generate @xmath870 . finally , we have a semi - direct product because the submonoids @xmath878 and @xmath853 are parabolic with a trivial intersection . indeed , the left greedy normal of every element in the intersection of the parabolic subgroups @xmath879 and @xmath872 has to be trivial . in order to state the next result , we recall that an artin - tits group of spherical type is said to be _ indecomposable _ when it is not the direct product of two of its proper standard parabolic subgroups . [ lastprop ] let @xmath146 be a monoidal garside structure such that @xmath8 is an artin - tit monoid of spherical type , with @xmath137 as set of atoms and @xmath1 as associated artin - tits group . assume that @xmath434 is a subset of @xmath137 . the subgroup @xmath870 is an artin - tits group of spherical type with @xmath838 as associated atom set . in particular , if @xmath449 are distinct in @xmath817 , the associated relation ( [ deregalite ] ) of the presentation of @xmath3 stated in theorem [ theorpreseconjplusa ] is @xmath880 for some positive integer @xmath881 . when @xmath882 is indecomposable with @xmath883 , the subgroups @xmath870 of @xmath1 where @xmath884 are classified by the _ tables of relations r2 _ of ( * ? ? ? * appendix 5 ) . by theorem [ thgarssubg ] , we know that @xmath870 is a garside group with @xmath838 as associated atom set . now , consider the decomposition @xmath885 of @xmath818 that starts with @xmath683 in @xmath733 ( _ cf . _ proposition [ antelastprop ] ) . by assumption , @xmath886 . by the argument of the proof of proposition [ antelastprop ] , we get @xmath887 , then @xmath888 , and step - by - step that @xmath889 . similarly @xmath890 . finally , @xmath891 because otherwise we could decompose @xmath892 as a product @xmath893 with @xmath894 , which is impossible . then @xmath870 is an artin - tits group of spherical type with @xmath838 as associated atom set . the proof of the remaining is a case by case computation , which is the same as in @xcite . theorem [ thgarssubg ] of the previous section claims that the group @xmath870 is a semi - direct product of two of its subgroups . in order to established that result , we use lemma [ lemstructconjnu ] @xmath418 . if we were dealing with groups instead of categories in the latter , we would have use the notion of a group semi - direct product in order to state it . the notion of a _ semi - direct product of categories _ is not so easy to defined , even though one can find a definition in @xcite , or more recently in @xcite . the notion of a _ zappa - szp product _ , as covered in @xcite , appears as more natural . * lemma 3.2 and 3.9 ) let @xmath6 be a category and @xmath895 be two subcategories of @xmath6 . we say that @xmath6 is an ( _ internal _ ) _ zappa - szp product _ of @xmath896 and @xmath897 if every morphism @xmath435 of @xmath6 can be uniquely decomposed as a product @xmath898 where @xmath505 is a morphism of @xmath896 and @xmath793 is a morphism of @xmath897 . in this case , we write @xmath899 . then , lemma [ lemstructconjnu]@xmath418 claims that @xmath900 where @xmath901 be the totally disconnected category whose objects are the parabolic submonoids of @xmath8 and whose vertex monoid at @xmath433 is @xmath902 . clearly , in our context , we have more than a zappa - szp product . let us introduce the following definition : let @xmath6 be a category and @xmath895 be two subcategories of @xmath6 such that @xmath899 . assume @xmath896 is equal to a totally disconnected category that is the union of its vertex monoids @xmath903 . we say that @xmath6 is a _ semi - direct product _ of @xmath896 and @xmath897 when for every @xmath435 in @xmath904 one has @xmath905 . in this case , we write @xmath906 . so , for instance one has @xmath907 . this definition extends the standard definition of a semi - direct product of monoids , and is a special case of the definition considered in @xcite . we finish this article with the following result which generalizes ( * ? ? ? * prop . 4 ) and provides a nice description of the connection between the standard parabolic subgroups of a garside group . [ theodescnormal]let @xmath146 be a monoidal garside structure such that @xmath8 has a @xmath2-structure . denote by @xmath1 the group of fractions of @xmath8 . + ( i ) let @xmath908 be the totally disconnected groupoid whose objects are the parabolic subgroups of @xmath1 and whose vertex group at @xmath10 is @xmath909 . then , @xmath910 ( ii ) let @xmath9 be the groupoid whose objects are the parabolic subgroups of @xmath1 , whose morphism set @xmath911 from @xmath10 to @xmath11 is @xmath12 , and whose composition is the product in @xmath1 . let @xmath13 be the totally disconnected groupoid whose objects are the parabolic subgroups of @xmath1 and whose vertex group at @xmath10 is @xmath10 . then , @xmath912 note that @xmath9 is the full subgroupoid of @xmath913 ( see definition [ defngc ] ) whose objects are the parabolic subgroups of @xmath1 . its vertex groups are normalizers of standard parabolic subgroups . this result holds in particular when @xmath1 is an artin - tits group of spherical type . @xmath418 the unique not obvious fact is that @xmath4 is equal to the zappa - szp product @xmath914 . this is a consequence of the zappa - szp product decomposition of @xmath3 , since @xmath4 , @xmath908 and @xmath14 are the groupoids of fractions of the garside categories @xmath3 , @xmath915 and @xmath790 , respectively ( where @xmath915 is the totally disconnected category whose objects are the parabolic submonoids of @xmath8 and whose vertex monoid at @xmath433 is @xmath916 ) . + @xmath175 again the unique not obvious fact is that conj(a ) = @xmath917 . the unicity of the decomposition follows easily from lemma [ lemstructconjnu]@xmath175 . let @xmath435 lie in @xmath1 such that @xmath918 . we show there exist @xmath359 in @xmath10 and @xmath919 in @xmath920 such that @xmath921 . there exists @xmath922 and @xmath505 in @xmath8 such that @xmath923 . furthermore , there exists a parabolic submonoid @xmath924 of @xmath8 such that @xmath591 belongs to @xmath925 . in particular , we have @xmath926 . we can write @xmath927 such that @xmath928 belongs to @xmath479 , @xmath929 belongs to @xmath707 , no elements of @xmath434 left - divides @xmath930 and no element of @xmath467 right - divides @xmath930 . now , let @xmath48 belongs to @xmath434 and @xmath931 be the left greedy normal form of @xmath932 in @xmath1 . there exists @xmath673 in @xmath8 such that @xmath933 and @xmath934 . since @xmath932 belongs to @xmath935 , the elements @xmath358 and @xmath168 belongs to @xmath707 . by the assumption on @xmath930 , this implies @xmath936 and @xmath937 . therefore @xmath938 is included in @xmath707 and by symmetry , @xmath939 . therefore , @xmath930 lies in @xmath481 and we have @xmath940 with @xmath941 in @xmath479 and @xmath942 in @xmath943 . using @xmath418 , we are done . garside categories , periodic loops and cyclic sets . . hermann , 1968 . on the zappa - szp product . ( 2005 ) 393424 normalizers of parabolic subgroups in coxeter groups . , 2 ( 1999 ) , 323351 . the language of geodesics for garside groups . ( 2004 ) 495509 . groupe de garside . ( 2002 ) 267306 . left - garside categories , self - distributivity , and braids . . aussian groups and garside groups , two generalisations of artin groups . ( 1999 ) 569604 . garside and locally garside categories . . phd thesis , universit de picardie - jules verne , 2001 . normalisateur et groupe dartin de type sphrique . ( 2003 ) 263274 . garside subgroups of garside groups parabolic subgroups of garside groups ( 2007 ) 116 .

many astrophysical observations indicate the existence of dark matter ( dm ) . a good example of such observations is the measurement of rotation curves for stars and gas in spiral galaxies . on the other hand , the nature of dm remains to be unveiled . in this regard , many dm candidates have been proposed with the proposal of weak interacting massive particles ( wimps ) @xcite the most popular among all candidates . the detections of dm shall test the idea of wimps . the detections of dm can be categorized into direct and indirect approaches . the former approach proceeds by observing the nucleus recoil as dm interacts with the target nuclei in the detector . the latter approach relies on detecting final state particles resulting from dm annihilations or decays . in this article , we focus on the indirect detection of dm through observing neutrinos produced by dm annihilations or decays in the galactic halo . the search for neutrinos coming from dm annihilations in galactic halo has been performed by icecube @xcite . data obtained from icecube 22-string configuration set the 90% c.l . upper limit on dm annihilation cross section @xmath4@xmath5s@xmath6 for @xmath7 channel at @xmath8 tev @xcite , while the preliminary result of icecube 40-string galactic - center analysis improves the above limit to @xmath9@xmath5s@xmath6 @xcite . based upon icecube 40-string data selection for searching diffuse flux of astrophysical muon neutrinos @xcite , a comparable constraint on the cross section of @xmath7 annihilations , which could take place in the core of earth , is derived @xcite for tev range dark - matter masses . furthermore , upper limits on the annihilation cross sections of @xmath10 and @xmath11 channels are also obtained . it is interesting to compare these upper limits with the required annihilation cross sections for the same channels for explaining the pamela data on positron fraction excess @xcite and fermi - lat @xmath12 fluxes measurement @xcite . as shown in ref . @xcite , the icecube upper limit on @xmath11 annihilation cross section is comparable to the required @xmath11 annihilation cross section for explaining pamela and fermi - lat data . on the other hand , the icecube upper limit on @xmath10 annihilation cross section is still too high to test the idea of using this mode to account for pamela and fermi - lat data . the icecube sensitivity on dm search is expected to improve with the data from all 86 strings analyzed . the analysis of deepcore array data will further enhance the sensitivity . the deepcore array @xcite is located in the deep center region of icecube detector . this array consists of @xmath13 densely instrumented strings plus the nearest standard icecube strings . the installation of deepcore array significantly improves the rejection of downward going atmospheric muons in icecube and lowers the threshold energy for detecting muon track or cascade events to about @xmath14 gev . this muon rejection is crucial for icecube to observe dm induced neutrino signature from galactic halo . in fact , it has been pointed out that the parameter range for @xmath10 channel preferred by pamela and fermi - lat data could be stringently constrained @xcite(see also discussions in ref . @xcite ) by the data from icecube detector augmented with deepcore array . in this work , we shall however focus on the low - mass dm instead of dm with its mass in the range preferred by pamela and fermi - lat . we note that previous analyses on deepcore sensitivity @xcite have set the threshold energy at @xmath15 gev for both track and cascade events . however , to take the full advantage of deepcore array , it is desirable to include neutrino events in the energy range @xmath16 . we have initiated such a study for track events @xcite . in this work , we generalize the previous study to cascade events in icecube deepcore detector . since we are only interested in low - mass dm , we only consider channels @xmath0 , @xmath1 , and @xmath2 for annihilations and channels @xmath3 and @xmath1 , and @xmath2 for decays . the neutrino fluxes generated through dm annihilations or decays into @xmath17 , @xmath18 and @xmath19 final states are not included in this analysis . this paper is organized as follows . in sec . ii we describe neutrino fluxes from dm annihilation / decay at the galactic halo for different halo profiles . in sec . iii we briefly describe our results on the background atmospheric neutrino fluxes taking into account neutrino oscillations . in sec . iv we present our results on the projected five year sensitivity of deepcore array on cascade events induced by dm annihilations and decays . in addition , we shall also compare our results with the up - to - date indirect detection limits from fermi - lat gamma ray data and cosmic microwave background ( cmb ) observations . specifically , we shall compare the deepcore sensitivities on dm annihilation cross section @xmath20 and dm decay time with corresponding constraints obtained from gamma ray observations @xcite(see also analysis in ref . @xcite ) and those obtained from cmb anisotropy @xcite based on one or both of the recent wmap 7-year @xcite and act 2008 @xcite data . discussions and conclusions are given in sec . the differential neutrino flux from the galactic dark matter halo for neutrino flavor @xmath21 can be written as @xcite @xmath22 for the case of annihilating dm , and @xmath23 for the case of decaying dm , where @xmath24 is the distance from the galactic center ( gc ) to the solar system , @xmath25 is the dm density in the solar neighborhood , @xmath26 is the dm mass , @xmath27 is the dm decay time and @xmath28 is the neutrino spectrum per annihilation or decay for a given annihilation or decay channel @xmath29 with a corresponding branching fraction @xmath30 . for @xmath31 channel , we assume that two neutrinos are produced per annihilated dm pair and all neutrino flavors are equally populated . thus the neutrino spectrum per flavor is a monochromatic line with @xmath32 . on the other hand , the neutrino spectrum per flavor for @xmath33 channel is @xmath34 . the neutrino spectra @xmath28 for other channels are summarized in refs @xcite . the quantity @xmath20 is the thermally averaged annihilation cross section , which can be written as @xmath35 with a boost factor @xmath36 @xcite . we set @xmath37 , which is the typical annihilation cross section for the present dark matter abundance under the standard thermal relic scenario @xcite . we treat the boost factor @xmath36 as a phenomenological parameter . the dimensionless quantity @xmath38 is the dm distribution integrated over the line - of - sight ( l.o.s ) and averaged over a solid angle @xmath39 , i.e. , @xmath40 + where @xmath41 is the dm density at a specific location described by the coordinate @xmath42 with @xmath43 the distance from the earth to dm and @xmath44 the direction of dm viewed from the earth with @xmath45 corresponding to the direction of gc . the distance @xmath46 is the distance from gc to dm . we note that the above definition of @xmath47 include the constant factors @xmath48 , @xmath49 and @xmath50 in the denominator . this definition differs from that adopted in papers by icecube and fermi - lat collaborations ( see , for example , refs . @xcite and @xcite ) where the above - mentioned constant factors are not included . we computed the values of @xmath38 with darksusy @xcite . for the galactic dm distribution , we consider navarro - frenk - white ( nfw ) @xcite , einasto @xcite and isothermal profiles @xcite . the functional forms of these profiles are given by @xmath51 @xmath52\right\ } , \,\alpha = 0.17,\end{aligned}\ ] ] @xmath53 with values of @xmath54 and @xmath55 given in table i. in all cases we impose the normalization @xmath25 , which is at @xmath56 kpc . ccc dm halo model & @xmath57 in kpc & @xmath58 in @xmath59 + + nfw & 20.0 & 0.260 + einasto & 21.5 & 0.0538 + isothermal & 3.50 & 2.07 + the comparison of different dm halo models is shown in fig . 1 . neutrinos are significantly mixed through oscillations when they travel a vast distance across the galaxy . the observed flavor ratio of dm induced neutrinos is related to the flavor ratio at the source through the probability matrix @xmath60 @xcite . in particular , the exact form of @xmath61 in terms of mixing angles @xmath62 and cp phase @xmath63 is given in ref . the mixing angles @xmath64 and @xmath65 has been well measured while newest results for @xmath66 from accelerator @xcite and reactor experiments @xcite are also available . we take @xmath67 , @xmath68 , @xmath69 and @xmath70 , which are best fit values of neutrino mixing parameters from a recent global fitting @xcite in the case of normal mass hierarchy . therefore , @xmath71 where @xmath72 and @xmath73 are neutrino fluxes at the source and on the earth respectively . while the best fit neutrino mixing parameters differ slightly in the case of inverted mass hierarchy , they do not produce noticeable change in the above relation . we follow the approaches in @xcite to compute the intrinsic atmospheric neutrino background fluxes . the @xmath74 flux arising from @xmath75 decays reads @xmath76 where @xmath77 is the neutrino energy , @xmath78 is the slant depth in units of g/@xmath79 , @xmath80 is the zenith angle in the direction of incident cosmic - ray nucleons , @xmath81 , @xmath82 is the pion decay length in units of g/@xmath79 , @xmath83 is the nucleon interaction length while @xmath84 is the corresponding nucleon attenuation length , and @xmath85 is the primary cosmic - ray spectrum . we have @xmath86 with @xmath87 the atomic number of each nucleus . the spectrum of each cosmic - ray component is parametrized by @xcite @xmath88)^{- \alpha},\end{aligned}\ ] ] in units of m@xmath89s@xmath6sr@xmath6gev@xmath6 . the fitting parameters @xmath90 depend on the type of nucleus . they are tabulated in ref . the function @xmath91 is the probability that a charged pion produced at the slant depth @xmath92 survives to the depth @xmath78 ( @xmath93 ) @xcite . @xmath94 is the normalized inclusive cross section for @xmath95 , and is given in ref . the kaon contribution to the atmospheric @xmath74 flux has the same form as eq . ( [ atm - nu ] ) with an inclusion of the branching ratio @xmath96 and appropriate replacements in kinematic factors as well as in the normalized inclusive cross section . the three - body muon decay contribution to the atmospheric @xmath74 flux is also included . the details are discussed in ref . @xcite . after summing the two - body and three - body decay contributions , we obtain the total intrinsic atmospheric muon neutrino flux . from the fig . 1 of ref . @xcite , we note that angle - averaged atmospheric muon neutrino flux obtained by our calculation and that obtained by honda et al . @xcite both agree well with amanda - ii results @xcite . , as a function of distance from the galactic center as described by the nfw , einasto and isothermal halo models.,title="fig:",scaledwidth=85.0% ] 0.5 cm the intrinsic atmospheric @xmath97 flux due to @xmath98 decays can be obtained by solving cascade equations @xcite . one obtains @xmath99 where @xmath100 , @xmath101 and @xmath102 are the z moments defined in our previous work @xcite . finally , the atmospheric @xmath74 flux taking into account the neutrino oscillation effect is given by @xmath103,\end{aligned}\ ] ] where @xmath104 is the @xmath105 oscillation probability . sub - leading contributions to atmospheric @xmath74 flux arising from @xmath106 and @xmath107 oscillations are not included in the above equation . we can write down the atmospheric @xmath97 flux in the similar way . in icecube deepcore , the track event rate for contained muons is given by @xmath108 while the cascade event rate is given by @xmath109 where @xmath110 is the density of ice , @xmath111 is avogadro s number , @xmath112 is the effective volume of icecube deepcore array for muon track events @xcite and @xmath113 is that for cascade events @xcite , @xmath114 is the neutrino flux arrived at icecube , which is the sum of dm induced flux and the background atmospheric neutrino flux , @xmath115 is taken as @xmath26 for annihilation and @xmath116 for decay , @xmath117 and @xmath118 are the threshold energies for track events and cascade events respectively , @xmath119 is the differential cross section of neutrino - nucleon charged - current scattering , and @xmath120 is the differential cross section for showers produced by neutrino - nucleon charged - current and neutral - current scatterings . in this work , we use differential cross sections @xmath121 and @xmath120 given by ref . @xcite with cteq6 parton distribution functions . the atmospheric part of @xmath122 is taken from ref . we also set @xmath123 gev . it should be noted that the value for @xmath124 is an average effective volume based on the energy dependent @xmath124 discussed in ref . @xcite , while the value for @xmath125 is just the instrumental volume of the deepcore detector . the updated effective volumes of deepcore detector for track and cascade events are available in ref . while we adopt constant effective volumes for evaluating deepcore sensitivities , we shall also estimate how much the updated effective volumes affect our results . , @xmath126 , @xmath127 , and @xmath128 channels , respectively . we adopt nfw profile for obtaining the results in this figure and those in the subsequent figures except fig . 4 . , title="fig:",scaledwidth=85.0% ] 0.5 cm as mentioned earlier , we consider neutrino fluxes generated through the annihilation channels @xmath129 , @xmath1 and @xmath130 , and the decay channels @xmath131 , @xmath1 and @xmath130 . by computing the cascade event rates , we present in fig . 2 the required dm annihilation cross section as a function of @xmath26 such that the neutrino signature from dm annihilations can be detected at the @xmath132 significance in five years . the @xmath132 statistical significance is defined as @xmath133 where @xmath134 and @xmath135 are numbers of signal and background events , respectively . in fig . 2 , we take nfw profile for dm density distribution in the galactic halo , the shower threshold energy @xmath136 and the cone half - angle @xmath137 . non - detection of dm neutrino signature would then exclude the parameter region above the curve at the @xmath132 level . we have presented results corresponding to different annihilation channels . it is seen that the required annihilation cross section for @xmath132 detection significance is smallest for @xmath138 channel and largest for the channel @xmath139 . at this juncture , it is desirable to estimate effects of the updated effective volume on our sensitivity calculations . the energy dependent effective volume @xmath140 as given in ref . @xcite is roughly 3 times smaller than the value @xmath141 adopted in our calculation for @xmath142 gev . on the other hand , @xmath140 increases monotonically with energy with @xmath140 greater than @xmath141 for @xmath143 gev . for the annihilation process @xmath7 , one has @xmath144 . hence the deepcore sensitivity to this process should be better than that presented in fig . 2 , which starts from @xmath145 gev . for @xmath10 mode , @xmath146 gev corresponds to @xmath147 gev . at this energy , the value @xmath148 overestimates the effective volume by about a factor of 2 , thus the deepcore sensitivity to @xmath149 should be corrected by a factor of @xmath150 . the correction factor gradually reduces to @xmath151 as @xmath26 approaches to about @xmath152 gev , which corresponds to @xmath153 gev . for @xmath154 gev , the deepcore sensitivity calculated with @xmath140 is better than that presented in fig . 2 . the correction factor for @xmath11 is similar to that for @xmath10 . the hadronic mode @xmath155 requires different correction factor for the same @xmath26 . however we shall not address such a correction here since deepcore detector is relatively insensitive to @xmath155 . as a function of @xmath26 such that the cascade events induced by neutrinos from dm annihilations can be detected at the @xmath132 significance in five years . results corresponding to different @xmath156 are presented . for comparison , we also show the result with @xmath157 gev and @xmath158 @xcite . , title="fig:",scaledwidth=85.0% ] 0.5 cm next , we show how the deepcore sensitivity on dm annihilation cross section varies with the chosen cone half - angle and threshold energy for nfw dm density profile . here we take @xmath10 channel for illustration . 3 shows the required dm annihilation cross section @xmath149 for a @xmath132 detection in five years for different cone half - angle @xmath159 . it is seen that the sensitivity is improved as @xmath159 increases from @xmath160 to @xmath161 while it turns weaker as @xmath159 increases further . in the latter case , the signal increases slower than the background does . we should point out that the choice of @xmath159 depends on the angular resolution of the experiment . the current angular resolution in icecube for cascade events is @xmath162 . however , an improvement on such a resolution is expected @xcite . in fact , new reconstruction method which can achieve a @xmath163 angular resolution for cascade events in large - scale neutrino telescopes has been proposed @xcite . hence results shown in fig . 3 with @xmath164 can be realized in the near future . in this figure , we also show the result for a higher threshold energy @xmath165 with a cone half - angle @xmath158 for comparison . this result is taken from ref . @xcite where @xmath158 is identified as the most optimal cone half - angle for constraining dm annihilation cross section at that threshold energy . for large @xmath26 , lowering @xmath166 from @xmath167 gev to @xmath168 gev does not affect much the signal rate while increases significantly the atmospheric background event rate . hence , the sensitivity on dm annihilation cross section becomes worse by choosing @xmath136 . one expects the situation turns opposite for @xmath26 approaching to the threshold energy . in fact , for @xmath169 gev , one can see that the sensitivity on dm annihilation cross section obtained with @xmath136 is always better than that obtained with @xmath170 . we note that the deepcore sensitivities on other annihilation channels have similar cone half - angle and threshold energy dependencies . from the above discussions , we observe that , in contrast to the main concern of this article , raising the shower threshold energy gains sensitivity for probing heavier dm . in fact , as the threshold energy approaches to @xmath171 gev , one enters into the operative energy range of the full icecube array such that the effective volume of the detector increases more rapidly with the threshold energy than the case with deepcore detector alone @xcite . the same situation holds for probing dm annihilation with track events . for einasto @xcite , isothermal @xciteand nfw @xcite dm density profiles.,title="fig:",scaledwidth=85.0% ] 0.5 cm after discussing how the deepcore sensitivities on dm annihilation cross section vary with the chosen cone half - angle and threshold energy for nfw profile , we study the variation of deepcore sensitivities on @xmath10 with @xmath159 for different dm density profiles . first , we present in fig . 4 our expected deepcore sensitivities corresponding to different @xmath159 for einasto dm density profile . similar to nfw profile , einasto dm density distribution also has a cusp in the central dm region . we refer to this class of dm density profiles as the cusped profile . therefore the deepcore sensitivity on @xmath149 with einasto profile becomes poorer as @xmath159 increases . in addition , we can see from fig . 1 that the dm density of einasto profile is higher than that of nfw profile between 0.1 kpc and 5 kpc from the gc , and both density profiles are almost identical beyond this range of distances . hence the deepcore sensitivity on @xmath149 with einasto profile is better than that with nfw profile for @xmath137 . next , we also present in fig . 4 the expected deepcore sensitivities for different @xmath159 with isothermal dm density profile . because there is a core in the central dm density distribution of isothermal profile , we refer to this class of profile as the cored profile . we note that the sensitivity to dm annihilation cross section with isothermal profile improves as @xmath159 increases from @xmath160 to @xmath162 while the sensitivity with einasto profile behaves oppositely . this is because that the dm density distribution of isothermal profile maintains flat for a much longer distance from the gc as compared to the cusped profile . however , as @xmath159 increases further , the deepcore sensitivity to dm annihilation cross section becomes poorer even for isothermal profile since the factor @xmath172 is proportional to the square of dm density . finally , deepcore sensitivities to dm annihilation cross section with isothermal profile are poorer than those with cusped profiles for the same @xmath159 . this results from the fact that dm densities of cusped profiles around gc are much larger than that of isothermal profile in the same region . as a function of @xmath26 such that the neutrino signature from dm annihilations can be detected at the @xmath132 significance in five years for track and cascade events.,title="fig:",scaledwidth=85.0% ] 0.5 cm having discussed the effect of dm density profiles on the derived deepcore sensitivities , we present in fig . 5 the comparison of deepcore sensitivities to @xmath149 obtained by measuring cascade events and track events , respectively . we note that nfw profile is adopted for results in this figure as well as results presented in the remaining figures of this article . we point out that @xmath74 is the dominant flavor of atmospheric neutrinos above few tens of gev and the neutrino - nucleon cross section are almost the same for all flavors . therefore , comparing with the result of track events , the signal to background ratio is enhanced in cascade events because @xmath74 only produces the cascade events through the neutral - current interaction , which is lower in cross section than that of charged - current interaction . hence cascade events in general provide better sensitivities to dm annihilation cross section than those provided by track events . we like to point out that the comparison between cascade and track events in fig . 5 is based upon the current angular resolution of icecube detector . if the angular resolution for cascade events is improved in the future , the advantage of measuring cascade events would be even more significant as one can see from fig . the dot - dashed line is the @xmath173 upper bound on annihilation cross section for low - mass wimp @xcite obtained by wmap7 data . the dot - dot - dashed line is the cmb constraint obtained by using wmap7 + act data at @xmath174 c.l . the dashed line is the dsphs constraint at @xmath174 c.l . the dotted line is the constraint due to fermi - lat observations on the region @xmath175 plus a @xmath176 square region centered at gc , assuming the nfw profile @xcite . the thick and thin solid lines are the expected @xmath132 sensitivities of deepcore detector with @xmath177 and @xmath178 , respectively.,title="fig:",scaledwidth=85.0% ] 0.5 cm it is interesting and essential to compare our results with constraints obtained from gamma - ray astronomy and cosmology . fermi large area telescope ( fermi - lat ) @xcite is a pair - conversion telescope that explores the gamma - ray sky in the 20 mev to 300 gev range with unprecedented sensitivity . in a recent work , fermi - lat collaboration derive constraints on wimp annihilation or decay into various final states which produce a continuous photon spectrum @xcite . these constraints are based upon the measured inclusive photon intensity spectrum from 4.8 gev to 264 gev obtained from two years of fermi - lat data over the region @xmath175 plus a @xmath176 square region centered at gc with point sources removed . in fig . 6 , the dotted line is cross section upper limit on dm annihilation channel @xmath10 from the diffuse gamma - ray spectrum with nfw profile . it is taken from ref . @xcite with a rescaling factor @xmath179 applied since we have adopted a local density of @xmath25 while fermi - lat analysis uses @xmath180 . our expected @xmath132 sensitivities on @xmath10 annihilation cross section by the deepcore detector with @xmath137 and @xmath158 are plotted for comparison . we can see that our expected @xmath132 sensitivity with @xmath137 is slightly stronger than this fermi - lat constraint , even if we calculate the sensitivity with @xmath140 given by ref . furthermore , the expected @xmath132 sensitivity with @xmath158 in the deepcore detector is almost an order of magnitude stronger than this fermi - lat constraint for small @xmath26 . . the dashed line is the dsphs constraint at @xmath174 c.l . the dotted line is the constraint due to fermi - lat observations for the region @xmath175 plus a @xmath176 square region centered at the gc , assuming the nfw profile @xcite . the solid line is the expected @xmath132 sensitivity by deepcore detector with @xmath181 and @xmath158 , which is weaker than both fermi - lat constraints . the deepcore sensitivity with @xmath137 is not shown since it is even less competitive.,title="fig:",scaledwidth=85.0% ] 0.5 cm constraints on dm annihilation cross section were also obtained from cosmology and gamma - ray observations on dwarf spheroidal satellite galaxies ( dsphs ) of the milky way . energy injection from dm annihilation at redshift @xmath182 affects the cmb anisotropy . this is because the injected energy can ionize the thermal gas and modify the standard recombination history of the universe . the updated cmb constraints on dm annihilation cross sections are derived in refs . @xcite and @xcite , where the former work is based on the recent wmap 7-year data @xcite while the latter one combines wmap 7-year and act 2008 @xcite data . we present the above two cmb constraints on @xmath10 annihilation cross section in fig . one can see that the thick solid line is higher than the dot - dot - dashed line ( @xmath174 c.l . ) by roughly a factor of 2 . thus the expected sensitivity of icecube deepcore detector with @xmath137 is slightly weaker than the cmb constraint . however , the deepcore sensitivity is comparable to the cmb constraint with @xmath158 . it should be noted that the dm annihilation cross section could be velocity dependent . hence , a model dependent extrapolation on dm annihilation cross section might be required to compare the constraint on @xmath183 at redshift @xmath182 to that at the present day universe @xcite . , @xmath184 , @xmath185 , and @xmath186 , respectively.,title="fig:",scaledwidth=85.0% ] 0.5 cm as a function of @xmath26 such that the neutrino signature from dm decays can be detected at the @xmath132 significance in five years for cascade events . results corresponding to different @xmath156 are presented . for comparison , we also show the result with @xmath157 gev and @xmath137 @xcite.,title="fig:",scaledwidth=85.0% ] 0.5 cm dwarf spheroidal satellite galaxies of the milky way are dm - dominated systems which do not have active star formation or detected gas content @xcite . satellite galaxies are among the best targets to search for dm signals in gamma rays because of small background from astrophysical sources and a favorable signal to noise ratio . in ref . @xcite , fermi - lat collaboration derive upper limits on dm annihilation cross sections by applying a joint likelihood analysis to 24 months of data from 10 satellite galaxies with uncertainties on the dark matter distributions in the satellite galaxies taken into account . we present the dsphs constraint on @xmath10 annihilation cross section at @xmath174 c.l . in fig . 6 . for @xmath187 , the expected deepcore sensitivity curves are below that set by dsphs constraint for both @xmath137 and @xmath158 . for @xmath188 gev , the dsphs constraint is comparable to the expected deepcore sensitivity with @xmath158 . thus it is stronger than the expected deepcore sensitivity with @xmath137 in this dm mass range . as a function of @xmath26 such that the neutrino signature from dm decays can be detected at the @xmath132 significance in five years for track and cascade events.,title="fig:",scaledwidth=85.0% ] 0.5 cm we also present fermi - lat s dsphs constraint and galactic gamma ray constraint on @xmath155 annihilation cross section in fig . 7 . by comparing figs . 6 and 7 , we note that fermi - lat data gives more stringent constraint on @xmath155 mode than its constraint on @xmath10 mode . this is in contrast to the deepcore case as one can see from fig . 2 . while the neutrino spectrum through @xmath10 is harder than that through @xmath155 , the gamma ray spectra through the above annihilations behave differently . the gamma ray spectrum from @xmath155 dominates over that from @xmath10 for most of the range of @xmath189 @xcite . the expected @xmath132 sensitivity on @xmath155 channel by the deepcore detector with @xmath158 is also shown in fig . one can see that the expected deepcore sensitivity on this channel is weaker than all existing constraints presented here . and @xmath190 channels due to fermi observations for the region @xmath175 plus a @xmath176 at gc , assuming the nfw profile @xcite . the solid line and dotted line are our expected @xmath132 sensitivities for @xmath191 and @xmath190 channels with @xmath181 and @xmath137 . , title="fig:",scaledwidth=85.0% ] 0.5 cm in addition to studying deepcore sensitivities on dm annihilation channels , we also study sensitivities on dm decay time for @xmath3 , @xmath1 and @xmath130 channels . figure 8 shows the required dm decay time for reaching @xmath132 detection significance in five years on cascade events from each channel with the threshold energy @xmath181 and a cone half - angle @xmath137 . non - detection of such a signature would then exclude the parameter region below the curve at the @xmath132 level . we shall see later that the deepcore sensitivity to dm decay time is not improved by considering @xmath159 smaller than @xmath162 . comparing various dm decay modes , one can see that @xmath186 channel requires the lowest decay rate ( longest decay time ) to avoid a @xmath132 detection significance in five years of deepcore data taking . we note that the energy - dependent effective volume @xmath140 @xcite downgrades the sensitivity to @xmath192 decay time by roughly a factor of 1.5 at @xmath193 gev , which corresponds to @xmath194 gev . however , the sensitivity to @xmath192 obtained with @xmath140 is better than that presented in fig . 8 for @xmath195 gev , which corresponds to @xmath196 gev . for @xmath191 mode , the sensitivity at @xmath193 gev is lower by a factor of @xmath197 by applying @xmath140 , since such a @xmath26 corresponds to lowest possible neutrino energy @xmath142 gev . however , @xmath140 shall enhance the sensitivity for @xmath198 gev , which corresponds to @xmath199 gev , assuming @xmath200 . the correction factor for @xmath201 is similar to that for @xmath191 , while the hadronic mode @xmath190 requires different correction factor for the same @xmath26 . once more , we do not address such a correction as the deepcore detector is relatively insensitive to @xmath190 . next , we show how the deepcore sensitivity on dm decay time varies with the chosen cone half - angle and threshold energy . we use @xmath191 channel to illustrate these effects . in fig . 9 , we present the required dm decay time @xmath202 as a function of @xmath26 for reaching @xmath132 detection significance in five years for different cone half - angle @xmath159 . the sensitivity curve rises as @xmath159 increases from @xmath160 to @xmath162 . as @xmath159 increases in this cone half - angle range , the dm event rate increases faster than that of atmospheric background . however , the sensitivity is not further improved by increasing @xmath159 from @xmath162 to @xmath203 . we also show the required dm decay time for a @xmath132 detection significance in five years with @xmath204 and @xmath205 for comparison . it has been pointed out in ref . @xcite that @xmath205 gives rise to the highest sensitivity on dm decay time for @xmath170 . we note that the sensitivity on dm decay time is improved by lowering @xmath166 from @xmath167 gev to @xmath168 gev for @xmath206 gev . it is of interest to compare sensitivities on dm decay time given by cascade events and track events . in fig . 10 , one can see that cascade events provide better sensitivity on dm decay time than that given by track events for the same threshold energy @xmath207 and @xmath159 . in the same figure we also show the sensitivity given by track events for @xmath208 . such a @xmath209 renders the best sensitivity for track events . however this sensitivity is still poorer than those given by cascade events . in fig . 11 , we present decay time lower limits for @xmath191 and @xmath190 channels obtained from the diffuse gamma - ray spectrum for the region @xmath175 plus a @xmath176 square region centered at gc , assuming the nfw profile . they are taken from ref . @xcite , with a rescaling factor @xmath210 applied since we have adopted a local density of @xmath25 while fermi - lat analysis uses @xmath180 . in the same figure , we also show the expected @xmath132 sensitivities to @xmath191 and @xmath190 decay time by the deepcore detector with @xmath137 for comparisons . for the deepcore detector , the sensitivity to the @xmath191 decay time is better than that to the @xmath190 decay time , since the neutrino spectrum in the former channel is harder than the one in the latter channel . on the other hand , the fermi - lat data in general gives more stringent constraint on @xmath190 decay time than its constraint on the decay time of @xmath191 . this is because that the gamma ray spectrum from the former channel dominates over the one from the latter channel for most of the range of @xmath189 , as the dm annihilation case . if dm decays predominantly into @xmath1 , one can see that the decay time sensitivity expected at deepcore is comparable to the constraint given by fermi - lat data . this conclusion is not altered by adopting the energy - dependent effective volume @xmath140 @xcite . on the other hand , if dm decays predominantly to @xmath211 , the expected deepcore sensitivity is much poorer than the constraint from fermi - lat data . in this paper , we have evaluated sensitivities of icecube deepcore detector to neutrino cascade events induced by dm annihilations and decays in galactic halo . we focus on the scenario of small dm mass and the threshold energy for the cascade events is taken to be @xmath168 gev . the event rate of background atmospheric neutrinos is calculated with @xmath212 oscillations taken into account for neutrino energies less than @xmath213 gev . the signal event rate is calculated by taking nfw profile for dm density distribution in the galactic halo . among all dm annihilation and decay channels , the annihilation mode @xmath7 and the decay mode @xmath192 provide the best search sensitivity , while the search sensitivity provided by the annihilation mode @xmath155 and the decay mode @xmath214 is the poorest . it is important to compare the expected sensitivities of deepcore detector to dm annihilation cross section and decay time with the existing constraints on the same quantities obtained by cosmology and fermi - lat gamma - ray observations . it is seen that the fermi - lat constraints on @xmath155 and @xmath214 are much stronger than the expected deepcore sensitivities to the same channels . hence , if dm predominantly annihilates or decays into @xmath211 , the deepcore detector is not expected to observe neutrino signature induced by dm in the galactic halo . for leptonic final states , one can see from fig . 6 that the expected deepcore sensitivity to @xmath10 annihilation cross section is stronger than the fermi - lat constraint on the same channel based upon gamma - ray data from the galactic halo . however , the former sensitivity with the current angular resolution of cascade events is slightly weaker than both the dsphs constraint and the constraint obtained from wmap and act results on cmb anisotropy . on the other hand , the deepcore sensitivity can be improved with an improved angular resolution for cascade events . in fig . 11 , we also see that the expected deepcore sensitivity to @xmath191 decay time is comparable to the fermi - lat constraint based upon gamma - ray data from the galactic halo . from the above comparisons , there remain slight possibilities to observe dm induced neutrino signature from the galactic halo provided dm annihilates or decays predominantly into leptons . it should be noted that fermi - lat and cmb data does not directly set limits on @xmath7 or @xmath192 modes with monochromatic neutrinos . there exist models @xcite in which @xmath7 and @xmath192 are dominant annihilation and decay modes , respectively . in the annihilation case @xcite for example , the dm candidate can be the lightest right - handed ( rh ) sneutrino in a @xmath215 extension of the minimal supersymmetric standard model . rh sneutrinos annihilate into a pair of rh neutrinos . each of these rh neutrinos then decays into ordinary left - handed ( lh ) neutrino and a neutral higgs boson while the decay of rh neutrino into charged final states @xmath216 is typically forbidden in such a model . in the case that the mass difference between rh sneutrinos and rh neutrinos is small , rh neutrinos are produced non - relativistically by dm annihilations . hence lh neutrinos produced by the decays of rh neutrinos are approximately monochromatic with an energy around half of the dm mass . therefore the sensitivity of icecube deepcore to this type of models can be read off from the @xmath7 curve in fig . 2 with the shift @xmath217 . before closing , we comment on the detection of neutrino signature induced by dm in the galactic halo with neutrino telescopes in the northern hemisphere . generally a neutrino telescope in such a location has advantages in detecting track - like neutrino events . first , the telescope s effective volume for upward going track events is enhanced by the muon range . this effect is particularly significant for energetic track events originated from high - energy muon neutrinos . second , a neutrino telescope in the northern hemisphere naturally suppresses atmospheric muon background while icecube needs to use optical modules located at the outer region for vetoing the same background @xcite . following these arguments , it is interesting to see if a detector array in the northern hemisphere with the size of deepcore detector has a significantly better sensitivity than the current deepcore detector surrounded by icecube strings . since the latter also has a good veto capability , the former can gain only in the effective volume expected to be enhanced by the muon range . we note that the deepcore array aims at detecting neutrinos in the energy range @xmath218 . muons induced by muon neutrinos in this energy range only travel around @xmath167 m for @xmath219 gev and @xmath220 m for @xmath221 gev . this does not significantly enhance the detector s effective volume in most cases since the height of the deepcore detector is already around @xmath222 m. therefore , given the existence of icecube detector augmented by deepcore in the south pole , it is clear that neutrino telescopes in the northern hemisphere only have advantages in detecting neutrino signature from heavier dm . in this regard , the acceptance for track events in km3net as a function of neutrino energy has been estimated @xcite . with this acceptance , we also calculate the sensitivity of km3net to dm annihilation cross section @xmath149 in the galactic halo . for @xmath223 and @xmath224 gev , the km3net @xmath132 sensitivity to @xmath149 in 5 years is comparable to the cmb constraint on this channel @xcite . however , for @xmath8 tev , cmb constraint gives @xmath149 no greater than @xmath9 @xmath5s@xmath6 whereas the km3net @xmath132 sensitivity on the same annihilation channel can reach to @xmath225@xmath5s@xmath6 in 5 years . in conclusion , we have made detailed comparisons between icecube deepcore sensitivities and other existing constraints on various dm annihilation and decay channels . the prospect for deepcore detector to observe neutrino signature induced by dm in the galactic halo has been discussed . we have also mentioned the expected performance of km3net on this observation . we thank y .- h . lin for useful discussions and c. rott for bringing ref . @xcite to our attention . f.f.l . is supported by the grant from research and development office , national chiao - tung university , g.l.l . is supported by national science council of taiwan under grant no . 99 - 2112-m-009 - 005-my3 , y.s.t . is funded in part by the welcome program of the foundation for polish science . 99 g. jungman , m. kamionkowski , and k. griest , phys . rep . * 267 * , 195 ( 1996 ) ; l. bergstrom , rep . . phys . * 63 * , 793 ( 2000 ) . g. bertone , d. hooper , and j. silk , phys . rep . * 405 * , 279 ( 2005 ) . a. achterberg , _ et al . _ [ icecube collaboration ] , astropart . phys . * 26 * , 155 ( 2006 ) . r. abbasi _ et al . _ [ icecube collaboration ] , phys . d * 84 * , 022004 ( 2011 ) [ arxiv:1101.3349 [ astro-ph.he ] ] . r. abbasi _ et al . _ [ icecube collaboration ] , arxiv:1111.2738 [ astro-ph.he ] . r. abbasi _ et al . _ [ icecube collaboration ] , phys . d * 84 * , 082001 ( 2011 ) [ arxiv:1104.5187 [ astro-ph.he ] ] . i. f. m. albuquerque , l. j. beraldo e silva and c. perez de los heros , phys . rev . d * 85 * , 123539 ( 2012 ) . o. adriani _ et al . _ [ pamela collaboration ] , phys . lett . * 106 * , 201101 ( 2011 ) [ arxiv:1103.2880 [ astro-ph.he ] ] . m. ackermann _ et al . _ [ fermi lat collaboration ] , phys . rev . d * 82 * , 092004 ( 2010 ) [ arxiv:1008.3999 [ astro-ph.he ] ] . e. resconi ( icecube collaboration ) , nucl . methods phys . res . , sect . a * 602 * , 7 ( 2009 ) . c. wiebusch , for the icecube collaboration , [ arxiv:0907.2263 [ astro-ph.im ] ] . t. deyoung [ icecube collaboration ] , arxiv:1112.1053 [ astro-ph.he ] . d. spolyar , m. r. buckley , k. freese , d. hooper and h. murayama , arxiv:0905.4764 [ astro-ph.co ] . m. r. buckley , d. spolyar , k. freese , d. hooper and h. murayama , phys . rev . d * 81 * , 016006 ( 2010 ) . s. k. mandal , m. r. buckley , k. freese , d. spolyar and h. murayama , phys . d * 81 * , 043508 ( 2010 ) . l. covi , m. grefe , a. ibarra and d. tran , jcap * 1004 * , 017 ( 2010 ) . a. e. erkoca , m. h. reno , and i. sarcevic , phys . rev . d * 82 * , 113006 ( 2010 ) . f. f. lee and g. l. lin , phys . d * 85 * , 023529 ( 2012 ) . m. ackermann _ et al . _ ( fermi - lat collaboration ) , arxiv:1205.2739 [ astro - ph ] . m. ackermann _ et al . _ ( fermi - lat collaboration ) , phys . . lett . * 107 * , 241302 ( 2011 ) . m. cirelli , p. panci and p. d. serpico , nucl . b * 840 * , 284 ( 2010 ) . g. hutsi , j. chluba , a. hektor and m. raidal , astron . astrophys . * 535 * , a26 ( 2011 ) . s. galli , f. iocco , g. bertone and a. melchiorri , phys . d * 84 * , 027302 ( 2011 ) . e. komatsu _ et al . _ , astrophys . ser . * 192 * , 18 ( 2011 ) . j. w. fowler _ et al . _ , astrophys . j. * 722 * , 1148 ( 2010 ) . j. hisano , m. kawasaki , k. kohri , and k. nakayama , phys . d * 79 * , 043516 ( 2009 ) . a. e. erkoca , g. gelmini , m. h. reno , and i. sarcevic , phys . d * 81 * , 096007 ( 2010 ) . a. sommerfeld , ann . ( leipzig ) * 403 * , 257 ( 1931 ) ; j. hisano , s. matsumoto and m.m . nojiri , phys . lett . * 92 * , 031303 ( 2004 ) ; j.m . russell , s.m . west , d. cumberbatch and d. hooper , jhep * 0807 * , 058 ( 2008 ) ; n. arkani - hamed , d.p . finkbeiner , t.r . slatyer and n. weiner , phys . d * 79 * , 015014 ( 2009 ) ; i. cholis , g. dobler , d.p . finkbeiner , l. goodenough and n. weiner , phys . rev . d * 80 * , 123518 ( 2009 ) ; j.m . russell and s.m . west , phys . b * 676 * , 133 ( 2009 ) ; s.m . koushiappas and m. kamionkowski , phys . lett . * 103 * , 121301 ( 2009 ) ; m. kamionkowski , s.m . koushiappas and m. kuhlen , phys . d * 81 * , 043532 ( 2010 ) ; m. lindner , a. merle and v. niro , arxiv:1005.3116 ; d. suematsu , t. toma and t. yoshida , phys . d * 82 * , 013012 ( 2010 ) ; s. hannestad and t. tram , arxiv:1008.1511 , j.l . feng , m. kaplinghat and h .- b . yu , arxiv:1005.4678 ; c. arina , f .- x . josse - michaux and n. sahu , phys . b * 691 * , 219 ( 2010 ) . gondolo , j. edsjo , p. ullio , l. bergstrom , m. schelke and e. a. baltz , jcap * 0407 * , 008 ( 2004 ) . j. f. navarro , c. s. frenk and s. d. m. white , astrophys . j. * 462 * , 563 ( 1996 ) . a. w. graham , d. merritt , b. moore , j. diemand and b. terzic , astrophys . j. * 132 * , 2685 ( 2006 ) . j. f. navarro _ et al . _ , arxiv : astro - ph/0810.1522 . j. diemand , m. kuhlen , p. madau , m. zemp , b. moore , d. potter and j. stadel , nature ( london ) 454 , 735 ( 2008 ) j. n. bahcall and r. m. soneira , astrophys . j. suppl . * 44 * , 73 ( 1980 ) . j. g. learned and s. pakvasa , astropart . * 3 * , 267 ( 1995 ) . h. athar , m. jezabek and o. yasuda , phys . d * 62 * , 103007 ( 2000 ) ; l. bento , p. keranen and j. maalampi , phys . b * 476 * , 205 ( 2000 ) . see also k. c. lai , g. l. lin and t. c. liu , phys . d * 80 * , 103005 ( 2009 ) . k. abe _ et al . _ [ t2k collaboration ] , phys . lett . * 107 * , 041801 ( 2011 ) . y. abe _ et al . _ [ double - chooz collaboration ] , phys . * 108 * , 131801 ( 2012 ) [ arxiv:1112.6353 [ hep - ex ] ] . f. p. an _ et al . _ [ daya - bay collaboration ] , arxiv:1203.1669 [ hep - ex ] . j. k. ahn _ [ reno collaboration ] , arxiv:1204.0626 [ hep - ex ] . g. l. fogli , e. lisi , a. marrone , d. montanino , a. palazzo and a. m. rotunno , phys . d * 86 * , 013012 ( 2012 ) . t. k. gaisser , astropart . * 16 * , 285 ( 2002 ) . f. f. lee and g. l. lin , astropart . phys . * 25 * , 64 ( 2006 ) . t. k. gaisser and m. honda , ann . nucl . part . * 52 * , 153 ( 2002 ) . m. honda , t. kajita , k. kasahara , and s. midorikawa , phys . d * 70 * , 043008 ( 2004 ) . p. lipari , astropart . * 1 * , 195 ( 1993 ) . m. honda , t. kajita , k. kasahara , s. midorikawa , and t. sanuki , phys d * 75 * , 043006 ( 2007 ) . r. abbasi _ et al . _ , astropart . phys . * 34 * , 48 ( 2010 ) . t. k. gaisser , _ cosmic rays and particle physics _ ( cambridge university press , cambridge , uk , 1992 ) . a. b. mcdonald , c. spiering , s. .schonert , e. t. kearns and t. kajita , rev . instrum . * 75 * , 293 ( 2004 ) . r. gandhi , c. quigg , m. h. reno , and i. sarcevic , astropart . phys . * 5 * , 81 ( 1996 ) ; phys . d * 58 * , 093009 ( 1998 ) . r. abbasi _ et al . _ [ icecube collaboration ] , astropart . phys . * 35 * , 615 ( 2012 ) . e. middell , j. mccartin and m. dagostino , _ proceedings of the 31st icrc , lodz 2009_. r. auer , nucl . instrum . methods phys . res . , sect . a * 602 * , 84 ( 2009 ) . w. b. atwood _ et al . _ ( fermi - lat collaboration ) , astrophys . j. * 697 * 1071 ( 2009 ) . for discussions on this issue , see , for example , j. hisano _ et al . d * 83 * , 123511 ( 2011 ) . m. mateo , ann . astron . astrophys . * 36 * , 435 ( 1998 ) . j. grcevich and m. e. putman , astrophys . j. * 696 * , 385 ( 2009 ) . the spectral shapes of gamma ray and neutrinos from dm annihilations are summarized in m. cirelli , g. corcella , a. hektor , g. hutsi , m. kadastik , p. panci , m. raidal and f. sala _ et al . _ , jcap * 1103 * , 051 ( 2011 ) , where original references are listed . r. allahverdi , s. bornhauser , b. dutta and k. richardson - mcdaniel , phys . d * 80 * , 055026 ( 2009 ) . a. falkowski , j. juknevich and j. shelton , arxiv:0908.1790 [ hep - ph ] . r. abbasi _ et al . _ [ icecube collaboration ] , arxiv:1210.3557 [ hep - ex ] . technical design report , km3net collaboration , http://www.km3net.org/tdr/tdrkm3net.pdf .

when dissolved in aqueous solvent , lipids self - assemble into an amazing variety of different structures as a function of concentration and temperature @xcite . most prominent is the lamellar phase , which consists of a stack of lipid bilayers separated by layers of water . in fact , the lipid bilayer constitutes the basic building block of biological membranes ; in eucaryotic cells , it not only envelopes the cell itself , but also separates its nucleus and its organelles from the cytosol , and builds up the endoplasmic reticulum , the golgi apparatus and the uni - lamellar vesicles which serve as reaction chambers and transport vehicles . the predominance of the lamellar phase at ambient temperatures derives from the fact that in contrast to surfactants , which usually have large headgroups and form micelles , lipids have rather bulky hydrocarbon chains . a so - called _ bilayer lipid _ has a chain region cross section which is of similar size as the area per head group ; thus the spontaneous curvature is small and the lipid forms a lamellar phase . if temperature is increased , the size of the head groups hydration shell decreases , spontaneous curvature increases and the lipid monolayers tend to curve towards the water regions . since the tendency of lipid monolayers to curve is frustrated in the lamellar phase , it is often found to transform into a inverse hexagonal phase which consists of a two - dimensional array of inverse cylindrical micelles . spontaneous curvature can also be increased by changing molecular architecture , by adding lipids with bulkier chains or by increasing salt concentration ( since salt screens electrostatic repulsion between charged headgroups ) @xcite . lipids with spontaneous curvature are often called _ non - bilayer lipids_. for large spontaneous curvature , a cubic phase of inverse spherical micelles is sometimes observed @xcite . the inverse phases have been denoted _ type ii ( water - in - oil ) phases _ in contrast to _ type i ( oil - in - water ) phases _ which dominate surfactant phase behavior @xcite . although there are some lipids which also form type i phases , we focus here on type ii phases since these are the ones formed by the majority of biologically relevant lipids . one of the intriguing aspects of the polymorphism of lipid - water mixtures is the existence of yet another structural type , which often occurs between the lamellar and the inverse hexagonal phase . like the lamellar phase , these _ inverse bicontinuous cubic phases _ ( ibcps ) consist of lipid bilayers . however , now a _ lipid bilayer extends throughout the whole of space , and divides it into two disconnected but interpenetrating labyrinths of cubic symmetry , which both are filled with water . thus the structure can also be considered to be built from water channels ( that is inverse cylindrical micelles ) as is the inverse hexagonal phase , only that now the channels meet in vertices with a coordination number related to the topology of the cubic phase . although the ibcps unite the structural elements of the lamellar and the inverse hexagonal phases , in contrast to them they are optically isotropic and highly viscous . like the inverse hexagonal phase , the ibcps are stabilized by spontaneous curvature . therefore they are formed by non - bilayer lipids and their stability can be controlled by changing the lipids spontaneous curvature . the generic occurrence of ibcps in lipid - water mixtures and their property of dividing space into two interwoven aqueous compartments lends itself readily to speculations on their biological relevance , e.g. on their possible occurrence in the endoplasmic reticulum or the golgi apparatus @xcite . a recent survey of a vast number of tem pictures suggests that ibcps are indeed ubiquitous in biological systems @xcite . the abundance of non - bilayer lipids in biological systems is a longstanding puzzle , and several studies have explored the possible relationship between lipid polymorphism and biological membrane function ( for reviews see refs . @xcite ) . in particular , it was shown that certain cells like _ e. coli _ regulate the relative amount of bilayer and non - bilayer lipid in order to control membrane morphology and function of membrane proteins . _ e. coli _ mutants lacking the non - bilayer lipid phosphatidylethanolamine were found to be severely impaired in regard to protein transport across the plasma membrane @xcite . luzzati and coworkers have pointed out that if any ibcp had a biological function , it should be stable at physiological temperatures and in excess water @xcite . ibcps are also relevant for biotechnological applications , e.g. they have been used recently as artificial matrix which enables membrane proteins to crystallize in a three - dimensional array @xcite . since ibcps consist of a lipid bilayer separating two labyrinths of water , is it reasonable to assume that the two monolayers are arranged in a way which is _ locally _ symmetrical about their mid - surface . in differential geometry , a surface which curves to both sides in the same way is known as a _ minimal surface _ , i.e. a surface for which the mean curvature vanishes at each of its points . therefore the mid - surfaces of the ibcp s lipid bilayers are often modeled as _ triply periodic minimal surfaces _ ( tpms , sometimes also denoted ipms for _ infinite periodic minimal surfaces _ ) of cubic symmetry @xcite . indeed it has been confirmed by a thorough analysis of electron density maps derived from x - ray data that the mid - surfaces of these structures can be very close to minimal surfaces @xcite . note that the stronger requirement of the monolayers _ global _ symmetry about the mid - surface leads to a certain subclass of cubic tpms which is called _ balanced _ since then the two water labyrinths have to be congruent to each other . since the seminal work of schoen @xcite , it is known that there exists a large variety of different cubic tpms . before schoen s work , only p , d and c(p ) were known from the 19th century work of schwarz and his students ; then schoen described g , f - rd , i - wp , o , c - to and c(d ) . more cubic tpms have been found later by karcher and polthier @xcite , fischer and koch @xcite and others , but most of them seem to be too complicated as to be of physical relevance . in figs . [ figureparallelsurfaces1 ] and [ figureparallelsurfaces2 ] we depict surfaces corresponding to the seven different tpms treated in this work . until now , only g , d and p have been established to be realized as ibcp in lipid - water mixtures @xcite . one system for which all three of these phases is stable is 2:1 lauric acid ( la)/dilauroyl phosphatidylcholine ( dlpc ) and water ( with p coexisting with excess water ) @xcite . the only other system for which g , d and p are known to be stable is didodecylphosphatidylethanolamine and water @xcite ; however , here g exists with hardly no hydration at all , and the regions for d and p are very difficult to access experimentally , thus this system should not be compared with the theoretical results for fully hydrated phases presented here . in all other systems with ibcps , only one or two of them are stable ( for example , for monoolein and water , g and d are stable , with d coexisting with excess water @xcite ) . in this work we present a model which can explain why only g , d and p are observed in lipid - water mixtures and why their actual occurrence seems to depend on the specific system under investigation . the phase diagram predicted by our model agrees well with the experimental phase diagram for 2:1 la / dlpc and water . exact ( _ weierstrass _ ) representations are only known for g , d , p and i - wp @xcite . however , it has been shown in recent studies @xcite that these and other cubic tpms can be generated as isosurfaces of a density field , which minimizes the free - energy functional of a simple ginzburg - landau model for amphiphilic systems @xcite . using the fourier representations of ref . @xcite , it is now possible to investigate also the physical properties of the ibcps , for whose minimal mid - surfaces no exact representations are known . in particular , the question of whether these structures might correspond to stable ibcps in lipid - water mixtures can now be addressed quantitatively . in this work we study g , d , p , i - wp , s , c(p ) and f - rd , of which g , d , p , s and c(p ) are balanced . the most important contribution to the free energy of amphiphilic interfaces is the curvature energy , which dominates all other contributions if the radii of curvature are large compared with molecular length scales @xcite . for the inverse phases in lipid - water mixtures , it is generally assumed that another important contribution to the free energy is the stretching energy of the hydrocarbon chains @xcite . in order to attain optimal free energy , the neutral surfaces of the two monolayers should realize at the same time constant mean curvature @xmath0 and constant distance @xmath1 to the minimal mid - surface . since this is not possible geometrically , the free energy will always be frustrated @xcite . the relative importance of interface ( bending ) and bulk ( stretching ) contributions in self - assembled interfaces is a long - debated issue . for amphiphilic systems , for which the free energy contributions come mainly from the interfaces , the main problem is of a mathematical nature : the overall solution can be investigated only in a model which allows the interfacial positions to adjust themselves freely . in contrast , in diblock copolymer systems bulk contributions are much more important and models based on random polymer coils and interacting monomer densities are more adequate . self - consistent mean - field theory for diblock - copolymer systems suggests that in this case , interfacial tension and stretching are equally important @xcite . recently , this framework has been adapted to treat lipid - water mixtures @xcite . this approach is complementary to the membrane approach ; it is most useful for the investigation of systems , in which the internal structure of the bilayer is not uniform . however , self - consistent field theory for large molecules is not expected to describe amphiphilic systems well at large lattice constants , when the membranes internal structure is essentially preserved and the main contributions to the free energy depends on their shape . in particular , only in the membrane approach does it becomes possible to address the role of thermally activated membrane undulations . in order to study interface - dominated systems , it has been suggested in ref . @xcite to consider the two classes of surfaces , which can completely relax one of the two relevant contributions . _ parallel surfaces _ to the minimal midsurface realize the optimal chain length @xmath1 and are frustrated only in their bending energy , while _ constant - mean - curvature surfaces _ ( cmc - surfaces ) realize the given spontaneous curvature , but are frustrated in their stretching energy . the mathematical properties of parallel surfaces are well known in terms of the properties of the underlying minimal surface ; the parallel - surface model has therefore been the main tool for the investigation of ibcps @xcite . up to now , the main difficulty of this approach has been to take into account the variation of gaussian curvature over the minimal mid - surface . in earlier work , the curvature energy was therefore expanded in moments of the distribution of the gaussian curvature . in ref . @xcite this was done up to eighth order for d , and in ref . @xcite up to second order for p , d and g. cmc - surfaces @xcite have the advantage that they offer an unified picture for all relevant phases lamellar , inverse bicontinuous cubic , inverse hexagonal and inverse micellar cubic phases are modelled by planes , cubic cmc - surfaces , cylinders and spheres , respectively . however , since no exact representations are known , the stretching frustration for ibcps has so far been evaluated numerically for d only @xcite . we will show below that in the case of short chains ( or large lattice constants ) , stretching energy is prohibitively large , and the two monolayers of a ibcp in lipid - water systems can be modeled by parallel surfaces . in the case of long chains ( or small lattice constants ) , we expect stretching and bending to be equally important ; this agrees with the results for diblock copolymers , which are in the long chain limit . our reasoning implies that cmc - surfaces are not a good representation of ibcps in lipid - water systems for physical reasons . however , for small distances to the minimal mid - surface , their mathematical properties are very similar to the ones of parallel surfaces @xcite , thus many results from the parallel surface model will carry over to the cmc - model . in general , however , the parallel surface model is more appropriate both for physical and mathematical reasons . moreover , there is strong experimental evidence for it , both from a detailed reconstruction of electron densities @xcite and from analysis of swelling data @xcite . in contrast to earlier work , we evaluate the curvature energy in the parallel surface model without any expansion and to high numerical accuracy by first calculating the distribution of gaussian curvature over the minimal mid - surface . for g , d , p and i - wp this can be done from their weierstrass - representations . we also consider the structures s , c(p ) and f - rd by calculating the distribution from the representations which we obtained from ginzburg - landau theory . our main result is that the free - energy densities of the structures g , d and p are considerably lower than those of the other investigated structures due to their narrow distribution of gaussian curvature over the minimal midsurface . we show that this result persists when thermal membrane undulations and higher order terms in the bending energy are considered , and argue that potential additional contributions to the free energy are unlikely to change it . this explains why only g , d and p have been observed in lipid - water mixtures . in fact , the calculated phase diagram agrees nicely with the experimental one for 2:1 la / dlpc and water , for which these three phases coexist . we show that due to the existence of a bonnet transformation between g , d and p , these phases coexist along a triple line in our model . simultaneously , p coexists with an excess excess water due to a mechanism called _ emulsification failure_. the bonnet transformation also implies that the free - energy densities of g , d and p scale as a function of concentration with a universal geometrical quantity , which we term _ topology index_. since the topology index decreases from g to d to p , the gyroid g is most prominent , followed by smaller regions of stability for d and p at higher water concentrations . any additional contribution to the free energy , which introduces a new length scale , is expected to change the delicate balance between these three phases ; this includes stretching contributions , van der waals and electrostatic interactions . since such additional effects are specific to a given experimental system , this result explains qualitatively why usually only one or two of these phases are observed in lipid - water mixtures . the paper proceeds as follows . in sec . [ sec : model ] , we specify the free - energy expression which has to be evaluated for each type of minimal mid - surface , and discuss some immediate consequences of the model . the calculation of the distribution of gaussian curvature is explained in sec . [ sec : weierstrass ] for p , d , g and i - wp from their weierstrass representation . results for the gaussian - curvature distributions for s , c(p ) and f - rd are obtained in sec . [ sec : ginzburglandau ] from the representations derived recently from a simple ginzburg - landau model . these results are combined in sec . [ sec : phasediagrams ] to predict phase behavior and to numerically calculate phase diagrams . in sec . [ sec : symmetry_breaking ] , we discuss which physical mechanisms will break the bonnet symmetry between p , d and g. finally , we compare our results with experiments and conclude in sec . [ sec : discussion ] . the bending energy of one lipid monolayer is described by the _ canham - helfrich hamiltonian _ @xcite @xmath2 where the integration extends over the monolayer s neutral surface , which is at a distance @xmath1 from the minimal midsurface of the bilayer . in the framework of the curvature model , bending and stretching contributions decouple at the neutral surface @xcite ; this allows to disregard area - stretching contributions . the neutral surface has been shown experimentally to be located close to the polar - apolar interface @xcite . in the curvature model ( [ helfrichhamiltonian ] ) , the neutral surface is characterized by mean curvature @xmath3 , gaussian curvature @xmath4 and differential area element @xmath5 ; the superscript @xmath1 indicates that the neutral surface is a parallel surface . the model parameters are the spontaneous curvature @xmath0 , the bending rigidity @xmath6 and the saddle - splay modulus @xmath7 of the monolayers . for lipids , the bending rigidity @xmath6 for a monolayer is of the order @xmath8 . usually the saddle - splay modulus @xmath7 is assumed to have a small negative value ; in fact the curvature model without any additional constraints is well - defined only for @xmath9 @xcite . the spontaneous curvature @xmath0 can be considered to result from the mismatch between optimal head group area , volume incompressibility of the chains and optimal chain length . for non - ionic lipids , it depends mainly on temperature , which changes the head group hydration and therefore optimal head group area . since lipids typically form lamellar and inverse phases , where the monolayers prefer to bend towards their polar sides , we define positive curvature to be towards the water regions . we assume that spontaneous curvature scales with temperature as @xmath10 , since such a linear relationship seems to hold very well for ternary surfactant systems @xcite . the stretching energy can be assumed to be harmonic about the average chain length @xmath1 , so that @xmath11 where @xmath12 is the local chain length . the stretching modulus @xmath13 is itself a function of the average chain length @xmath1 . the relevant scaling law has to be obtained from a microscopic model ; here , we model the lipid monolayer as a gaussian polymer brush grafted to the polar - apolar interface , but more realistic models give similar results @xcite . the stretching energy per chain reads @xmath14 , where @xmath15 is the ( microscopic ) kuhn length and @xmath16 the number of kuhn segments . enforcing the volume constraint @xmath17 ( where @xmath18 is the grafting density ) , and transforming from energy per chain to energy per area , results in a potential of the form with @xmath19 . thus the stretching modulus @xmath13 scales inversely with average chain length @xmath1 stretching becomes very difficult for short chains @xcite . although the bending energy is well known not to depend on microscopic details , for our purpose it is important to note that the bending rigidity @xmath6 also depends on the average chain length @xmath1 . starting from the stretching energy of and taking into account how the chain volume varies with curvature , results in the bending energy of with @xmath20 @xcite . the result that the bending rigidity @xmath6 scales with the third power of film thickness @xmath1 is well known also from the elasticity of thin solid sheets . we now can estimate the relative importance of bending and stretching contributions . it is known that for both cmc - surfaces and parallel surfaces , the normalized standard deviations of @xmath12 and @xmath3 , respectively , are almost independent of hydrocarbon volume fraction @xcite , so that @xmath21 and @xmath22 , where @xmath23 is the lattice constant . therefore @xmath24 . thus , for small chain length @xmath1 ( large lattice constant @xmath23 ) , the stretching term is much larger and the chains can be assumed to be of constant length . this corresponds to the well - known result that the curvature energy is the relevant contribution to the free energy as long as the curvature radii ( that is the lattice constants ) are much larger than any molecular length @xcite . only for large chain length @xmath1 ( small lattice constant @xmath23 ) , the bending energy becomes comparable and both terms have to be considered . experimentally , this corresponds to the fact that for small water content ( i.e. chain length @xmath1 of comparable size as lattice constant @xmath23 ) , the bicontinuous cubic phases become unstable with respect to the hexagonal phase , whose geometrical properties require larger chain stretching . note that the relative importance of stretching and bending is asymmetric : since the chain length @xmath1 can not be larger than the lattice constant @xmath23 , there is no regime in which bending can be neglected . thus cmc - surface will never be a good approximation for monolayers in lipid - water systems . however , the stretching frustration can be relaxed by swelling the bilayer with oil ; therefore cmc - surfaces are good approximations for amphiphilic monolayers in ternary systems with oil , water and amphiphile @xcite . similar considerations of the relative importance of stretching and bending can be found also in the context of large membrane inclusions ( like transmembrane proteins ) where chains have to stretch in order to decrease the hydrophobic mismatch at the inclusion boundary @xcite . for the rest of this paper , we proceed within the framework of the parallel surface model . we assume that the interfaces between the tail and head group regions of the two lipid monolayers in a icbp are located at distances @xmath25 away from a given cubic tpms ; typically @xmath1 has values between one and two nanometers . note that for a tpms the mean curvature satisfies @xmath26 everywhere ( where @xmath27 and @xmath28 are the two principal curvatures ) , while the gaussian curvature @xmath29 varies as function of the position and is only restricted to satisfy @xmath30 . the differential area element @xmath31 , the mean curvature @xmath32 and the gaussian curvature @xmath33 on the parallel surfaces then follow as functions of @xmath1 and the quantities @xmath5 and @xmath4 on the minimal mid - surface @xcite , with @xmath34 note that these formulae are special cases of steiner s theorem of integral geometry ; they are therefore no approximations for small @xmath1 , but exact . since minimal surfaces have @xmath30 , positive mean curvature is defined here to correspond to positive @xmath1 . the effective bending energy of the lipid bilayer is only a function of the gaussian curvature @xmath4 of the minimal midsurface since its mean curvature @xmath3 vanishes . in an expansion for small chain length @xmath1 , it can be expressed as an infinite series in powers of @xmath4 . using and yields up to order @xmath35 @xmath36 therefore the effective saddle - splay modulus for the lipid bilayer , @xmath37 , is not just the sum of the monolayer s moduli , but is corrected to higher positive values due to the presence of the spontaneous curvature @xmath0 @xcite . we conclude that as long as @xmath38 , the ibcps are favored over the lamellar phase , since the preferred curvature of the monolayers translates into a topological advantage of saddle - type bilayer structures . the third term in favors saddle - type structures in any case , and also removes the degeneracy which minimal surfaces might experience under deformations which preserve the minimal surface property . this has been noted in a seminal work by bruinsma @xcite ; in his notation , we have @xmath39 for the elastic modulus of the @xmath40-term . the volume fraction occupied by the hydrocarbon can be calculated as @xmath41 where we have used the gauss - bonnet theorem @xmath42 . @xmath43 and @xmath44 denote surface area and euler characteristic of the given cubic tpms in the conventional unit cell with unit lattice constant . for the following , it is useful to introduce the _ topology index _ @xmath45 . this quantity is independent of scaling and choice of unit cell and characterizes a given tpms s topology in an universal way ; the higher its value , the smaller is the porosity and the larger the specific surface area of a tpms . it is the only quantity which characterizes a two - dimensional minimal surface in three - dimensional space independently of lattice constant and choice of unit cell , and its relevance for the description of bicontinuous cubic phases in amphiphilic systems has been discussed before @xcite . the topology index of the various tpms is of order @xmath46 and is given in tab . [ tableindex ] . the gyroid g is expected to have the highest value since it divides space into labyrinths with threefold coordinated vertices . s has a similar value for @xmath47 , since it features a mixture of three- and fourfold coordinated vertices . then comes the double diamond structure with fourfold coordinated vertices , while all other structures have higher values . in tab . [ tableindex ] we collect the values of @xmath43 , @xmath44 and @xmath47 for the cubic tpms investigated in this work as obtained from our numerical representations in the framework of a simple ginzburg - landau theory ( see below ) . exact values are known for g , d , i - wp , p and c(p ) and given in tab . [ tablenumbers ] . comparing numerical and exact values shows that for most structures , our numerical results are quite accurate . due to numerical limitations for the more complicated structures , the values for s and f - rd are less precise . in particular , we expect s to have a smaller value for @xmath47 than g. equation ( [ hydrocarbonvolume ] ) can be inverted numerically to give the dimensionless function @xmath48 , the lattice constant @xmath23 in units of the chain length @xmath1 , as a function of hydrocarbon volume @xmath49 . for small @xmath49 we find @xmath50 in fig . [ figurelatticeconstant ] we plot the first - order approximation @xmath51 of for all structures considered ; it can be hardly distinguished from the full curve over the full range of @xmath49 . by visual inspection of our numerical representations ( see below ) , we find that for all structures considered , parallel surfaces do not self - intersect as long as @xmath52 ; ibcps can accommodate large amounts of hydrocarbon since their geometry is so close to that of the lamellar phase . a typical value for @xmath1 is @xmath53 , so that for @xmath54 we have @xmath55 , @xmath56 , @xmath57 , @xmath58 , @xmath59 , @xmath60 and @xmath61 . note that at a given @xmath49 , the ratio of the lattice constants of two coexisting ibcps is simply the ratio between their scaled surface areas @xmath43 . in figs . [ figureparallelsurfaces1 ] and [ figureparallelsurfaces2 ] we depict for each of the structures considered one of the two monolayers for the hydrocarbon volume fraction @xmath54 . then the volume fraction of each of the two labyrinths is @xmath62 . we are now in a situation to rewrite the curvature energy from for our purpose . we consider a typical experimental situation , in which temperature is controlled at constant volume . as explained above , changing temperature @xmath63 amounts to changing spontaneous curvature @xmath0 . under the assumption of incompressibility of both components , there is only one degree of freedom for composition , which we take to be the hydrocarbon volume @xmath49 ( which is equivalent to the amount of interfacial area ) . dividing the free energy by the constant volume , we arrive at the free energy per _ unit volume _ as a function of @xmath49 . however , @xmath49 is not a controlled quantity , since at two phase coexistence , it is not clear how it will distribute into the two phases . the controlled quantity can be considered to be the chemical potential coupled to @xmath49 ( that is the derivative of free energy or free energy volume density for hydrocarbon volume fraction or interfacial area ) , and therefore phase transitions have to be calculated by using the maxwell construction for the free energy density as a function of @xmath49 ( which amounts to requiring the same chemical potential for both phases ) . we want to remark parenthetically that the free energy per _ unit area _ can not be used to calculate phase coexistences , since an exchange of area , controlled by the chemical potential of the lipid , has to be allowed between coexisting phases . we write the free energy volume density dimensionless by using a factor @xmath64 . for the lamellar phase the hydrocarbon volume fraction @xmath65 , thus its dimensionless free - energy density follows from to be @xmath66 the larger @xmath49 , the smaller @xmath23 and the more frustrated curvature energy per volume accumulates . for ibcps , we again use and in order to write the dimensionless free - energy density as a function of hydrocarbon volume fraction @xmath49 and the properties of the bilayers minimal mid - surface : @xmath67 where again we have used the gauss - bonnet theorem and defined @xmath68 . as mentioned above , the curvature model without any additional constraints is well - defined for @xmath69 and bicontinuous phases are favored for @xmath70 . @xmath43 , @xmath71 and @xmath44 refer to a conventional unit cell . using to lowest order , @xmath51 , and defining @xmath72 , we rewrite as @xmath73 in eq . ( [ freeenergydensitybicont2 ] ) and below , spontaneous curvature @xmath0 is measured in units of @xmath74 . since for the minimal midsurface both @xmath71 and @xmath44 are negative , @xmath75 is a positive quantity which varies over the surface . note that @xmath76 due to the gauss - bonnet theorem . the free - energy density @xmath77 of the lamellar phase , eq . ( [ freeenergydensitylamellar ] ) , is a special case of @xmath78 , with @xmath79 . in this work , we will treat the free - energy density without any further approximations . previously , similar expressions have been expanded in small @xmath49 ( which is equivalent to small @xmath1 ) @xcite , similarly as discussed in . in fact , each term @xmath80 in an expansion for small @xmath1 corresponds to a term @xmath81 in an expansion for small hydrocarbon volume @xmath49 . since @xmath82 , higher - order terms in would result in a free - energy density which has a similar structure as . as we will see below , this means that our results will remain valid in the case that higher - order terms are included in the curvature energy of the monolayers . it is well known that thermal fluctuations can contribute significantly to the free - energy of amphiphilic systems . since we consider the regime where the stretching energy dominates , undulations are the favored fluctuation modes . for the lamellar phase , they give rise to steric repulsion @xcite , which in our case can be written as @xmath83 where @xmath84 from monte carlo simulations @xcite and field - theoretical calculations @xcite . for an ibcp , steric repulsion is hardly relevant since the lateral correlation length @xmath85 is of the order of the lattice constant @xmath23 . therefore , the perpendicular correlation length @xmath86 , and few membrane collisions should occur . however , fluctuations of the lipid bilayer lead to a renormalization of @xmath6 and @xmath7 @xcite for all phases . since @xmath6 multiplies the average mean curvature squared , which vanishes for the mid - plane , only the renormalization of @xmath87 at length scale @xmath88 has to be taken into account . we identify the typical length scale of a cubic structure with @xmath89 . this implies @xmath90 , so that @xmath91 in eq . ( [ freeenergydensitybicont2 ] ) gets renormalized , with @xmath92\ .\ ] ] since the renormalization increases @xmath7 , this effect increases the topological advantage of the saddle - shaped structures as does the spontaneous curvature . note that for the lamellar phase , these effects might favor the formation of wormholes @xcite . here , we neglect this aspect , as well as the acoustic modes of the bicontinuous structures at large wavelengths , which have been discussed by bruinsma @xcite . note also that the thermal fluctuations also give a contribution to the free energy , which is linear in the membrane area . we have omitted this term here , since it can be absorbed into the chemical potential of the lipid . having defined the model , we now turn to its phase behavior . we first note that the dimensionless mean curvature averaged over the parallel surface follows from as @xmath93 where again @xmath94 and the gauss - bonnet theorem has been used . it is plotted in fig . [ figuregeometry1]a . note that all curves fall on a universal curve when the hydrocarbon volume fraction @xmath49 is scaled with the topology index @xmath47 . it reaches the value @xmath0 at @xmath95 thus , @xmath0 is realized in a sequence given by increasing topology index ; in particular , the gyroid structure g reaches @xmath0 at the highest value of @xmath49 . we conclude that if all considered structures were stable , they would appear in the sequence g - s - d- i - wp - p etc with increasing water concentration . the stability of the different phases is determined by the free - energy density given in . the second ( topological ) term is easy to understand . the topological properties of the parallel and mid - surfaces are the same : for negative saddle - splay modulus @xmath7 ( @xmath96 ) , the different ibcps are favored according to their values of the topology index @xmath47 . here the g - structure performs best since it has the lowest porosity . the first term in is more complicated , since the mean - curvature properties of the parallel surfaces translate into the gaussian - curvature properties of the minimal mid - surface in a quite complicated way . as discussed above , the spontaneous curvature of the monolayers implies a topological advantage of the bicontinuous phases . for a more detailed analysis , which includes all terms of the expansion , we note that the first term in measures the standard deviation of the parallel surface s mean curvature from the spontaneous curvature . since it follows from that the distribution of mean curvature over the parallel surface , @xmath97 , is proportional to the distribution of gaussian curvature over the minimal surface , @xmath98 , we expect that those structures will be more favorable which have small standard deviations for their gaussian curvature distributions @xcite . in order to quantify this concept , we now turn to the distributions of gaussian curvature which later will allow us to evaluate without any further approximations . weierstrass representations are known for p , d and g @xcite and i - wp @xcite . for each of these tpms , a fundamental domain can be identified , so that the rest of the surface follows by replicating it with the appropriate space group symmetries ( @xmath99 , @xmath100 , @xmath101 and @xmath99 , respectively ) . the weierstrass representation is a conformal mapping of certain complicated regions within the complex plane onto the fundamental domain : @xmath102 where @xmath103 are the internal ( and conformal ) coordinates of the minimal surface . the geometrical properties of such a surface follow as @xmath104 with @xmath105 . obviously the ( isolated ) poles of @xmath106 correspond to the flat points ( @xmath107 ) of the minimal surface . only few choices of @xmath106 yield embedded minimal surfaces . the ones for d and p have been known since the 19th century from the work of schwarz : for d it is @xmath108 . p corresponds to the same region and follows simply by the bonnet transformation @xmath109 with @xmath110 . equation ( [ weierstrassdarstellung2 ] ) implies that p and d have the same metric and the same distribution of gaussian curvature . however , since they map differently into embedding space , they have different space groups and lattice constants . the gyroid g was discovered in 1970 by schoen @xcite as another bonnet transformation of d , with @xmath111 . the weierstrass representation for i - wp was found only recently @xcite . if one of its poles is chosen to be at infinity , one has @xmath112 . in fig . [ figureweierstrass ] we show the fundamental domains in the complex plane of the different weierstrass representations . in tab . [ tablenumbers ] we collect the values for scaled surface area @xmath43 and lattice constant @xmath23 as they follow from the weierstrass representations . we also give the number @xmath16 of replications of the fundamental domain needed to build up the surface in one conventional unit cell and the values for the euler characteristic @xmath44 which follows from the topology generated in this process . note that for d one can choose another conventional unit cell which is contained @xmath113 times in the one chosen here . this has been done in refs . @xcite ; then @xmath44 , @xmath114 and @xmath4 scale as @xmath115 , @xmath116 and @xmath117 , respectively . in particular , one then has @xmath118 and @xmath119 . our choice is motivated by the fact that it is the relevant one also for ternary systems . we define the distribution function of gaussian curvature , @xmath120 over some surface parametrized by internal coordinates @xmath103 , where in our case @xmath121 and @xmath122 are determined by . for the generating functions @xmath123 detailed above , it is not possible to calculate @xmath124 analytically . therefore , we evaluate it numerically by considering a histogram @xmath125 for a set of discrete gaussian curvature values @xmath126 spaced equidistantly with @xmath127 . we cover the @xmath103-region corresponding to the fundamental domain with a square grid of @xmath128 points . for each point , we calculate @xmath5 and @xmath4 from and add @xmath5 to the @xmath129 which corresponds to the @xmath130 with @xmath131 $ ] . in order to obtain results for one conventional unit cell , the values @xmath132 are then multiplied by the number @xmath16 of replications and the values @xmath126 are scaled with @xmath133 ( both @xmath16 and @xmath23 are given in tab . [ tablenumbers ] ) . in practice we use @xmath134 and @xmath135 ; the latter value amounts to about @xmath136 bins for the histogram . by this procedure , we have @xmath137 the moments @xmath138 of the distribution @xmath139 follow as @xmath140 in particular , @xmath141 and @xmath142 . the procedure described here was carried out to obtain @xmath143 for d and i - wp . since p and g are related to d by a bonnet transformation , their distributions can be obtained simply by appropriately rescaling . for example , for the transformation @xmath144 , the @xmath130 have to be rescaled with @xmath145 and the @xmath129 with @xmath146 , as given in tab . [ tablenumbers ] . in fig . [ figuredistributions1]a , we plot the four different distributions obtained . p , d and g feature one peak around a ( relatively large ) k - value , that means they are quite uniformly curved , although they have flat regions , too ( which correspond to the poles of @xmath106 ) . i - wp is bimodal , that is it has prominently both flat and strongly curved regions . the strong curvature of d is somehow artificial due to our choice of unit cell . if one corrects for this by a factor @xmath147 for the k - values , it becomes clear that i - wp has a much broader distribution than p , d and g. this can also be inferred from tab . [ tablemoments ] , where we list the first moments @xmath138 of the distributions for d and i - wp . the ones for d are identical to the ones given in ref . @xcite after appropriate rescaling @xmath138 with @xmath148 due to the different choice in unit cell ( the numerical deviation is less than 0.02 % even for @xmath149 ) . the moments for p and g can be calculated from the ones for d due to the bonnet - transformation between them ; e.g. @xmath150 with the values for @xmath16 and @xmath23 given in tab . [ tablenumbers ] . in particular , @xmath151 , @xmath152 and @xmath153 as can be verified from tab . [ tablenumbers ] . it follows immediately that the quantity @xmath72 defined for is invariant under a bonnet transformation . in order to quantify the variance of the different distributions , we calculate standard deviation , @xmath154 ( here @xmath155 means area average ) . this quantity is independent of scaling , choice of unit cell and bonnet - transformation , and given in the last row of tab . [ tablemoments ] . as expected , its value for i - wp is higher than for p / d / g . in order to investigate the curvature properties of s , c(p ) and f - rd , we use the representations as isosurfaces to fields @xmath156 which we obtained recently as local minima of a simple ginzburg - landau theory for ternary amphiphilic systems @xcite . the same model can also be employed for binary systems if the scalar order parameter field @xmath156 is interpreted to be the local concentration difference between water in the first labyrinth and water in the second labyrinth . then the mid - surface of the lipid bilayer is identified with the surface @xmath157 . the free - energy functional is given by @xmath158 = \int\ d\textbf{r}\ \left\ { ( \delta \phi)^{2 } + g(\phi ) ( \nabla \phi)^{2 } + f(\phi ) \right\}\ .\ ] ] a reasonable choice for @xmath124 and @xmath159 has been found to be @xmath160 for given model parameters @xmath161 , the free - energy functional is minimized in fourier space by implementing the correct space group symmetry and minimizing for fourier amplitudes and lattice constant with conjugate gradients . for balanced tpms , an additional black and white symmetry has to be implemented . the resulting fields @xmath156 correspond to the various local minima . we showed in ref . @xcite that by tuning @xmath161 appropriately , the @xmath162-isosurfaces can be made to become very close to minimal surfaces , and generated representations for nine different tpms : g , d , p , s , c(p ) , c(d ) , c(y ) , i - wp and f - rd . k - distribution can be obtained from these representations as follows . by triangulating the isosurfaces with the marching cube algorithm , we numerically obtain a metric on the surface . from the fourier representations of the fields @xmath156 , it is possible to calculate exactly the k - values on the vertices of the triangulation . using a histogram technique like in the weierstrass case , it is thus possible to obtain @xmath143 even for structures like s , c(p ) and f - rd , which might be of physical relevance but for which no exact representations are known . the distributions obtained in this way for @xmath163 , @xmath164 and @xmath165 are plotted in fig . [ figuredistributions1]b . all three of them are roughly trimodal and feature much stronger curved parts than the ones discussed above . this reflects their complicated shape , which makes them more difficult to access numerically than the ones which have been generated by weierstrass representations ; it also explains the difficulty to obtain smooth data for the k - distributions . in tab . [ tablemoments ] , we give their first moments @xmath138 and the measure @xmath166 of their variance . all of them have broader distributions than i - wp , in the sequence s - f - rd - c(p ) . therefore we can conclude already from the distributions that the ibcps should become less favorable in the sequence p / d / g - i - wp - s - f - rd - c(p ) . given the distribution of gaussian curvature as histogram @xmath143 , it is straightforward to evaluate the first term of the curvature energy , , of the bicontinuous phases . since it is an area average over a complicated function of @xmath71 , we have to replace @xmath167 and @xmath71 by @xmath168 and @xmath130 , respectively . in fig . [ figuregeometry1]b , we plot the quantity @xmath169 as a function of hydrocarbon volume fraction @xmath49 for @xmath170 . in contrast to @xmath171 , compare , no simple formula exists , and it has to be calculated numerically as @xmath172 since @xmath173 is invariant under a bonnet transformation , @xmath174 is the same for p , d and g , except for a rescaling of @xmath49 with @xmath47 . in particular , the curves reach the same minimal values and the minima are shifted to the right with increasing topology index @xmath47 . thus the structures p , d and g perform equally well , but at different values of @xmath49 . from fig . [ figuregeometry1]a , we see that all other structures can not reach this low value of frustration since they have much broader distributions than p , d and g ( compare tab . [ tablemoments ] ) . note that the scaling with @xmath175 is valid in for all structures and in only for those related by a bonnet transformation . except for a factor of @xmath176 , the free - energy density ( [ freeenergydensitybicont2 ] ) for @xmath177 is identical with @xmath178 ( see ) . in fig . [ figurefreeenergies ] , we show the free - energy density of all considered phases for @xmath170 , @xmath177 and @xmath179 . > from the graph , we infer the phase sequence l - g - d - p with increasing water content ; the other phases are not stable since they can not reach that small an amount of bending frustration as can p , d and g. as discussed above , implies that the sequence g - d - p is determined by the topology index . since the free - energy density for p rises again for small @xmath49 , it would be more favorable for it to coexist with excess water . this mechanism is known as _ emulsification failure _ in surfactant systems @xcite and allows to extend the sequence to include a p+w coexistence . the emulsification failure can be interpreted as a maxwell construction of p with the pure water phase with @xmath180 at @xmath181 . it follows from and the scaling of @xmath178 with @xmath175 that the free - energy density @xmath78 for p , d and g can be written in the scaling form @xmath182 with a universal function @xmath183 . for small @xmath91 , this function has @xmath184 , a minimum at @xmath185 , and diverges at @xmath186 . a maxwell construction shows that these structures always lie on a triple line , irrespective of the values for @xmath0 , @xmath91 and @xmath6 ( compare inset of fig . [ figurefreeenergies ] ) . this means that if we calculated the whole phase diagram with the maxwell construction , d and p would be stable only along a line . in the following we identify phase transitions between g , d and p with the intersections of the free - energy curves , since under experimental conditions , small additional contributions to the free energy are certain to destroy the delicate free - energy balance resulting from the bonnet transformation , and then will lead to extended one - phase regions for d and p ( or remove one or both of them from the overall phase diagram ) . in we use this reasoning to construct phase diagrams as a function of @xmath49 and @xmath0 for different values of @xmath91 and @xmath187 . with increasing water concentration @xmath188 we always obtain the sequence l - g - d - p - w+p . the most stable ibcp is the gyroid structure g close to the lamellar phase at high values of @xmath49 since it has both a narrow distribution of gaussian curvature and the highest topology index . at lower concentrations of @xmath49 , one finds narrow regions of stability for d and p since they have narrow distributions , too , but increasingly lower values of @xmath47 . increasing @xmath0 shifts the loci of least frustration to higher @xmath49 as evident from fig . [ figuregeometry1]a . decreasing @xmath7 ( i.e. increasing @xmath91 ) disfavors the ibcps and increases the stability region for the lamellar phase , as can be seen by comparing a and [ figurephasediagram]b , where @xmath91 is increased from @xmath189 to @xmath190 , respectively . however , for @xmath191 there always will be a region of stability for the ibcps , even for @xmath192 . this stands in marked contrast to similar modelling for ternary systems where bicontinuous cubic phases are expected to disappear for @xmath193 @xcite . the explanation lies in the fact that changing the spontaneous curvature @xmath0 of the monolayers amounts to changing the effective saddle - splay modulus of the bilayer , compare eq . ( [ eq : effectivebilayerbending ] ) . thus , increasing @xmath0 can counteract the effect of making @xmath194 more negative . the bicontinuous phases are also favored by lowering @xmath6 without changing @xmath91 , since this increases the fluctuation effects which favor ibcps ( positive renormalization of @xmath7 ) and disfavor the lamellar phase ( increase of steric repulsion ) . this can be seen by comparing b and [ figurephasediagram]c , where @xmath195 is decreased from @xmath196 to @xmath197 , respectively , while @xmath91 is kept constant at @xmath190 . it is clear from our discussion in sec . [ sec : model ] that all contributions to the free energy , which can be written in the scaling form @xmath198 , will preserve the triple - line coexistence of g , d and p. obviously , an interaction which introduces an additional length scale , in addition to @xmath1 and @xmath89 , is needed to break the bonnet symmetry and to lift the triple - line coexistence . here we want to discuss van der waals interactions and chain stretching . the non - retarded van der waals interaction at distance @xmath199 has the general form @xmath200 where @xmath201 is the hamacker constant , and @xmath202 and @xmath203 are the number densities of atoms in the two interacting bodies . the total interaction energy between two colloidal particles of volumes @xmath204 and @xmath205 is then given by @xmath206 unfortunately , the energy density of the van der waals interaction can not be calculated easily in a cubic bicontinuous phase . however , it is clear from eq . ( [ eq : vdw_energy ] ) that it is the embedding of the tpms in three - dimensional space , which is important for the van der waals energy . distances between different parts of the bilayer are not determined uniquely by the @xmath4-distribution , which describes the internal geometry of a surface . this is evidenced by the self - intersections of the parallel surfaces for @xmath207 , which do _ not _ affect any of the considerations of the curvature energy described in the previous sections , but would give a divergent contribution to the van der waals energy . thus , we conclude that the van der waals interaction must break the bonnet symmetry . we illustrate this point in appendix [ app : fcc ] , where we approximate the tpms by a fcc array of spherical shells . in order to estimate the frustration energy due to chain stretching , it has been suggested in ref . @xcite to calculate the variance of the distance between constant - mean - curvature companions of a given tpms . for the d - surface , the variance @xmath208 has been shown @xcite to be almost independent of @xmath49 , with @xmath209 and @xmath210 , where eq . ( [ latticeconstant ] ) has been employed in the second equality . similar results can be expected for g and p , with prefactors @xmath211 and @xmath212 , respectively . we introduce a function @xmath213 , which is defined such that @xmath214 , @xmath215 and @xmath216 for the three values of @xmath43 for the p , d and g phases . then , the stretching - energy density has the scaling form @xmath217 where @xmath218 is the stretching energy from and has been used for @xmath219 . note that this energy density has been made dimensionless with the same factor as the curvature energy in sec . [ sec : model ] , i.e. it is normalized to the free - energy density of the lamellar phase [ compare eq . ( [ freeenergydensitylamellar ] ) ] . these results make the scaling argument of sec . [ sec : model ] more precise . equation ( [ deltal ] ) shows very clearly that the lattice constant @xmath23 appears as new length scale , and that the bonnet symmetry is broken in general by the stretching contributions . it is interesting to consider the possible functional dependencies of @xmath213 in more detail . first , if @xmath220 is a monotonically increasing ( decreasing ) function of @xmath43 , then p is stabilized ( destabilized ) by chain stretching due to its small value of @xmath43 , while d is disfavored ( favored ) . in the case of increasing @xmath220 , d would not appear in the phase diagram , while in the case of decreasing @xmath220 , p can not be present . second , if @xmath221 const , the bonnet symmetry is restored . finally , @xmath220 can of course be a non - monotonic function of @xmath43 , which stabilizes the phase with the smallest value of @xmath220 . in any case , the difference between the three phases can be expected to be small , since the prefactor @xmath222 ( and presumably also @xmath211 and @xmath212 ) is very small ; nevertheless , the stretching energy might account for small differences , like the ones needed to produce the phase diagram of 2:1 la / dlpc and water , which shows no triple - line symmetry , but something close to it . a more precise statement requires the calculation of @xmath219 for g and p in the framework of the cmc - model . other estimates of the stretching energy have been discussed in ref . @xcite ; they also indicate a breaking of the bonnet symmetry . in this work , we have investigated inverse bicontinuous cubic phases ( ibcps ) in lipid - water mixtures . we systematically included all ibcps which might be expected to be of physical relevance , and found that only the ones with a narrow distribution of gaussian curvature over the minimal mid - surface are stable , i.e. p , d and g. while this result has been anticipated in ref . @xcite , we were able to prove it quantitatively for the first time by calculating the distributions of gaussian curvature over the minimal mid - surfaces of all relevant structures ; from these we calculated the variance @xmath166 of the @xmath4-distributions given in which is the same for p , d and g due to bonnet transformations and higher for all other structures . the bonnet transformation also implies that the free - energy densities for p , d and g have the scaling form @xmath223 with hydrocarbon volume fraction @xmath49 and topology index @xmath47 . the topology index is a dimensionless inverse euler characteristic which is independent of scaling and choice of unit cell ; the higher its value , the less porous the structure . since it is highest for the gyroid structure g , it has the largest region of stability . d and p have progressively lower values for @xmath47 , and have therefore smaller regions of stability at lower values for @xmath49 . we also used the concept of the emulsification failure ( which is well established in surfactant science ) to explain the stability of ibcps in excess water . a maxwell construction with the excess phase shows that p , d , g and excess water lie on a triple line for all values of @xmath0 and @xmath91 as a consequence of the bonnet transformation . this bonnet symmetry should not be interpreted as a degeneracy between g , d and p , since their different topologies ensures that their regions of stability are arranged according to their respective values of the topology index . it should be noted that the free energy per _ unit area _ ( rather than per unit volume ) has the scaling form @xmath224 ; this might be interpreted as some kind of degeneracy ( compare refs . @xcite ) , since now all three structures attain the same minimal values . our results agree nicely with experimental observations . at fixed temperature , it is always the sequence g - d - p which is observed with increasing water content , with g having the largest region of stability and only d or p coexisting with excess water @xcite . for example , for monoolein and water at @xmath225 , with increasing water concentration the sequence l - g - d has been found , where g has the largest region of stability and d coexist with excess water @xcite . our phase diagram for @xmath226 and @xmath179 ( a ) agrees particularly well with the one for 2:1 la / dlpc and water @xcite ( compare ) . we do not treat the inverse hexagonal phase , which experimentally occurs at large @xmath49 and large @xmath0 , but we showed above that ibcps will occur at large @xmath0 and _ small _ @xmath49 , and argued that at large @xmath49 stretching energy should become comparable to bending energy . this , in fact , favors the inverse hexagonal phase , whose geometry requires chain stretching . finally , we can discuss the role of temperature for 2:1 la / dlpc and water by using the ansatz @xmath227 , with a balanced temperature @xmath228 and a phenomenological constant @xmath229 . using @xmath230 , we find @xmath231 and @xmath232 . since the chain melting temperature is known to be @xmath233 , we conclude that the balanced temperature @xmath228 is well below the main transition . this seems to be the generic case for lipid systems ; it might be related , e.g. , to the anomalous swelling of pc - bilayers when the main transition is approached from above @xcite . it is interesting to note that the constant @xmath229 obtained here for 2:1 la / dlpc and water is very close to the corresponding value @xmath234 for monolayer of @xmath235 in ternary microemulsions @xcite . in this paper , we only considered inverse bicontinuous phases with cubic symmetry . a similar analysis for non - cubic phases should be straight - forward , since here weierstrass representations are known for the main structures , especially for the rhombohedral and tetragonal variants of g , d and p @xcite . however , we expect that these phases will not be more favorable than g , d and p for the same reason as for the other cubic phases . in order to make further progress in understanding their phase behavior , it seems more important to investigate the different mechanisms which might break the bonnet symmetry between g , d and p. in our view , two calculations are needed now : ( i ) of the van der waals energy of g , d and p and ( ii ) of the variance of the distance distribution of a constant - mean - curvature model of g and p. for a rough estimate of the energy density due to the van der waals interaction in cubic bicontinuous phase , we use a fcc array of spherical shells of radius @xmath236 , thickness @xmath237 and lattice constant @xmath23 . here , the lattice constant @xmath23 corresponds to the conventional unit cell and is taken to be the same as for the corresponding bicontinuous phase , and @xmath236 is determined by the requirement of equal area @xmath114 per unit cell , so that @xmath238 . thus we have @xmath51 again . the distance of closest approach between two neighboring spheres is @xmath239 ; in order to avoid contact between them , @xmath240 is necessary , which implies @xmath241 ; this inequality is satisfied for all tpms studied here , compare tab . [ tableindex ] . for the crystal , we consider the interaction between nearest - neighbor spheres only . the van der waals energy between two hollow spheres is described by the girifalco potential , which is often used to model the van der waals interaction between two buckyballs @xmath242 @xcite . for our purpose it is sufficient to use the derjaguin approximation for the girifalco potential , which describes the leading contributions for close approach correctly , and shows that the van der waals attraction for two hollow spheres scales as @xmath243 . we thus find for the van der waals energy density @xmath244 this energy density has again been made dimensionless with the same factor as the curvature energy in sec . [ sec : model ] . its scaling shows that _ two _ ratios of length scales , @xmath245 and @xmath246 , determine the van der waals energy . it is intersting to note that the curvature - energy density in this model of spheres scales as @xmath247 ^ 2\ .\ ] ] since @xmath248 ( for fixed @xmath44 ) , is consistent with the scaling function ( [ eq : scaling ] ) . this indicates that our identification of the sphere radius and lattice constant gives a consistent approximation for both van der waals and bending energies . , e. ; rummel , g. ; rosenbusch , j. p. ; , e. m. _ science _ * 1997 , * _ 277 , _ 16761681 . rummel , g. ; hardmeyer , a. ; widmer , c. ; chiu , m. l. ; nollert , p. ; locher , k. p. ; pedruzzi , i. ; landau , e. m. ; rosenbusch , j. p. _ j. struct . biol . _ * 1998 , * _ 121 , _ 8291 . templer , r. h. ; seddon , j. m. ; warrender , n. a. ; syrykh , a. ; huang , z. ; winter , r. ; erbes , j. _ j. phys . chem . b _ * 1998 , * _ 102 , _ 72517261 . templer , r. h. ; seddon , j. m. ; duesing , p. m. ; winter , r. ; erbes , j. _ j. phys . b _ * 1998 , * _ 102 , _ 72627271 . .euler characteristic @xmath44 , scaled surface area @xmath43 ( both in the conventional unit cell ) and topology index @xmath249 for all tpms considered , as obtained from local minima of the free - energy functional of a simple ginzburg - landau theory . exact values are known for g , d , i - wp , p and c(p ) and given in . the structures are ordered according to decreasing @xmath47 ; the value for s is probably too large , but difficult to improve numerically . [ cols="<,<,<,<,<,<,<,<",options="header " , ]

ca@xmath0ruo@xmath1 ( cro ) is the mott insulating@xcite end member of the series ca@xmath2sr@xmath3ruo@xmath1 , which possess a rich diversity of structural , magnetic and transport properties@xcite . sr@xmath0ruo@xmath1 , the other end member , is a spin - triplet superconductor with broken time - reversal symmetry @xcite . the metal - insulator transition in cro at @xmath4 k @xcite goes along with severe structural distortions @xcite , the ruo@xmath5 octahedrons flattens , which is also apparent in jumps of the lattice constants in the order of 0.1 . these structural distortions are enhanced till the onset of antiferromagnetic order @xcite . the nowadays widely used picture assumes that an orbital ordering is associated with the structural changes @xcite . the 4@xmath6 orbitals become doubly occupied and the 4@xmath7 singly occupied resulting in flattened octahedrons and a @xmath8 state . in the past the nature of the mott transition of this multiband system with four @xmath9 electrons on the ru site was intensively discussed @xcite . recently , it was proposed , that spin - orbit coupling ( soc ) in this 4@xmath9 system is strong enough to change the multiplet structure and couples @xmath10 and @xmath11 to @xmath12 resulting in a nonmagetic @xmath13 ground state . the occurring of a magnetic order was proposed to be of a singlet magnetism type ( see e.g. sec . 5.5 . in ref . @xcite ) , which was called excitonic magnetism in ref . @xcite , and a special type of magnon dispersion was predicted in this theory @xcite . the main branches of the dispersion obtained by inelastic neutron scattering ( ins ) experiments , however , could be successfully described with a conventional heisenberg model @xcite and strongly disagree with @xmath13 calculations @xcite . nevertheless , neutron scattering experiments revealed features and additional magnetic scattering not describable with a simple square - lattice antiferromagnet . the main in - plane transverse modes exhibit a large anisotropy gap , underlining the impact of broken tetragonal symmetry in combination with soc . furthermore , an additional signal was detected in the neutron scattering experiments at low and at high energies @xcite , that can not be explained by the two transverse magnon branches . here we present a study of the magnon dispersion using ins with and without polarization analysis on crystals containing 0 , 1 , and 10 % ti combined with density functional theory ( dft ) calculations . the full dispersion of magnetic excitations in cro reveals a very strong impact of soc , and oxygen moments that lead to an additional branch . ( color online ) crystal and magnetic structure of cro and the associated reciprocal space . in ( a ) one layer of the crystal and magnetic structure the ru ( blue balls ) sit in octahedrons of oxygen ( red balls ) . the ab - plane is indicated in light blue . in black and green the orthorhombic and tetragonal cells are shown . the pink arrows denote the magnetic moments on the ru and o sites . at the tip of the low right ru moment , the directions of the different polarizations of the magnon modes are indicated by black arrows and some labellings of atoms and atom distances are given . in ( b , c ) the ( hkl ) planes in reciprocal space are illustrated for l=0 and l=1 , respectively . the orthorhombic and tetragonal cells are drawn in black and red , respectively . green circles denote tetragonal zone centers and pink circles zone boundaries . black squares ( stars ) mark antiferromagnetic zone centers of the @xmath14- ( @xmath15)-centered antiferromagnetic order . ] crystal growth of insulating ca@xmath2sr@xmath3ruo@xmath1 is severely hampered by the metal - insulator transition occurring in pure cro slightly above room temperature , @xmath4 k. the space group does not change at this first - order phase transition , but there are sizable jumps in the lattice - parameters , in particular for @xmath16 , with @xmath17 @xcite . therefore , upon cooling the crystals after the growth process , they tend to crack and only small pieces of mm@xmath18 size can be recovered . because inelastic neutron scattering experiments require samples of @xmath19cm@xmath18 size , we substituted a small amount of ru by ti in order to broaden the phase transition and to be able to cool the crystals through the metal - insulator transition without breaking them into small pieces . crystals with a mass of 0.5 g to 1 g were obtained with only 1 % ti substitution . we were also able to obtain two larger pieces of pure cro , one with a mass of 0.45 g and one with 0.3 g , but these crystals posses a bad mosaic spread of 5@xmath21 , which most likely is the reason why they were not destroyed upon cooling . nevertheless , these crystals are suitable for some inelastic neutron scattering studies . the ca@xmath0ru@xmath3ti@xmath22o@xmath1 crystals with @xmath23% used in this neutron scattering study ( labeled 0ti , 1ti , 10ti , respectively ) were grown by the floating - zone method in a canon machinery inc . sc1-mdh11020-ce furnace equipped with two 2000 w halogen lamps and a coldtrap following the procedure described in references @xcite . caco@xmath24 , ruo@xmath0 and tio@xmath0 were mixed in stoichiometric ratios and a ru excess of 32 % was added . the powder was mixed and reacted for 24 h with an intermediate grinding . then a rod was pressed and sintered at 1350 @xmath21c . a growth speed of 17 mm / h and a feed - rod speed of 20 mm / h were used and the atmosphere contained 90 % ar and 10 % o at a total pressure of 9 bar . phase purity was checked by x - ray powder diffraction from crushed parts of the single crystals , which indicates no impurity phases . the lattice constants at room temperature were obtained by lebail fits of these powder pattern using the fullprof suite@xcite . single - crystal x - ray diffraction experiments were performed on a bruker x8 apex diffractometer equipped with a charge - coupled - device detector . details of this structure determination and a detailed list of the refined parameters can be found in the appendix a. magnetization measurements were performed with a superconducting quantum interference device ( squid ) magnetometer from quantum design , and the resistivity was measured with a standard four - point method . spin wave calculations were performed using spinw @xcite . elastic and inelastic neutron scattering experiments were performed with the cold triple - axis spectrometer ( tas ) 4f1 at the llb , with the thermal tas in3 at the ill and with the polarized thermal tas in22 at the ill . the polarized neutron scattering experiments were performed using heusler ( 111 ) monochromator and analyzer crystals . a pg filter was inserted on the scattered beam before the neutron - spin flipper and the monitor was put on the incoming beam between monochromator and sample . a set of helmholtz coils was used to produce the guide field and the sample was zero field cooled ( less than 2 g ) in an orange - type cryostat . a mounting of 3 crystals with a total mass of 2.5 g containing 1 % ti ( 1ti ) was used for the polarized neutron scattering experiment . the sample was twinned with non - equal twinning - fractions of 2.5:1 determined by scanning the orthorhombic ( 200 ) and ( 020 ) reflections . the @xmath25 plane [ see fig . [ explanations](a ) ] was chosen as the scattering plane in order to efficiently integrate the inelastic signal along the vertical direction , where the resolution is poor . we used the standard coordinate system in polarized neutron scattering @xcite . * x * is parallel to the scattering vector ( * q * ) , * y * and * z * are perpendicular to * q*. while * y * is in the scattering plane , * z * is perpendicular to it . therefore , * y * lies in the @xmath25 plane of the crystal and * z * parallel to the @xmath16 axis . in neutron scattering only magnetic moments perpendicular to * q * intervene . therefore , the magnetic scattering intensities sense a geometry factor sin@xmath26(@xmath27 ) with @xmath27 being the angle between * q * and the magnetic moment , which corresponds to either the static ordering moment in diffraction or to the oscillating moment in a magnon . with longitudinal neutron polarization six different channels can be analyzed : three spin - flip ( sf ) and three non - spin - flip ( nsf ) channels . while phonons always contribute to the nsf channel there is an additional selection rule for magnetic scattering . the magnetic component parallel to the direction of the neutron polarization contributes to the nsf channel , while the components perpendicular to the neutron polarization generate sf scattering . by combining this polarization rule with the geometry factor one can distinguish the different magnetic components . in the geometry we use for the polarized neutron scattering experiment we see the @xmath16 polarized modes in the sfy channel and they do not loose intensity due to the geometry factor , because the scattering vector is always parallel to the @xmath25 plane and thus perpendicular to @xmath16 . in contrast the sfz channel contains the in - plane modes , the transverse and longitudinal ones , but weighted with the geometry factor . for example at @xmath28 the geometry factor for the transverse mode is sin@xmath29 and that for the longitudinal mode sin@xmath30 . at @xmath31 this ratio is inverted [ see fig . [ explanations](b , c ) ] . respecting the twinning ratio of 2.5:1 , we expect that the intensities at @xmath28 and @xmath31 have a ratio of 0.6 for the transverse mode . for the longitudinal mode this ratio is inverted . thus , it is possible to distinguish between transverse modes and longitudinal modes by comparing scattering at properly chosen * q*. because the polarization of the neutron beam is not perfect , one has to correct the intensities for the finite flipping - ratio ( fr ) @xcite . thereby , the magnetic signals are obtained from the intensities of the different sf channels corrected for the fr : @xmath32\ ] ] the fr of our experiment on in22 amounts to 12 , which is determined by comparing the signals of rocking scans on the ( 200 ) bragg - reflection in the sf and nsf channels . on the cold tas 4f1 crystals with different ti content were used , 0ti and 10ti . 0ti has a mass of 0.45 g and a twinning ratio of 1:1 and 10ti a mass of 0.93 g and a twinning ratio of 9:1 . the scattering plane for both samples was chosen to be the [ 010]/[001]-plane . so the scattering plane for the other twin domains was the [ 100]/[001]-plane . for all 4f1 scans a pyrolytic graphite monochromator and analyzer were used , a cooled be - filter was put on * * k**@xmath33 to suppress higher order contaminations . for the cooling of the samples an orange - type cryostat was used and @xmath34 was set to 1.55 @xmath35 for all scans . with the in3 tas we analyzed the magnetic order of the crystal containing 1% ti used in the previous neutron scattering study@xcite ( 1tib ) . this crystal was found to exhibit a majority twin domain of 95% and the experiment was performed in the [ 010]/[001 ] orientation used in the inelastic neutron scattering experiment . ( color online ) temperature dependence of physical properties of cro with different ti content . ( a ) field - cooled magnetization curves in 0.1 t measured on heating . ( b ) resistivity on heating in dark color and on cooling in bright color . only the 10 % ti curves do not coincide . ] the crystal structure of cro @xcite is similar to the structure of sr@xmath0ruo@xmath1 of k@xmath0nif@xmath1 type @xcite , where the ru atoms sit in oxygen - octahedron cages , which are corner shared in the @xmath25 plane . in cro the octahedrons are rotated and tilted and they also become severely distorted . as a consequence , the tetragonal symmetry of sr@xmath0ruo@xmath1 is reduced to the orthorhombic space group pbca@xcite . one layer of the crystal structure is drawn in fig . [ explanations ] , which also depicts the magnetic order with antiferromagnetic moments on ru pointing along @xmath36 . the orthorhombic unit cell is rotated by @xmath37 with respect to the tetragonal one and enhanced to : @xmath38 , @xmath39 , @xmath40 . unless otherwise specified all notations refer to the orthorhombic lattice of the majority twin orientation . our crystals usually posses two twin orientations , which are obtained by interchanging the @xmath41 and @xmath42 directions . all crystals were examined by x - ray diffraction , magnetization and resistivity measurements in addition to the neutron scattering experiments described below . [ magnetization ] shows the magnetization and electric resistance data of the three ti concentrations . the temperature of the metal - insulator transition is 4 k lower for the sample containing 1 % ti than for the pure compound . the sample containing 10 % ti does not show indications for a phase transition upon cooling down to 80 k where the experimental limit of the increasing resistivity is reached . the absolute values of the resistivity curves have a large uncertainty because of the first - order structural phase transition . there the crystals tend to crack , which prohibits the current to flow through the hole sample . this effect has been frequently observed during a single measurement cycle . the resistivity is enhanced by a multiplication factor after passing the structural transition . as the samples have already passed this transition after the crystal growth , there are some cracks inside the sample , which is evident from measuring several pieces of the same crystal without obtaining the same room - temperature specific resistivity . the magnetic and insulating properties in ca@xmath2sr@xmath3ruo@xmath1 are closely related to the crystal structure . cro is heavily distorted with respect to sr@xmath0ruo@xmath1 , which possesses the ideal structure of k@xmath0nif@xmath1 type without structural distortions , but which already is close to such a structural instability@xcite . in the layered ruthenates the distortions can be described as octahedron rotation around the @xmath16 direction and tilting around an in - plane axis . these distortions and the associated structural phase transitions result from bond - length mismatch , and the distortions in cro are induced by the chemical substitution of sr by the isovalent , but smaller ca , which is not able to fill the space between the oxygen octahedrons like sr does . so the octahedrons start to rotate and tilt in order to reduce the coordination volume around ca . by chemically substituting the smaller ti for ru , one expects a small decrease of these deformations , which is indeed realized . the tilt and rotation angles of 10ti are significantly smaller than the values determined in the samples containing only 0 and 1 % ti , see appendix a , tab . [ tab : char ] . temperature and ti doping dependence of the ruo@xmath5 octahedron distortion obtained by single crystal x - ray diffraction analyses ; ( a , b ) presents the ratios of in - plane to out - of - plane ruo bond distances ( flattening ) and of the two octahedron edges parallel to @xmath36 and @xmath43 , respectively . both entities are enhanced in the insulating phase at low temperature due to the lifting of orbital degeneracy . ] most interesting is the impact of the ti doping on the deformation of the ruo@xmath5 octahedron , which is a fingerprint of the lifting of orbital degeneracy@xcite . the structural change at the metal - insulator transition in pure cro is characterized by a jump of the @xmath16 lattice - constant and by a flattening of the octahedron as it is visible in the ru - o bond distance ratio@xcite . upon cooling into the insulating phase this octahedron flattening continues until it saturates at the antiferromagnetic phase transition . slightly below room temperature the octahedron shape passes from elongated to flattened @xcite . in all samples studied here , this crossover occurs below room temperature , thus at 100 k the octahedron shape is flattened for all ti substitutions , see fig . [ ruooo ] . due to flattening of the lattice and octahedron the @xmath6 orbital shifts down in energy compared to the @xmath44,@xmath45 orbitals . a similarly strong and anomalous temperature dependence is also observed for the ratio of the two o - o octahedron edge lengths parallel @xmath43 and @xmath36 which is related to the orthorhombic splitting [ @xmath46 @xcite . at room temperature the octahedron is longer along @xmath43 , while it is elongated along @xmath36 at low temperature . all these effects can be attributed to a temperature dependent orbital ordering @xcite . the considerable elongation along @xmath36 at low temperature agrees with soc and the alignment of the magnetic moment mainly along the @xmath36 direction @xcite . figure [ ruooo ] illustrates that these two distortions of the ruo@xmath5 octahedron become suppressed by the 10 % ti substitution . this underlines the orbital ordering character of these distortions , which is obviously suppressed by ti@xmath47 with an empty @xmath48 shell . while the @xmath25 plane ( or the average in - plane parameter ) increases with high ti content , the @xmath16 lattice - constant decreases at room temperature . the reduction of the @xmath16 lattice - constant , in hand with an increase of the @xmath25 plane , is the structural signature of the insulating state of cro . with 1 % of ti substitution there is only a small decrease in @xmath49 ( fig . [ magnetization ] ) , with 10 % of ti substitution there are drastic effects . in the resistivity of 10ti there are no indications for a sharp metal - insulator transition on cooling from 700 down to 80 k , where the upper experimental limit of the resistance experiment is reached . note that the absolute values of the resistivity curves are prone to a large uncertainty because of the first order structural phase transition which causes cracks in the crystals . but the orders of magnitude larger resistivity of the 10ti sample and the absence of the metal - insulator transition is unambiguous . this remarkable stabilization of the insulating state by only small amounts of ti substitution is also seen in ca@xmath24ru@xmath0o@xmath50 @xcite . since the ionic radii of ti and ru are very similar , this suppression of the metallic state seems to originate from the very effective suppression of the hopping . also in sr@xmath0ruo@xmath1 ti substitution has a strong impact : it stabilizes spin - density wave ordering associated with the fermi - surface nesting of the pure material for only 2.5% ti @xcite . also in sr@xmath24ru@xmath0o@xmath50 ti substitution stabilizes a spin - density wave magnetic instability @xcite . ( color online ) elastic neutron scattering scans with polarization analysis . the orthorhombic splitting is clearly visible in this elastic scans . ( a , b ) present the values of the nsfx ( sfx ) channel in magenta ( blue ) of longitudinal scans across the strong nuclear reflections ( 200 ) and ( 020 ) , respectively . the values of the nsfx channel are divided by the fr . these elastic longitudinal scans show that both twins are well separated . ( c - f ) present data of the sfx ( blue ) , sfy ( red ) and sfz ( black ) channels of scans across the magnetic ( c ) ( 100 ) , ( d ) ( 010 ) , ( e ) ( 210 ) , and ( f ) ( 2 - 10 ) reflections . ] the magnetic structure of 1ti was studied on in22 with polarization analysis . the analysis of the magnetic structure of 1tib on in3 is described in appendix b. scans along the ( 200 ) and ( 020 ) reflections ( fig . [ polel ] ) reveal that crystal 1ti exhibits both twin - domain orientations in the ratio 2.5:1 . [ polel](c - f ) show the sf intensities in the three channels for scans across the ( 100 ) , ( 010 ) , ( 120 ) and ( -120 ) magnetic reflections for crystal 1ti . the magnetic structure of cro has been previously determined by neutron powder diffraction @xcite . as illustrated in fig . [ explanations](a ) the magnetic moments are essentially aligned along the @xmath36 direction ( corresponding to the tilt axis ) with antiferromagnetic alignment between nearest neighbors . due to the low symmetry of the crystal structure ( pbca ) the dzyaloshinski - moriya interaction is finite and results in a canting of the moments along the @xmath43 direction . the canting of the moments along @xmath43 is parallel for all ru sites in a single layer resulting in a net ferromagnetic moment per layer . there are two different magnetic structures reported for cro @xcite , which differ by the stacking of the single layer arrangement shown in fig . [ explanations](a ) . the magnetic structure of the main antiferromagnetic @xmath36 component is either @xmath14 or @xmath15-centered . in the @xmath14-centered phase the two magnetic moments at ( 0,0,0 ) and ( 0,0.5,0.5 ) are parallel , in the @xmath15-centered one the two moments at ( 0,0,0 ) and ( 0.5,0,0.5 ) are parallel . the magnetic space groups are @xmath51 ( @xmath14-centered ) and @xmath52 ( @xmath15-centered ) . while in @xmath51 the net ferromagnetic canted moments per layer cancel due to an antiferromagnetic stacking , the @xmath15-centered @xmath52 structure results in a total ferromagnetic moment , that can be measured with a magnetometer . there is also canting along @xmath16 direction in both magnetic structures . this canting corresponds to an antiferromagnetic @xmath16 component that is @xmath15-centered in @xmath51 and @xmath14-centered in @xmath52 . this moment should , however , be small because the main part of the ordered moment points along the tilt axis , and only microscopic methods can detect such an antiferromagnetic @xmath16 component . the analysis of the magnetic bragg peaks with polarization analysis first confirms that moments point along the @xmath36 direction and show a dominating @xmath15-centered scheme for 1ti and an almost exclusive @xmath15-centered scheme for 1tib , see appendix b. in contrast , small pure cro samples show only the @xmath14-centered scheme @xcite . in fig . [ magnetization](a ) the magnetization on heating of samples with different ti content is presented . the magnetic field of 0.1 t is applied in the @xmath53 plane . the weak ferromagnetic component parallel to the @xmath43 direction , resulting from the canting of the spins , dominates the magnetization below @xmath54 . the fact that the magnetization of the lowest temperature is the highest , points to the @xmath15-centered phase , in which the ferromagnetic component does not cancel among two neighboring layers @xcite . thus all samples contain a dominating @xmath15-centered phase in agreement with the neutron diffraction studies . the small crystal used for the magnetization measurement with 1 % ti content is essentially untwinned , while the sample containing no ti is partially twinned with nonidentical twin fractions , in contrast to the much larger pure sample used for neutron scattering , for which the twinning ratio amounts to 1:1 . if the field is applied along [ 1,1,0 ] , the observable ferromagnetic component is reduced by a factor @xmath55 but both twin domain orientations contribute ( sample containing 10% ti ) . the highest total ferromagnetic component is observed in the pure sample , and the sample 1ti already exhibits a reduced ferromagnetic component possibly stemming from a slightly reduced fraction of the @xmath15-centered phase . we can conclude that cro exhibits sizable moment canting resulting in an ordered magnetic component along the @xmath43 direction of @xmath56 ( canting angle @xmath57 deg ) . the canting of the magnetic moment arises from a strong dzyaloshinski - moriya interaction that in turn stems from the strong soc . minimizing the static energy of a pair sensing only the heisenberg interaction @xmath58 and the dzyaloshinski - moriya interaction , @xmath59 , yields the condition tan@xmath60 , and thus a rough estimate of the dzyaloshinski - moriya interaction in cro : @xmath61 . the magnetization curves further show a decrease of the nel temperature with increasing ti content , see tab . [ tab : char ] in appendix a , which can be explained by the dilution of the magnetic lattice by non - magnetic ti . ( color online ) energy scans at various * q * values with polarization analysis . raw data of different ( a ) nsf , and ( b , c , e - g ) sf channels with the polarization parallel to * x * ( * y*),(*z * ) in blue ( red ) , ( black ) at different scattering vectors . ( d , h , i ) show the results of the polarization analysis with eq . ( 1 ) , in red m@xmath62 , and m@xmath63 in blue ( green ) at * * q**=(2,1,0 ) ( * * q**=(1,2,0 ) ) . in ( d ) the blue line depicts the modeled asymmetric shape resulting from the folding of the magnon dispersion with the resolution function , see ref . @xcite , which is also used for the modeling in ( i ) , but scaled with the expected correction factors for geometry and twinning fractions . the inset in ( i ) shows a zoom into the data . ] with the alignment of the antiferromagnetic moment along @xmath36 one expects non - degenerate transverse magnon modes corresponding to polarization along @xmath43 or @xmath16 @xcite . with our previous unpolarized inelastic neutron scattering experiments @xcite only the @xmath43 polarized mode could be clearly identified . the gap of the @xmath16 polarized transverse mode could either be hidden in the shoulder of the in - plane signal , which inevitably arises from the folding of the instrument resolution with the steep spin - wave dispersion , or appear at higher energies where weak signals were detected @xcite . with the new polarized experiment the in - plane and @xmath16 polarized transverse magnons can be easily separated in the polarization and it was the first aim to search for the @xmath16 polarized mode . figure [ modes](a - c ) shows energy scans through the anisotropy gap at the antiferromagnetic zone center . the nsf channels shown in fig . [ modes](a ) have a larger background than the sf channels [ fig . [ modes ] ( b - c ) ] and are thus less informative but confirm the main conclusions . the @xmath16 polarized mode must entirely contribute to the sfy and nsfz channels with a complete geometry factor . spin - wave calculations show that the two non - degenerate transverse modes exhibit an intrinsic signal strength inversely proportional to their zone - center energy @xcite . therefore , the @xmath16 polarized mode can be excluded in the asymmetric peak of the in - plane transverse mode . these scans confirm the in - plane character of the signal in the range 14 to 20 mev , see fig . [ modes](d ) . a special effort was laid on the analysis of the signal maximum at the in - plane magnon gap , e=14 mev , see fig . [ modes](e , f ) . with the partially detwinned crystal one expects this signal from the transverse magnon to be reduced by a factor 0.6 for @xmath28 , which agrees with the measured ratio of the intensities at 14 mev between @xmath28 and @xmath31 of 0.55(10 ) . we also measured the backfolded mode at the ferromagnetic zone - center , which according to the spin - wave calculations , see below , exhibits a @xmath16 polarization . indeed such a @xmath16 polarized signal can be determined at @xmath64 and e=14 mev [ fig . [ modes](g ) ] by counting for a very long time ( 45 min for each data point ) . in view of the recent observation of longitudinal magnetic excitations in cro @xcite it seemed worthwhile to search for such modes first at the antiferromagnetic zone centers ( 2,1,0 ) and ( 1,2,0 ) . in fig . [ modes](i ) the fitted background from the in - plane polarized transverse mode from fig . [ modes](d ) is used to separate a weak signal peaking at 29 mev . the signals for the two different * q * values are assumed to posses the same shape , but are scaled with the expected factors for geometry and twinning . around 30 mev the stronger signal for @xmath28 compared to @xmath31 thus points to a longitudinal mode in the sense that it is polarized parallel to the static moment . the energy of this mode corresponds fairly well to the double of the gap of the in - plane transverse branch and the strength is below 10 % of that signal . therefore it seems most likely that this longitudinal signal completely stems from the two - magnon excitation , which is expected to appear in the longitudinal polarization channel @xcite . a similar discussion has also been initiated for the parent material of feas - based superconductors where again the two - magnon explanation seems most likely @xcite . the 30 mev longitudinal signal is thus not anomalous . the @xmath16 polarized antiferromagnetic zone - center mode is detected at higher energy , where the unpolarized experiment found some evidence for additional scattering @xcite , but where the signal strength is expected to be rather small @xcite . in order to cover the energies of the order of 40 mev we needed to use a larger value of @xmath65 @xmath35 , which also allows one to avoid the contamination appearing at e=44 mev for the standard value of @xmath66 @xmath35 . the magnetic signals polarized along @xmath67 and along * y * at large energies are shown in fig . [ modes](h ) and show that the zone - center @xmath16-polarized magnon possesses a large energy of 45.5(1.5 ) mev . its signal strength is in rough agreement with the signal strength of the in - plane transverse mode with the energy being enhanced by a factor three . our conclusion of @xmath16 polarized modes at the ferromagnetic and antiferromagnetic zone centers agrees with the interpretation of similar polarized neutron scattering experiments performed with twinned crystals in a different scattering geometry @xcite . this large @xmath16 anisotropy is remarkable , as it strongly deforms the magnon dispersion in cro with respect to a simple isotropic model . the splitting of the two antiferromagnetic zone - center magnons , 14 and 45.5 mev , is larger than the dispersion of the in - plane branch to the zone - boundary . as will be discussed below , the entire dispersion of transverse branches is , nevertheless , well described with the @xmath8 spin - wave model using the holstein - primakoff transformation . the in - plane and @xmath16-polarized branches exchange their character : while the @xmath16-polarized mode is the high - energy mode at the antiferromagnetic zone center , it appears at the lower energy of the in - plane transverse magnon at the ferromagnetic zone center , in accordance to the data shown in fig . [ modes](g ) . the in - plane polarized branch thus starts at 14 mev at the antiferromagnetic zone center , exceeds to the zone boundary at 41 mev and then continues to stiffen till 45.5 mev at the ferromagnetic zone center . the out - of - plane polarized branch just exhibits the opposite dispersion . by comparing the intensities at the ferromagnetic zone centers @xmath68 and @xmath69 with an energy transfer of 45 mev we may confirm the in - plane transverse character of this high - energy mode appearing at the ferromagnetic zone center . the expected ratio of the signals taking into account the different twinning fractions and geometry factors amounts to 1:2.5 for a transverse and to 2.5:1 for a longitudinal magnon . we obtain a ratio of 0.1(2 ) indicating the transverse in - plane character . evidence for a longitudinal high energy branch could not be obtained in our experiment @xcite . magnon dispersion calculated with the spinw program using the values given in the text . the blue circles denote the fit values of the dispersion obtained in this and the previous @xcite neutron scattering studies . the black lines are the calculated dispersion and the color code denote the calculated neutron scattering intensity of the convoluted spectra . in each panel black stands for maximum intensity and white for none . in ( b , c ) only the in - plane- and out - of - plane polarized transverse modes are shown , respectively , illustrating the opposite dispersion of these branches . ] the anisotropy gap of the @xmath16-polarized transverse magnon of 45.5 mev leads to the uncommon feature in the magnon dispersion , that the transverse in - plane branch continues to increase in energy between the zone boundary and a ferromagnetic zone center . this peculiarity can , however , be well described with a rather conventional model . the hamiltonian , which is used for the description of the magnon dispersion is given by : @xmath70 the sum runs over pairs of magnetic ions , so that each pair or bond appears twice and @xmath10 is set to 0.67 following the experimental results in ref . the spin - wave calculations using the spinw program and the parameters @xmath71 mev , @xmath72 mev , @xmath73 mev , @xmath74 mev and @xmath75 mev , give a good description of the magnon dispersion obtained in this and the previous @xcite neutron scattering studies [ fig . [ spinw](a ) ] . panels fig . [ spinw](b , c ) show the in - plane and c - polarized branches , respectively , starting at an antiferromagnetic zone center and proceeding to a ferromagnetic one . in this figure the color code denotes maximum neutron scattering intensity of the convoluted spectra with black color , and zero intensity with white . the energy resolution is set to 1 mev . the spin - wave calculation thus perfectly describes not only the energy dispersion but also the polarization of the magnon branches . note that this rather uncommon dispersion is described with strong single - ion parameters arising from soc . these strong anisotropy terms interfere with the heisenberg interaction parameters in contrast to models invoking only weak anisotropy . therefore , the @xmath58 parameter differs from that obtained by fitting only the lower part of the dispersion with a small uniaxial anisotropy @xcite . figure [ spinw](a ) shows that the non - sinusoidal parts of the dispersion are perfectly described , clearly better than without the strong anisotropy terms . it is , however , important to note that the model is not unique . it is possible to obtain similar fitting by partially reducing the single - ion anisotropy and by inducing an anisotropic nearest - neighbor interaction . the coupling parameter j@xmath76 acting between neighboring layers splits the magnon modes into two by introducing a finite dispersion perpendicular to the planes . whether the lower or upper mode is seen at @xmath77 is determined by the sign of @xmath78 . with the partially detwinned crystals ( in particular with the experiment on 1tib ) it is possible to determine the sign of this interaction . here it is chosen to couple spins at ( 0,0,0 ) and ( 0,0.5,0.5 ) . the dominant @xmath36 components of these spins are parallel for the @xmath14-centering and antiparallel for the @xmath15-centering . if @xmath78 is positive ( negative ) , corresponding to an antiferromagnetic ( ferromagnetic ) coupling , the upper ( lower ) mode is seen at @xmath77 . the experimental data reveal that the upper mode is seen at @xmath77 , so the coupling is ferromagnetic , stabilizing the @xmath14-centered phase . this coupling contradicts the observation of a @xmath15-centered structure as the main magnetic scheme , see subsection iii.b . cro thus exhibits the uncommon situation that the minimum magnon energy does not occur at the magnetic bragg peaks . this observation can be attributed to an anisotropic @xmath78 which differs for the @xmath43 and @xmath36 spin components . the linearized augmented plane wave ( lapw ) method as implemented in the wien2k package@xcite was used for the dft calculations . the exchange correlation potential for the generalized gradient approximation ( gga ) was chosen to be in the form proposed in ref . the soc was treated in a second variational way . the parameter of the plane - wave expansion was chosen to be @xmath79 , where @xmath80 is the smallest atomic sphere radii ( @xmath81 a.u . , @xmath82 a.u . , @xmath83 a.u . ) and @xmath84 the plane - wave cutoff . we used a mesh consisting of 800 @xmath85points . the on - site hubbard repulsion and intra - atomic hund s exchange were taken to be @xmath86 ev and @xmath87 ev@xcite in the gga+u@xcite and gga+u+soc calculations . the exchange parameters were recalculated via total energies of three different collinear magnetic configurations ( a ferromagnetic and two , in the @xmath16 direction differently stacked , antiferromagnetic configurations ) using the gga+u approach . we found that @xmath88 mev , which agrees with the results of the spin - wave calculations . in order to estimate the single - ion anisotropy we add soc to the calculation scheme and computed the energies of two configurations , where all spins are either directed along @xmath16 or lie in the @xmath25 plane . the antiferromagnetic order was assumed in these calculations . the lowest total energy corresponds to the configuration , where all spins are directed predominantly along the @xmath36 axis . the single - ion anisotropy is found to be very large , @xmath89 mev , again in good agreement with experiment . very recently microscopic magnetic parameters were calculated for similar @xmath90 and @xmath91 values finding a nearest - neighbor interaction of 3 - 6mev and a large single ion anisotropy in agreement with our results @xcite . the spin moment on the ru ions was calculated to be @xmath92 ( i.e. @xmath93 ) , while the orbital moment is @xmath94 . we also find a sizable magnetic moment on the apical o @xmath95 , while planar ligands stay nonmagnetic . the total moment , @xmath96 , agrees well with the powder neutron diffraction result , and the small value of the orbital moment indicates that cro is not close to a @xmath13 state driven by soc as it is also deduced from the calculations in reference @xcite . ( color online ) influence of ti substitution on the anisotropy gap . ( a ) shows energy scans through the anisotropy gap at the antiferromagnetic zone center using the crystal 10ti , l=2 ( blue ) and l=3 ( black ) and 0ti , l=2 ( red ) . in ( b ) the energy of the lower zone - center mode of crystals with different ti content is shown , the data for 1 % ti are from ref . @xcite . ] in fig . [ ti ] the influence of different ti substitutions on the anisotropy gap is addressed by comparing constant * q * scans across the in - plane gap at the antiferromagnetic zone center . the modes are slightly split due to finite inter - layer interaction , see subsection iii.d . even @xmath97 corresponds to the higher and odd to the lower modes , respectively . the lower zone - center magnon energy is displayed as function of ti doping in fig . [ ti](b ) . the gap clearly diminishes with increasing ti content . since the non - magnetic ti dilutes the magnetic lattice a general softening can be expected , as it is visible in the anisotropy gap , which in first view corresponds to the square root of exchange and anisotropy energies . in addition , ti also perturbs the lifting of orbital degeneracy , as it is shown in fig . [ ruooo ] and subsection iii . a. in consequence also the single - ion anisotropy will be reduced with increasing ti content . the impact of a small ti content of the order of one percent , which is used to favor the synthesis of large crystals , can however , safely be neglected . 10ti , which is essentially untwinned , shows a @xmath97 dependence of the anisotropy gap [ fig : [ ti](a ) ] consistent with the previous study @xcite and the discussion given above . ( color online ) additional magnon branch near 5 mev . ( a , b ) show scattering maps of energy versus * q*. ( c ) presents some characteristic scans included in ( b ) for an energy transfer of 3 mev ( 5mev ) , ( 8 mev ) in blue ( red ) , ( green ) . the data in ( a - c ) are taken from ref . ( d ) shows * q * scans with an energy transfer of 5 mev for 10ti ( blue ) and 0ti ( red ) . the intensities are scaled with the maximum intensities of the scans in fig . [ ti ] . scaled in the same way are the energy scans in ( e ) at @xmath98 using the crystals 10ti , @xmath99 ( blue ) , @xmath100 ( red ) and 0ti , @xmath100 ( black ) . ( f ) presents the energy of this additional mode at * * q**=(0,1,0 ) for crystals containing different amounts of ti . ] ( color online ) polarization analysis of the additional mode . sf intensities of energy scans at ( a , d ) @xmath101 , and ( b , e ) @xmath77 for the sfx ( sfy ) , ( sfz ) chanel in blue ( red ) , ( black ) . ( c ) shows the results of the polarization analysis with eq . ( 1 ) , in red m@xmath62 and in blue m@xmath63 . ( d , e ) present only the 5 mev data of the two * q * positions , which are measured with higher statistics . ] the previous unpolarized experiments gave evidence for an additional scattering at the lower energy of 5mev that could not be explained by the magnon dispersion expected for a simple square - lattice antiferromagnet . the data , which are presented in the mappings shown in fig . [ omode2](a , b ) , were obtained on the thermal tas in8 @xcite . panel ( c ) shows some characteristic scans included in ( b ) . the 5 mev signal is much weaker than the in - plane transverse mode in particular when considering the by a factor three reduced energy . furthermore , it is not possible to connect this low - energy feature with a dispersing branch with a slope comparable to that of the transverse branch . instead the intensity of this feature seems to be rather localized in * q * space and to exhibit a flat dispersion . the signal could also be followed along the vertical direction without any indication for finite dispersion . in ref . @xcite it was speculated that this extra mode could arise from the orbital disorder induced by the ti substitution , a scenario which can now be ruled out by the comparison of ti - containing and ti - free cro . [ omode2](d - e ) show * q * and energy - scans across the 5 mev signal on samples with 0 and 10% of ti . most interestingly there seems to be no difference in the strength of the low - energy signal when normalized to that of the in - plane gap mode at @xmath1914 mev . the additional signal is thus intrinsic to the magnetic order in cro , and it is not induced by disorder , which must be significantly enhanced in ti10 . the ti content furthermore does not change the energy of the additional scattering , see the fitted energy maxima displayed in fig . [ omode2](f ) . the minor variation between measurements on the same concentration on different spectrometers can be attributed to the energy calibration . the polarization of the additional mode was determined with polarized neutron scattering following the same procedures as those used to determine the character of the dispersion at higher energies . the polarization analysis directly shows that the additional excitation possesses an in - plane character , because it is not observed in the sfy channel ( or in @xmath102 ) , see fig . [ omodepol ] . comparing the intensities at @xmath101 and @xmath77 one may furthermore deduce that the additional mode is @xmath43 polarized , which is transverse with respect to the static moment . the expected ratio of the intensities at 5 mev between @xmath77 and @xmath101 based on the twinning fractions and the geometry factor is 1:2.5 for an in - plane transverse mode and 2.5:1 for an in - plane longitudinal mode , respectively . the measured ratio is 0.25(21 ) characterizing this mode as an in - plane transverse mode . the simple antiferromagnetic square lattice exhibits a magnon dispersion consisting of two transverse branches , which were both observed at larger energy in cro , see above . in order to explain the additional mode the model needs to be extended , and the weakness and localization of the signal point to a smaller moment with smaller coupling . the gga+u calculations on cro reveal a small magnetic moment situated at the two apical oxygens of 0.11@xmath103 each . such a polarization of the oxygen results from the strong hybridization between ru @xmath104 and o @xmath105 orbitals . dft calculations on several ferromagnetic or nearly ferromagnetic ruthenates find a sizable amount of magnetic moment on o , which in total sum up to about 30 % of the entire magnetization @xcite . experimentally this sizable magnetization of the o states has been observed by polarized neutron diffraction on ca@xmath106sr@xmath107ruo@xmath1 where indeed 33 % of the magnetization stems from o states @xcite . cro , however , exhibits antiferromagnetic order , in which the o bridging two antiparallel ru moments can not be polarized , see fig . [ explanations](a ) . therefore , there is no moment on the in - plane o sites , which is also confirmed by the dft calculations of cro . however , the apical o is connected only to a single ru site and can be polarized . there are two principal contributions to the exchange between ru and o moments due to hopping between o 2@xmath108 and ru 4@xmath9 orbitals in the scenario where the ru @xmath6 orbital is doubly occupied and the @xmath44 and @xmath45 singly occupied , see e.g. ref . one is a hopping from the @xmath109 and @xmath110 to the @xmath44 and @xmath45 orbitals , respectively . there , because of pauli exclusion principle , only antiparallel spins can hop , leaving parallel spins in the now singly occupied o @xmath108-orbitals . in total this yields a ferromagnetic coupling : @xmath111 where @xmath112 is the hopping integral between @xmath44 and @xmath109 , and @xmath45 and @xmath110 orbitals of ru and oxygen . @xmath113 and @xmath114 are the centers of the ru @xmath115 and apical o @xmath108 bands . the factor two in eq . ( [ t2u ] ) reflects the fact that there are two processes ( virtual hoppings from @xmath109 to @xmath116 and from @xmath110 to @xmath45 orbitals ) , contributing to this ferromagnetic ru - o exchange . the second contribution to the ru - o exchange comes from the hopping from the @xmath117 orbital of o to the empty @xmath118 orbital of ru . because of hund s rule , predominantly parallel spins hop , leaving anitparallel spins on the o @xmath117 orbital . this yields an antiferromagnetic coupling between ru and o moments : @xmath119 here @xmath120 is the hopping matrix element between a @xmath117 orbital of o and an empty @xmath121 orbital of ru , cf . @xcite . the @xmath122 hopping is larger than the @xmath123 one @xmath124 ) @xcite , but the antiferromagnetic exchange ( eq . [ t2jd2 ] ) occurs through the @xmath125 band , which lies much higher in energy than @xmath115 . our gga+u calculations show that @xmath126 ev , while @xmath1274.7 ev . there is only one such antiferromagnetic exchange path , in contrast to two for ferromagnetic exchanges . [ t2u ] ) and hund s exchange on the ru , @xmath128 ev , also slightly reduces this exchange ( eq . [ t2jd2 ] ) . therefore , one should expect ferromagnetic coupling to dominate , so that the moment of oxygen would be parallel to the moment of neighboring ru . our dft calculations support this conclusion . ( color online ) magnon dispersion calculated with the spinw program including the additional mode . the labeling corresponds to the one in fig . [ spinw ] . ] we have extended the spin - wave model by a small moment on the apical oxygens , 0.11 @xmath103 , coupled ferromagnetically to the next ru moment , see fig . [ explanations](a ) . this extension necessitates to reduce the single - ion anisotropy parameter @xmath129 in order to keep the anisotropy gap of the in - plane transverse mode at the correct energy . the additional mode can not be obtained at an energy near 5 mev at the antiferromagnetic zone center without introducing an anisotropy in terms of a single ion anisotropy for the magnetic moment on the o sites , or an anisotropic coupling parameter . because the single ion anisotropy stems from soc which is small in o , the first scenario seems to be less likely . a better choice is an anisotropic coupling , because the ru with its larger soc is also involved in this process . therefore we used @xmath130 mev to fit the experimental results . with this anisotropic coupling , @xmath129 has to be reduced to 0.5 mev to keep the in - plane transverse mode at the experimental value . with this extension we can perfectly describe the additional feature , see fig . [ omodedisp ] . the flat dispersion of the additional branch results from the lack of coupling between two o moments and perfectly describes the experimental finding . also the localization of the signal strength at the antiferromagnetic zone center is well reproduced by this model . we can therefore conclude that the additional moment , situated on the apical o site and ferromagnetically coupled to the ru , can explain the additional low - energy magnetic signal in cro . we also tried to reproduce the data with the o moments being antiferromagnetically coupled to ru , but such a scenario can not explain the location of the scattering in q - space . the significant polarization of the o site is another consequence of the strong hybridization of the ru @xmath104 and o @xmath108 orbitals resulting in large hopping parameters , which can already be deduced from the closeness of the metallic phase of cro . it furthermore yields a convincing explanation of the reduced moment , because the in - plane oxygen ion , which should carry a similar amount of charge carriers , can not be magnetically polarized in the antiferromagnetic structure reducing the ordered moment from the simple @xmath8 expectation . the ordered moment may thus also oscillate in length by transferring moment between ru and in - plane o , which can give rise to a longitudinal mode , as an alternative explanation for the observation in reference @xcite . various neutron scattering experiments on partially untwinned crystals of ca@xmath0ru@xmath3ti@xmath22o@xmath1 were performed to determine the magnetic structure and magnon dispersion in the insulating state of cro . most remarkable is the large magnetic anisotropy , which results in a splitting of the two transverse zone - center modes that exceeds the full dispersion . the entire dispersion of transverse magnon branches , however , is perfectly described by a conventional spin - wave model , in which the strong impact of the soc is reflected by large anisotropy parameters . dft calculations within the gga+u approximation yield an orbital moment of only 0.13@xmath103 , which indicates that cro is not close to a @xmath13 state , which can be seen in the discrepancy of the experimental magnon dispersion with the @xmath13 calculation @xcite . the calculated magnetic interaction and anisotropy parameters reasonably well agree with those obtained by fitting the dispersion , and also the ordered moment is correctly reproduced in the dft calculation . the dft study finds a sizable amount of magnetic moment on the apical oxygen reminiscent of previous reports on ferromagnetic ruthenates . this additional moment explains an additional signal appearing in the neutron scattering experiments at lower energy and limited to the antiferromagnetic zone center . extending the spin - wave model to the weakly ferromagnetic coupled oxygen moments describes the flat dispersion and the limited appearance in * q * space of this signal . the spin - wave dispersion in cro is thus dominated by the impact of strong soc and by the presence of magnetic moments on the oxygen sites . this work was supported by the deutsche forschungsgemeinschaft through crc 1238 projects no . a02 and no . b04 and by the russian foundation of the basic research via program 16 - 02 - 00451 , russian president council on science through md-916.2017.2 , faso ( theme `` electron '' 01201463326 ) and mon ( project 236 ) . complete crystal structure analyses were performed with an x8-apex by bruker axs single - crystal diffractometer with a goniometer in kappa - geometry and x - ray radiation from a molybdenum anode with a wavelength of @xmath131 . the distance between the sample and the detector was set to 50 mm . structure refinements were carried out using jana2006 @xcite . a type i extinction correction was applied during the refinements and the data were corrected for absorption . the thermal parameters for ru1 and ti1 were constrained to be equal , and the total occupancy of this site was fixed to 1 . the results of the structural refinements are given in table [ tab : ref ] . table [ tab : char ] presents further characteristics of the crystal structure as well as the metal - insulator- and antiferromagnetic transition temperatures . @xmath54 is determined from the magnetization curves presented in fig . [ magnetization](a ) by finding the zero point of the second derivation . @xmath49 is determined in the same way from the resistance curves presented in this figure . [ cols= " < , < , < , < , < , < , < " , ] [ tab : char ] ( color online ) elastic neutron scattering scans of 1tib across ( a ) @xmath69 , ( b ) @xmath77 , ( c ) @xmath134 , ( d ) @xmath135 , ( e ) @xmath135 . blue color denotes scans taken in the magnetic phase at 10 k and red color scans taken in the non magnetic phase at 150 k. ] the crystal 1tib used in the previous neutron scattering study @xcite was further investigated on the thermal tas in3 at the ill . the crystal was mounted in the [ 010]/[001 ] orientation into a ill orange cryostat , @xmath34 was set to 2.662 @xmath35 for all scans and a pyrolitic graphite filter was used to suppress higher order wave length . rocking scans revealed the mosaic spread to be less than 0.5@xmath21 . with the good resolution of the in3 it is easily possible to resolve the bragg peaks from the two twins present in the crystal . the twinning ratio amounts to 20:1 , which is deduced from rocking scans of strong nuclear bragg reflections at the 2@xmath136 scattering angles of the ( 020 ) and ( 200 ) reflections . figure [ magn - str](a ) shows a longitudinal scan over the ( 020 ) reflection , the smaller twin is barely seen in the tail . scans across the positions , where magnetic scattering can be expected , are shown in ( b - e ) . there several scans of the kind ( 01l ) are compared at 10 k , in the antiferromagnetic phase , and at 150 k , in the non magnetic phase . it is clearly seen , that the minority twin gains intensity for an odd @xmath97 and not for an even @xmath97 . the majority twin gains intensity for even l and much less , but also clearly detectable , for odd l. the ( 012 ) and ( 014 ) reflections are magnetic bragg peaks for the @xmath15-centered phase but not for the @xmath14-centered one . so the strong gain of intensity of these reflections in the antiferromagnetic phase reveals an almost exclusive @xmath15-centered phase . in opposite the ( 011 ) and ( 013 ) peaks are bragg peaks in the @xmath14-centered phase , but not in the @xmath15-centered one . so the gain of intensity at low temperatures of those peaks points to the @xmath14-centered phase , which is however much weaker than the gain of the bragg peak intensities of the @xmath15-centered phase . so the main stacking scheme of this crystal is the @xmath15-centered one . 53ifxundefined [ 1 ] ifx#1 ifnum [ 1 ] # 1firstoftwo secondoftwo ifx [ 1 ] # 1firstoftwo secondoftwo `` `` # 1''''@noop [ 0]secondoftwosanitize@url [ 0 ] + 12$12 & 12#1212_12%12@startlink[1]@endlink[0]@bib@innerbibempty link:\doibase 10.1143/jpsj.66.1868 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.84.2666 [ * * , ( ) ] link:\doibase 10.1103/physrevb.63.174432 [ * * , ( ) ] @noop * * , ( ) link:\doibase 10.1103/physrevlett.93.146401 [ * * , ( ) ] @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) @noop * * , ( ) link:\doibase 10.1103/physrevb.60.r8422 [ * * , ( ) ] link:\doibase 10.1103/physrevb.58.847 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.87.077202 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.104.226401 [ * * , ( ) ] @noop ( ) , link:\doibase 10.1103/physrevlett.91.056403 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.88.017201 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.81.2747 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.98.216403 [ * * , ( ) ] http://stacks.iop.org/0295-5075/63/i=1/a=097 [ * * , ( ) ] link:\doibase 10.1140/epjb / e20020021 [ * * , ( ) ] @noop _ _ ( , , ) link:\doibase 10.1103/physrevlett.111.197201 [ * * , ( ) ] link:\doibase 10.1103/physrevb.90.035137 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.115.247201 [ * * , ( ) ] @noop ( ) , \doibase http://dx.doi.org/10.1016/s0921-4526(99)00989-8 [ * * , ( ) ] link:\doibase 10.1002/crat.201600020 [ * * , ( ) ] \doibase http://dx.doi.org/10.1016/0921-4526(93)90108-i [ * * , ( ) ] http://stacks.iop.org/0953-8984/27/i=16/a=166002 [ * * , ( ) ] , ed . , @noop _ _ ( , ) link:\doibase 10.1103/physrevb.86.060410 [ * * , ( ) ] @noop * * , ( ) link:\doibase 10.1103/physrevb.57.1236 [ * * , ( ) ] link:\doibase 10.1103/physrevb.76.014505 [ * * , ( ) ] \doibase http://dx.doi.org/10.1016/s0921-4534(96)00637-5 [ * * , ( ) ] link:\doibase 10.1103/physrevb.84.201102 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.88.197002 [ * * , ( ) ] link:\doibase 10.1103/physrevb.79.054422 [ * * , ( ) ] link:\doibase 10.1103/physrevb.72.094104 [ * * , ( ) ] link:\doibase 10.1103/physrevb.24.3939 [ * * , ( ) ] link:\doibase 10.1103/physrevx.3.041036 [ * * , ( ) ] link:\doibase 10.1140/epjb / e2016 - 60588 - 6 [ * * , ( ) ] @noop ( ) , @noop @noop _ _ ( , , ) http://www.ncbi.nlm.nih.gov/pubmed/10062328 [ * * , ( ) ] link:\doibase 10.1038/nmat1605 [ * * , ( ) ] link:\doibase 10.1103/physrevb.86.064429 [ * * , ( ) ] http://iopscience.iop.org/0953-8984/9/4/002 [ * * , ( ) ] g. zhang and e. pavarini , phys . rev . b * 95 * , 075145 ( 2017 ) . link:\doibase 10.1103/physrevb.56.2556 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.89.087202 [ * * , ( ) ] @noop _ _ ( , , ) p. @noop * * , ( )

solar - type stars are currently divided into `` exoplanets - host stars '' ( ehs ) , which are known to harbor planets , and `` non - exoplanets - host stars '' ( nehs ) , for which no planet has yet been detected with the available techniques and instruments . this dichotomy is obviously related to exoplanet detection bias , as stars without detected planets may very well host planets in the presently non - detectable range . in this framework , the studies concerning the observed differences between these two categories of solar - type stars must be taken with caution . however , taking this restriction into account , these studies give interesting information , at least on the observed ehs . here we present a theoretical study concerning the fate of the metal - rich ( planetary ) material which may fall onto ehs during their early lives , and we analyze the consequences of such events on the abundances of light elements . we compare our results with the presently available observations , and give some predictions and discussion for future work . the observed ehs are , on average , metal - rich compared to nehs @xcite , although some very metal poor ehs have also been observed @xcite . this average metal excess was a subject of discussion during the last decade , as the star could either has started with a higher metal content than average , or be related to metal - rich accretion onto the star . this last explanation is now ruled out for several reasons including the fact that the metal - rich accreted matter can not remain inside the outer convective zone but falls down inside the star due to thermohaline convection . @xcite gave a first discussion of this specific process . she showed how the accreted matter was mixed down inside the star and left only a very small @xmath0-gradient at the end of the mixing process . it is now proved that this remaining @xmath0-gradient is too small to account for the observed overmetallicity . contrary to @xcite , @xcite suggested that the induced mixing region could go down to the lithium destruction zone . independently , @xcite performed precise computations of thermohaline convection in a static @xmath1 model , with the assumption of accretion of one jupiter planet , using the thermohaline coefficient derived by @xcite ( see below , section 5.7 ) . she confirmed that the remaining metal excess was much too small to account for the observed overabundances , and found that the mixing region extended below the lithium burning region . the efficiency of the thermohaline mixing and the resulting abundances strongly depend on the atomic diffusion processes at work in the stars . for example in the case of carbon enhanced metal poor stars ( cemp ) , which accrete matter while they have already spent some time on the main sequence , the thermohaline mixing is strongly limited by the pre - existing stable @xmath0-gradient , induced by helium settling @xcite . this is not expected to occur in the case of ehs which accrete matter at the very beginning of their lives , when atomic diffusion did not yet have time to built important helium gradients . however , as will be seen below , even in this case atomic diffusion is not negligible and has to be taken into account . we precisely analyze the consequences of metal - rich accretion onto solar - type stars , for various stellar masses and various assumptions of accreted masses . we simulate the induced thermohaline mixing using various prescriptions for the mixing coefficient , including the most recent 2d simulations by @xcite and 3d simulations by @xcite and we compare the results . we show that the situation may be quite complex . in any case , if some metal - rich layer can remain in the stellar outer regions , it is much too small to be observable . this confirms the primordial origin of the average overmetallicity observed in ehs . on the other hand , lithium and beryllium destruction layers are reached in most cases and these elements may be partially destroyed . in the present computations , we do not add any other mixing process which can also destroy lithium and beryllium , like rotation - induced mixing or mixing by internal waves . the reason is that we want to disentangle the various processes acting on the light elements and look at the effects of thermohaline convection only . in the future , these other mixing processes should be added for a complete description of light element evolution in solar - type stars . in section 2 , we give a review of the possible accretion phenomena expected on ehs and their observed evidences ; we also discuss in this section the induced thermohaline process . a review of the observed element abundances in ehs versus nehs stars is given in section 3 . section 4 is devoted to the modeling techniques . section 5 presents our computations and results ; the effects of the various parameters like stellar mass , accretion mass , different prescriptions for the mixing , etc . are compared in detail in this section . finally the consequences of these results are discussed in section 6 , followed by the conclusion . it is now well - established that stars may experience episodic accretion along their evolution through different processes . the planet formation period which extends from the star birth to the end of the pms phase @xcite is a favorable epoch for such events . the interactions of a young planet with its surrounding disk affect the planet s orbital energy and angular momentum . as a consequence the planet may migrate radially towards or away from the central star @xcite . at the end of the protoplanetary disk phase , the planetary system architecture is not necessarily set . current models for planetary system evolution indicate that various instabilities may commonly arise within the first gigayear after the system birth , as it moves towards a stable configuration . during this phase , the system experiences quasi - stable periods interrupted by more or less violent restructuring episodes . these events which may result in orbital reconfigurations , collisions , gravitational captures or ejections , may be favoured and intensified by the presence of asteroids and debris belts . during these catastrophic events , planetesimals , asteroids or debris may be pushed towards the central region and fall onto the inner objects and/or the central star . evidences for the previously described evolution are present in our solar system and comes from a variety of sources . planetary formation theories suggest that the giant planets formed on circular and coplanar orbits different from their current path @xcite . the asteroid and kuiper belts are believed to have been formed with considerably more mass than their present content @xcite , their radial location may have changed over time @xcite and the dynamical structure of the kuiper belt provides evidence for sculpting by planet migration @xcite . the petrology and cratering records on the moon and on telluric planets suggest that a cataclysmic spike in the cratering rate occurred nearly 700 million years after the planets formed @xcite . this event which is referred to as the late heavy bombardment ( lhb ) , is characterized by an intense period of planetesimal bombardment probably originated from the existing asteroid belt @xcite and spreading throughout the inner solar system @xcite . this episode would have last 50 to 200 million years . models and theories have been proposed to explain these observations @xcite . according to the most popular ones , at the end of the protoplanetary disk phase , the gas giant planets were surrounded by a rich and interacting asteroid belt imposing slow planet migrations . as a consequence of these migrations , the occurrence of a 1:2 orbital resonance reached by jupiter and saturn would have been the catalyst for a catastrophic episode . this phenomenon would have lead to a dramatic destabilization and a rapid restructuring of the outer solar system , involving the giant planets and the asteroid belt . as a result of this process , the giant planets would have evolved to their current orbits while most of the components of the asteroid belt would have been ejected from the planetary system or simply pushed towards outer regions ( forming the current kuiper belt ) . however some objects of the belt would have experiment high orbital eccentricity motions allowing them to enter the inner solar system and impact with telluric objects , resulting in the lhb @xcite . the solar system history probably reveals a typical scenario for planetary system evolution . various models indicate that the instabilities encountered by the solar system may be a common occurrence for planetary systems and that this can occur after up to 1 gyr of quasi stable evolution . the presence of solid bodies belts may interfere in the stabilization process of planetary systems . in particular , as stated in the previous section for the solar system case , perturbations of asteroid and/or debris disks may result in the scattering of planetesimals throughout the planetary system , favouring the occurrence of intense impact episodes . as most stars are supposed to form surrounded by protoplanetary disks , they are expected to present remnant belts ( seen as asteroids or debris disks ) resulting from a partial clearing process of their original disks . indeed massive circumstellar disks are observed around several hundred main sequence stars and preferentially around exoplanet host stars ( hereafter ehs ) @xcite . we consequently expect planetary systems to offer preferential conditions for violent bombardment episodes and consequently for accretion event onto the central star . the accretion of planetary ( metal rich ) material onto a star produces a brutal increase of its surface metal content . the metal - rich accreted material is rapidly mixed within the surface convective zone : a metal - enriched zone may then overlay lighter layers which produces an unstable weight gradient in the external layers of the star . a similar situation is often encountered in oceanography , when warm salted water lies upon cool unsalted one . large inverse salinity gradients make warm salted blobs diffuse rapidly downward in spite of the stabilizing temperature gradient . heat diffusing more rapidly than salt , the warm salted blob falling down in cool fresh water experiences a temperature decrease before the salt has time to diffuse out ; the blob continues falling because of its weight , while unsalted matter rises around . this special mixing process which takes the form of elongated cells , the salt fingers " , is called thermohaline convection " . it proceeds until the fingers mix with the surroundings and a stable salinity gradient is restored . a similar type of convection can grow inside stars . it occurs in a stable stratification that satisfies the ledoux criterion for convective stability : @xmath2 but where the molecular weight decreases with depth : @xmath3 with the classical notations , @xmath4 and @xmath5 are respectively the usual adiabatic and local gradients ( @xmath6 ) , @xmath7 and @xmath8 . in the stellar case , the role of the aquatic salinity is then played by the molecular weight . when a blob of heavy matter ( i.e with high molecular weight ) begins to fall down , heat is exchanged with the surroundings more rapidly than particles so that it goes on falling like a finger , forcing light matter to go up instead . this happens every time heavy element accumulation occurs above lighter layers . it leads to short episodes of deep thermohaline mixing which proceed until the molecular - weight gradient becomes too small to trigger the instability . this can be summarized with the following condition ( see @xcite and @xcite ) : @xmath9 where @xmath10 represents the ratio : @xmath11 and @xmath12 is the inverse lewis number , ratio of the molecular diffusivity to the thermal diffusivity . a more complete description of the methods used to model thermohaline convection is given in section 4.2 . thermohaline instabilities due to inverse @xmath0-gradients are expected to arise in stellar interiors in a variety of already identified contexts . this is the case for instance when helium or carbon rich material is deposited at the surface of a star in a mass - transferring binary @xcite . it also develops in stars where radiative levitation produces heavy element accumulation regions @xcite . the thermohaline mixing has also been proposed as a possible mechanism to account for the photospheric composition of low and intermediate mass giant @xcite . in such stars an inverse @xmath0-gradient is created by the @xmath13he(@xmath13he,2p)@xmath14he reaction in the external wing of the hydrogen burning shell . an inverse @xmath0-gradient may also arise after the ignition of @xmath14he burning in a degenerate shell ( i.e. core helium flash ; see @xcite ) . up to now the effects of accretion - induced thermohaline instabilities on ehs have been little studied . they have been discussed by @xcite and more recently by @xcite , but no detailed stellar evolutionary models have yet been computed to investigate the effects of accretion and thermohaline convection on the stellar lithium , which is the aim of the present paper . stars hosting planets compared with other stars of the same spectral types clearly show a metallicity excess of a factor 2 on average . two scenarios have been proposed to explain the metal enrichment of ehs . the first one , the accretion hypothesis , considers that ehs formed with normal abundances but experienced during the planet formation period metal - rich material accretion @xcite . the ingestion of planets , planetesimals , asteroids , comets , dust would then produce a local surface overmetallicity . the second scenario , the primordial hypothesis , assumes that central stars of planetary system form out from metal rich molecular clouds @xcite : the overmetallicity of the protostellar gas would be a necessary condition for planets to form around stars . in this case , the stars would be metal - rich from the center to their surface . several studies have tried to differentiate between the two possible explanations using different techniques including asteroseismology ( e.g. * ? ? ? today a consensus seems to be achieved which favors the primordial hypothesis . several arguments have been invoked which rule out the accretion scenario . it was first thought that the dilution of metal - rich material in stellar external convective zones should lead to metallicity variations depending on the stellar mass , which were not observed @xcite . another argument was that the overmetallicity observed in ehs would require the ingestion of several tens of earth , which is not completely excluded but seems quite high @xcite . more recently , as already discussed in previous section , it was recognized that the overmetallic matter diluted in the surface convective zone would rapidly sink in the radiative interior due to thermohaline instability effects , so that only a small fraction of the accreted matter could remain in the outer layers ( see @xcite and @xcite ) . the potential discrepancy between lithium abundances in ehs and nehs is a controversial subject which has divided the observers community for more than 10 years . @xcite first suggested that ehs present smaller lithium abundances than field stars . they were soon contradict by @xcite and later by @xcite who stated that lithium abundances of stars with detected exoplanets are indistinguishable from those of stars without detected planets . @xcite , who revisited this topic , reported a lithium deficiency in ehs within the effective temperature range 5600k-5850k . their results were confirmed by @xcite and @xcite but contradict by @xcite . more recently @xcite , @xcite and @xcite confirmed an enhanced lithium depletion in sun like stars with orbiting planets but once more these results were contradict by @xcite who attributed the lithium depletion derived by @xcite to a wrong interpretation of observational data or to observational bias . the situation is still more difficult for beryllium . @xcite and @xcite have determined the beryllium abundances in a sample of solar - type stars and tried to determine a difference between exoplanet - host stars and stars without detected planets . they may find an extra beryllium depletion for the coolest stars ( t@xmath15@xmath16 k ) , but this result still has to be confirmed . considering the potential lithium deficiency observed in ehs as true , theoreticians proposed several explanations to account for this `` extra '' depletion . some of them suggest that the lithium destruction and the planet formation both results from special conditions surrounding some stars ( e.g. overmetallicity , long - lived disk ) . others consider that the destruction is directly related to the presence of planets . [ [ metallicity - effects . ] ] metallicity effects . + + + + + + + + + + + + + + + + + + + + @xcite computed stellar evolutionary models including atomic diffusion and rotation - induced mixing . in this framework they investigated the effects of various primordial overmetallicities on the lithium depletion . they showed that the ehs - lithium deficiency can not be attributed to overmetallicity effects : extra - mixing is required below the surface convective zone to account for the li observations . [ [ stellar - rotational - history . ] ] stellar rotational history . + + + + + + + + + + + + + + + + + + + + + + + + + + + @xcite investigated a possible connection between a lithium deficiency , the presence of planets and the rotational history of solar type stars . thanks to rotational evolution models , he showed that contrary to zams fast rotating stars , the slow zams rotators develop strong differential rotation between their surface convective zone and their radiative interior , which leads to hydrodynamical instabilities responsible for enhanced lithium depletion . he concluded that ehs were slow rotators on the zams and that the protoplanetary long - lived disk probably responsible for their slow rotation rate could also be a necessary condition for the planet formation . this could explain the correlation observed between the presence of planets and large lithium destructions . [ [ formation - migration - and - accretion - of - planets . ] ] formation , migration and accretion of planets . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + several authors suggest that the strong lithium depletion observed in ehs could be directly related to the mechanism of formation or migration of planets . this idea is supported by observations from @xcite which show that overmetallic stars without detected planet do not present strong lithium depletion . @xcite suggest that protoplanetary disks lock a large amount of angular momentum during the planet formation mechanism thus inducing some rotational braking in the host star during the pms . the mixing resulting from the induced differential rotation could lead to a large lithium depletion . @xcite also suggests that the lithium destruction could be associated with a late migration of giant planets at the end of the pms . angular momentum transfer from the planetary disk to the outer layers of the star and/or tidal effect of the giant planet on the host star could induce differential rotation and then mixing in the host stars . if the migration process is overly efficient , planets may be accreted onto the star . @xcite investigate the effects of the accretion of jupiter - like planets on the lithium isotopic abundances in main sequence stars , using either standard stellar models or models with a light mixing process induced by gravity waves . they do not take into account any thermohaline convection , so that the stellar outer layers are not mixed with the lithium depleted regions . as a consequence , they obviously find an increase of the lithium abundance and the 6/7 isotopic ratio due to the planet chemical composition . baraffe & chabrier ( 2010 ) explores the effects of early episodic accretion ( i.e. occurring during the pms ) on the internal structure of young ( fully or nearly fully convective ) low mass and solar type stars . they conclude that an early protostar accretion history can affect the internal structure of the star and can severely enhanced the li - depletion in these object . we computed stellar models of solar type stars , using the toulouse - geneva evolution code ( tgec ) . the code is described in detail in @xcite and @xcite . all our models include atomic diffusion computed as described in @xcite and @xcite . the selective diffusion of a large number of individual elements important for the stellar structure is computed with respect to hydrogen , including the effects of the local pressure gradient , temperature gradient , concentration gradient and radiative transfer . at each time step , all the abundances are renormalized to account for the correlated motion of hydrogen and the induced modifications of the opacities are taken into account . when macroscopic motions occur , which can be described in terms of a macroscopic diffusion coefficient , the mixing effect is added as a new term in the diffusion equation , acting on the abundance gradient , as described in many papers ( e.g. @xcite , @xcite , @xcite , @xcite , @xcite , @xcite , etc . ) . in the tgec code the main isotopes of the following elements are treated separately : he , li , be , b , c , n , o , ne , mg , ca , fe . the other species are treated together . for each separate isotope i , the diffusion equation is written : @xmath17 where c@xmath18 is the concentration of the isotope and @xmath19 the diffusion flux . in atomic diffusion theories , @xmath20 is traditionally called the `` diffusion velocity '' . the expression is deduced from detailed theories of transport processes in multicomponent non - uniform gases ( e.g. * ? ? ? * ) . for test atoms , it may be written : @xmath21 where the first term is the classical diffusion term and the three other ones are respectively the pressure , thermal and radiative diffusion terms ( e.g. eq . 2 of * ? ? ? d@xmath18 is the atomic diffusion coefficient and d@xmath22 the mixing diffusion coefficient . in the cases where atomic diffusion is negligible , these equations reduce to the classical diffusion ( mixing ) equation . in the following , we discuss how accretion and thermohaline convection are introduced in the code . we assume that stars could swallow planetary material during their early main sequence life . the accretion of a given amount of matter is modeled through an increase of the metal content of the surface convective region . the accreted mass is expressed in jupiter mass ( hereafter @xmath23 ) or earth mass unit ( hereafter @xmath24 ) ; its composition is assumed similar to the solar mixture determined by @xcite but devoid of hydrogen and helium . the composition of the surface convective zone is computed again after each accretion episode : the mass fractions of the chemical elements are computed taking into account the newly accreted matter but the structure of the star is supposed unaltered by the accreted material . in the case of large amounts of accreted matter , the @xmath0-gradient is large enough at the beginning to destabilize the medium by dynamical convection . this is modelled by a very large diffusion coefficient simulating complete mixing . thermohaline mixing begins when the @xmath0-gradient is small enough for the medium to be stable against dynamical convection . the prescriptions used to describe this mixing are discussed below . the effects of thermohaline convection as a mixing process in stars are far from trivial . @xcite and @xcite ( hereafter referred to as krt ) proposed to model the thermohaline mixing process in terms of a diffusion coefficient proportional to the inverse @xmath0-gradient : @xmath25 which can simply be written : @xmath26 where @xmath27 is thermal diffusivity , @xmath28 is the pressure scale height , @xmath29 the radiation constant , @xmath30 the speed of light , @xmath31 the temperature , @xmath4 the adiabatic gradient , @xmath32 the specific heat at constant pressure , @xmath33 the opacity . the variable @xmath34 is a factor involving the aspect ratio ( length / width ) of the fingers . krt and @xcite considered different geometries for the fingers and , as a consequence , they deduced different @xmath34 values ( 12 for krt and 658 for @xcite ) . in their computations of red giant stars , @xcite considered this parameter as free and adjusted it ( @xmath35 ) to reproduce the abundance observations in evolved stars . on the other hand @xcite used the original krt value with @xmath36 for the computations of diffusion induced thermohaline convection . a large step forward has recently been acomplished owing to the numerical simulations performed by @xcite ( 2d ) and @xcite ( 3d ) . they both deduced from their simulations of fingering convection a near - unit aspect ratio of the fingers much smaller than that used by @xcite . in the intermediate regime , their different expressions for the thermohaline mixing coefficients lead to values very close to the krt ones , if computed with @xmath34 = 12 . important differences appear only for the boundary values of the @xmath10 ratio , when @xmath37 , which corresponds to the case of very small @xmath0-gradients , or @xmath38 , which corresponds to the largest possible @xmath0-gradients before dynamical convection . the @xcite coefficient may be written as : @xmath39 which reduces to the krt one for @xmath40 , but vanishes for @xmath41 and goes to infinity when @xmath42 . @xcite ( hereafter tgs ) have performed 3d simulations of thermohaline convection with various sets of physical parameters , and found that the ratio of the turbulent to the diffusive fluxes settled onto a single universal profile for all parameter values , which seems a robust result . they derive a completely empirical expression for the mixing coefficient which , at first sight , looks quite different than the previous ones , but present also some similarities : @xmath43 where , in stellar conditions for which @xmath44 , @xmath45 . after some manipulations , we find that in the intermediate regime this mixing coefficient may be approximated as @xmath46 , which is slightly smaller than the krt one with @xmath34 = 12 . once again , important differences only appear for the behavior in the extreme cases . here @xmath47 vanishes for @xmath41 but it reaches a finite value when @xmath42 . this last point will be discussed in section 5.7 . an illustration of this discussion may be found in [ coefficients ] where we present a graph of the krt and tgs mixing coefficients precisely computed in one of our models . in the following , we first present results obtained with the krt prescription using @xmath48 = 12 , and we also do computations with the assumption @xmath48 = 1000 for comparison purposes . we show that in this case , the resulting lithium and beryllium depletions are too high to be realistic , which confirms that this coefficient is overestimated . then we present computations obtained with the new tgs empirical expression and we compare it with the previously obtained results . we made computations for stars with masses between 0.70 and 1.30m@xmath49 . we tested various initial metal contents with metallicity values ranging from [ fe / h]=-0.20 to + 0.20 . we computed models with a solar initial helium mass fraction and models with an initial helium abundance varying with metallicity ( y@xmath50=y@xmath51 ) according to the galaxy chemical evolution law @xcite . one to fifteen accretion events were introduced in our models with various accretion rates for a total accreted mass ranging from 1m@xmath52 to 1.5m@xmath53 . we present here a choice of our models , representative for all the computations done . in this section we describe in detail the accretion / thermohaline mixing process , and present results obtained with the krt coefficient ( @xmath34 = 12 ) . as a first example , we present the result obtained for a 1.10 m@xmath49-model with y@xmath50=y@xmath51=0.29 and [ fe / h]=0.20 . this model experiences a single accretion event of 0.03m@xmath53 two million years after its arrival on the zams . this accretion rate is of the order of a giant planet core , which represents a realistic value for the mass of heavy material collected by the star through an impact with a jupiter like object . figure [ mudkip1]a displays the molecular weight profile in the outer regions of the model at various epochs during the `` accretion / mixing period '' ( which refers to the period starting with the accretion episode and ending when the thermohaline mixing stops ) . the instantaneous dilution of a 0.03m@xmath53 planetesimal in the external convective zone produces an enhancement of the surface metal content of the model . the dotted ( purple ) line on figure [ mudkip1]a represents the theoretical metallicity increase produced by the dilution in the convective zone of the impacting object . this metallicity increase creates a sharp positive @xmath0-gradient at the transition between the radiative interior and the convective envelope which triggers thermohaline instabilities . the induced mixing acts to soften the unstable @xmath0-stratification . it occurs simultaneously with the atomic diffusion and in particular the downward migration of helium which also contributes to restore a stable @xmath0-gradient . when a flat @xmath0-gradient is reached below the convective zone , thermohaline convection becomes inefficient and a stable @xmath0-gradient induced by helium settling begins to take over . thermohaline convection still occurs below this stable region . as will be seen in forthcoming sections , this effect is less important for larger accretion rates ( section 5.3 ) . figure [ mudkip1]b displays the diffusion coefficient used to parametrized the thermohaline mixing at the same epochs as presented on figure [ mudkip1]a . the mixing starts right after the accretion event , when an inverse @xmath0-gradient is created at the transition between the radiative and convective zones . it mixes a localized region restricted to the unstable @xmath0-stratified layers lying below the convective zone . this process smooths down the unstable @xmath0-gradient in the mixed zone , it decreases the diffusion coefficient but it also continuously widens the unstable region . the bottom of the mixed zone deepens until a stable stratification can be reached : the mixing then stops and the diffusion coefficient drops to zero . the phases during which the accretion mixing proceeds have strong consequences for light elements like li and be as discussed in the following sections . we now examine the case of multiple successive accretion events . we present the results obtained for two 1.10m@xmath49 ( with the same @xmath54 and [ fe / h ] values as previously ) . in both cases the models experience ten accretion events . the first accretion episode occurs 2 myr after the arrival on the zams , and the following ones occur successively every 2 myr . in the first model , the amount of accreted matter is 1m@xmath52 ( @xmath550.003m@xmath53 ) at each episode . in the second model , the amount swallowed each time is 0.03m@xmath53 , of the order of a giant planet core , which represents a realistic value for the mass of heavy elements collected by the star through an impact with a jupiter like object . figure [ dkip100a ] ( a and b ) and figure [ dkip100b ] ( a and b ) present @xmath0 and d@xmath56 profiles inside our models at various steps along the accretion / mixing period . as already discussed , the first accretion event induces a surface @xmath0 increase which triggers the thermohaline instabilities . the resulting mixing tends to smooth down the unstable stratification before the next accretion event occurs . the next successive accretion episodes continuously increase the inverse @xmath0-gradient , which feeds the mixing and increases the width of the mixed zone . after 2 accretion events , the helium settling becomes unable to counteract and freeze the thermohaline mixing before the next event occurs . after the last accretion episode , the thermohaline convection still proceeds , until a stable @xmath0-gradient is restored . during that process , a small stable region develops just below the convective zone . this is the reason why , on the top right graph , one curve , drawn after 51 myr , present a @xmath57 which vanishes slightly below the convective zone . the time required to restore a stable stratification depends on the accretion rate : the higher the accretion rate , the longer the mixing period and the larger the mixed region . the model experiencing the 1m@xmath52 accretion rate recovers a stable @xmath0-stratification at 39 myr ( i.e. 37 myr after the first impact ) while the model with the 0.03m@xmath53 accretion rate is stable at 51 myr ( 49 myr after accretion beginning ) . as shown on figures [ dkip100a ] and [ dkip100b ] , the top of the li - depleted region is close to the bottom of the surface convective zone . the thermohaline mixing then allows connecting the stellar surface with the lithium nuclear burning region and is therefore expected to produce a li depletion . figures [ dkip100a]c and [ dkip100b]c represent the lithium profiles inside the two considered models . in both models the first accretion episode does not deplete significantly the lithium surface abundance : the mixing - induced lithium destruction is compensated or even exceeded by the lithium amount brought by the impacting object . the next events increase the unstable @xmath0-gradient , as a result the diffusion coefficient is enhanced and the mixed region extends towards deeper regions : as shown on figures [ dkip100a]c and [ dkip100b]c , this leads to large lithium destructions . we will see in the next section that similar features are obtained for be ( the be depletion is however expected to be smaller than the li one since its nuclear burning region lies deeper in the stellar interior ) . in this section , we test the effects of various accretion rates on the surface metallicity and the li and be depletions . in this framework , figure [ fesurh ] displays the variations of the surface metallicity over the accretion / mixing period for the two series of models ( including the two models presented in the previous section ) . all these models have the same mass ( 1.10m@xmath49 ) and initial parameters ( in particular the same chemical composition ) but they experience various accretion scenarios : 1 , 5 or 10 accretion events of 1m@xmath52 ( 0.003m@xmath53 ) for the first series of models and 1 , 5 or 10 accretion events of 0.03m@xmath53 for the second series . each accretion episode induces a metallicity increase which obviously depends on the accretion rate . for the considered accretion rates and number of impacts , the overmetallicity does not exceed 0.05 dex . after the end of the accretion period , the overmetallicity decreases under the combined effects of thermohaline mixing and atomic diffusion which makes heavy elements diffuse out of the surface convective zone . when the mixing stops ( i.e. before 60 myr for all the presented models ) , the remaining metallicity increase is smaller than 0.02 dex and is unable to explain the average overmetallicity of 0.2 dex observed in ehs . figure [ libesurf ] shows the evolution of the li and be surface abundances over the accretion / mixing period . the results demonstrate that the accretion of planetesimals may decrease the surface li and be abundances , with a depletion rate depending on the accretion rate and the number of impact events ( the li and be depletion increasing with the accretion rate and/or the number of impacts ) . the be nuclear depleted zone being deeper that the li one , the obtained be destructions are smaller than the li ones . models . @xmath58 $ ] is the difference between the current surface metallicity and its initial value . dashed lines are models experiencing ( from bottom to top ) : 1 , 5 or 10 accretion events of 1m@xmath52 . solid lines are models experiencing ( from bottom to top ) : 1 , 5 or 10 accretion events of 0.03m@xmath53.,title="fig:",scaledwidth=2.0% ] models . @xmath58 $ ] is the difference between the current surface metallicity and its initial value . dashed lines are models experiencing ( from bottom to top ) : 1 , 5 or 10 accretion events of 1m@xmath52 . solid lines are models experiencing ( from bottom to top ) : 1 , 5 or 10 accretion events of 0.03m@xmath53.,title="fig:",scaledwidth=48.0% ] in this section we present models with various masses and initial metallicities . all the presented models undergo 5 accretion episodes of 0.03m@xmath53 following the same scheme as previously discussed ( same time intervalle between the zams and the first impact and then between successive accretion episodes ) . as in previous sections , the thermohaline mixing is still computed using the krt coefficient with @xmath59 . figure [ lisurfz ] presents the surface li abundance during the accretion / mixing period in two sets of models computed with different initial metal contents . for each metallicity several masses are presented . for a given metallicity , the li - depletion induced by the combined effects of accretion and mixing decreases with increasing mass . several competing effects lead to this result . the mass of the surface convective zone decreases as the stellar mass increases , as a result the dilution of a metal rich planetesimal in the external convective zone should lead to larger unstable @xmath0-gradients and then to more efficient mixing and larger li - depletion in more massive stars . however other competing effects lead to a different result . the li amount brought through the ingestion of a metal - rich planetesimal increases the li surface abundance of the target star . this increase is more significant in more massive stars ( due to their narrow convective zone ) . as a result more massive stars enter the mixing period with higher li surface abundance . moreover as the mass increases the top of the li - depleted region moves away from the bottom of the convective zone . larger mixed regions are then needed to connect the surface and the li - nuclear burning zone . both these effects ( the accretion - induced li increase and the growing distance between the convective zone and the li burning region ) make the li depletion smaller in more massive stars . figures [ mz1]a and b compare the li depletion in models with the same mass but with different metallicities . it shows that the accretion / mixing induced li depletion increases with the initial metal content . for a better understanding of this result we present on figure [ mmec ] the surface convective mass as a function of the stellar mass for zams models with various metallicities . it shows that the mass of the surface convective zone increases with metallicity . this feature confirms that several competing effects leads to larger depletion in stars with larger surface convective zones . up to now , we presented results obtained using the prescription proposed by krt with the adjustment @xmath59 . as already discussed , this coefficient is in much better agreement with the results of recent numerical computations @xcite than the one proposed by @xcite , using @xmath60 . we nevertheless computed models using this large value of @xmath34 and give the results below for comparison . figure [ dkip1a ] present the results obtained with @xmath60 for a 1.10m@xmath49 model ( with y@xmath50=0.29 and [ fe / h]=0.20 as previously presented models ) , undergoing ten accretion episodes of 0.03m@xmath53 . the accretion sequence is the same as previously : it starts 2 myr after the zams beginning and the successive accretion events are separated by 2 myr . the models presented on figures [ dkip100b ] and [ dkip1a ] differ only through the value used for the @xmath61 factor . a larger @xmath61 value increases the diffusion coefficient and consequently produces a larger mixing . the 2 myr - mixing occurring in - between two successive accretion events is more efficient in smoothing the unstable stratification . consequently the succession of several accretion episodes produces smaller unstable gradients in models with larger @xmath61 values . this feature is indeed observed ( cf . figures [ dkip100b]a and [ dkip1a]a ) . a retroactive effect of smaller @xmath0-gradients is to decrease the diffusion coefficient ( cf . [ dth ] ) . the mixing efficiency then results from the competition between two opposite effects . the larger @xmath34 tends to increase the diffusion coefficient and the mixing efficientcy , thereby leading to smaller @xmath0-gradient , but on the other hand these smaller unstable gradients tend to decrease it . for models computed with @xmath60 . dashed lines are models experiencing ( from bottom to top ) : 1 , 5 or 10 accretion events of 1m@xmath52 . solid lines are models experiencing ( from bottom to top ) : 1 , 5 or 10 accretion events of 0.03m@xmath53.,title="fig:",scaledwidth=2.0% ] for models computed with @xmath60 . dashed lines are models experiencing ( from bottom to top ) : 1 , 5 or 10 accretion events of 1m@xmath52 . solid lines are models experiencing ( from bottom to top ) : 1 , 5 or 10 accretion events of 0.03m@xmath53.,title="fig:",scaledwidth=48.0% ] figure [ dkip1fesurh ] illustrates the variations of the accretion - induced overmetallicity during the accretion / mixing perio , using the large diffusion coefficient ( for comparison with figure 4 ) . figure [ dkip1blibesurf ] shows the lithium and beryllium depletion due to accretion and mixing . as previously discussed they depend on the accretion rate and the number of accretion events but for a given accretion scenario , the obtained destructions are larger than those obtained with a smaller diffusion coefficient ( cf . figure [ libesurf ] ) . in previous sections , we discussed either the cases of single accretion events , or the cases of several accretion events occuring every 2 myr . when a new event occurs , the element abundances are first modified inside the convective zone due to the chemical composition of the accreted matter , and then they go on varying due to the thermohaline mixing . in the case of lithium , each new event leads to a rapid lithium increase followed by a slower decrease induced by thermohaline convection and nuclear burning . obviously the final results depend on the amount of accreted matter but also on the time scale between two events . if another accretion occurs before the complete settling of the previous one , the effects overlap . in this section , we discuss the effects on lithium of changing the time interval between successive accretion events . we have tested , for accretion amounts of 0.03m@xmath53 , the cases of events separated by 1 myr and 4 myr . the results are presented in figure [ mumdkipdt ] for comparison with figure [ dkip100b ] . figure [ lisurfdt ] shows the surface lithium abundance variation with time , together with the previously studied case of 2 myr . the left graph displays the results for five events and the right graph for ten events . in each case , two series of models are computed : one with @xmath34=12 and the other one with c@xmath62=1000 . the final results are not trivial . for larger time intervals , the @xmath0-gradients before every new accretion event are smaller , so that the diffusion coefficients and thus the mixing efficiencies are also smaller , leading to less lithium destruction . on the other hand , the total mixing episodes last longer , in favor of more lithium destruction . figure [ lisurfdt ] shows that these two combined effects can lead to opposite results according to the circumstances . for @xmath34=12 , the results obtained with the three different time intervals are very similar , although lithium is slightly less depleted for increasing intervals . for @xmath34=1000 , the situation is different if accretion stops after five or ten events . for five events , the results are also very similar but this time the case of events separated by 1 myr lead to slightly less depletion than the other cases . on the other hand , for ten events , the situation is similar to the case described for @xmath34=12 , except that the difference is more important . an interesting point in this case is that the end of the accretion period is clearly visible in the graph . it corresponds to the time when no fresh lithium is brought anymore , so that the lithium abundance variation is only due to nuclear destruction . as already mentionned , the expressions for the diffusion coefficients deduced from recent numerical simulations lead to results close to the krt ones with @xmath36 , except for the boundaries of the thermohaline regime . here we are basically interested in the @xmath42 limit , which corresponds to the boundary between the dynamical convection and thermohaline mixing . in the krt prescription , for which @xmath47 becomes very large at this limit , there is a simple continuity between the two regimes . after accretion , if @xmath63 , the region below the standard convective zone becomes dynamically unstable , which leads to complete mixing and may be modelled by a very large diffusion coefficient . as the @xmath0-gradient decreases , @xmath64 increases . when it reaches unity , thermohaline convection begins . one thus expects a continuity between the diffusion coefficients , which means that when @xmath64 becomes slightly larger than one , @xmath47 should rapidly decrease from a very large value to the value obtained in the intermediate regime . this description is well fitted by the krt coefficient , but not by the tgs one where @xmath47 remains finite as it reaches unity . however , a close look at their figure 2 , which presents the turbulent heat and compositional fluxes as a function of their rescaled density ratio @xmath65 , reveals a clear tendency for the numerical simulation points to go up to large values for @xmath66 , corresponding to @xmath42 . we thus infer that their should be a small transition region between the tgs value and the very large one needed for @xmath63 , consistently with the simulations . on the other hand , the value of the mixing coefficient as given by tgs drops to zero at the other limit , for @xmath67 , corresponding to @xmath41 . this feature , which is not present in the krt coefficient , is consistent with the fact that @xmath47 must become zero for @xmath68 . figure [ coefficients ] displays the krt and the tgs coefficients precisely computed as a function of @xmath10 in one of our models ( m=1.10m@xmath49 , [ fe / h]=0.20 at 4 myr ) . this graph is similar to figure 3 of @xcite . the difference comes from the fact that tgs used constant @xmath69 , @xmath27 and @xmath70 in their computations , whereas we use the exact values computed in our model . these values vary with depth and with age , as well as the @xmath0-gradient . the two coefficients are very similar in the region between @xmath71 and @xmath72 , i.e. between @xmath73 and @xmath74 , with @xmath75 . for @xmath10 close to one , @xmath76 has a finite value which would produce a discontinuity with the case @xmath63 . for this reason , we have added a transition region in our computations ( solid black line , label @xmath57 ) . for the limit where @xmath10 becomes close to @xmath77 , the tgs coefficient decreases rapidly , contrary to the krt one . also shown on this graph is the coefficient used by @xcite , labelled @xmath78 . here we present results obtained with the tgs diffusion coefficient , with a very small transition region below @xmath79 , similar to the krt one , to preserve the continuity with the dynamical regime . figure [ garaud ] displays the lithium and beryllium abundance variations with time for various accretion scenarios , which must be compared with figure 5 . we can see that the obtained lithium and beryllium depletions are very similar , although slightly smaller , that those obtained with the krt prescription with @xmath80 . in the worst case , the relative difference is of order @xmath81 . on the contrary , the depletions obtained with @xmath82 are much too large . .s@xmath83 ) computed in a 1.10m@xmath49 model with [ fe / h]=0.20 at 4 myr . long dashed - small dashed ( green ) line : krt coefficient ; dashed ( blue ) line : tgs coefficient ; solid ( black ) line : tgs coefficient including a transition region for @xmath10 close to one ; dotted line : coefficient used by @xcite.,scaledwidth=50.0% ] we have investigated the effects of planetesimal accretion onto solar - type stars . we have shown that the dilution of metal - rich material in the surface convective zone of young main sequence stars creates an unstable @xmath0-gradient at the transition between the convective and radiative zones . the induced double diffusive instabilities , the so - called thermohaline convection , result in a mixing which dilutes the metal - rich matter until the @xmath0-stratification becomes stable . the thermohaline convection effects combined with the stabilizing action of atomic diffusion softens the inverse @xmath0-gradient on a few millions or tens of million years , depending on the considered accretion scenario . note that this time scale is much larger than the one suggested by @xcite . the reason is that , following krt , she used as a characteristic length the evaluated fingers dimension , whereas the real length to be used is that of the whole mixed region . at the end of the accretion / mixing period , only a very small metallicity increase may remain , much smaller than the average overmetallicity observed in exoplanet - host stars . this result is consistent with a primordial origin of this overmetallicity . on the other hand , as suggested by @xcite , the thermohaline mixing allows connecting the stellar surface with the li - nuclear burning region , which may lead to rapid li - depletion depending on the accretion rate and the number of impacts experienced by the star . the li destruction depends on many different parameters linked to the stellar structure and to the rate , frequency and amount of the accretion episodes . concerning the stellar structure , the lithium depletion is obviously larger for deeper convective zones , which is related to the stellar mass and metallicity . it also depends on the age at which the accretion occurs , because the diffusion - induced @xmath0-gradients act as stabilizers for the thermohaline convection . concerning the accretion process itself , we have shown that different results are obtained not only for different accreted amounts , but also according to the way it is accreted : all at ones , or in successive episodes . different stellar characteristics and various accretion scenarios produce , at the end of the accretion / mixing period , a significant dispersion in the li abundances . the results also depend on the computational parameters used for the treatment of convection . in the present computations , the convective zone is computed whithin the framework of the mixing length theory , with a mixing length ratio adjusted on the solar models ( @xmath84 = 1.8 ) . no overshooting nor extra - mixing except thermohaline convection is added . thermohaline convection is treated using the diffusion approximation . we have tested several prescriptions for the diffusion coefficient . recent 3d numerical simulations by @xcite allowed real improvements in the modelling . we have shown that the corresponding prescription for the diffusion coefficient leads to results very similar to the krt one , provided that the aspect ratio to the fingers , that is the ratio of their length to their width , be taken equal to one as suggested by the simulations ( see also @xcite ) . this corresonds to a numerical coefficient @xmath36 in the krt expression . as the diffusion coefficients are larger for larger @xmath0 gradients , which themselves decrease faster when the diffusion processes are stronger , the final lithium depletion is not simply related to the initial diffusion coefficient value . here we have tested three cases : the krt coefficient with @xmath36 , the same coefficient with @xmath35 , as suggested by @xcite for the rgbs and the coefficient deduced from numerical simulations by tgs . we confirmed that the krt , @xmath36 coefficient gives results similar to the tgs one . @xcite performed computations of thermohaline convection in a @xmath1 static model , using the tgs coefficient . no atomic diffusion was introduced in her computations . we can see from her results , and particularly from the values of the diffusion coefficient and the mean molecular weigth presented in her figure 3 , that our results and hers are similar , of the same order of magnitude . this convergence is quite encouraging . the general trend of lithium depletion with age in various galactic clusters may be accounted for by rotation - induced mixing and/or internal waves in a somewhat satisfactory way ( e.g.@xcite ) . the large dispersion observed in stars of similar intrinsic parameters like mass , age , rotation velocity , chemical composition needs a more complicate process to be accounted for @xcite . as discussed by @xcite , the details of the rotation history of stars can help explaining these different features . here we suggest that the accretion / thermohaline mixing process may also be an important parameter in the description of the lithium dispersion in solar - type stars , and that it should be taken into account . \1 ) if one assumes that accretion of planetary matter onto solar type stars is a common phenomenon , thermohaline convection should often occur . on the other hand , if most of the lithium abundance behavior in solar type stars is well accounted for by rotation - induced mixing and related processes , as generally believed , it leads to constraints on the accretion efficiency . using the realistic tgs or krt thermohaline coefficients , we find that the stars must not accrete more than ten earth masses after their arrival on the main sequence , to keep their original lithium value within 0.1 dex . \2 ) stars which accrete more than ten earth masses are expected to be subject to stronger lithium depletion . as already pointed out by @xcite and @xcite , this process could be an elegant explanation for the observations of stars in which strong lithium depletion is found . note that this does not only apply to those stars for which planets have been detected , as stars may well host planets even if they have not yet been detected because of observational bias . \3 ) in any case , the observed overmetallicity in exoplanets - host stars is not due to accretion , and it is not possible either to explain any increase of lithium or of the @xmath85li/@xmath86li ratio by accretion , because of the thermohaline dilution . \4 ) the solar case deserves a more precise analysis . whether accretion at the period of the late bombardment has participated in the presently observed lithium depletion or not depends on the amount of matter swallowed by the sun , and also on the diffusion - induced helium gradient at the time of the event . \5 ) finally , a complete study of lithium depletion in solar - type stars needs a combined computation of this process with other depletion processes like rotation - induced mixing . complete evolutionary sequences including accretion , thermohaline mixing and rotational mixing will be investigated and presented in a forthcoming paper . baumann , p. , ramirez , i. , melendez , j. , asplund , m. & lind , k. 2010 , , 519 , a87 bazot , m. & vauclair , s. 2004 , , 427 , 975 bouvier , j. 2008 , , 489 , l53 bryden , g. , beichman , c. a. , carpenter , j. m. , rieke , g. h. , stapelfeldt , k. r. , werner , m. w. , tanner , a. m. , lawler , s. m. , wyatt , m. c. , trilling , d. e. , su , k. y. l. , blaylock , m. & stansberry , j. a. 2009 , , 705 , 122 cantiello , m. & langer , n. 2008 , in iau symposium , eds . l. deng & k. l. chan , vol . 252 , 103 castro , m. , vauclair , s. , richard , o. & santos , n. c. 2009 , , 494 , 663 chapman , s. & cowling , t. g. 1970 , the mathematical theory of non - uniform gases . an account of the kinetic theory of viscosity , thermal conduction and diffusion in gases , cambridge : university press charbonnel , c. , vauclair , s. , maeder , a. , meynet , g. & schaller , g. 1994 , , 283 , 155 charbonnel , c. & zahn , j .- p . 2007 , , 476 , l29 charbonnel , c. & lagarde , n. , 2010 , , 522 , a10 chen , y. q. & zhao , g. 2006 , , 131 , 1816 chiang , e. i. , lithwick , y. , murray - clay r. , buie , m. , grundy , w. & holman , m. 2007 , in protostars and planets v , p. 895 delgado mena , e. , israelian , g. , gonzalez hernandez , j.i . , santos , n.c . & rebolo , r. 2011 , , 728 , 148 denissenkov , p. a. 2010 , , 723 , 563 eggleton , p. p. , dearborn , d. s. p. & lattanzio , j. c. 2006 , science , 314 , 1580 eggleton , p. p. , dearborn , d. s. p. & lattanzio , j. c. 2008 , , 677 , 581 ford , e. b. & chiang , e. i. 2007 , , 661 , 602 galvez - ortiz , m.c . , delgado mena , e. , israelian , g. , gonzalez hernandez , j.i . , santos , n.c . , rebolo , r. & ecuvillon , a. 2011 , arxiv : 1104.0512 garaud , p. 2011 , , 728 , 30 gomes , r. , levison , h. f. , tsiganis , k. & morbidelli , a. 2005 , , 435 , 466 gonzalez , g. 1998 , , 334 , 221 gonzalez , g. 2008 , , 386 , 928 gonzalez , g. & laws , c. 2000 , , 119 , 390 gonzalez , g. , laws , c. , tyagi , s. & reddy , b. e. 2001 , , 121,432 grevesse , n. , & noels , a. 1993 , in origin and evolution of the elements , eds kubono , s. and kajino , t. , 14 hartmann , w. k. , ryder , g. , dones , l. & grinspoon , d. 2000 , in origin of the earth and moon , edited by r.m . canup and k. righter and 69 collaborating authors . tucson : university of arizona press . , p.493 - 512 hui - bon - hoa , a. 2008 , , 316 , 55 ida , s. & lin , d. n. c. 2008 , , 673 , 487 israelian , g. , santos , n. c. , mayor , m. & rebolo , r. 2004 , 414 , 601 israelian , g. , delgado mena , e. , santos , n. c. , sousa , s. g. , mayor , m. , udry , s. , domnguez cerdea , c. , rebolo , r. & randich , s. 2009 , 462 , 189 - 191 izotov , y. i. & thuan , t. x. 2004 , , 602 , 200 kippenhahn , r. , ruschenplatt , g. , & thomas , h .- c . 1980 , , 91 , 175 king , j.r . , schuler , s.c . , hobbs , l.m . & pinsonneault , m.h . 2010 , 710 , 1610 korn , a.j . , grundahl , f.,richard , 0 . 2006 , 442 , 657 kring , d. a. & cohen , b. , a. 2002 , journal of geophysical research ( planets ) 107 , 5009 lubow , s. h. & ida , s. 2010 , astroph - arxiv:1004.4137 to appear in `` exoplanets , '' ed . s. seager , univ . arizona press luck , r. e. & heiter , u. 2006 , 131 , 3069 michaud , g. , richer , j. , fochard , o. 2005 , , 623 , 442 montalbn , j. & rebolo , r. 2002 , , 386 , 1039 morbidelli , a. , levison , h. f. , tsiganis , k. & gomes , r. 2005 , , 435 , 462 morbidelli , a. , levison , h. f. & gomes , r. 2008 , in the solar system beyond neptune , m. a. barucci , h. boehnhardt , d. p. cruikshank , and a. morbidelli ( eds . ) , university of arizona press , tucson , 592 pp . , p.275 - 292 murray , n. , chaboyer , b. , arras , p. , hansen , b. & noyes , r. w. 2001 , , 555 , 801 obrien , d. p. , morbidelli , a. & bottke , w. f. 2007 , 191 , 434 pinsonneault , m. h. , in light elements in the universe , eds . c. charbonnel , m. tosi , f. primas & c. chiappini 2011 , proceedings of the international astronomical union , iau symposium , volume 268 pinsonneault , m. h. , depoy , d. l. s. & coffee , m. 2001 , , 556 , 59 richard , o.,vauclair , s. , charbonnel , c. , dziembowski , w. 1996 , , 312 , 100 richard , o. , thado , s. , & vauclair , s. 2004 , soph , 220 , 243 ryan , s. g. 2000 , , 316 , l35 santos , n. c. , israelian , g. & mayor , m. 2001 , , 373 , 1019 santos , n. c. , israelian , g. , mayor , m. , rebolo , m. & udry , s. 2003 , , 398 , 364 setiawan , j. , klement , r. , j. , henning , t. , rix , h .- w . , rochau , b. , rodmann , j. & schulze - hartung , t .. 2010 , science , 330 , 1642 siess , l. 2009 , , 497 , 463 sousa , s.g . , fernandes , j. , israelian , g. & santos , n.c . 2010 , 512 , l5 stancliffe , r. j. , glebbeek , e. , izzard , r. g. & pols , o. r. 2007 , , 464 , l57 stancliffe , r. j. 2010 , , 403 , 505 stern , a. a. 1996 , 112 , 1203 stothers , r & simon , n. r. 1969 , , 157 , 673 strom , r. g. , malhotra , r. , ito , t. , yoshida , f. , kring , d. a. 2005 , science 309 , 1847 tagle , r. 2008 , 36th annual lunar and planetary science conference , march 14 - 18 , 2005 , in league city , texas , abstract no.2008 takeda , y. & kawanomoto , s. 2005 , , 57 , 45 takeda , y. , kawanomoto , s. , honda , s. , ando , h. & sakurai , t. 2007 , , 468 , 663 talon , s. & charbonnel , c. 2005 , 440 , 981 terquem , c. e. j. m. l. j. 2010 , eas publications series , eds t. montmerle , d. ehrenreich , & a .- m . lagrange , vol . 41 , pp.209 thado , s. & vauclair , s. 2003a , , 587 , 784 thado , s. & vauclair , s. 2003b , , 587 , 795 thado , s. , vauclair , s. , alecian , g. & leblanc , f. 2009 , , 704 , 1262 thado , s. , bohuon , e. & vauclair , s. 2010 , in light elements in the universe , eds . c. charbonnel , m. tosi , f. primas , c. chiappini , proceedings of the international astronomical union , iau symposium , volume 268 , 427 thomas , h .- c . 1967 , , 67 , 420 thomas , h .- c . 1970 , , 6 , 400 thommes , e. w. , bryden , g. , wu , y. & rasio , f. a. 2007 , 675 , 1538 thommes , e. w. , duncan , m. j. & levison , h. f. 1999 , 402 , 635 thompson , i. b. , ivans , i. i. , bisterzo , s. , sneden , c. , gallino , r. , vauclair , s. , burley , g. s. , shectman , s. a. & preston g. w. 2008 , 677 , 556 traxler , a. , garaud , p. & stellmach , s. 2011 , , 728 , 29 tsiganis , k. , gomes , r. , morbidelli , a. & levison , h. f. 2005 , 435 , 459 ulrich , r. k. 1972 , , 172 , 165 vauclair , s. 1988 , , 335 , 971 vauclair , g. , vauclair , s. , michaud , g. 1978 , 223 , 920 vauclair , s. , vauclair , g. 1982 , , 20 , 37 vauclair , s. , thado , s. 2003 , , 587 , 777 vauclair , s. 2004 , , 605 , 874 wyatt , m. c. , clarke , c. j. & greaves , j. s. 2007 , 380 , 173

large protein - protein interaction data sets @xcite and functional information about many proteins are increasingly available . this allows one to investigate the patterns in protein - protein interactions that enable proteins to act concertedly to carry out their functions . in particular , considerable recent attention has been given to the modularity of the cell s functional organisation @xcite . a module is often thought of as a group of components that carry out a functional task fairly independently from the rest of the system . it is thought that such modules yield robust and adaptable systems @xcite . there is also much suggestive evidence that modules within the cell are themselves the building blocks of a higher level of structural organisation ( e.g. @xcite ) . within the networks literature a great many algorithms have been proposed that locate dense regions in a network , often called communities ( reviewed in @xcite ) . a community is loosely defined as a group of nodes that are more closely associated with themselves than with the rest of the network . such communities are potentially good candidates for functional modules , and many studies report running one of the myriad algorithms for detecting community structure on protein interaction networks @xcite . having located communities , such studies then attempt to assess their functional homogeneity by searching for terms in a structured vocabulary usually the gene ontology ( go , @xcite ) or munich information centre for protein sequences categories ( mips , @xcite)that are significantly over - represented within communities . if such terms exist , the identified communities are said to be ` enriched ' for biological function . in many studies such enriched communities are found , and hence are plausible candidates for biological modules . recently there has been an acknowledgement that many community detection algorithms in particular all those that rely on optimising the quality function known as modularity impose an artificial _ resolution limit _ on the communities detected @xcite . such algorithms return communities found at one particular resolution i.e. at one particular scale within the network whereas there are many scales of potential functional relevance within the protein interaction network . for example , one might expect to find smaller communities embedded inside progressively larger ones @xcite . there are now algorithms available that include a ` resolution parameter ' , which allow one to uncover structure at many different resolutions @xcite . however , no study to our knowledge has systematically applied such an algorithm and analysed the results across different resolutions in protein interaction networks ( one study reports testing more than one value of a parameter akin to the resolution on a protein interaction network , in order to select an optimal value for their purposes @xcite ) . in this study , we probe the functional relevance of communities at multiple resolutions ( scales ) in the yeast protein interaction network , for two main biological reasons . first , considering the whole proteome , it is possible to view how the network breaks into communities ( hierarchically or otherwise ) , and to investigate whether some scales of organisation are of more relevance than others biologically . second , the relationship of multi - scale community structure to a particular protein is of interest : it is possible to see which other proteins co - occur with it at different resolutions perhaps it co - occurs robustly with a small group of proteins at high resolution but also with a larger set of proteins at a lower resolution . both groups are of potential interest in understanding what role the protein plays . this is particularly pertinent for poorly annotated proteins , as their patterns of potential function can be revealed through clustering into communities @xcite . although it is already thought that communities have some relationship to functional modules , here we expand on previous work to assess the functional relevance of communities in four main ways . first , assessing functional relevance by counting over - represented terms amongst a group of proteins is not a sufficiently stringent test of functional relevance when the group of proteins in question is a community . this is because two proteins that interact are functionally more similar than a randomly chosen pair of proteins , so one must control for the number of interactions when assessing the biological relevance of a community ( which will necessarily include more interacting pairs than a randomly selected group of proteins ) . we therefore control for the number of interacting proteins found in a community . second , instead of assessing functional homogeneity on a term by term basis we use all the annotations available within a given ontology . third , go and mips are subjective by their nature , both in the definition of the sets of terms themselves and in the process of annotation of terms to proteins . due to their role in a particular process , a protein might well be both annotated more fully and have a higher probability of having had protein interaction experiments performed on it . therefore , in addition to using go and mips as protein functional characterizations , we use a single high - throughput experiment on the growth rates of gene knock - out strains under various conditions ( using data from @xcite ) . fourth , protein interactions are of two fundamentally different types . the molecular interactions ontology @xcite recognises two distinct types of interactions : physical associations ( henceforth denoted @xmath0 ) and associations ( henceforth denoted @xmath1 ) . the main experimental type for the former are yeast - two - hybrid screens ( e.g. @xcite ) . the main type of experiment to fall under the latter are based on tandem affinity purification ( tap , e.g. @xcite ) . these interaction types are known to have very different properties @xcite . additionally , the networks constructed using these two types of interactions have quite different global properties ( see table @xmath2 ) . we thus investigate the two networks , based on type @xmath1 and type @xmath0 interactions , independently . we identify communities at multiple resolutions in these two fundamentally different interaction networks . we then use novel tests to determine the communities functional homogeneity using three different characterisations of function . as the functional knowledge of proteins is far from complete ( even for well characterised organisms such as yeast ) , we also search for topological properties of communities that are correlated with functional homogeneity . in our study we find many functionally homogeneous communities at multiple network resolutions . almost all proteins are in functionally homogeneous communities at some resolution ( @xmath3 of @xmath4 proteins in the @xmath1 network , and @xmath5 of @xmath6 proteins in the @xmath0 network ) . the resolution that places most proteins in functionally homogeneous communities is beyond the ` resolution limit ' , or standard resolution , discussed above . at this maximum , @xmath7 out of @xmath4 proteins are in functionally homogeneous communities according to our go similarity measure in the @xmath1 network . communities at this resolution have mean size @xmath8 , compared to mean size @xmath9 at the standard resolution . we find similar numbers for the @xmath0 network . additionally , we find a high degree of overlap between communities judged functionally homogeneous using three separate quantifications of functional similarity . through a further characterization of the communities using @xmath10 topological properties , we identify the mean clustering coefficient of a community as a good predictor of functional homogeneity , with a true positive rate of @xmath11 achievable with a false positive rate of @xmath12 . in addition to these proteome - scale results , we demonstrate via examples how this approach can be used to predict groups of proteins likely involved in similar processes to a particular protein of interest . additional files can be found at http://www.stats.ox.ac.uk/research/proteins/resources . here we use the biogrid ( www.thebiogrid.org , downloaded january 2010 , @xcite ) , intact ( www.ebi.ac.uk/intact , downloaded january 2010 , @xcite ) and mint databases ( mint.bio.uniroma2.it/mint , downloaded january 2010 , @xcite ) to assemble our protein interaction networks . we use only interactions between proteins that have an sgd identification ( saccharomyces genome database , www.yeastgenome.org ) . we divide interactions on the basis of their type ( @xmath1 or @xmath0 ) and hence assemble the two networks ( see additional file @xmath2 for details ) . of the potential @xmath13 proteins in the yeast proteome ( www.yeastgenome.org ) , there are @xmath14 proteins connected by @xmath1 type interactions , and @xmath15 connected by @xmath0 type interactions . here we only study the largest connected component of these networks , leaving @xmath4 proteins in the @xmath1 network and @xmath6 in the @xmath0 network . some summary statistics for the two amalgamated networks are shown in table @xmath2 . the @xmath1 network is denser , and has higher clustering . there are @xmath16 interactions in common between the @xmath1 and the @xmath0 networks . we apply the potts method @xcite . it partitions the proteins into communities at many different values of a resolution parameter , thus finding communities at different scales within the network . the method seeks a partition of nodes into communities that minimises a quality function ( ` energy ' ) : @xmath17 where @xmath18 is the community of node @xmath19 , @xmath20 is the kronecker delta , @xmath21 is the resolution parameter , and the interaction matrix @xmath22 gives an indication of how much more connected two nodes are than one would expect at random ( i.e. , in comparison to some null hypothesis ) . the energy @xmath23 is thus given by a sum of elements of @xmath24 for which the two nodes are in the same community . optimising @xmath23 is known to be an np - complete problem @xcite , so one must use a computational heuristic . here we use the greedy algorithm discussed in @xcite and freely available ( www.lambiotte.be/codes.html ) , which performs well against various benchmark tests @xcite . the interaction matrix @xmath24 has elements @xmath25 where the matrix @xmath26 with elements @xmath27 is the adjacency matrix . in this case @xmath28 if proteins @xmath19 and @xmath29 interact , and @xmath30 otherwise . the matrix @xmath31 with elements @xmath32 defines a null model , against which we are comparing the network of interest . here we choose the standard _ newman - girvan _ null model @xcite , which has the property that it preserves the node degree sequence . that is , @xmath33 where @xmath34 is the degree of node @xmath19 , and @xmath35 is the number of edges in the network . when @xmath36 , @xmath23 is the standard newman - girvan modularity quality function , upon which many community detection algorithms are based @xcite . we hence refer to this value of the resolution parameter as the standard resolution . values of @xmath37 probe the network at resolutions above the resolution limit . we investigate partitions of the network in the range @xmath38 , and sample at intervals of @xmath39 on a logarithmic scale ( we hence report results for @xmath40 ) . at @xmath41 , all nodes in our set will be assigned to the same community . as we increase @xmath21 , communities split and become smaller . if we allow @xmath21 to increase until all of the entries in @xmath42 are negative , then each node will be assigned to its own community . it is impossible to uniquely quantify similarity in biological function . here we rely primarily on the go ( www.geneontology.org ) , which provides the most comprehensive available database of functional annotations . we use the biological process sub - ontology annotations to yeast , which are maintained by the sgd consortium @xcite . terms are related to each other through a directed acyclic graph ( dag ) ( see additional file @xmath2 figure s@xmath2 for a visualisation of this structure ) . proteins are annotated with the most specific terms that are known about them . it is then possible to add to this set their parent terms by following the structure of the dag , up to the root node . well - characterised proteins are those annotated with terms far from the root node . of the @xmath43 yeast proteins in the go annotation set , @xmath44 have biological process annotations ( excluding the root node ) . we carried out the same tests using the molecular function and cellular component sub - ontologies , which gave similar results . we also use mips terms ( www.helmholtz-muenchen.de/en/ibis , @xcite ) , which are a useful double check on our results from go , and have the added advantage that the terms are all found at the same level within the hierarchy of terms . here we only use the top level of the mips hierarchy . following @xcite , we quantify the functional similarity between two proteins @xmath19 and @xmath29 by finding the set of go terms annotated to both proteins and counting the total number of proteins , @xmath45 , that share that set of terms . we then define a similarity measure between proteins @xmath19 and @xmath29 as @xmath46 where @xmath47 is the total number of proteins . if both proteins are annotated with a set of terms that few proteins share , then they will be judged as functionally similar under this measure . unlike many other measures , @xmath48 does not penalise proteins for lack of annotation when judging their similarity . this is desirable , as we know that the go annotations ( even for the well - characterised _ s. cerevisiae _ ) are far from complete . the quantity @xmath49 is similarly defined through equation [ eq : sim_measure ] for the mips annotations . the benefit of using a pairwise similarity measure that takes into account the full set of functional information available , rather than examining enrichment of function on a term by term basis , is that the measure has the potential to capture more general functional similarities between a pair of proteins . we also define a similarity between two proteins from a single high - throughput experiment via the growth rates of knock - out strains under a range of different conditions . using the data in @xcite , we define @xmath50 , the correlation in growth rates of the strain with gene @xmath19 knocked out to the strain with gene @xmath29 knocked out under @xmath51 different conditions : @xmath52 where the elements of the vector @xmath53 are @xmath54 the parameter @xmath55 is the mean growth rate of strain @xmath19 under different control conditions , and @xmath56 is the growth rate under one of the @xmath51 treatment conditions . we use the results from the homozygous strains . because many gene deletions are lethal , there is only data available for @xmath57 proteins , of which @xmath58 are in the @xmath1 network and @xmath59 are in the @xmath0 network . as mentioned previously , a fair test of the functional homogeneity of a community must take into account the fact that a pair of proteins that interact will be more similar than a randomly chosen pair . standard enrichment tests do not take this into account , as they compare enrichment in a group of proteins , in this case a community , to what one would expect to attain from a randomly chosen set of proteins @xcite . a community necessarily contains many more interacting pairs than a randomly chosen set . we thus compare the pairwise functional similarities of all interacting pairs of proteins in a community to the same measure for all interacting pairs in the network , thereby controlling for the number of interacting pairs . to capture the pairwise similarity between two proteins that interact @xmath60 , we use @xmath61-scores : @xmath62 where @xmath63 stands for one of our three similarity measures ( based on go , @xmath64 , mips , @xmath65 , or correlated growth rates , @xmath66 ) , @xmath67 is the mean and @xmath68 the standard deviation of all of the elements of @xmath63 for which proteins @xmath19 and @xmath29 interact in the network of interest ( @xmath1 or @xmath0 ) . a desirable quality for our test of functional homogeneity is the ability to compare communities found at different resolutions in an even handed manner . it is inherent in the nature of a statistical test that the significance of the test statistic under consideration ( for example , the difference between the sample mean and the population mean ) depends on the sample size : if one has a larger sample size , one can judge smaller differences to be ` significant ' . to determine the aggregate @xmath61-score , @xmath69 , for the mean of a set of individual @xmath61-scores , @xmath70 , one calculates @xmath71 , where @xmath47 is the number of @xmath70s and @xmath72 is their mean @xcite . so , given a @xmath72 , a larger and hence more significant @xmath69 is achieved for a larger sample size ( i.e. larger @xmath47 ) . in order to separate out the effects of the number of interactors in the community from functional homogeneity , we thus choose to base assessment of functional homogeneity on the @xmath72 , in our case @xmath73 ( @xmath74 is defined in equation [ eq : zs ] ) . we judge as ` significant ' all those communities that have @xmath73 above @xmath75 , and call such communities `` functionally homogeneous '' . we stress that this is not strictly an assessment of statistical significance , as we are choosing to ignore sample size . the value of @xmath75 would be judged to be significant at the @xmath76 significance level for any community with @xmath77 or more interacting pairs . we investigate @xmath10 topological properties of the identified communities and assess whether any of these can be used to identify functionally homogeneous communities . examples include mean clustering coefficient , betweenness measures , and network diameter . any topological properties that correlate well with functional homogeneity can then be used to predict functionally homogeneous communities . we use each topological property as a classifier by predicting communities as functionally homogeneous when the value of that property is above a threshold , which we vary to construct a receiver operating characteristic ( roc ) curve . an roc curve plots the number of communities correctly predicted as functionally homogeneous versus the number falsely predicted @xcite . we calculate the area under the roc curve ( auc ) for each metric at each value of @xmath21 , and report the mean of this quantity over resolutions between @xmath78 ( we exclude @xmath79 , as the results are very noisy due to the small number of communities present ) . an auc of @xmath80 would be expected from a random classifier . aucs of greater than @xmath80 imply that higher values of the metric are predictive of functional homogeneity . aucs of less than @xmath80 imply predictive power if _ below _ a threshold of that particular property was used ( i.e. that the property and functional homogeneity are negatively correlated ) . community structure , if of any biological relevance , should uncover patterns that are more than the sum of effects from pairs of interacting proteins . in table @xmath81 we show the pairwise similarity of proteins in each network under our three different measures of functional similarity ( based on go , mips , and correlated growth rates ; see methods ) . the similarity of pairs known to interact with either @xmath1 or @xmath0 type interactions is much higher than a randomly chosen pair of proteins under all three measures . this both helps motivate the investigation of the connection between functional similarity of proteins and the topology of the network , and demonstrates the necessity of taking into account pairwise properties when assessing any additional information that one can gain by studying communities . and @xmath0 networks*. communities identified in the yeast protein interaction network for interactions of a ) type @xmath1 and b ) type @xmath0 . when the resolution parameter @xmath21 is very small , all nodes are assigned to the same community ( which is analogous to viewing the network at a great distance ) . as @xmath21 is increased ( viewing the network at progressively closer distances ) , more structure is revealed . the figures on the right hand side show visualisations of the networks partition into communities at three different values of @xmath21 . each circle represents a community , with size proportional to the number of proteins in that community , positioned at the mean position of its constituent nodes . ( these positions were determined via a standard force directed network layout algorithm @xcite . ) the shade of the connecting lines is proportional to the number of links between two communities . the main figure shows the communities that we find as we vary the resolution . we identify communities as the same through changing resolution parameter , and hence colour them the same , according to a convention described in additional file @xmath2 ( only communities of size @xmath82 or more are shown ) . note that the ordering of proteins is not the same in the two figures . , width=453 ] figure @xmath2 shows the communities that we find in the @xmath1 and @xmath0 yeast networks as the resolution parameter @xmath21 is varied . as @xmath21 increases , more and smaller communities are found ( see table @xmath83 ) . at @xmath84 ( i.e. @xmath85 ) , which corresponds to standard newman - girvan modularity @xcite , most communities contain a few hundred proteins . by @xmath86 however , almost all proteins are in communities of size three or smaller . as shown in figure @xmath2 , some sets of nodes are classified in the same community through large changes in the resolution parameter and hence represent particularly inter - connected parts of the network . figure @xmath2 should be contrasted with figures s@xmath81 in additional file @xmath2 , which are similar calculations on a random network and a network designed to possess strong communities . in the former , not much structure is present , in the latter , there are very distinct blocks . figure @xmath81a illustrates for the @xmath1 network the number of communities of size four or more as the resolution changes , and figure @xmath81b shows how many proteins are in those communities . ( figure s@xmath83 in additional file @xmath2 is the same plot for the @xmath0 network , and shows similar behaviour ) . the two networks , @xmath1 and @xmath0 , contain very different types of interactions , and they can therefore be used to identify different aspects of the cell s functional organisation . the @xmath1 network is also much denser than the @xmath0 network . @xmath1 interactions would therefore dominate the clustering into communities , thereby making it very hard to pick out any structures given by @xmath0 type interactions ( as occurs in @xcite ) . when considering a particular protein or set of proteins , comparisons between communities found in the @xmath1 and @xmath0 networks can be made , see the examples section . global comparisons between the partitions of the @xmath1 and @xmath0 networks at a particular resolution are not necessarily meaningful as , for example , the size of communities depends both on the size and other properties of the network . data files containing the @xmath1 and @xmath0 networks and the community membership of proteins at multiple resolutions are available at http://www.stats.ox.ac.uk/research/proteins/resources . network a ) the number of communities of size four or more and b ) the number of proteins in such communities and the fraction of these that are judged functionally homogeneous . * a ) the number of communities with changing resolution parameter ( solid black curve ) b ) the number of proteins @xmath87 in communities of size four or more ( solid black curve ) . also shown are the numbers of communities / proteins in such communities judged to be functionally homogeneous according to the go similarity measure ( green curves ) , the mips measure ( dot - dashed blue curves ) and the correlated growth similarity measure ( dashed red curves ) . at values of @xmath88 , relatively few proteins are in communities judged to be functionally homogeneous . the equivalent figure for the @xmath0 network is given in additional file @xmath2 figure s@xmath83.,width=453 ] we now assess how many communities are judged functionally homogeneous , looking in particular at how our results vary with resolution parameter . figure @xmath81a illustrates the number of communities judged to be functionally homogeneous , and figure @xmath81b shows the number of proteins in communities judged to be functionally homogeneous . both are for the @xmath1 network . we find that the large communities present at small values of the resolution parameter @xmath21 are not judged to be functionally homogeneous . as @xmath21 is increased , larger numbers of proteins occur in functionally homogeneous communities , peaking in the range @xmath89 . at @xmath90 , the mean community size is @xmath8 proteins , and the majority of proteins , @xmath7 of @xmath4 , are in functionally homogeneous communities as judged by our go similarity measure . the shapes of the curves of both figure @xmath81a and b for all three similarity measures are very similar . indeed , we find that the overlap between the communities judged to be functionally homogeneous between any two of the three measures is high ( see figure s@xmath91 in additional file @xmath2 ) ; for example , it is @xmath11 between the go and correlated growth rates measure over almost the entire range of the resolution parameter in both @xmath1 and @xmath0 networks . given that the correlated growth similarity measure represents a very different data type to the go and mips annotations , this agreement gives us confidence in the similarity measure we use for go and mips . as we use only the top level of the mips functional annotations , we capture less information than the go measure , so it is unsurprising that fewer communities are found to be functionally homogeneous under this measure . the @xmath0 network ( see figure s@xmath83 in additional file @xmath2 ) shows a similar pattern to the @xmath1 network . one difference is that communities start to be judged as functionally similar at a slightly lower resolution . this is most likely due to the different topological properties of the @xmath0 network . that there are comparably many functionally homogeneous communities in the @xmath0 network as the @xmath1 network is of interest , as communities found in @xmath0 networks are found to be poor choices for predicting function on the basis of enrichment of terms @xcite . for almost all proteins , there is some value of the resolution parameter that assigns them to a functionally homogeneous community . in fact @xmath3 out of @xmath4 @xmath1 proteins and @xmath5 and of @xmath6 @xmath0 proteins are in such communities at some value of the resolution parameter . for a given protein , it may not be that it interacts most closely with proteins involved in the same process . indeed it is often necessary to look at a larger scale , placing the community in a bigger community in order to identify the biological processes it participates in . whether or not this is the case , and which network scale ( resolution ) is most indicative of the processes a protein is involved in , will depend on the particular protein one is interested in . this demonstrates the biological motivation for investigating community structure at multiple resolutions , and suggests the desirability of a method to easily identify those communities most likely to be functionally homogeneous . we might expect proteins involved in particular processes to show different propensities to lie in functionally homogeneous communities . we focus on a small but broad set of protein types , which are the go biological process terms within the yeast go slim @xcite that are annotated to at least @xmath92 yeast proteins . there are @xmath93 such terms , which are listed in additional file @xmath2 , as well as the numbers of proteins annotated to each . we investigate what fraction of each type of protein lie in communities judged functionally homogeneous under the go measure through changing resolution parameter . figure @xmath83 illustrates for the @xmath1 network these percentages for four particular processes . ( figure s@xmath94 in additional file @xmath2 shows the same figure for all @xmath93 terms for the @xmath1 network and separately for the @xmath0 network ) . proteins of some types are far more likely to be found in functionally homogeneous communities than others . for example , for both the @xmath1 and @xmath0 networks , proteins involved in chromosome organisation are far more likely to be found in functionally homogeneous communities than proteins involved in lipid metabolism . in addition , there are some indications that the resolutions of most interest can depend on the type of protein under investigation . as can be seen in figure @xmath83 , proteins involved in rna metabolic processes are more likely to be found in functionally homogeneous communities at @xmath95 , where the mean size of communities is @xmath77 . in contrast , proteins involved in vesicle - mediated transport are found in greater numbers in functionally homogeneous communities at @xmath96 , where the mean size of communities is @xmath97 . , of particular types that are in functionally homogeneous communities in the @xmath1 network , with changing resolution parameter . with changing resolution parameter proteins of particular types have consistent differences as to how often they are found in functionally homogeneous communities . for example , proteins involved in chromosome organisation are far more likely to be in functionally homogeneous communities than proteins involved in metabolism . there are also some features that suggest ` good ' resolutions for particular processes . for example , a good resolution for proteins involved in vesicular mediated transport would be @xmath98 ( for which the mean size of communities is @xmath97 ) , whereas for proteins involved in rna metabolic processes , @xmath95 would be better ( the mean size of communities is @xmath77).,width=453 ] network at resolution parameter value @xmath99 , with nodes ( proteins ) coloured according to the partition of this community at @xmath100 . the colours are the same as for figure @xmath2a , where this group of proteins has labels roughly in the range @xmath101 . almost all of the nodes have some relationship to the ribosome . the proteins in the yellow community are mostly ribosomal subunits , those in the red community are mostly pre - cursors to and processors of the small ribosomal subunit , and those in the blue community have similar roles to those in the red community but for the large subunit . the shading of the links has no significance ; its purpose is to ease visualisation . black nodes are not located in one of the three largest communities discussed in the text . b ) a representation of a community at @xmath100 , with nodes ( proteins ) coloured according to the partition of this community at @xmath102 . the proteins identified at the lower resolution almost all play some role in transcription initiation . at the higher resolution , more structure is revealed : the pink community consists mostly of proteins from the rna polymerase ii mediator complex and the green community mostly consists of proteins from the tfiid and saga complexes . c ) the partition at a higher resolution ( @xmath103 ) . the green community from b ) has split into the saga complex ( green ) and the tfiid complex ( orange ) . the names and descriptions of the proteins in these example communities are given in additional file @xmath81 . the node positions for visualisation were computed in the same way as for figure @xmath2.,width=453 ] consider the community at @xmath85 that is marked as the blue block in figure @xmath2 for the @xmath1 network ( over node labels approximately @xmath104 to @xmath105 ) . this contains @xmath106 proteins and consists largely of proteins with some relationship to the ribosome ( based on short protein descriptions found on the sgd website ) . figure @xmath91a shows this community , where we have coloured nodes according to the community partition at the later partition @xmath107 . the colours red , yellow , and blue are the same as in figure @xmath2 , where most of the community present at @xmath99 has split into three communities at @xmath100 . the blue community consists of @xmath108 proteins , which are largely precursors to and processors of the large ribosomal unit . the red community consists of @xmath109 proteins , which have a similar function but for the small ribosomal subunit . the yellow community has @xmath110 proteins , @xmath111 of which are constituents of the ribosome and the remainder of which are either of unknown function or associate to the ribosome . we give short descriptions of the proteins in these communities in additional file @xmath81 . an illustration of the biological relevance of community structure at three partitions is given in figures @xmath91b and c. we show a community of @xmath112 proteins at @xmath100 , and display its partition into communities at b ) @xmath102 and c ) @xmath103 . almost all of the proteins in the community at @xmath100 play some role in transcription initiation . at @xmath102 this community has split into two main smaller communities : the pink community contains constituent proteins of the rna polymerase ii mediator complex and the green community contains components of the closely related saga and tfiid complexes . at @xmath103 , this second community has split into the saga and tfiid complexes . multi - resolution community detection and characterisation is relevant both from the global viewpoint , where one can investigate the aggregate functional organisation of the proteome , and from the local perspective , where the community membership of particular proteins can be traced through changing resolution parameter . we thus now consider a protein - centred view of multi - resolution community detection . we consider , for an example protein , the properties of the communities to which it is assigned through changing resolution parameter , see figure @xmath94 . the size of the communities , their mean similarity under the @xmath64 and @xmath66 measures , and the mean clustering coefficient are shown . the protein is a member of the escrt - i complex . ( figure s@xmath113 in additional file @xmath2 gives a further four examples . ) note the very robust properties of the communities in the @xmath1 network over resolution parameter values of approximately @xmath114 , despite the tendency for them to be partitioned as @xmath21 increases . at these resolutions , the protein is in the same community as other members of the complex , as well as a few other very closely associated proteins . beyond @xmath115 , the complex is broken up , as reflected in the drop in mean similarity values . the community present over @xmath116 in the @xmath0 network contains many proteins associated to the complex ( in addition to the complex itself ) . above the step observable at @xmath117 , only members of the complex are present . in additional file @xmath81 , we give the names and brief functional descriptions of proteins that occur in some of the same communities for this example , and the four other examples given in additional file @xmath2 . these five examples all show the following behaviour . * in general , as would be expected , the size of the community to which a protein is assigned decreases with increasing resolution . there is often a large range of resolutions over which the community has constant size ( which we have observed in practice to entail the same community across multiple resolutions ) . such communities are particularly resilient to being split up at increasing resolutions , despite the tendency for them to be partitioned . * the community similarity under the @xmath64 , @xmath66 and @xmath65 measures often shows a close correlation . * at higher resolutions , there tends to be a higher community similarity , as might be expected of a hierarchically organised system . this is , however , not always the case : community similarity can decrease at higher resolutions . in these instances , a group of proteins has been partitioned beyond the point at which function is shared , possibly through the exclusion of proteins involved in the same processes that do not necessarily directly interact with each other . * there is often a large overlap between the community membership in the @xmath1 and @xmath0 networks , but it can also be quite different . for example , in additional file @xmath2 figure s@xmath113c , the protein occurs with other proteins in the same complex in the @xmath1 network , whereas in the @xmath0 network it occurs with non - complex members which are nonetheless involved in the same process . the functional homogeneity of communities can also be different : sometimes the protein occurs in many functionally homogeneous community in the @xmath1 network and not the @xmath0 , and sometimes vice versa . this is unsurprising given the very different nature of @xmath1 and @xmath0 interactions . by treating them separately , we are able to pick out both types of pattern . -score under the go measure ( solid green curve ) , and correlated growth measure ( dashed red curve ) with changing resolution for the @xmath1 network ( top ) and @xmath0 network ( bottom ) . long plateaus in these properties represent robust communities . we give further examples in additional file @xmath2 figures s@xmath113.,width=453 ] almost all proteins are in functionally homogeneous communities at some value of the resolution parameter , and we therefore devise a method to swiftly identify these resolutions , especially if there is a dearth of functional information . we investigate whether any easily - calculated topological properties of the communities can act as indicators of functional homogeneity . given a protein of interest we can then use such measures to quickly identify good resolutions , without the need to assess functional homogeneity . we tested @xmath10 topological properties for their ability to predict functional homogeneity using the auc metric ( see methods ) , and show our results in table @xmath91 . in general , the aucs for the @xmath0 network are lower than those for the @xmath1 network , perhaps because there is more potentially usable information in the @xmath1 network as it is significantly denser ( see table @xmath2 ) . .*topological metrics tested and aucs . * the network topology measures tested and their associated aucs . we report the results for using each of these as a predictor for functional homogeneity as judged under the three measures of functional similarity ( go , @xmath64 , correlated growth rates , @xmath66 , and mips , @xmath65 ) for both the @xmath1 and @xmath0 networks . the aucs are given as the average performance over the range @xmath78 . the clustering coefficient ( definition given in the text , equation [ eq : clustering ] ) is the best predictor in all cases . ( the topological properties were computed from code developed by gabriel villar . ) [ cols="<,^,^,^,^,^,^ " , ] we find that clustering coefficient is the most useful of the topological properties tested in the prediction of functional homogeneity for all three similarity measures and in both the @xmath1 and @xmath0 networks . the clustering coefficient of a network is a measure of the mean local clustering around nodes : a node has a high clustering coefficient , @xmath118 , if its neighbours are also neighbours of each other @xcite . it is defined for each node as @xmath119 where @xmath120 is the number of triangles of which the node is a member , and @xmath121 is the number of connected triples of which the node is a member . ( a connected triple is a single node with edges running to an unordered pair of other nodes . ) figure @xmath113 shows the roc curve for using the mean clustering coefficient of nodes in a community as a predictor of functional homogeneity for each of the three similarity measures in the @xmath1 network . ( see methods for a description of the construction . the corresponding figure for the @xmath0 network is given in additional file @xmath2 , figure s@xmath122 . ) network . * the receiver operating characteristic ( roc ) curve for using mean clustering coefficient as a predictor of functional homogeneity under the go measure ( solid green curve ) , mips measure ( dot - dashed blue curve ) and correlated growth measure ( dashed red curve ) . we plot the false positive rate ( fpr ) versus the true positive rate ( tpr ) . a random classifier would give the solid black line . for the go measure , a true positive rate of @xmath11 is achievable with a false positive rate of @xmath12 . the best predictive ability is achieved for the go measure , and the worst for the mips measure ( see table @xmath91 for areas under the curves ( aucs)).,width=453 ] there is some element of discretion for annotating @xmath1 type interactions , i.e. deciding which pairs to list interactions between following experiments , with the principle competing models referred to as ` matrix ' and ` spoke ' @xcite . this choice could cause artefactual topological features , so the extent to which we find particular topological features correlating with functional homogeneity could be sensitive to annotation choice . we are therefore encouraged that the same trends in predictive ability are evident in the @xmath0 network , for which there is no such element of discretion . as can be seen from figure @xmath94 and the figures in additional file @xmath2 figure s@xmath113 , clustering appears to be a good proxy for functional homogeneity when looking at individual proteins , and in the absence of much functional information could guide which resolution(s ) should be targeted for investigation . if protein interaction networks are to aid understanding of how biological function emerges from the concerted action of many proteins , then it is crucial to explore connections between network structure and biological function . in this paper we investigate how the function of sets of proteins varies with network community structure of yeast at multiple resolutions . we find that community structure does indeed help identify sets of proteins that act together , and that this connection between network structure and biological function depends on what network scales are probed . we do not expect there to be any single scale of interest in this middle - scale structure of the protein interaction network ; although previous studies have applied community detection algorithms to protein interaction networks , no study to our knowledge has investigated this structure at multiple resolutions . we find that @xmath3 of @xmath4 proteins in the @xmath1 network , and @xmath5 of @xmath6 proteins in the @xmath0 network , are in functionally homogeneous communities at some value of the resolution parameter as judged under the go similarity measure . the number of proteins in functionally homogeneous communities peaks at about @xmath123 for the @xmath1 network ( which is beyond the standard ` modularity ' resolution of @xmath84 ) . for the @xmath0 network the peak is less pronounced , with the actual maximum occurring at @xmath124 ( i.e. @xmath125 ) . these findings emphasise that there are different scales of interest in the community structure of protein interaction networks , and that the one of primary interest will depend on which proteins and processes one is investigating . for some protein types , there are natural resolutions , at which more proteins of that type are assigned to functionally homogeneous communities . we also find that proteins involved in some processes are much more likely to be in functionally homogeneous communities than others . for example we find for both networks and across a range of resolutions that approximately @xmath126 of proteins involved in chromosome organisation compared to @xmath127 involved in lipid metabolism are in functionally homogeneous communities . having a good measure of functional homogeneity is central for our analysis . we approach this issue by using three different characterisations of functional similarity : two based on the go and mips structured vocabularies respectively and one based on the growth rates of gene knock - out strains under different chemical conditions @xcite ( an independent and objective characterization of biological function ) . the prevalent method in the literature for assessing functional homogeneity of a group of proteins is inappropriate for communities , as the number of interacting pairs in a group must be taken into consideration . by defining similarity at the pairwise level , we have developed a fair test of functional homogeneity through a comparison of interacting pairs . we also capture the aggregate functional similarity of two proteins , overcoming the need to assess functional homogeneity on a term by term basis ( although this is , of course , also possible once communities of particular interest have been identified ) . our tests of functional homogeneity ( which are not statistical tests in the conventional sense because of our desire to exclude the effects of sample size ) using the three measures of similarity show a high level of agreement with each other , giving us confidence in our chosen measures of functional similarity . throughout this study , we have investigated two separate yeast protein interaction networks : that based on associations ( the @xmath1 network ; mostly tap - like data ) , and that based on physical associations ( the @xmath0 network ; mostly yeast - two - hybrid data ) . we find that the two networks have similar properties with respect to their community structure , despite their very different global topological properties . rather than regarding the yeast - two - hybrid data as of an inferior quality @xcite , we start from the basis that it is of a fundamentally different type and should thus be treated separately . we find similar percentages of functionally homogeneous communities in both networks . as we have found a connection between network communities and biological function , we can use observed community structure to predict aspects of biological function . we find in particular that communities with a high mean clustering coefficient are far more likely to be functionally homogeneous than those with a lower one . the mean clustering coefficient of nodes within a community can therefore be used to predict that a group of proteins is functionally homogeneous , even in cases where our current knowledge does not allow us to infer this on the basis of functional annotations alone . these results give insights into the relationships between the structural and functional organisation of the cell considering the whole proteome . we have also illustrated the utility of our framework for biologists who are interested in a particular protein . in a chosen interaction network , one can determine the community membership of the protein of interest at multiple resolutions . even in the dearth of functional information , the easily - calculated clustering coefficient can be computed to suggest resolutions of particular interest . in conclusion , we have linked the community structure of a protein interaction network with biological function by probing different scales of network structure . the identified communities are candidates for biological modules within the cell . we have also illustrated how this connection can be used to select groups of proteins that likely participate in similar biological functions . all four authors conceived of the study , and acfl carried it out . we thank gesine reinert , simon myers , sumeet agarwal , dan fenn , and peter mucha for useful discussions . we thank gabriel villar for implementing the network measures , and amanda traud for implementation of the kamada - kawai visualisation code ( which we modified for use here ) , which can be found at http://netwiki.amath.unc.edu/viscomms/viscomms . tarassov k , messier v , landry cr , radinovic s , molina mm , shames i , malitskaya y , vogel j , bussey h , michnick sw : * an in vivo map of the yeast protein interactome*. _ science _ 2008 , * 320*(5882):14651470 . yu h , braun p , yildirim ma , lemmens i , venkatesan k , sahalie j , hirozane - kishikawa t , gebreab f , li n , simonis n , hao t , rual jf , dricot a , vazquez a , murray rr , simon c , tardivo l , tam s , svrzikapa n , fan c , de smet as , motyl a , hudson me , park j , xin x , cusick me , moore t , boone c , snyder m , roth fp , barabsi a - l , tavernier j , hill de , vidal m : * high - quality binary protein interaction map of the yeast interactome network*. _ science _ 2008 , * 322*(5898):104110 . han jdj , bertin n , hao t , goldberg ds , berriz gf , zhang lv , dupuy d , walhout ajm , cusick me , roth fp , et al . : * evidence for dynamically organized modularity in the yeast protein - protein interaction network*. _ nature _ 2004 , * 430*(6995):8893 . bu d , zhao y , cai l , xue h , zhu x , lu h , zhang j , sun s , ling l , zhang n , et al . : * topological structure analysis of the protein - protein interaction network in budding yeast * . _ nucleic acids research _ 2003 , * 31*(9):24432450 . ashburner m , ball ca , blake ja , botstein d , butler h , cherry jm , davis ap , dolinski k , dwight ss , eppig jt , et al . : * gene ontology : tool for the unification of biology*. _ nature genetics _ 2000 , * 25*:2529 . mewes hw , frishman d , guldener u , mannhaupt g , mayer k , mokrejs m , morgenstern b , munsterkotter m , rudd s , weil b : * mips : a database for genomes and protein sequences * . _ nucleic acids research _ 2002 , * 30*:3134 . hillenmeyer me , fung e , wildenhain j , pierce se , hoon s , lee w , proctor m , st onge rp , tyers m , koller d , et al . : * the chemical genomic portrait of yeast : uncovering a phenotype for all genes*. _ science _ 2008 , * 320*(5874):362 . hermjakob h , montecchi - palazzi l , bader g , wojcik j , salwinski l , ceol a , moore s , orchard s , sarkans u , von mering c , et al . : * the hupo psi s molecular interaction format a community standard for the representation of protein interaction data*. _ nature biotechnology _ 2004 , * 22*(2):177183 . li s , armstrong cm , bertin n , ge h , milstein s , boxem m , vidalain po , han jd , chesneau a , hao t , goldberg ds , li n , martinez m , rual jf , lamesch p , xu l , tewari m , wong sl , zhang lv , berriz gf , jacotot l , vaglio p , reboul j , hirozane - kishikawa t , li q , gabel hw , elewa a , baumgartner b , rose dj , yu h , bosak s , sequerra r , fraser a , mango se , saxton wm , strome s , van den heuvel s , piano f , vandenhaute j , sardet c , gerstein m , doucette - stamm l , gunsalus kc , harper jw , cusick me , roth fp , hill de , vidal m : * a map of the interactome network of the metazoan c elegans*. _ science _ 2004 , * 303*(5657):540543 . collins mo , choudhary js : * mapping multiprotein complexes by affinity purification and mass spectrometry*. _ current opinion in biotechnology _ 2008 , * 19*(4):324330 , [ http://www hubmed org / display cgi?uids=18598764][[http://www hubmed org / display cgi?uids=18598764 ] ] . kerrien s , alam - faruque y , aranda b , bancarz i , bridge a , derow c , dimmer e , feuermann m , friedrichsen a , huntley r , et al . : * intact open source resource for molecular interaction data*. _ nucleic acids research _ 2007 , * 35*(database issue):d561 . boyle ei , weng s , gollub j , jin h , botstein d , cherry jm , sherlock g : * go : : termfinder open source software for accessing gene ontology information and finding significantly enriched gene ontology terms associated with a list of genes*. _ bioinformatics _ 2004 , * 20*(18):3710 . hong el , balakrishnan r , dong q , christie kr , park j , binkley g , costanzo mc , dwight ss , engel sr , fisk dg , et al . : * gene ontology annotations at sgd : new data sources and annotation methods*. _ nucleic acids research _ 2008 , * 36*(database issue):d577 .

finding the full automorphism group is one of the fundamental objectives when dealing with symmetry properties of mathematical objects , such as for example vertex - transitive graphs , see @xcite . many of these objects naturally display certain inherently obvious symmetries . it is often the case , however , that certain additional symmetries , though hidden or difficult to grasp , are present . the goal is to find a reason for their existence and a method for describing them . along these lines the above question reads as follows : _ given a ( transitive ) group @xmath0 acting on the set of vertices of a graph , determine whether @xmath0 is its full automorphism group or not . when the answer is no , find a method to describe the additional automorphisms . _ in other words , decide whether the group @xmath0 imbeds into a larger group @xmath1 preserving the structure of the graph in question . in this paper such group `` extensions '' are considered by studying the existence of automorphisms that act as odd permutations on the vertex set of a graph . such an automorphism will be referred to as _ odd automorphism_. analogously , an automorphism acting as an even permutation on the vertex set will be referred to as an _ even automorphism_. the implications are more far - reaching than one would expect from the simplicity of the concept of even and odd automorphisms alone . some arguments supporting this claim are given in the paragraphs following formulation of problems [ problem - graph ] and [ problem - group ] . for example , when @xmath0 consists of even automorphisms only , a partial answer to the above question could be given provided the structure of the graph in question forces existence of odd automorphisms . we propose the following problem . [ problem - graph ] which vertex - transitive graphs admit odd automorphisms ? it is convenient to consider problem [ problem - graph ] in the framework of orbital ( di)graphs of transitive groups . namely , isomorphism classes of vertex - transitive ( di)graphs are in a one - to - one correspondence with orbital ( di)graphs of transitive group actions , as seen below . a transitive group @xmath0 acting on a set @xmath2 induces an action of @xmath0 on @xmath3 - the corresponding orbits are called _ orbitals _ of @xmath0 on @xmath2 . given a union @xmath4 of orbitals not containing the diagonal orbital @xmath5 , the _ orbital ( di)graph _ @xmath6 of @xmath0 on @xmath2 with respect to @xmath4 has vertex set @xmath2 and edge set @xmath4 . clearly , @xmath7 . if @xmath4 consists of a single orbital the orbital ( di)graph @xmath6 is said to be _ the problem below is a somewhat restricted reformulation of problem [ problem - graph ] . [ problem - group ] given a transitive group @xmath0 , possibly consisting of even permutations only , is there a basic orbital ( di)graph of @xmath0 admitting an odd automorphism ? transitive groups for which the answer to problem [ problem - group ] is positive are referred to as _ orbital - odd_. not all transitive groups consisting solely of even permutations admit orbital - odd imbeddings . for example , the group @xmath8 acting transitively on the set of cosets of @xmath9 consists of even permutations only , and is isomorphic to the full automorphism group of every basic orbital ( di)graph associated with this action . hence , @xmath10 is not orbital - odd . on the other hand , the alternating group @xmath11 acting transitively on the set of cosets of @xmath12 ( which also consists of even permutations only ) is orbital - odd . namely , the full automorphism group of its two basic orbital graphs - the petersen graph and its complement - contains odd automorphisms ( and is isomorphic to the symmetric group @xmath13 ) . more generally , as observed in @xcite , every vertex - transitive graph of order twice a prime admits an odd automorphism , and hence every transitive group of degree twice a prime is orbital - odd . an essential ingredient in the proof of this result is the fact that the above actions of @xmath11 and @xmath13 are the only simply primitive groups of degree twice a prime a direct consequence of the classification of finite simple groups ( cfsg ) . it would be of interest to obtain a cfsg - free proof of the fact that every vertex - transitive graph of order twice a prime admits an odd automorphism . combining this together with a classical wielandt s result on simply primitive groups of degree twice a prime being of rank @xmath14 ( see @xcite ) currently the best available cfsg - free information on such groups one would move a step closer to obtaining a cfsg - free proof of non - existence of simply primitive groups of degree @xmath15 , @xmath16 a prime ( see also @xcite ) . observe that , if a graph has an odd automorphism then it must have an odd automorphism whose order is a power of @xmath17 . in fact , an odd automorphism exists if and only if there is one in a sylow @xmath17-subgroup of the automorphism group , as observed in @xcite . consequently , for some classes of graphs the existence of odd automorphisms is easy to determine . for instance , in cayley graphs the corresponding regular group contains odd permutations if and only if its sylow @xmath17-subgroup is cyclic . when a sylow @xmath17-subgroup is not cyclic , the search for odd automorphisms has to be done outside the regular group , raising the complexity of the problem . studying `` extensions '' via odd automorphisms is also essential in obtaining new approaches for solving certain long - standing open problems in algebraic graph theory . for example , in the semiregularity problem @xcite , which asks for the existence of semiregular automorphisms in vertex - transitive ( di)graph , knowing that the graph admits odd automorphisms would be helpful in the case when the underlying transitive group @xmath0 consisting of even automorphisms is elusive ( that is , without semiregular elements ) . one could therefore hope for the existence of semiregular automorphisms in the odd part of the `` extension '' . for example , the group @xmath18 acting on the set of @xmath19 lines of the affine plane @xmath20 is elusive and consists of even permutations only . its @xmath17-closure , however , contains semiregular as well as odd permutations , in fact it contains an odd semiregular permutation of order @xmath21 . further , in the hamiltonicity problem @xcite as well as in the snark problem for cubic cayley graphs @xcite , important progress was recently obtained by combining algebraic methods with the theory of regular maps on surfaces @xcite . knowing the graphs admit odd automorphisms will most likely be helpful in our quest to solve the remaining open cases . our ultimate goal is to build a theory that will allow us to decide whether a given graph , admitting a transitive group action , does or does not have odd automorphisms . in this respect the first step in our strategy consists in finding necessary and sufficient conditions for the existence of odd automorphisms for as large as possible class of graphs , admitting a transitive action of a group , with certain suitably imposed constraints . it is natural to expect that the solution to the question of existence of odd automorphisms will have some degree of arithmetic flavor . as may be suggested by the results of this paper this is true , but only up to a certain extent . in approaching problems [ problem - graph ] and [ problem - group ] in a systematic manner it seems natural to start with the analysis of vertex - transitive graphs of smallest possible ( non - trivial ) valency @xmath14 . such graphs fall into three classes : graphical regular representations of groups , vertex - transitive graphs with two edge orbits , and symmetric graphs . as already mentioned , graphs in the class of ( cubic ) graphical regular representations of groups admit odd automorphisms if and only if sylow @xmath17-subgroups of their automorphism groups are cyclic . cubic vertex - transitive graphs with two edge orbits ( together with the corresponding tetravalent locally imprimitive symmetric graphs ) will be considered in a separate paper . in this paper we give a complete solution to problem [ problem - graph ] for cubic symmetric graphs . ( as a consequence a partial solution to problem [ problem - group ] is obtained . ) this class of graphs was first studied by foster @xcite and has been the source of motivation for various research directions in algebraic graph theory . in foster census @xcite , a list of such graphs of order up to @xmath22 was produced . with the availability of advanced computational tools @xcite the list was recently expanded to graphs of order up to @xmath23 , see @xcite . many well - known graphs are cubic symmetric graphs , arising in connection with certain open problems in graph theory ( such as the hamiltonicity problem @xcite ) , and more generally as an important concept in other areas of mathematics ( such as finite geometry @xcite ) . for example , the petersen graph @xmath24 and the coxeter graph @xmath25 are the only known symmetric graphs of order greater than @xmath17 without a hamilton cycle . next , all bipartite cubic symmetric graphs of girth at least @xmath26 are levi graphs of @xmath27 configurations . for example , the heawood graph @xmath28 , the moebius - kantor graph @xmath29 , the pappus graph @xmath30 and the desargues graph @xmath31 are levi graphs of @xmath32 , @xmath33 , @xmath34 and @xmath35 configurations , respectively . ( hereafter the notation @xmath36 , @xmath37 , etc . will refer to the corresponding graphs in the foster census @xcite . ) there are @xmath38 possible types of cubic symmetric graphs ( see table [ tab : cubic - types ] in subsection [ sec : cubic - sym ] ) , and the existence of odd automorphisms depends on these types and orders of graphs in question . results of this paper do imply that the solution to problem [ problem - graph ] is to some degree arithmetic . however , some special cases sprouting out in the solution for cubic symmetric graphs suggest that the even / odd question is likely to uncover certain more complex structural properties of graphs ( and combinatorial objects in general ) , that go beyond simple arithmetic conditions . the following is the main result of this paper . [ thm : main ] let @xmath39 be a cubic symmetric graph of order @xmath40 . in table [ tab : odd ] a full information on existence of odd automorphisms in @xmath39 is given . in particular , the following statements hold : a. if @xmath39 is of type @xmath41 , @xmath42 , @xmath43 , @xmath44 , @xmath45 , @xmath46 , @xmath47 , @xmath48 or @xmath49 then it has odd automorphisms if and only if @xmath50 is odd . b. if @xmath39 is of type @xmath51 , @xmath52 , @xmath53 or @xmath54 then it has odd automorphisms if and only if @xmath50 is odd and @xmath39 is bipartite . c. if @xmath39 is of type @xmath55 then it has odd automorphism if and only if either @xmath50 is odd , or @xmath56 , @xmath57 and @xmath39 is a @xmath58-cayley graph on a cyclic group of order @xmath59 . d. if @xmath39 is of type @xmath60 , @xmath61 or @xmath62 then it has no odd automorphisms . @xmath63 as a direct consequence of theorem [ thm : main ] we have the following corollary about transitive groups with a self - paired suborbit of length @xmath14 . [ cor : groups ] let @xmath0 be a transitive group with a suborbit of length @xmath14 giving rise to an orbital graph satisfying conditions for the existence of odd automorphisms from table [ tab : odd ] . then @xmath0 is orbital - odd . the paper is organized as follows . in section [ sec : prelim ] notation and terminology is introduced and certain results on cubic symmetric graphs , essential to the strategy of the proof of theorem [ thm : main ] , are gathered . in section [ sec : cells ] the concept of rigid subgraphs and rigid cells are introduced which will prove useful in the study of the even / odd question for cubic symmetric graphs . in section [ sec : main ] the proof of theorem [ thm : main ] is given following a series of propositions in which the @xmath38 particular types of cubic symmetric graphs are considered . throughout this paper graphs are finite , simple , undirected and connected , unless specified otherwise . given a graph @xmath39 we let @xmath64 , @xmath65 , @xmath66 and @xmath67 be the vertex set , the edge set , the arc set and the automorphism group of @xmath39 , respectively . for adjacent vertices @xmath68 and @xmath69 in @xmath39 , we denote the corresponding edge by @xmath70 . if @xmath71 then @xmath72 denotes the set of neighbors of @xmath68 and @xmath73 denotes the set of vertices at distance @xmath74 from @xmath68 . a graph @xmath39 is said to be _ cubic _ if @xmath75 for every vertex @xmath71 . a sequence @xmath76 of distinct vertices in a graph is called an _ @xmath77-arc _ if @xmath78 is adjacent to @xmath79 for every @xmath80 . for @xmath81 we let @xmath82 $ ] denote the induced subgraph of @xmath39 on @xmath83 . for a partition @xmath84 of @xmath64 , we let @xmath85 be the associated _ quotient graph _ of @xmath39 _ relative to _ @xmath84 , that is , the graph with vertex set @xmath84 and edge set induced by the edge set @xmath65 in a natural way . when @xmath84 consists of orbits of a subgroup @xmath0 of @xmath67 we denote @xmath85 by @xmath86 , and by @xmath87 when @xmath88 is a cyclic group generated by an automorphism @xmath89 . a subgroup @xmath90 is said to be _ vertex - transitive _ , _ edge - transitive _ and _ arc - transitive _ provided it acts transitively on the sets of vertices , edges and arcs of @xmath39 , respectively . a graph is said to be _ vertex - transitive _ , _ edge - transitive _ , and _ arc - transitive _ if its automorphism group is vertex - transitive , edge - transitive and arc - transitive , respectively . an arc - transitive graph is also called _ symmetric_. a subgroup @xmath91 is said to be _ @xmath77-regular _ if it acts transitively on the set of @xmath77-arcs and the stabilizer of an @xmath77-arc in @xmath1 is trivial . a _ quasi - dihedral group _ @xmath92 ( sometimes also called a _ semi - dihedral group _ ) is a non - abelian group of order @xmath93 with a presentation @xmath94 to end this subsection let us recall that if a transitive permutation group @xmath1 on a set @xmath2 contains an odd permutation then the intersection @xmath95 of @xmath1 with the alternating group @xmath96 on @xmath2 is its index @xmath17 subgroup . this will be used in several places in subsequent sections . in @xcite tutte proved that every finite cubic symmetric graph is @xmath77-regular for some @xmath97 . a further deeper insight into the structure cubic symmetric graphs is due to djokovi and miller @xcite who proved that a vertex stabilizer in an @xmath77-regular subgroup of automorphisms of a cubic symmetric graph is isomorphic to @xmath98 , @xmath99 , @xmath100 , @xmath101 , or @xmath102 depending on whether @xmath103 or @xmath104 , respectively . consequently , the automorphism group of a cubic symmetric graph of order @xmath50 is of order @xmath105 . djokovi and miller @xcite also proved that for @xmath106 there is just one possibility for edge stabilizers , while there exists two possibilities for @xmath107 , see table [ tab : vertex - and - edge - stabilizers ] . in particular , for @xmath108 the edge stabilizer is either isomorphic to @xmath109 or @xmath110 , and for @xmath111 the edge stabilizer is either isomorphic to the dihedral group @xmath112 of order @xmath113 or to the quasi - dihedral group @xmath114 of order @xmath113 . @xmath115 the automorphism group of any finite symmetric cubic graph is an epimorphic image of one of the following seven groups:@xmath116 this implies that an arc - transitive subgroup of a cubic symmetric graph is a quotient of one of these seven groups by some normal torsion - free subgroup . in particular , an @xmath77-regular subgroup of automorphisms of a cubic symmetric graph is a quotient group of a group @xmath117 , if @xmath106 , and of a group @xmath118 , @xmath119 , if @xmath120 . moreover , in @xcite a complete characterization of admissible types of cubic symmetric graphs according to the structure of arc - transitive subgroups is given . for example , a cubic symmetric graph @xmath39 is said to be of type @xmath42 if its automorphism group is @xmath14-regular , admitting two @xmath17-regular subgroups , of which one is a quotient of the group @xmath121 and one of @xmath122 , and admitting also a @xmath123-regular subgroup . all possible types are summarized in table [ tab : cubic - types ] ( for details see @xcite ) . @xmath124 finally , in section [ sec : main ] the following result about the girth of cubic symmetric graphs of types @xmath60 and @xmath125 , extracted from ( * ? ? ? * theorems 2.1 2.2 ) will be needed . ( the _ girth _ of a graph is the length of a shortest cycle contained in the graph . ) [ pro : girth ] a cubic symmetric graph , which is either of type @xmath60 or @xmath125 , has girth greater than @xmath126 . given a graph @xmath39 and an automorphism @xmath127 of @xmath39 let @xmath128 denote the set of all vertices of @xmath39 fixed by @xmath127 . with the assumption that @xmath129 we call the subgraph @xmath130 $ ] induced on @xmath128 the _ rigid subgraph of _ @xmath127 or , in short , the @xmath127-_rigid subgraph_. every component of @xmath130 $ ] is referred to as an _ @xmath127-rigid cell_. this concept will prove useful in the study of the `` even / odd question '' for cubic symmetric graphs . for this purpose let us fix some notation and terminology . we will use the terms @xmath131-_tree _ , @xmath0-_tree _ , @xmath132-_tree _ , @xmath133-_tree _ and @xmath134-_tree _ for the graphs given in figure [ fig : trees ] below . more precisely , we will denote these graphs by @xmath135 , @xmath136 , @xmath137 , @xmath138 and @xmath139 , respectively . let @xmath39 be a cubic graph of girth greater than or equal to @xmath104 . for a vertex @xmath69 in @xmath64 let @xmath140 be the set of neighbors of @xmath69 , and let @xmath141 be the second neighborhood of @xmath69 , where @xmath142 , @xmath143 . similarly , for an edge @xmath70 of @xmath39 let @xmath144 and @xmath145 be the remaining two neighbors of @xmath69 , and let @xmath146 and @xmath147 be the remaining two neighbors of @xmath68 . for a vertex @xmath69 of @xmath39 we therefore have @xmath148 $ ] , @xmath149 $ ] and @xmath150 $ ] , where @xmath143 . for short , we will use the notation @xmath139 for any of the trees @xmath151 , @xmath152 and @xmath153 ( and call it a @xmath134-tree ) . similarly , for two adjacent vertices @xmath68 and @xmath69 of @xmath39 we have @xmath154 $ ] and @xmath155 $ ] . note that @xmath156 . @xmath131-tree , @xmath0-tree , @xmath132-tree , @xmath133-tree and three @xmath134-trees . , width=340 ] we will now restrict ourselves to cubic symmetric graphs . the structure of vertex stabilizers in cubic symmetric graphs implies that only automorphisms of order @xmath17 , @xmath14 , @xmath21 and @xmath26 can fix a vertex ( see section [ sec : cubic - sym ] ) . in the propositions below all possible rigid cells for such automorphisms are listed . ( the exclusion of @xmath132-trees as rigid cells for automorphisms of order @xmath21 is proved in proposition [ pro : cell3 ] and uses proposition [ pro : cell25 ] . ) for convenience we let @xmath157 denote the set of all involutions of a cubic symmetric graph @xmath39 which fix some vertex of @xmath39 , and we let @xmath158 denote the set of all semiregular involutions of @xmath39 . [ pro : cell1 ] let @xmath39 be a cubic symmetric graph and let @xmath159 be an automorphism of @xmath39 fixing a vertex . a. if @xmath127 is of order @xmath14 or @xmath26 then the only @xmath127-rigid cells are isolated vertices . b. if @xmath127 is of order @xmath21 then the only possible @xmath127-rigid cells are @xmath131-trees and @xmath132-trees . the proof of part ( i ) is straightforward and is omitted . as for part ( ii ) , if an @xmath0-tree , a @xmath134-tree or an @xmath133-tree was a rigid cell of an automorphism of order @xmath21 then the square of this automorphism ( a non - identity element ) would fix a path of length greater than @xmath21 . given a cubic @xmath77-regular graph @xmath39 , @xmath160 , we call an involution @xmath161 _ canonical with respect to a pair of adjacent vertices _ @xmath68 and @xmath69 of @xmath39 if it fixes the tree @xmath135 for @xmath108 and if it fixes the tree @xmath136 for @xmath111 . similarly , we call it _ canonical with respect to a vertex _ @xmath69 of @xmath39 if it fixes the tree @xmath162 for @xmath163 and if it fixes the tree @xmath138 for @xmath164 . the next four propositions give a full information on rigid cells of involutions . the proof of the first of these propositions is straightforward and is omitted . [ pro : cell212 ] let @xmath39 be a cubic @xmath77-regular graph , @xmath165 and let @xmath159 be an involution . then the following statements hold : a. if @xmath166 then @xmath167 , that is , @xmath127 is semiregular . b. if @xmath108 then the only possible @xmath127-rigid cells are @xmath131-trees . [ pro : cell23 ] let @xmath39 be a cubic @xmath14-regular graph and let @xmath168 be a non - semiregular involution of @xmath39 . then the possible @xmath127-rigid cells are @xmath131-trees and @xmath132-trees , with both types of cells possibly occurring simultaneously only when the graph @xmath39 is of type @xmath52 . assume that @xmath39 is not of type @xmath52 . then @xmath67 contains a @xmath17-regular subgroup @xmath169 ( of index @xmath17 ) . because of @xmath14-regularity of @xmath39 given any vertex @xmath170 there exists a canonical involution @xmath171 fixing the tree @xmath137 point - wise . we now show that the only rigid cell of canonical involutions are @xmath132-trees . suppose , on the contrary , that there exists a canonical involution @xmath172 with both the @xmath131-tree and the @xmath132-tree as rigid cells of @xmath130 $ ] . then there exist a vertex @xmath170 and adjacent vertices @xmath173 such that @xmath137 and @xmath174 are rigid cells of @xmath127 . clearly , there also exists an involution @xmath175 whose action on @xmath176 coincides with the action of @xmath127 . it follows that the product @xmath177 acts trivially on @xmath178 , thus fixing a path of length @xmath14 . but @xmath39 is @xmath14-regular , and so @xmath179 . consequently , @xmath180 , a contradiction . [ pro : cell24 ] let @xmath39 be a cubic @xmath21-regular graph and let @xmath168 be a non - semiregular involution of @xmath39 . then the only possible @xmath127-rigid cells are @xmath0-trees . note that theoretically the possible rigid cells are the @xmath131-tree , the @xmath132-tree and the @xmath0-tree . we now show that the first two possibilities can not occur . let @xmath68 and @xmath69 be adjacent vertices of @xmath39 . because of @xmath21-regularity of @xmath39 there exists an involution @xmath181 with @xmath136 as its rigid cell . suppose that there exists an involution @xmath182 with @xmath132-tree as a rigid cell . since vertex stabilizers are conjugate subgroups there exists a conjugate @xmath183 of @xmath182 with @xmath137 as a rigid cell . obviously , @xmath183 is different from @xmath184 since they have different rigid cells containing @xmath69 . but since @xmath184 and @xmath183 both interchange the remaining two neighbors @xmath185 and @xmath186 of @xmath144 and the remaining two neighbors @xmath187 and @xmath188 of @xmath145 it follows that @xmath189 fixes point - wise the @xmath134-tree @xmath139 , and thus a path of length @xmath21 . since @xmath111 it follows that @xmath190 , a contradiction . therefore no involution can have the @xmath132-tree as a rigid cell . suppose now that there exists an involution @xmath191 with @xmath131-tree as a rigid cell . then a conjugate @xmath192 of @xmath191 has @xmath193 as a rigid cell . obviously , @xmath192 is different from @xmath184 since they have different rigid cells containing @xmath69 and @xmath144 . observe that @xmath184 and @xmath192 both interchange the remaining neighbors of @xmath185 and @xmath186 . the actions of @xmath184 and @xmath192 on these remaining neighbors of @xmath185 and @xmath186 are either different or identical . in the first case , @xmath194 is a rigid cell of @xmath195 . by previous paragraph @xmath196 , and moreover it fixes a path of length greater than or equal to @xmath21 , which is not possible in a @xmath21-regular graph . this completes the proof of proposition [ pro : cell24 ] . [ pro : cell25 ] let @xmath39 be a cubic @xmath104-regular graph and let @xmath168 be a non - semiregular involution of @xmath39 . then the only possible @xmath127-rigid cells are @xmath0-trees and @xmath133-trees , with both types of cells possibly occurring simultaneously only when the graph @xmath39 is of type @xmath54 . the stabilizer @xmath197 of a vertex @xmath170 is isomorphic to @xmath102 . hence the center of @xmath197 is isomorphic to @xmath198 , and moreover there are precisely @xmath199 involutions in @xmath197 . because of @xmath104-regularity of @xmath39 there exists an involution @xmath184 in @xmath197 with either @xmath138 or @xmath139 as a rigid cell . in what follows we essentially translate ( into our language ) the arguments from ( * ? ? ? * lemma 1 ) where it is shown that the central involution in @xmath197 has @xmath138 as a rigid cell . suppose that @xmath184 has a @xmath134-tree @xmath139 as a rigid cell . then for conjugacy reasons there exist involutions @xmath200 , @xmath201 and @xmath202 in @xmath197 whose respective restrictions to @xmath138 are : @xmath203 , @xmath204 and @xmath205 . their products @xmath206 , @xmath207 , @xmath208 and @xmath209 are also involutions in @xmath197 , and so the stabilizer of the @xmath17-arc @xmath210 , where @xmath211 and @xmath212 , is an elementary abelian group @xmath213 . since @xmath39 is @xmath104-regular , stabilizers of @xmath17-arcs in @xmath39 are conjugate subgroups , and so the stabilizer of any @xmath17-arc in @xmath39 is isomorphic to @xmath213 . let @xmath214 be such that @xmath215 and @xmath216 . ( because of @xmath104-regularity such an element clearly exists . ) then @xmath217 and @xmath218 , implying that @xmath219 . but this is a contradiction since @xmath201 and @xmath182 both belong to the stabilizer of the @xmath17-arc @xmath220 , which is an abelian group . this shows that the @xmath134-tree can not occur as a rigid cell in @xmath39 . moreover , @xmath184 has the whole of @xmath138 as a rigid cell . in other words , @xmath184 is the canonical involution for @xmath69 . this means that @xmath184 is in the kernel of the restriction of @xmath197 to @xmath138 . let us also remark that the restrictions of involutions in @xmath197 to @xmath221 are @xmath222 , @xmath223 and @xmath224 . next , we show that the @xmath132-tree can not be a rigid cell . assume on the contrary that there exists an involution @xmath225 with the @xmath132-tree as a rigid cell . then there is an involution @xmath226 ( a conjugate of @xmath225 ) whose rigid cell is @xmath194 . obviously , @xmath226 is different from @xmath184 since they have different rigid cells containing @xmath69 . but then @xmath227 either fixes a path of length @xmath104 or it has @xmath228 as a rigid cell . if the first possibility occurs then @xmath229 , contradicting our assumption on @xmath226 . if the second possibility occurs then a non - identity element @xmath227 has the @xmath134-tree as a rigid cell . but this is not possible by proposition [ pro : cell1 ] if @xmath227 is of order @xmath21 , and by the argument in the previous paragraph if @xmath227 is an involution . we now show that the @xmath131-tree can not be a rigid cell . assume on the contrary that there exists an involution @xmath191 with the @xmath131-tree as a rigid cell . then there is an involution @xmath192 ( a conjugate of @xmath191 ) whose rigid cell is @xmath230 . obviously , @xmath192 is different from @xmath184 since they have different rigid cells containing @xmath144 . but then it can be shown that @xmath195 either has @xmath231 or @xmath232 as a rigid cell or it fixes a path of length greater than @xmath21 . each of these cases leads to a contradiction by the previous arguments with the exception of the case where @xmath195 is of order @xmath21 and has @xmath231 as a rigid cell . there exists a conjugate @xmath233 of @xmath195 with @xmath194 as a rigid cell . but then @xmath234 either has @xmath228 as a rigid cell or it fixes a path of length greater than @xmath21 . both possibilities clearly lead to a contradiction . this shows that the only possible rigid cells in @xmath39 are @xmath0-trees and @xmath133-trees . to finish the proof assume that @xmath39 is not of type @xmath54 . then @xmath67 contains a @xmath21-regular subgroup @xmath235 ( of index @xmath17 ) . because of @xmath104-regularity of @xmath39 it is clear that @xmath133-trees occur as rigid cells of non - semiregular involutions in @xmath236 . in order to prove that all rigid cells of such involutions are @xmath133-trees suppose , on the contrary , that an involution @xmath172 has both the @xmath0-tree and the @xmath133-tree as rigid cells of @xmath130 $ ] . then there exist a vertex @xmath170 and adjacent vertices @xmath173 such that @xmath138 and @xmath178 are rigid cells of @xmath127 . clearly , there also exists an involution @xmath175 whose action on @xmath178 coincides with the action of @xmath127 . it follows that the product @xmath177 fixes a path of length greater than @xmath21 . but then @xmath179 , and so @xmath180 , a contradiction . this completes the proof of proposition [ pro : cell25 ] . as an almost immediate consequence of proposition [ pro : cell25 ] we can now determine rigid cells of automorphisms of order @xmath21 . [ pro : cell3 ] let @xmath39 be a cubic @xmath77-regular graph and @xmath127 an automorphism of @xmath39 of order @xmath21 fixing a vertex . then the only possible @xmath127-rigid cells are the @xmath131-trees . the assumption on the existence of an automorphism @xmath127 of order @xmath21 fixing a vertex implies that either @xmath111 or @xmath164 ( see table [ tab : vertex - and - edge - stabilizers ] ) . by proposition [ pro : cell1 ] the only possible candidates for @xmath127-rigid cells are @xmath131-trees and @xmath132-trees . suppose that @xmath127 has a rigid cell isomorphic to a @xmath132-tree . then @xmath237 has a rigid cell isomorphic to an @xmath133-tree , and thus @xmath39 is @xmath104-regular . let @xmath170 with neighbors @xmath144 , @xmath145 and @xmath238 . by assumption there exists a conjugate @xmath184 of @xmath127 such that @xmath194 is an @xmath184-rigid cell . let @xmath239 be the canonical involution of @xmath69 . it follows that the @xmath134-tree @xmath240 $ ] is contained in a rigid cell of a non - identity automorphism @xmath241 . clearly this rigid cell does not contain @xmath145 and @xmath238 , and so it is either a @xmath134-tree or a rigid cell containing a path of length greater than @xmath21 . both possibilities lead to a contradiction . we conclude that the only possible @xmath127-rigid cells are @xmath131-trees . the proof of theorem [ thm : main ] is carried out through a series of propositions ( propositions [ pro : cubic-1-regular ] [pro:5 - 4-oddx ] ) , each dealing with particular types of cubic symmetric graphs . the first of these propositions gives the proof of theorem [ thm : main ] for cubic symmetric graphs of type @xmath41 . in fact it is a slightly more general for it gives a necessary and sufficient conditions on existence of odd automorphisms in any cubic symmetric graph admitting a @xmath123-regular subgroup in its automorphism group . [ pro : cubic-1-regular ] let @xmath39 be a cubic symmetric graph of order @xmath40 admitting a @xmath123-regular subgroup @xmath242 . then there exists an odd automorphism of @xmath39 in @xmath1 if and only if @xmath50 is odd . let @xmath170 . by ( [ g1 ] ) the group @xmath1 is generated by @xmath243 , @xmath170 , and an involution @xmath89 interchanging two adjacent vertices . automorphisms in the vertex stabilizer @xmath244 are all even automorphisms as they are of odd order . hence we can conclude that @xmath1 admits odd automorphisms if and only if @xmath89 is an odd automorphism . since @xmath89 is semiregular the result follows . the next three propositions give the answer regarding existence of odd automorphisms in cubic @xmath77-regular graphs , @xmath160 , having no @xmath245-regular subgroup of automorphisms . the first proposition in the series shows that all automorphisms of cubic symmetric graphs of type @xmath60 are even . first , we note that , by proposition [ pro : girth ] , @xmath246 suppose first that @xmath39 is not bipartite . then it must have only even automorphisms for otherwise there would be an intransitive index @xmath17 subgroup of even automorphisms in @xmath67 given rise to a bipartition of @xmath39 . we may therefore assume that @xmath39 is bipartite . by ( [ g22 ] ) the automorphism group @xmath67 is generated by an automorphism @xmath247 of order @xmath14 and an edge reversing automorphism @xmath89 of order @xmath21 . since @xmath247 is of odd order it is an even permutation . to complete the proof we need to show that @xmath89 is an even permutation too . consider the orbits of @xmath248 : they are either of length @xmath17 or @xmath21 . orbits of @xmath248 of length @xmath17 are of two types : those with an inner edge ( we refer to them as orbits of _ type 1 _ ) and those with no inner edge ( we refer to them as orbits of _ type 2 _ ) . finally , orbits of @xmath248 of length @xmath21 will be referred to as orbits of _ type 3_. let us remark that orbits of type 3 have no inner edges because such an edge would either give rise to a @xmath21-cycle in @xmath39 ( contradicting ( [ girth ] ) ) or would give rise to an edge flipped by the involution @xmath249 , which is not possible in cubic symmetric graphs of type @xmath60 . observe that the numbers of orbits of type 1 and type 2 are of the same parity when the bipartition sets are of even cardinality ( that is , for @xmath50 even ) , and are of different parity when the bipartition sets are of odd cardinality ( that is , for @xmath50 odd ) . let us now consider the quotient graph @xmath87 of @xmath39 with respect to the set of orbits of @xmath248 . a type 1 orbit in @xmath87 can not be adjacent to another type 1 orbit , for this would give a cycle of length @xmath21 in @xmath39 , contradicting ( [ girth ] ) . further , a type 1 orbit can not be adjacent to two type 2 orbits , for this would imply that the involution @xmath249 has a rigid cell containing the @xmath0-tree as a subgraph , which , by proposition [ pro : cell212 ] , is impossible in a @xmath17-regular graph . the only remaining possibility therefore is that an orbit of type 1 is adjacent to an orbit of type 3 . in other words , each orbit of type 1 is paired off with an orbit of type 3 . we now turn to orbits of type 2 . such an orbit can not be adjacent to three other orbits of type 2 , for otherwise the involution @xmath249 would have a @xmath132-tree as a rigid cell , contradicting proposition [ pro : cell212 ] . the only remaining possibility is that an orbit of type 2 is adjacent to an orbit @xmath250 of type 3 and another orbit of type 2 , the latter being adjacent also to an orbit of type 3 which has to be different from @xmath250 in view of ( [ girth ] ) . in conclusion , orbits of type 2 come in pairs , implying that @xmath251 moreover , each such pair of orbits of type 2 has two neighbors in @xmath87 , namely a pair of orbits of type 3 . we now consider orbits of type 3 . recall that they have no inner edges . further , by ( [ girth ] ) two adjacent orbits of type 3 must be joined by a single matching . we conclude that in @xmath87 an orbit of type 3 is either adjacent to three orbits of type 3 or to two orbits of type 3 and an orbit of length @xmath17 ( either of type 1 or of type 2 ) . it follows that there are only five possible local adjacency structures of orbits in @xmath87 , those shown in figure [ fig:22 ] . the rest of the argument depends on the parity of @xmath50 . suppose first that @xmath50 is even . since in this case orbits of type 1 and type 2 are of the same parity , the number of orbits of type 1 is even by ( [ eq : orbits ] ) . consequently , the number of orbits of type 3 adjacent to some orbit of length @xmath17 is therefore also even . to see that the total number of all orbits of @xmath248 is even , it suffices to show that the number of orbits of type 3 which are only adjacent to orbits of type 3 is even , too . we do this by modifying the quotient graph @xmath87 with the removal of all orbits of length @xmath17 , and edges incident with them . we obtain a graph which is a subdivision of a cubic graph . such a graph must have an even number of vertices of valency @xmath14 meaning that the number of orbits of type 3 not adjacent to orbits of length @xmath17 was even to start with . in summary , the number of orbits for each of the three types is even . it follows that @xmath89 is an even automorphism . suppose now that @xmath50 is odd . then there is an odd number of orbits of type 1 , and hence also an odd number of orbits of type 3 adjacent to these orbits . recall that by ( [ eq : orbits ] ) we have an even number of orbits of type 2 , and consequently an even number of orbits of type 3 as their neighbors . as in the case @xmath50 even , we remove all orbits of length @xmath17 ( an odd number of orbits ) from @xmath87 . we obtain a graph which is a subdivision of a cubic graph in which there is an odd number of vertices of valency @xmath17 and an even number of vertices of valency @xmath14 . therefore @xmath248 has an odd number of orbits of length @xmath17 and an odd number of orbits of length @xmath21 , in total an even number of orbits . this implies that @xmath89 is an even automorphism , completing the proof of proposition [ pro : cubic-2 - 2 ] . as in the proof of proposition [ pro : cubic-2 - 2 ] we note that , by proposition [ pro : girth ] , @xmath252 for the same reasons as in the proof of proposition [ pro : cubic-2 - 2 ] we may assume that @xmath39 is bipartite . by ( [ g42 ] ) @xmath67 is generated by an element @xmath247 of order @xmath14 , an edge - flipping automorphism @xmath89 of order @xmath21 , and three conjugate involutions @xmath253 , @xmath254 , and @xmath255 ( of which @xmath253 is a square of @xmath89 ) . clearly proposition [ pro : cubic-4 - 2x ] will be proved provided we show that @xmath89 is an odd automorphism if and only if @xmath50 is odd . for this purpose we now count the orbits of @xmath248 . in doing so we use an approach analogous to that used in the proof of proposition [ pro : cubic-2 - 2 ] . the orbits of @xmath248 are of three types : orbits of length @xmath17 with an inner edge ( type 1 ) , orbits of length @xmath17 without inner edges ( type 2 ) , and orbits of length @xmath21 ( type 3 ) . the latter have no inner edges as an edge would either give rise to a @xmath21-cycle in @xmath39 ( which is impossible by ( [ girth1 ] ) ) or would give rise to an edge flipped by the involution @xmath249 , contradicting non - existence of edge - flipping involutions in cubic symmetric graphs of type @xmath125 . as in the proof of proposition [ pro : cubic-2 - 2 ] we observe that the numbers of orbits of type 1 and type 2 are of the same parity when the bipartition sets are of even cardinality ( that is , for @xmath50 even ) , and are of different parity when the bipartition sets are of odd cardinality ( that is , for @xmath50 odd ) . let us now analyze the quotient graph @xmath87 of @xmath39 with respect to the set of orbits of @xmath248 and the possible adjacencies between the orbits of the three types . first we show that an orbit of type 1 must be adjacent to two orbits of type 2 . it certainly can not be adjacent to an orbit of type 3 , for otherwise the involution @xmath249 would have an @xmath131-tree as a rigid cell , contradicting proposition [ pro : cell24 ] . further , an orbit of type 1 can not be adjacent to another orbit of type 1 , for this would give rise to a cycle of length @xmath21 in @xmath39 , contradicting ( [ girth1 ] ) . hence , an orbit of type 1 is adjacent to two orbits of type 2 . the remaining two neighbors in @xmath87 of these two orbits of type 2 must be orbits of type 3 . namely , if one of these remaining neighbors was an orbit of type 2 then @xmath249 would fix a path of length @xmath21 in @xmath39 , which is clearly not possible in a @xmath21-regular graph . consider now orbits of type 2 . if such an orbit is adjacent to three orbits of type @xmath17 and if each of these three orbits has an orbit of type 3 as a neighbor then @xmath249 has the @xmath132-tree for a rigid cell , contradicting proposition [ pro : cell24 ] . therefore at least one of these three orbits is adjacent to two additional orbits of length @xmath17 ( which must be of type 2 , of course ) . in fact , precisely one of these three orbits has such neighbors , as otherwise @xmath249 would fix a path of length @xmath21 in @xmath39 , contradicting @xmath21-regularity of @xmath39 . these six orbits of type 2 give rise to an @xmath0-tree in @xmath87 . the remaining neighbors of the four vertices of valency @xmath123 in this @xmath0-tree must all be orbits of type 3 , for otherwise @xmath249 would fix a path of length greater than or equal to @xmath21 . it follows that there are only two possible local adjacency structures of orbits of length @xmath17 in @xmath87 those shown in figure [ fig:42 ] . consequently , the number of adjacencies between orbits of type 3 with orbits of type 2 is even ( and there are no adjacencies between orbits of type 1 and orbits of type 3 ) . consider now orbits of type 3 . recall that there are no edges inside an orbit of type 3 . further , by ( [ girth1 ] ) , two adjacent orbits of type 3 must be joined by a single matching . we can conclude that an orbit of type 3 is either adjacent to three orbits of type 3 in @xmath87 or to two orbits of type 3 and one orbit of type 2 . we will now modify the quotient graph @xmath87 by removing all orbits of length @xmath17 , obtaining thus a subdivision of a cubic graph . of course the number of vertices of valency @xmath14 in this graph is even . also , we know that the number of vertices of valency @xmath17 ( corresponding to orbits of type 3 with one neighbor being an orbit of type 2 ) is even . consequently , the number of orbits of type 3 is even . since we know that the number of orbits of length @xmath17 is even if and only if @xmath50 is even , it follows that it is the parity of @xmath50 that determines whether the number of orbits of @xmath89 is even or odd . consequently , @xmath89 is an odd permutation if and only if @xmath50 is odd . [ remark]there are no graphs of type @xmath125 in conder s list of all cubic symmetric graphs up to order @xmath23 @xcite . however , such graphs do exist @xcite . known examples in the literature are of order @xmath256 . marston d. e. conder has kindly pointed out to us that there exists a graph of order @xmath257 . in particular , the group @xmath258 has a quotient @xmath259 of order @xmath260 that is an extension of an elementary abelian normal @xmath14-subgroup of order @xmath261 by @xmath262 ( the point stabilizer in the mathieu group @xmath263 ) . the group @xmath259 act @xmath21-regularly on a cubic @xmath104-arc - transitive graph @xmath132 ( of order @xmath264 ) , but it has two normal subgroups of order @xmath265 that are interchanged under conjugation by an element of the full automorphism group of @xmath132 ( not lying in @xmath259 ) . factoring out one of these normal subgroups gives a quotient @xmath266 of order @xmath267 , which is the full automorphism group of a cubic @xmath21-regular graph of type @xmath125 of order @xmath268 . this graph is a regular @xmath269-cover of tutte s @xmath270-cage @xmath271 . by proposition [ pro : cubic-4 - 2x ] it admits odd automorphisms . the following proposition considers all remaining types of cubic @xmath77-regular graphs , @xmath160 , without @xmath245-regular subgroups in their automorphism groups . it shows , for example , that amongst the graphs @xmath272 , @xmath273 and @xmath274 , the three smallest examples of cubic symmetric graphs of type @xmath52 , only the graph @xmath274 has no odd automorphisms . namely , all these three graphs are of order @xmath275 , but @xmath274 is not bipartite whereas the other two graphs are bipartite . [ pro : cubic - s - regular - without - s-1 ] let @xmath39 be a cubic symmetric graph of order @xmath40 admitting an @xmath77-regular subgroup @xmath276 , which is an epimorphic image of a group different from @xmath122 and @xmath258 , with no @xmath245-regular subgroup , where @xmath160 . then there exists an odd automorphism of @xmath39 in @xmath1 if and only if @xmath50 is odd and @xmath39 is bipartite . assume that there exists an odd automorphism of @xmath39 in @xmath1 . then there exists a subgroup @xmath277 of index @xmath17 consisting of even automorphisms . since an element of order @xmath14 from @xmath278 lies in @xmath235 and there is no @xmath245-regular subgroup of @xmath1 we can conclude that @xmath235 is intransitive with two orbits on @xmath64 , forcing @xmath39 to be bipartite . moreover , odd automorphisms in @xmath1 are precisely those automorphisms that interchange the bipartition sets @xmath39 . since @xmath1 is not an epimorphic image of either @xmath122 or @xmath258 there exists an involution @xmath279 flipping an edge in @xmath39 . this involution @xmath89 is clearly semiregular , and thus it is odd if and only if @xmath50 is odd . in what follows , propositions [ pro : cubic-2-regular-2mod4 ] and [ pro : cubic-2-regular-0mod4 ] combined together lay the ground for the proof of theorem [ thm : main ] for cubic symmetric graphs of type @xmath55 , while proposition [ pro : janko ] recalls a group - theoretic result needed in the proof of proposition [ pro : cubic-2-regular-0mod4 ] . given a group @xmath1 and a subset @xmath280 we let @xmath281 and @xmath282 be the normalizer and the centralizer of @xmath83 in @xmath1 , respectively . recall that in a transitive permutation group @xmath1 acting on a set @xmath2 , the number of points left fixed by @xmath283 , @xmath284 , is equal to the index @xmath285 of @xmath278 in its normalizer @xmath286 ( see , for example , ( * ? ? ? * theorem 3.5 ) ) . proposition [ pro : cubic-2-regular-0mod4 ] gives the answer to the question about existence of odd automorphisms in graphs of order @xmath287 which admit an epimorphic image of @xmath121 as a @xmath17-regular subgroup in their automorphism groups . in the proof of this proposition we will use the following janko s result ( which can be extracted from ( * ? ? ? * theorem 1.1 ) ) about finite @xmath17-groups having a small centralizer of an involution . [ pro : janko ] let @xmath1 be a finite non - abelian @xmath17-group containing an involution @xmath288 such that the centralizer @xmath289 is of the form @xmath290 , where @xmath291 is a cyclic group of order @xmath17 . then either @xmath1 is a dihedral group of order @xmath93 , @xmath292 , or @xmath1 is a quasi - dihedral group @xmath92 of order @xmath93 , @xmath293 . given integers @xmath294 and @xmath295 , a group of automorphisms of a graph is called @xmath296-_semiregular _ if it has @xmath297 orbits of length @xmath50 and no other orbit , and the action is regular on each orbit . an @xmath297-_cayley graph _ @xmath39 _ on a group _ @xmath0 is a graph admitting an @xmath296-semiregular subgroup of automorphisms isomorphic to @xmath0 . [ pro : cubic-2-regular-0mod4 ] let @xmath39 be a cubic symmetric graph of order @xmath298 , where @xmath57 and @xmath294 is odd ( thus , @xmath39 is of order @xmath256 ) , with a @xmath17-regular subgroup @xmath242 , which is an epimorphic image of @xmath121 . then @xmath1 contains an odd automorphism if and only if @xmath39 is non - bipartite , of type @xmath55 an @xmath297-cayley graph on a cyclic group of order @xmath59 . observe that with the assumption on @xmath299 and @xmath297 , clearly an @xmath297-cayley graph on a cyclic group of order @xmath59 admits odd automorphisms : such an automorphism is a generator of the cyclic group of order @xmath59 . to complete the proof of proposition [ pro : cubic-2-regular-0mod4 ] assume that @xmath39 has an odd automorphism in @xmath1 . since @xmath1 is @xmath17-regular we have @xmath300 . by ( [ g21 ] ) , @xmath1 is generated by an element of order @xmath14 , which can not be an odd automorphism , and by two involutions @xmath89 and @xmath253 . therefore @xmath39 has an odd automorphism in @xmath1 if and only if at least one of these two involutions is an odd automorphism . since @xmath39 is of order @xmath256 , an involution is odd if and only if it fixes an odd number of pairs of adjacent vertices . hence , let @xmath301 be an odd involution . then @xmath302 , where @xmath303 is odd , implying that @xmath304 . let @xmath266 be a sylow @xmath17-subgroup of @xmath1 containing @xmath288 . note that @xmath305 . since @xmath306 is an involution it can not be odd inside a cyclic subgroup of order @xmath21 . it follows that @xmath307 . if @xmath266 is abelian then @xmath308 , implying that @xmath309 , and thus @xmath310 , contradicting the assumption that @xmath305 . therefore , @xmath266 is non - abelian , and so proposition [ pro : janko ] implies that @xmath266 contains a cyclic subgroup @xmath291 of order @xmath311 . since @xmath312 is isomorphic to @xmath109 it follows that @xmath313 . in particular , since @xmath314 , @xmath170 and @xmath315 , a generator of @xmath291 is semiregular on @xmath64 . it follows that @xmath39 is an @xmath297-cayley graph on @xmath291 . to complete the proof we need to show that the existence of odd automorphisms in @xmath39 implies that @xmath39 is a non - bipartite graph of type @xmath55 . suppose on the contrary that @xmath39 is bipartite . then the odd involution @xmath301 , which is contained in some vertex stabilizer , fixes the two bipartition sets of @xmath39 . therefore any orbit of @xmath316 is contained in one of the two bipartition sets . further , @xmath317 splits equally into the two bipartition sets . namely , every vertex fixed by @xmath288 in one bipartition set has a neighbor in the other bipartition set which is also fixed by @xmath288 . in view of the fact that the two bipartition sets are of equal cardinality , this implies that the two bipartition sets contain the same number of orbits of @xmath318 of length @xmath17 also , implying that the total number of orbits of length @xmath17 is even , contradicting the fact that @xmath288 is an odd automorphism . therefore @xmath39 is non - bipartite . finally , proposition [ pro : cubic - s - regular - without - s-1 ] and the fact that graphs of types @xmath42 and @xmath43 are all bipartite ( see table [ tab : cubic - types ] ) combined together imply that @xmath39 is of type @xmath55 . [ pro:3 - 2-odd ] let @xmath39 be a cubic symmetric @xmath14-regular graph of order @xmath40 admitting a @xmath17-regular subgroup @xmath235 in its full automorphism group @xmath67 . then @xmath39 has odd automorphisms if and only if @xmath50 is odd . by proposition [ pro : cell23 ] there exists an involution @xmath321 with all of its rigid cells isomorphic to the @xmath132-tree . therefore the number of fixed vertices of @xmath225 is divisible by @xmath21 . consequently , @xmath225 is an odd permutation if and only if @xmath50 is odd . to complete the proof let us assume that @xmath50 is even , and suppose , by contradiction , that @xmath39 contains odd automorphisms . suppose first that @xmath39 contains an odd automorphism in @xmath197 , @xmath170 . this implies that @xmath322 consists of odd automorphisms and that @xmath323 consists of even automorphisms . but this contradicts the above statement about involutions in vertex stabilizers being odd if and only if @xmath50 is odd . consequently , odd automorphisms of @xmath39 can possibly only exist outside vertex stabilizers . let @xmath0 be the index @xmath17 subgroup of @xmath67 consisting of even automorphisms only . but note that the whole of @xmath197 is contained in @xmath0 , and so @xmath0 is an index @xmath17 intransitive subgroup of @xmath67 , forcing @xmath39 to be bipartite . furthermore , all automorphisms interchanging the two bipartition sets are odd permutations . by ( [ g3 ] ) , there exists an involution @xmath324 flipping an edge of @xmath39 and thus interchanging the bipartition sets . it follows that @xmath89 has @xmath50 orbits of length @xmath17 ( and no fixed vertices ) , and so is an even permutation , a contradiction . the last proposition in the series solves the problem of existence of odd automorphisms in cubic symmetric graphs of types @xmath61 , @xmath62 , @xmath325 and @xmath326 . since cubic symmetric graphs of types @xmath61 and @xmath62 are never bipartite ( see table [ tab : cubic - types ] ) , proposition [ pro:5 - 4-oddx ] implies that none of these graphs admits odd automorphisms . an example of a graph of type @xmath62 is the graph @xmath327 , the only cubic @xmath104-regular vertex - primitive graph @xcite . on the other hand , since cubic symmetric graphs of types @xmath325 and @xmath326 are always bipartite ( see table [ tab : cubic - types ] ) proposition [ pro:5 - 4-oddx ] implies that graphs of these two types admit odd automorphisms if and only if they are of order twice an odd number . [ pro:5 - 4-oddx ] let @xmath39 be a cubic symmetric @xmath104-regular graph of order @xmath40 admitting a @xmath21-regular subgroup @xmath235 in its full automorphism group @xmath67 . then @xmath39 has odd automorphisms if and only if @xmath39 is bipartite and @xmath50 is odd . suppose first that @xmath39 is bipartite and @xmath50 is odd . then the edge - flipping involution @xmath324 , which exists by ( [ g5 ] ) , is semiregular with @xmath50 orbits of length @xmath17 , and thus an odd automorphism . since @xmath39 contains odd automorphisms there exists an index @xmath17 subgroup of even automorphisms , which is either arc - transitive and hence an epimorphic image of @xmath330 or @xmath258 ( in which case the claim is clearly true ) or it is intransitive . if the latter occurs then the whole of @xmath197 , @xmath170 , consists of even automorphisms , and hence also the whole of @xmath329 consists of even automorphisms . this proves the above claim . by proposition [ pro : cell25 ] , for every vertex @xmath170 there exists a unique canonical involution @xmath331 for @xmath69 ( with @xmath138 as its rigid cell ) . note also that , by ( * ? ? ? * lemma 1 ) , @xmath171 is a central element of @xmath332 . let @xmath333 , @xmath334 , , @xmath335 be those vertices of @xmath39 for which @xmath336 , @xmath337 , are all @xmath225-rigid cells . consider the orbits of @xmath338 . orbits of @xmath338 are of length @xmath123 and @xmath17 . exactly @xmath339 orbits of @xmath338 are of length @xmath123 ( each @xmath225-rigid cell gives @xmath340 such orbits ) . orbits of @xmath338 of length @xmath17 are of two types : those with an inner edge ( we refer to them as orbits of _ type 1 _ ) and those with no inner edge ( we refer to them as orbits of _ type 2 _ ) . let @xmath341 be the number of orbits of @xmath338 of type 1 and @xmath342 be the number of orbits of @xmath338 of type 2 . then @xmath343 of @xmath39 with respect to the set of orbits of @xmath338 . let us modify the quotient graph @xmath344 with the removal of all orbits of length @xmath123 , and orbits of type 1 adjacent to these orbits ( note that there are @xmath345 such orbits as orbits of length @xmath123 form an @xmath133-tree ) . we obtain a graph which is a subdivision of a cubic graph . such a graph must have an even number of vertices of valency @xmath14 , meaning that @xmath346 , and thus also @xmath342 , is an even number . suppose first that @xmath347 . then , by ( [ eq : j ] ) , @xmath297 and @xmath50 are of the same parity , implying that @xmath225 has an even number of orbits of length @xmath17 , and is thus an even automorphism . since , by ( [ g5 ] ) , all involutions @xmath348 , @xmath254 , @xmath255 and @xmath77 in a generating set of @xmath67 are conjugate , we may conclude that @xmath197 , @xmath170 , consists of even automorphisms as claimed . we may therefore assume that @xmath349 . ( note that in this case @xmath39 is non - bipartite in view of the fact that there exist both a vertex fixed by @xmath225 and an edge flipped by @xmath225 . by table [ tab : cubic - types ] , @xmath39 is either of type @xmath62 or of type @xmath61 . ) by ( [ eq : j ] ) it suffices to show that @xmath341 is even . let @xmath350 be the set of @xmath341 edges flipped by @xmath225 . since @xmath225 is the central element in @xmath351 , one can easily see that @xmath352 for each @xmath353 . since @xmath354 the length of every orbit of @xmath197 on @xmath350 is a divisor of @xmath355 . further , since the stabilizer of an edge in @xmath39 is a @xmath17-group , the element of order @xmath14 in @xmath197 can not fix an edge in @xmath350 , and so the length of every orbit of @xmath197 on @xmath350 is divisible by @xmath14 . in fact the lengths of orbits of @xmath197 on @xmath350 are divisible by @xmath26 . suppose on the contrary that there exists an orbit of @xmath197 on @xmath350 of length @xmath14 . then the intersection @xmath356 of the edge stabilizer @xmath357 , @xmath358 with @xmath197 is a sylow @xmath17-group of @xmath197 , and thus isomorphic to @xmath359 . since @xmath225 flips the edge @xmath358 , it follows that half of the elements of @xmath356 fix @xmath360 point - wise whereas the other half interchange @xmath68 and @xmath361 . note however that the edge - stabilizer in a @xmath104-regular cubic graph is isomorphic to @xmath362 and has @xmath363 subgroups of order @xmath113 . amongst these @xmath363 subgroups only the stabilizer of the arc @xmath364 is isomorphic to @xmath359 , a contradiction . it follows that the length of every orbit of @xmath197 on @xmath350 is divisible by @xmath26 . we conclude that @xmath350 is of even cardinality , that is , @xmath341 is even . it follows by ( [ eq : j ] ) that @xmath297 and @xmath50 are of the same parity , and consequently @xmath225 has an even number of orbits of length @xmath17 , and is therefore an even automorphism . since , by ( [ g5 ] ) , all involutions @xmath348 , @xmath254 , @xmath255 and @xmath77 in a generating set of @xmath67 are conjugate , we may conclude that @xmath197 , @xmath170 , consists of even automorphisms as claimed . this completes the proof of claim 2 . we are now ready to prove that the existence of odd automorphisms in @xmath39 implies that @xmath39 is bipartite of order @xmath275 . so let us assume that there exists an odd automorphism in @xmath67 . in view of claim 2 vertex stabilizers consist of even automorphisms , and so @xmath67 contains an intransitive subgroup of index @xmath17 consisting of even automorphisms , forcing @xmath39 to be bipartite . in addition , all automorphisms interchanging the two bipartition sets are odd . in particular , the edge - flipping involution @xmath324 , which exists by ( [ g5 ] ) , is odd . it follows that the bipartition sets are of odd cardinality , that is , @xmath50 is odd . this completes the proof of proposition [ pro:5 - 4-oddx ] . let @xmath39 be a cubic symmetric graph . then @xmath39 is of one of @xmath38 possible types ( see subsection [ sec : cubic - sym ] ) . part ( i ) of theorem [ thm : main ] follows from proposition [ pro : cubic-1-regular ] for graphs of type @xmath41 , from proposition [ pro : cubic-4 - 2x ] for graphs of type @xmath125 , from proposition [ pro:3 - 2-odd ] for graphs of types @xmath42 , @xmath43 , @xmath44 and @xmath45 , and from proposition [ pro:5 - 4-oddx ] for graphs of types @xmath48 and @xmath49 . since every cubic symmetric graph of type @xmath46 is bipartite ( see table [ tab : cubic - types ] ) , part ( i ) of theorem [ thm : main ] for graphs of type @xmath46 follows from proposition [ pro : cubic - s - regular - without - s-1 ] . finally , part ( iv ) of theorem [ thm : main ] follows from proposition [ pro : cubic-2 - 2 ] for graphs of type @xmath60 , and from proposition [ pro:5 - 4-oddx ] for graphs of types @xmath61 and @xmath62 ( since none of graphs of types @xmath61 and @xmath62 is bipartite ) . the work of klavdija kutnar was supported in part by the slovenian research agency ( research program p1 - 0285 and research projects n1 - 0032 , n1 - 0038 , j1 - 6720 , j1 - 6743 , and j1 - 7051 ) , in part by woodwisdom - net+ , w@xmath365b , and in part by nsfc project 11561021 . the work of dragan marui was supported in part by the slovenian research agency ( i0 - 0035 , research program p1 - 0285 and research projects n1 - 0032 , n1 - 0038 , j1 - 5433 , j1 - 6720 , and j1 - 7051 ) , and in part by h2020 teaming innorenew coe .

resonant interaction of metal nanoparticles ( mnps ) with optical light provides a tool for the strong localization of electromagnetic field @xcite . intensity enhancements as high as @xmath0 can be achieved @xcite within the localized surface plasmon - polariton ( pp ) fields , in terms of coupled oscillations of surface electrons and the localized optical field @xcite . such orders of magnitude increase in the intensity leads to the emergence of optical nonlinearities @xcite , e.g. enhanced raman scattering @xcite , four wave mixing @xcite and second harmonic generation ( shg ) @xcite . emergence of nonlinear processes can be both desirable or unwanted depending on the operating properties of the fabricated device . as an example , plasmon - polariton mediated surface enhancement is successfully used to achieve raman imaging of materials @xcite . the nonlinear response of the media can also be utilized for optical switching @xcite . the shg process can enhance the absorption efficiency in photovoltaic devices @xcite , may increase the coherence time ( length ) of the field @xcite as well as being able to generate entangled photon pairs @xcite . despite such advantages , nonlinear conversion may be undesirable in other devices . raman scattering process in fiber - optic cables causes loses in the signal and limits the number of channels that could be used for a given bandwidth @xcite . similarly , nonlinear effects can decrease the quality factor of microwave cavities @xcite . in addition , one may require the operation of a device in the linear regime even for higher input powers . because nonlinearities may cause unexpected chaotic behavior for the long term operation @xcite . besides the emergence of nonlinearities , fano resonances analogous to electromagnetically induced transparency @xcite has also been observed in plasmonic excitations of mnps @xcite . the attachment of a quantum oscillator [ e.g. a molecule or a quantum dot ( qd ) ] to a mnp strongly modifies the optical response of the hybrid material @xcite . the presence of a quantum oscillator with small decay rate induces a weak hybridization . hybridization is weak in the sense that frequency splitting in the mnp resonance is very small compared to the spectral width of the resonance . this introduces two possible excitation paths for the absorption / polarization of the incident light , which are unresolvable . both excitation frequencies lie within the frequency window of mnp resonance and interfere destructively @xcite . there emerges a transparency window centered about @xmath1 , where polarization of mnp - quantum oscillator hybrid system is avoided @xcite . such resonances are observed as long as the decay / damping rate of one of the oscillators is significantly small compared to the second one @xcite . in this paper , we show that it is possible to manage the nonlinear behavior of a material using the path interference effects . as an example , we consider the following system . we place a quantum oscillator ( qd , molecule or a nitrogen vacancy center ) at the hot - spot of a mnp dimer ( see fig . [ fig1 ] , _ top _ ) which has a low decay rate ( @xmath2 ) compared to the mnp @xcite ( @xmath3 ) . the dimer has localized surface plasmon - polariton ( pp ) resonances @xmath4 and @xmath5 ( see fig.[fig1 ] , _ bottom _ ) for the polarization field . the drive frequency @xmath6 and the second harmonic ( sh ) frequency @xmath7 fall into the excitation range of the @xmath4 and @xmath5 polarization modes , respectively . without the presence of the quantum oscillator , resonance of the shg process occurs when @xmath8 and @xmath9 ( see the discussion in sec . [ sec : suppression ] ) . we show that , ( i ) even in the resonance condition for sh conversion ( that is @xmath8 , @xmath9 ) , the presence of coupling to the quantum oscillator ( emitter ) can suppress the nonlinear process several orders of magnitude . the factor of achievable suppression is inversely proportional to the square of the quantum decay rate and increases with the strength of the mnp - quantum oscillator coupling . a suppression factor of @xmath10 is possible when a high - quality ( small decay rate ) quantum oscillator ( a qd ) , with spectral width of @xmath11 hz , is coupled to the mnp ( see fig . [ fig2 ] ) . this effect is observed , because cancellation of the two excitation paths does not allow polarization in the @xmath5 pp mode of the mnp dimer . suppression is maximum when quantum level spacing is resonant to conversion frequency , @xmath12 . ( ii ) on the other hand , a similar path cancellation effect can be adopted to kill the nonresonant term [ ( @xmath13 ) ] that emerges when @xmath5 is not resonant to the shg frequency @xmath7 . without adjusting the resonances @xcite of the dimer ( @xmath4 , @xmath5 ) , the shg process can be carried closer to resonance ( fig . [ fig4 ] ) . these two effects together , enables the control over induction of the nonlinearities without the need for managing the properties of the material . ) drives the @xmath14 pp polarization mode ( resonance @xmath4 ) of the dimer . the intense localized polarization field of @xmath14 , oscillating with @xmath6 , gives rise to shg @xcite . this process induces oscillations ( @xmath15 ) in the second pp mode @xmath16 whose resonance is @xmath5 . the quantum oscillator ( level spacing @xmath17 ) interacts with the field of the @xmath16 polarization mode . quantum oscillator ( emitter ) is chosen to have no sh response @xcite to @xmath6 . the observation of the sh light occurs due to the radiative decay of the @xmath16 pp mode @xcite.,width=326 ] the effect of fano resonances on the nonlinear conversion processes has already been studied both theoretically ( using finite element simulations ) @xcite and experimentally @xcite . however , an explicit demonstration [ i.e. as in eq . ( [ alph2 ] ) below ] of how path interference can tune the conversion has not been examined , yet . in an exciting recent work @xcite , a theoretical model of enhancement of the nonlinear response originating from a quantum dot by help of plasmonic particles was studied . our work , in contrast , provides a theoretical model of how the nonlinear response originating from a plasmonic ( classical ) system is enhanced by a quantum oscillator . emergence of the suppression phenomenon necessitates presence of coupling to a quantum oscillator with a small decay rate . however , enhancement effect may also be observed for coupled plasmonic resonators with broad absorption ( emission ) bands @xcite . we verify the emergence of such an enhancement by 3-dimensional ( 3d ) boundary element method simulations using the mnpbem toolbox @xcite in matlab ( see fig . [ fig5 ] ) . our simulations take the retardation effect into account . the simple model can predict the both the emergence and the spectral position of the shg enhancement successfully . in a separate experiment of our research team @xcite , we observe the sh radiation from the hybrid system of composite mnps which are decorated with dye molecules . shg can originate only from the mnps since the eyfp molecule @xcite does not have a sh response to the drive frequency @xcite . our simple model can easily predict [ using eq.s ( [ eq : timea])-([eq : timed ] ) ] an enhancement factor of @xmath181000 in the experiment . taking this enhancement factor into account , sh signal from mnp clusters illuminated with a cw laser ( 40mw/@xmath19 at the sample) reaches the values attained in the typical experiments @xcite where samples are illuminated with high peak intensity ( 60gw/@xmath19 ) ultra - short lasers . in fact , the observed effect is the classical ( or semi - classical ) analog of nonlinear response enhancement obtained via electromagnetically induced transparency ( eit)-like atomic coherence @xcite . in difference , the present system does not necessitate a microwave drive , which makes it efficient considering energy consumption . here , we present the method for the second harmonic generation process in a mnp dimer . however , method can be generalized to other nonlinear frequency generation processes as long as the two modes of the material can be resolved ( see sec . [ sec : othernl ] ) . it is also possible to use the excitation modes of other nanoscale resonators @xcite . the paper is organized as follows . in sec . [ sec : hamiltonian ] , we describe the shg process in the coupled system of a mnp dimer and a quantum oscillator . we introduce the hamiltonian for the hybrid system . nonlinear frequency conversion process is included in the second quantized hamiltonian . we derive the equations of motion for the system using the density matrix formalism for the quantum oscillator . we include the damping and quantum decay rates and the source driving the mnp dimer . in sec . [ sec : suppression ] , we demonstrate that conversion process is suppressed for @xmath20 . in sec . [ sec : enhancement ] , we present a contrary effect . the cancellation of the nonresonant terms leads to enhanced production of the shg for the choice of @xmath21 . in sec . [ sec : mnpbem ] , we compare the results of our model with the 3d simulations which are based on the exact solutions of the maxwell equations . in sec . [ sec : othernl ] , we discuss how the model can be adopted to other nonlinear processes . in sec . [ sec : classification ] , we provide a classification for the types of fano resonances and their relevance with eit @xcite , for the sake of generalization of the enhancement phenomenon to other composite systems . [ sec : conclusion ] includes our conclusions . in this section , we describe the response of a coupled mnp dimer - quantum oscillator system to a driving electromagnetic field . we shortly mention about the nature of couplings in the hybrid system and the mechanism for shg on the mnp - dimer resonator . we give the effective hamiltonian for the system and drive the equations of motion for the fields of the plasmon - polariton modes together with the excitation of the quantum oscillator . we find the equations governing the steady state values of the excitations to obtain the linear behavior of the hybrid system . using these equations , we demonstrate the principle behind gaining control over the process of nonlinear frequency generation . we show that by choosing the appropriate level spacing ( @xmath1 ) for the quantum oscillator , one can either suppress and or enhance the non - linear frequency generation . we consider a system where a quantum oscillator ( e.g. quantum dot @xcite , molecule @xcite or a nitrogen - vacancy center @xcite ) is placed in the center of the mnp dimer . the two mnps can still be coupled to each other , and be said to have dimerized , due to the small dimensions of the quantum oscillator @xcite . the two resonances of the mnp dimer @xmath4 and @xmath5 are relevant to the incident ( @xmath6 ) and sh ( @xmath7 ) frequencies , respectively ( see fig . [ fig1 ] , _ bottom _ ) . the resonance frequency of the @xmath16 pp mode ( @xmath5 ) is about the sh frequency @xmath7 , but not necessarily resonant with it . the incident light , in the planewave mode with frequency @xmath6 , couples strongly to the @xmath14 plasmon - polariton mode of the dimer . the direct coupling of light to the quantum oscillator is of negligible strength compared to the plasmon . quantum oscillator couples to the localized plasmon - polariton field of the dimer . the hot - spots for the both plasmon - polariton modes emerge in the middle of the two mnps @xcite , where the quantum oscillator is tightly placed . the dynamics of the total system is as follows . the incident planewave mode field ( @xmath22 ) drives the first dimer mode @xmath14 ( resonance @xmath4 ) at the oscillation frequency @xmath6 . the polarization of the plasmon - polariton ( pp ) excitation yields a localized strong electromagnetic field mode ( @xmath14 ) between the two mnps . such an enhancement in the field gives rise to the emergence of nonlinear effect ( e.g. shg ) in the electron gas @xcite . explicitly ; the field ( oscillating at @xmath6 ) trapped in the @xmath14 pp polarization gives rise to second harmonic polarization oscillations ( @xmath7 ) @xcite in the @xmath16 pp polarization mode of the dimer . the quantum oscillator , whose level spacing is compatible with the sh oscillation frequency @xmath17 , interacts with the polarization field of the @xmath16 pp mode . the resonance of @xmath16 mode is @xmath5 . the field localization at the hot - spot provides strong interaction with the quantum oscillator . the sh light is observed through the radiative decay of the @xmath16 pp mode @xcite . we assume that the molecule does not have a shg response to the drive frequency @xmath6 . such molecules exists ; for example , see fig . 2 in ref . @xcite . here , we consider a simplified model for the hybrid system . we mainly aim to demonstrate the principles behind the control mechanism . in more realistic calculations @xcite , one has to consider complicating effects , such as the influence of the dielectric environment and exact spatial distribution of the fields . however , the oscillators model @xcite predicts the basic behavior of the mnps combined with the quantum oscillators @xcite . the total hamiltonian @xmath23 for the described system can be written as the sum of the energy of the quantum oscillator @xmath24 , energy of the plasmon - polariton oscillations @xmath25 of the mnp dimer @xmath26 , the interaction of the quantum oscillator with the plasmon - polariton modes @xcite @xmath27 @xmath28 as well as the energy transferred by the pump source @xmath29 , @xmath30 and the second harmonic generation process among the plasmon - polariton fields @xmath31 @xmath32 respectively @xcite . in eq . ( [ eq : h0 ] ) , @xmath33 ( @xmath34 ) is the excited ( ground ) state energy of the quantum oscillator . states ( @xmath35 ) , @xmath36 correspond to the ( excited ) ground levels of the quantum oscillator . @xmath14 , @xmath16 are the plasmon - polariton excitations induced on the mnp dimer and @xmath37 , @xmath38 are the corresponding energies for the oscillation modes . @xmath39 ( @xmath40 ) is the coupling matrix element between the field induced by the @xmath14 ( @xmath16 ) polarization mode of the mnp dimer and the quantum oscillator . ( [ eq : hp ] ) describes the interaction of the light source ( oscillates as @xmath41 ) driving the plasmon - polariton mode with smaller resonance frequency @xmath4 . in eq . ( [ eq : hsh ] ) , the fields of two excitations in the low - energy plasmon - polariton mode ( @xmath14 ) combine to generate the field of a high energy plasmon - polariton mode . stronger the second harmonic generated plasmon - polariton oscillations , the higher the number of emitted shg photons @xmath42 . because , @xmath16 mode radiatively decay to @xmath7 photon mode @xcite . energy is conserved in the input - output process . the parameter @xmath43 , in units of frequency , is proportional to the second harmonic susceptibility of the mnp dimer . we note that , one could also treat the shg process as originating directly from the incident field , e.g. @xmath44 . even though the following results would remain unaffected , physically such a model would be inappropriate . because , enhanced nonlinear processes emerge due to the electromagnetic field of the localized intense surface plasmon - polariton ( polarization ) mode @xcite . however , the mode of the incident field ( @xmath6 ) is planewave . we use the commutation relations ( e.g. @xmath45 $ ] ) in driving the equations of motions . after obtaining the dynamics in the quantum approach , we carry @xmath46 to classical expectation values @xmath47 . we introduce the decay rates for plasmon - polariton fields @xmath48 , @xmath49 . quantum oscillator is treated within the density matrix approach . since we restrict ourselves to the classical properties of the fields , e.g. we do not not take squeezing into account , we could alternatively take the plasmon fields to be of classical nature in eq.s ( [ eq : h0])([eq : hsh ] ) . we could derive the equations of motion ( [ eq : timea])-([eq : timeb ] ) by functional minimization . however , we choose to keep the operator notations for @xmath14 and @xmath16 quanta up to a certain derivation step in order to avoid incomplete modeling of the equations of motion . the equations of motion take the form @xmath50 , \label{eq : timed}\end{aligned}\ ] ] where @xmath51 , @xmath52 are the damping rates of the mnp dimer modes @xmath48 , @xmath49 . @xmath53 and @xmath54 are the diagonal and off diagonal decay rates of the quantum oscillator , respectively . to make a comparison , @xmath51,@xmath55hz for mnps @xcite while @xmath56hz for molecules @xcite and @xmath57 hz for quantum dots @xcite . the constraint on the conservation of probability @xmath58 accompanies eqs . ( [ eq : timea]-[eq : timed ] ) . in our simulations ( figs . [ fig2][fig4 ] ) , we time - evolve eqs . ( [ eq : timea]-[eq : timed ] ) numerically to obtain the long time behavior of @xmath59 , @xmath60 , @xmath48 , and @xmath49 . we determine the values to where they converge when the drive is on for long enough times . we perform this evolution for different @xmath5 frequency values with the initial conditions @xmath61 , @xmath62 , @xmath63 , @xmath64 . beside the time - evolution simulations , one may gain understanding about the linear behavior of eqs . ( [ eq : timea]-[eq : timed ] ) by seeking solutions of the form @xmath65 for the steady states of the oscillations . this form of solutions are valid under the following assumptions . when the level spacing of the quantum oscillator is about the sh frequency , @xmath66 , its interaction with the first pp mode becomes highly off - resonant as compared to the @xmath16 mode . in addition , mnp system can be chosen such that the hotspots of @xmath14 mode and @xmath16 mode emerge at different spatial positions . the quantum oscillator can be placed at the @xmath16 hot - spot . in this case , its interaction with @xmath14 mode can be neglected even without the need for the off - resonance assumption . in our numerical simulations governing the time - evolution of eqs . ( [ eq : timea]-[eq : timed ] ) , we check that the solutions indeed converge to the form of eq . ( [ eq : slowvary ] ) for long - time behavior . inserting eq . ( [ eq : slowvary ] ) into eqs . ( [ eq : timea]-[eq : timed ] ) , one obtains the equations for the steady state @xmath67\alpha_1+i2\chi^{(2)}\alpha_1^*\alpha_2=\epsilon_p , \label{eq : steadya } \\ [ i(\omega_2 - 2\omega)+\gamma_2]\alpha_2+i\chi^{(2)}\alpha_1 ^ 2=-i f_2\rho_{ge } , \label{eq : steadyb } \\ [ i(\omega_{eg}-2\omega)+\gamma_{eg}]\rho_{ge}=i f_2\alpha_2(\rho_{ee}-\rho_{gg } ) , \label{eq : steadyc } \\ \gamma_{ee}\rho_{ee}=i f_2(\alpha_2^*\rho_{ge}-\alpha_2\rho_{ge}^ * ) , \label{eq : steadyd}\end{aligned}\ ] ] where @xmath68 , @xmath69 , @xmath70 , @xmath71 are constants independent of the time . using eqs . ( [ eq : steadyb ] ) and ( [ eq : steadyc ] ) , one can obtain the steady state value of @xmath16 plasmon polariton mode field as @xmath72}\;\tilde{\alpha}_1 ^ 2 \label{alph2}\ ] ] for @xmath66 . here , @xmath73 is the steady state value of the population inversion . quantum oscillator couples with the @xmath16 mode into which sh conversion take place . since the sh intensity is weak , quantum oscillator only weakly excited . therefore , in eq . [ alph2 ] inversion usually takes on values very close to @xmath74 . even for enhanced sh conversion , as discussed in sec . [ sec : enhancement ] , we observe in our simulations that the value of @xmath75 does not rise above -0.9 . in our simulations , @xmath75 is not used as a fixed parameter with value @xmath76 . we always determine @xmath75 from the time evolution of @xmath77 . taking a closer look at the denominator of eq . ( [ alph2 ] ) , one can immediately realize that @xmath78 attains huge values on resonance @xmath12 . because , linewidth of the quantum oscillator ( @xmath2 ) is very small compared to all other frequencies . if @xmath79 , the largeness of the @xmath78 term dominates the denominator . this results in the suppression of the generation of the @xmath69 plasmon - polariton polarization field in the mnp dimer . in fig . [ fig2 ] , we demonstrate that the shg in the mnp dimer can be suppressed very effectively by coupling the mnp dimer to a quantum oscillator . we time evolve eqs . ( [ eq : timea]-[eq : timed ] ) to obtain steady state values for the excitations . plasmon - polariton ( pp ) mode from the @xmath14 mode . even at the presence of resonant conversion condition , @xmath8 and @xmath9 , the presence of quantum oscillator ( @xmath12 ) prevents the take place of the shg process . eit does nt allow the polarization in the @xmath16 pp mode . the resonant conversion is represented by unity in the figure . when @xmath12 , the nonlinear intensity can be suppressed 9-orders of magnitude with respect to the resonant value . decay rates are @xmath80 and @xmath81 . we use @xmath82 and @xmath83.,width=307 ] without the presence of a quantum oscillator , the shg would be maximum @xmath84 that is when the second plasmon - polariton mode is on resonance @xmath9 [ see eq . ( [ alph2 ] ) ] . in fig . [ fig2 ] , we observe that even at the presence of this resonance @xmath85 , eit suppresses the shg by 9 orders of magnitude . this effect arises simply because eit does nt allow the polarization of the second plasmon - polariton mode @xmath16 at @xmath7 . the two paths introduced in the mnp dimer due to the hybridization with the quantum oscillator for polarization transfer from the @xmath14 mode interfere destructively . this cancels the transfer of @xmath14 polarization ( oscillating at @xmath6 ) to @xmath16 polarization in the dimer ( oscillating at 2@xmath6 ) . [ fig3 ] shows the dependence of the shg intensity @xmath86 on the decay rate @xmath53 of the quantum oscillator that is coupled to the dimer . the slope of the graph implies a @xmath87 dependence . in fact , this can be easily inferred from eq . ( [ alph2 ] ) for the small values of @xmath2 when @xmath88 . . the log - log plot has a steepness of value @xmath892 , thus pointing out the relation @xmath90 . this relation can be easily inferred from eq . ( [ alph2 ] ) for small values of @xmath53 with @xmath12 . smaller quantum decay rate results in higher quality suppression.,width=307 ] contrary to suppression phenomenon , the interference effects can be arranged so that shg process can be carried closer to the resonance . in the denominator of eq . ( [ alph2 ] ) , the imaginary part of the first term @xmath91 $ ] can be arranged to cancel the @xmath92 expression in the second term of the denominator . this gives the condition @xmath93=0 . \label{eqcond}\ ] ] eq . ( [ eqcond ] ) has two roots @xmath94 the first ( smaller ) root @xmath95 is not useful for shg enhancement . because it enlarges the real part of the @xmath91 $ ] term . this is already the suppression condition for shg . since @xmath1 is not very close to @xmath7 for the second root @xmath96 , it does not cause the real part of the @xmath91 $ ] term to rapidly diverge . at the same time , @xmath96 is minimizes the absolute value of the denominator of eq . ( [ alph2 ] ) , that gives rise to the maximum shg . for the case of the suppression of shg , one can safely use the approximation @xmath97 , because excitations are suppressed in the hybrid system , @xmath98 , and this leads to @xmath99 . however , in the case of shg enhancement , one can not use the value @xmath100 for @xmath75 . we observe that at resonances @xmath75 can attain inversion values that are close to zero . nevertheless , eq . ( [ eqroot ] ) still serves at least as a guess value for the order of @xmath96 , where shg enhancement arises . ) is far - off resonant to the shg ( @xmath101 ) . the nonlinear process can be carried closer to resonance by arranging the quantum level spacing @xcite to @xmath102 . the conversion is enhanced upto 30 times compared to the off - resonant process . the conversion for off - resonant process ( with @xmath103 ) is represented by unity in the figure . for @xmath12 , nonlinear process is suppressed similar to fig . note that , the bandwidth for the enhancement of the nonlinear conversion is much wider than the suppression bandwidth . decay rate of the nanoscale resonator @xcite is chosen as @xmath104 . we use @xmath82 and @xmath83.,width=307 ] in fig [ fig4 ] , we depict the enhancement of the nonlinear frequency conversion for a nanoscale dimer whose pp mode ( @xmath101 ) is far off - resonant to the sh frequency . off - resonant shg conversion [ @xmath103 in eq . ( [ alph2 ] ) ] is represented by unity in fig . we observe a 30 times enhancement in the conversion intensity for the choice of @xmath105 . parameters are given in fig . [ fig4 ] . as it can be inferred from eq . ( [ alph2 ] ) ; relative enhancement efficiency can be grown much more with respect to the off - resonant value ( @xmath101 with @xmath103 ) if higher quality ( small @xmath52 ) resonators are used . quality factors of @xmath181300 can be achieved for micro - cavities operating at optical wavelength @xcite . in obtaining eqs . ( [ eq : steadya])-([eq : steadyd ] ) , we neglect the coupling between the @xmath14 pp mode with the quantum oscillator due to off - resonant behavior . without such a negligence , analytic results like eqs . ( [ eq : steadya])-([eq : steadyd ] ) can not be obtained since expressions in eq . ( [ eq : slowvary ] ) are not valid anymore . in case , when quantum level spacing @xmath1 is close to @xmath4 , steady state value of @xmath49 is obtained numerically by time evolution of eqs . ( [ eq : timea])-([eq : timed ] ) . surprisingly , enhancement factor obtained in this case can reach as high as @xmath181000 . such an enhancement factor , consistently , shown of being able to explain the observed sh signal , from noncentrosymmetric mnp clusters decorated with molecules by cw laser irradiation @xcite . for the suppression effect to emerge , the first term in the denominator of eq . ( [ alph2 ] ) must be sufficiently large enough . this necessitates the coupling of plasmonic resonator to a quantum oscillator which has a small decay rate @xmath2 . on the other hand , cancellation of the nonresonant terms in eq . ( [ eqcond ] ) does not require the presence of a high - quality oscillator with a sharp resonance . hence , nonlinear response enhancement may emerge even for coupled plasmonic resonators with broad emission bands . in fig . [ fig5 ] , we compare the predictions of our model [ eq.s ( [ eq : timea]-[eq : timed ] ) ] with the 3-dimensional boundary element simulations . simulations are performed with the mnpbem toolbox @xcite in matlab . we use the _ bemret _ environment @xcite which is based on the exact solutions of the maxwell equations using surface integral evaluations @xcite . simulations take the retardation effects into account . we use the experimental data , that is present in the toolbox , for the dielectric function of the gold nanoparticles . coupled oscillators model correctly predicts the emergence of the shg enhancement ( compare fig . [ fig5]b and [ fig5]d ) as well as its position in the spectrum . in the mnpbem simulation , we calculate the sh response of a gold nanoparticle of 70 nm diameter . [ fig5]a shows the emitted sh radiation in case the gold nanoparticle stands alone . the driving radiation @xmath106 is converted to the sh wavelength @xmath107 . in fig . [ fig5]b , we place a small particle ( blue ) , which has a tiny decay rate @xmath108 , near the gold nanoparticle . the small blue sphere , of 12 nm diameter , is filled with a dielectric medium which has a lorentzian response function @xmath109 , where @xmath110 determines the oscillation ( polarization ) strength . in this way , the small blue particle mimics the response of a dye molecule having a sharp resonance ( @xmath111 ) near @xmath1 ( @xmath112500 nm ) . [ fig5]b shows that shg is enhanced about 100 times near @xmath113 for @xmath114500 nm . the take place of such a phenomenon for @xmath115255 nm can be predicted within the simulations of our model . the shg intensity for the classical oscillator ( fig . [ fig5]c ) is enhanced about 40 times at the same spectral position ( fig . [ fig5]d ) . a careful glance at fig . [ fig5]b reveals that the normal shg remains at @xmath116240 nm . it is also worth noting that , a linear fano resonance @xcite reveals itself in the form of a dip at the center of the scattering cross - section peak ( at @xmath117 nm ) in bem simulation . the fact that position of the dip follows the resonance @xmath118 , ensures that this effect is indeed a fano resonance not a simple hybridization between the two particles . in fig . [ fig5 ] , we demonstrate the enhancement of shg for @xmath119 , in difference to the case we discuss in eq . ( [ alph2 ] ) and fig . [ fig3 ] where @xmath120 . mnpbem toolbox is limited in handling fano resonances for @xmath66 . it treats the second power of the electric field s normal component at the(inner ) particle boundaries as sh source and calculates the sh field intensity @xcite . since sh field ( @xmath7 ) is weak compared to the linear ( @xmath6 ) response and @xmath121 is off - resonant to @xmath7 , fano resonances near @xmath121 can be treated within this approach . on the contrary , fano resonances near @xmath66 can not be handled within this method . because , presence of a particle ( blue ) resonant to @xmath122 can not alter ( has no feedback on ) the magnitude of the generated sh sources which are only the second power of the linear ( @xmath6 ) electric field distribution . nevertheless , fig . [ fig5 ] reveals the reliability of our model which can explain the underlying physical phenomenon leading to control over the sh response . the fano resonance effects for @xmath66 , depicted in fig.s [ fig2 ] and [ fig3 ] , would be observed for a more complete treatment studied in ref . @xcite . here , we discuss a single mnp coupled to a high quality oscillator , in difference to 2 mnps configuration presented in fig . [ fig1 ] . the arrangement pictured in fig . [ fig1 ] provides stronger coupling between mnps and quantum oscillator . however , 2 mnps give two scattering resonances about @xmath4 , which complicates the comparison between mnpbem simulation and simple oscillators simulation . the double peak in fig . [ fig5]a is due to the emergence of double resonance near @xmath5 even for a single mnp . the coupled system is excited with an x - polarized ( electric field is along the line connecting the centers of the mnp and quantum dot ) plane wave . when y- or z - polarized plane wave excitation is used , no enhancement of sh conversion is obtained . in fig . [ fig6]a , we show the dependence of the shg enhancement factor to the size of the gap between the mnp and the quantum dot ( qd ) . in fig . [ fig6]b , variation of the corresponding maximum enhancement wavelength as a function of gap size is given . , and ( b ) corresponding excitation wavelengths @xmath123 where enhancement emerges . the system is the same with fig . [ fig5 ] . only the distance between mnp and quantum oscillator ( a qd ) is varied . , width=326 ] in case the control of the third harmonic generation ( thg ) is required , instead of shg , change of eq . ( [ eq : hsh ] ) to @xmath124 would be sufficient . ( [ alph2 ] ) should simply be modified to @xmath125}\;\tilde{\alpha}_1 ^ 3 . \label{eq : alpha2th}\ ] ] all of our considerations , stated above , works for eq . ( [ eq : alpha2th ] ) as well . in order to be able to use the introduced model , one has to be careful if the nonlinear conversion oscillates another pp mode which is resolvable from the first one . in some nonlinear conversion processes ( e.g. enhanced raman scattering @xcite ) , both the drive ( @xmath6 ) and the generated ( @xmath126 ) frequencies may excite the same pp mode ( @xmath14 ) . this may occur because , frequency spacing @xmath127 can be small compared to the decay rate ( @xmath51 ) of the pp polarization mode . in this case too , suppression of the nonlinear conversion will necessarily arise for @xmath128 . this is because ; coupling to a high quality factor oscillator prevents the polarization / absorption for frequency values around @xmath129 independent of the details of the conversion mechanism @xcite . the generated frequency mode will be suppressed if it coincides with @xmath1 . fig . 5 in ref . @xcite demonstrates that polarization cancellation emerges at @xmath130 for the coupled mnp - quantum oscillator system . however , enhancement of the nonlinearities depends on the physics of the conversion mechanism . fano resonances induced in coupled classical / quantum oscillators follow from a common mechanism . coupling of the resonant excitation to an auxiliary mode ( qd in our case ) introduces two possible absorption paths . the two paths counteract and avoid the excitation ( hence polarization and absorption ) at the resonance @xcite . the same mechanism is also responsible for the phenomenon of electromagnetically induced transparency ( eit ) @xcite in 3-level atoms . path interference effect can reveal itself in a variety of systems . therefore , a classification of such coupled systems becomes necessary , for the purpose of illuminating the possible extensions of the nonlinearity enhancement to other systems . the first class can be defined as follows . a classical / quantum oscillator can be coupled to a classical or quantum object . here , the word _ coupling _ implies an interaction which does not yield a strong hybridization that can change the entire spectrum severely . in other words , the splitting in the plasmon mode ( which has a wide spectral width ) is below the resolution limit . this class contains the following examples . ( i ) a mnp attached with a molecule / qd , ( ii ) a molecule exhibiting nonlinear response attached to a qd @xcite , or ( iii ) two weakly interacting mnps ( e.g. distance between two mnps is relatively long or a dielectric is placed in between ) . the second class is closer , in physical aspects , to the standard eit . coupling is strong enough to completely modify the spectrum , e.g. two closely placed hybridized mnps ( dimer ) . in this case , too , fano resonance can be induced if the bright ( dipole - like ) and dark ( higher orders ) modes overlap spectrally @xcite and spatially i.e . overlap integral for the interaction hamiltonian does not vanish ( similar to ref . this is a single system whose internal states interact . in this respect , this class is analog of eit . the difference is ; in eit , the third level ( dipole - forbidden transition ) first level is the ground state , second is the dipole - allowed excited state is coupled externally with the excited level using a microwave drive @xcite . in the plasmonic analog , coupling between the two excitations arises naturally due to the spectral overlap , which is absent in 3-level atoms . it is well demonstrated that the presence of a quantum oscillator , with a smaller decay rate , changes the optical response of mnps dramatically . due to the destructive interference of the ( hybridized ) absorption paths , mnp can not be polarized at the resonance frequency of the quantum oscillator . we demonstrate that a similar path interference effect can be adopted to both suppress and enhance the nonlinear conversion processes in a mnp dimer . a quantum oscillator is placed in the center of two hybridized mnps where hot - spot of both plasmon - polariton mode emerges . if the quantum oscillator is resonant to the second or third harmonic frequency , ( e.g. @xmath12 ) , this frequency conversion process is suppressed several orders of magnitude . because , eit does not allow the excitation of the @xmath7 oscillation in the second plasmon - polariton mode of the dimer ( @xmath5 ) . on the other hand , the similar interference effects can be used also to enhance the nonlinear frequency conversion . the level spacing of the quantum oscillator can be arranged @xcite so that the nonresonant [ e.g. ( @xmath13 ) ] terms cancel . it is worth noting that emergence of the enhancement phenomenon does not require coupling to a high - quality oscillator . a small decay rate ( @xmath2 ) is not necessary for nonresonant terms to cancel in eq . ( [ eqcond ] ) . hence , the effect can readily be observed also for two coupled low - quality mnps , if one chooses the resonances of the nanoparticles properly . we compare the predictions of our simple model with 3d mnpbem simulations which are based on the exact ( computational ) calculations of the maxwell equations . we show that our model successfully predicts the emergence of the shg enhancement as well as its position in the spectrum . we present our method for the engineering of shg process . however , the method can be used also for other nonlinear frequency generation processes . this happens as long as the converted frequency falls in the range of a different plasmon - polariton mode . contrary to the enhancement case , suppression phenomenon is independent of the conversion mechanism . if a quantum oscillator resonant to the generated frequency is coupled to the mnp system , the conversion process is prohibited . hence , it can be used to suppress undesired raman scattering processes which cause loses in the signal strength . m.e.t and d.t . acknowledge support from tbitak - kariyer grant no . this work was undertaken while one of the authors ( m.e.t ) was in residence at bilkent university with the support provided by ouz glseren . acknowledges support from bilim akademisi the science academy , turkey under the bagep program and metu bap-08 - 11 - 2011 - 129 grant . the research leading to these results has received funding from the european union s seventh framework program fp7/2007 - 2013 under grant agreement no . 270483 . k. thyagarajan , j. butet , and olivier j. f. martin , nano lett . * 13 * , 1847 ( 2013 ) . j. berthelot , g. bachelier , m. song , p. rai,1 g. c. des francs , a. dereux , and a. bouhelier , opt . express * 20 * , 10498 ( 2012 ) . s. wunderlich and u. peschel , opt . express * 21 * , 18611 ( 2013 ) . s. gao , k. uneo and h. misawa , accounts chem . res . * 44 * , 251 ( 2011 ) . m. s. singh , nanotechnology * 24 * , 125701 ( 2013 ) . g. f. walsh and l. d. negro , nano lett . * 13 * , 3111 ( 2013 ) . k. kneipp , y. wang , h. kneipp , l. t. perelman , i. itzkan , r. r. dasari , and m. s. feld , phys . lett . * 78 * , 1667 ( 1997 ) . a. r. chraplyvy , j. lightwave technol . * 8 * , 1548 ( 1990 ) . a. m. glass , d. j. digiovanni , t. a. strasser , a. j. stentz , r. e. slusher , a. e. white , a. r. kortan and b. j. eggleton , bell labs tech . * 5 * , 168 ( 2000 ) . j. toulouse , lightwave technol . * 23 * , 3625 ( 2005 ) . l. g. l. wegener , m. l. povinelli , a. g. green , p. p. mitra , j. b. stark and p. b. littlewood , physica d * 189 * , 81 ( 2005 ) . j. wiersig and m. hentschel , phys . a * 73 * , 031802 ( 2006 ) . a. manjavacas , j. garca de abajo , and p. nordlander , nano lett . * 11 * , 2318 ( 2011 ) . p. weis , j. l. garcia - pomar , r. beigang and m. rahm , opt . express * 19 * , 23573 ( 2011 ) . s. g. kosionis , a. f. terzis , s. m. sadeghi and e. paspalakis , j. phys . matter * 25 * , 045304 ( 2013 ) . r. d. artuso and g. w. bryant , phys . b * 82 * , 195419 ( 2010 ) . r. d. artuso and g. w. bryant , nano lett . * 8 * , 2106 ( 2008 ) e. waks and d. sridharan , phys . a * 82 * , 043845 ( 2010 ) . a. ridolfo , o. distefano , n. fina , r. saija , and s. savasta , phys . * 105 * , 263601 ( 2010 ) . w. zhang , a. o. govorov and g. w. bryant , _ phys . lett . _ * 97 * , 146804 ( 2006 ) . s. g. kosionis , a. f. terzis , v. yannopapas and e. paspalakis , j. phys . c * 116 * , 23663 ( 2012 ) . m. a. noginov , g. zhu , a. m. belgrave , r. bakker , v. m. shalaev , e. e. narimanov , s. stout , e. herz , t. suteewong and u. wiesner , nature * 490 * , 1110 ( 2009 ) . m. pfeiffer , k. lindfors , c. wolpert , p. atkinson , m. benyoucef , a. rastelli , o. g. schmidt , h. giessen and m. lippitz , nano lett . * 10 * , 4555 ( 2010 ) . l. zhao , t. ming , h. chen , y. liang and j. wang , nanoscale * 3 * , 3849 ( 2011 ) . p. anger , p. bharadwaj and l. novotny , phys . lett . * 96 * , 113002 ( 2006 ) . sh conversion of molecules can not be emerging due to the field enhancement effect of the mnps , as well . this is because , in ref . @xcite we note that no sh signal is detected from centrosymmetric mnp clusters decorated with molecules . c. ciraci , e. poutrina , m. scalora and d. r. smith , phys . b * 85 * , 201403 ( 2012 ) . m. finazzi and f. ciccacci , phys . b * 86 * , 035428 ( 2012 ) . m. i. dykman , d. g. luchinsky , r. mannella , p. v. e. mcclintock , s. m. soskin and n. d. stein , phys . e * 54 * , 2366 ( 1996 ) . g. y. panasyuk , j. c. schotland and v. a. markel , phys . lett . * 100 * , 047402 ( 2008 ) . w. hbner , k. h. bennemann and k. bhmer , phys . b * 50 * , 17597 ( 1994 ) . s. kielich , r. tanas and r. zawodny , j. mod . optic . * 34 * , 979 ( 1987 ) . g. bachelier , j. butet , i. russier - antoine , c. jonin , e. benichou and p. f. brevet , phys . rev . b , * 82 * 235403 ( 2010 ) . a. t. georges , j. opt . b * 28 * , 1603 ( 2011 ) . g. y. panasyuk , j. c. schotland , v. a. markel , phys . b * 84 * , 155460 ( 2011 ) . a. trgler , dissertation , university of graz , 2011 . in practice , arrangement of the level spacing of a quantum oscillator ( @xmath1 ) to high accuracy may not be possible for experimental demonstrations . however one may arrange the incident frequency @xmath6 such that @xmath7 coincides with the desired level spacing .

the powerful emission of active galactic nuclei ( agn ) is thought to originate from gas accretion onto supermassive black holes ( bhs ) . in particular , a large fraction of type - i ( unobscured ) agns are powered by super - eddington accreting massive black holes ( seambhs ) that are characterised by a large eddington ratio , @xmath0 , where @xmath1 is the bolometric luminosity and @xmath2 is the eddington luminosity ( nobuta et al . 2012 ; kelly & shen 2013 ; netzer & trakhtenbrot 2014 ) . in the local universe , most objects of this type are classified as narrow line seyfert 1 galaxies ( nls1s ) . they have prominent features of relatively narrow broad emission lines , strong lines , weak lines , and very strong soft x - ray continuum ( osterbrock & pogge 1985 ; boller 1996 ) . seambhs as central engines of nls1s are not entirely understood . it is thought that they contain slim accretion disks where photon trapping is very important ( e.g. , abramowicz 1988 ; wang & zhou 1999a ; ohsuga 2001 ) . the accelerated expansion of the universe has been extensively studied using type ia supernovae ( sne - ia ) ( riess 1998 ; perlmutter 1999 ) . much attention has been given to improve the measurements of sn - ia in two directions : 1 ) more detections of sn - ia at high redshifts and 2 ) high accuracy of sn - ia calibrated using cepheids ( freedman & madore 2010 ; riess 2011 ; freedman & madore 2013 ) . this method based on sn - ia , however , is limited to redshifts up to about @xmath3 , even for the next generation of extremely large telescopes both on the ground or in space ( hook 2012 ) . thus the exact dynamical state of the universe at higher redshifts is poorly known because of the lack of reliable luminous candles ( weinberg 2013 , kim et al . 2013 ) . the very high luminosity of agns makes them observable to very large distances , up to a redshift of about 7 ( mortlock et al . 2011 ) . in principle , they could be used to follow the expansion of the early universe , beyond the distance accessible by the sne - ia method . however , the diversity in their intrinsic properties ( ho 2008 ; netzer 2013 ) hampers the use of most agns as distance indicators . there were several recent attempts , based on reverberation mapping ( rm ) of dust and gas near the central bh , to test the idea of using agns as standard cosmological candles . one of them makes use of the known correlation between the broad line region size ( @xmath4 ) and the continuum luminosities at 5100 ( @xmath5 ) found in @xmath6 agns ( kaspi 2000 ; bentz 2013 ; peterson 2013 ) . since @xmath4 is empirically determined by measuring the time lag between the h@xmath7 line and the visual continuum , it was suggested to derive this cosmologically independent size by monitoring a large number of agns at various redshifts ( horne 2003 ; teerikorpi 2011 ; watson 2011 ; czerny 2013 ; bentz 2013 ; melia 2014 , elvis & karovska 2002 ) . x - ray variability correlated with bh mass or luminosity has also been suggested to estimate cosmic distances ( la franca et al . 2014 ) . similarly , dust rm ( correlating the variable rest frame @xmath8 and @xmath9 magnitudes ) was suggested to measure the innermost size of the dusty `` torus '' as an alternative size measure ( hoenig 2014 ; yoshii et al . 2014 ) . other suggestions involve radio megamasers ( humphreys et al . 2013 ) , and other agn components ( see marziani & sulentic 2013 for a review of such methods ) . the goal of this paper is to show that seambhs can be used as new probes of cosmological distances provided their mass is directly measured by methods such as rm . these objects show a unique dependence of their bolometric luminosity on the bh mass ( @xmath10 ) which , given a proper calibration , can be used to infer cosmological distances . this idea was first introduced by wang et al . ( 2013 , hereafter w13 ) who showed that selecting seambhs by the slope of their x - ray spectral energy distribution ( sed , wang et al . 2004 ) , and estimating their bh mass using the @xmath11 relation ( e.g. , kaspi et al . 2000 , 2005 , bentz et al . 2013 ) can be used to isolate a sub - group of such sources whose properties are suitable for measuring cosmological distances . however , w13 could not test their idea directly since accurate masses of seambhs were not available at the time . in 2012 we started a large rm campaign to measure bh mass in seambh candidates and to identify a large enough number of such sources that can be used to establish the method and to calibrate them as standard cosmological candles . the first paper in the series ( du et al . 2014 , hereafter paper i ) described our observing project that was carried out in lijiang , china . paper i provides detailed information about the sample selection , telescope and spectrograph , the light curves and cross correlation ( cc ) analysis , and the mass measurements . the observations reported in the present paper were obtained during the same observing season reported in paper - i . their analysis revealed the presence of of three confirmed seambhs and one source which probably belongs to this group but with only an upper limit on the bh mass .. combining the new confirmed seambhs with the three objects reported in paper - i , and with several other seambhs identified in earlier rm experiments , results in a sample which is large enough to test the idea that such objects can be used as standard cosmological candles . the paper is arranged as follows : 2 describes various types of accretion disks in agns and gives the necessary approximations to estimate the accretion rate . 3 presents new rm observations and their analysis . in 4 we list the newly obtained mass and accretion rates , and explain our unique method of identifying seambhs . 5 gives a full description of the new method to measure distances with seambhs by way of a rigorous analysis of the errors associated with the method . in the last section we draw some conclusions regarding the merit of the new method and how can we improve it in the future . we consider three main types of accretion flows with angular momentum onto bhs . the properties of all such flows are determined by the dimensionless accretion rate , @xmath12 , where @xmath13 , @xmath14 is the mass - to - radiation conversion efficiency and @xmath15 the mass accretion rate . here @xmath16 is equivalent to the eddington ratio defined by @xmath17 , where @xmath18 for solar composition gas . for @xmath19 , the accretion flow becomes advection - dominated in the radial direction and radiative cooling is inefficient and dominated by optically thin free - free emission ( narayan & yi 1994 ) . this situation probably applies to liners ( ho 2008 ) . for small to moderate @xmath20 , the flow can be described as optically thick geometrically thin accretion disk , with @xmath21 , where @xmath22 is the height of the disk at a radius @xmath23 ( shakura & sunyaev 1973 ; hereafter ss73 ) . in such cases , the radiative efficiency @xmath14 depends on the radius of the last stable orbit which is determined by the bh spin . standard optically thick geometrically thin accretion disk models assume keplerian rotation in almost perfect circular orbits and very slow inward drift velocity . blackbody emission is very efficient at all radii and the dissipated energy is released locally ( see ss73 ) . while recent studies of thin disks show that the sed can differ , substantially , from the ss73 approximation ( due to e.g. , comptonization , radiative transfer in the atmosphere , etc ; frank 2002 , kato 2003 ) , none of these effects changes significantly the total energy released locally by the disk . the ss73 model has a canonical spectrum whose ( bh mass dependent ) low frequency sed is given approximately by @xmath24 . over this part of the spectrum one can use the standard disk equations to estimate the mass accretion rate , @xmath25 ( see e.g. , frank 2002 ; netzer 2013 ) , where @xmath26 is the inclination angle of the accretion disk to the line of sight , @xmath27 and @xmath28 is the observed flux at @xmath29 . the corresponding @xmath16 by the following expressions ( slightly adopted from earlier works by using our preferred wavelength of 5100 ) , @xmath30 we use @xmath31 to refer to @xmath32 derived in this way ( collin et al . 2002 ) and assume an averaged @xmath33 for type - i agn . obviously @xmath31 and @xmath15 require the knowledge of the distance to the object ( the quantity we want to determine in this paper ) . this however is only required to justify that the objects we are selecting , at small redshift , are indeed seambhs and the associated uncertainties do not affect much the final results since the derived distances are insensitive to the exact values of @xmath16 ( see details below ) . lllllcrrc mrk 335 & 00 06 19.5 & @xmath3420 12 10 & 0.0258 & oct . , 2012 @xmath35 feb . , 2013 & 91 & @xmath36 & @xmath37 & yes + mrk 1044 & 02 30 05.5 & @xmath3508 59 53 & 0.0165 & oct . , 2012 @xmath35 feb . , 2013 & 77 & @xmath38 & @xmath39&yes + iras 04416 + 1215 & 04 44 28.8 & @xmath3412 21 12 & 0.0889 & oct . , 2012 @xmath35 mar . , 2013 & 92 & @xmath40 & @xmath41 & no + mrk 382 & 07 55 25.3 & @xmath3439 11 10 & 0.0337 & oct . , 2012 @xmath35 may . , 2013&123 & @xmath42 & @xmath43 & yes + mrk 142 & 10 25 31.3 & @xmath3451 40 35 & 0.0449 & nov . , 2012 @xmath35 apr . , 2013&119 & @xmath44 & @xmath45 & yes + mcg @xmath46 & 11 39 13.9 & @xmath3433 55 51 & 0.0328 & jan . , 2013 @xmath35 jun . , 2013 & 34 & @xmath47 & @xmath48 & yes + iras f12397 + 3333 & 12 42 10.6 & @xmath3433 17 03 & 0.0435 & jan . , 2013 @xmath35 may . , 2013 & 51 & @xmath49 & @xmath50 & yes + mrk 42 & 11 53 41.8 & @xmath3446 12 43 & 0.0246 & jan . , 2013 @xmath35 apr . , 2013 & 53 & @xmath51 & @xmath52 & no + mrk 486 & 15 36 38.3 & @xmath3454 33 33 & 0.0389 & mar . , 2013 @xmath35 jul . , 2013 & 45 & @xmath53 & @xmath54&yes + mrk 493 & 15 59 09.6 & @xmath3435 01 47 & 0.0313 & apr . , 2013 @xmath35 jun . , 2013 & 27 & @xmath55 & @xmath56 & yes / no the third type of flow is the one with large @xmath20 . in this case , radiation pressure dominates the flow geometry at almost all radii and the disk becomes slim or thick with @xmath57 . the concept of slim disks was originally suggested by paczynsky & bisnovatyi - kogan ( 1981 ) and muchotrzeb & paczynski ( 1982 ) to get rid of the singularity of gas density at the inner edge of the ss73 disks . such systems have been extensively studied using vertically - averaged equations ( matsumoto 1984 ; muchotrzeb - czerny 1986 ) . abramowicz ( 1988 ) used the equations to treat slim disks as transonic flows of super - eddington accretion onto black holes . some of the recent studies include include szuszkiewicz ( 1996 ) , beloborodov ( 1998 ) , wang & zhou ( 1999a , b ) , fukue ( 2000 ) , mineshige ( 2000 ) , watarai & mineshige ( 2001 , 2013 ) , sadowski ( 2011 , 2013 ) and mckenney et al . slim disks are characterized by sub - keplerian rotation and transonic radial motion , which results in a non - localized energy conservation . these three properties are very different from the ss73 disks and there is no simple analytical solution except for the case of extremely high accretion rates , which can be described by a self - similar solution ( wang & zhou 1999a ; wang & netzer 2003 ) . as a result of the fast radial transportation in slim disks , most photons will be trapped and advected into the bh before escaping . this results in inefficient emission of radiation and a big reduction in @xmath14 . some models suggest that in this case , @xmath58 ( wang & zhou 1999a ; mineshige 2000 ; sadowski et al . this leads to the so called `` saturated luminosity '' given by : @xmath59 where @xmath60~{\ergs}m_{\odot}^{-1}$ ] ( wang & zhou 1999a ; mineshige et al . thus @xmath61 over a large range of accretion rates around @xmath62 . given an accurately measured bh mass , one can deduce @xmath63 and hence the bh distance provided there is a way ( i.e. , a reliable bolometric correction factor ) to convert @xmath63 to the monochromatic luminosity of the disk at an accessible wavelength ( see 5 ) . detailed information of our rm campaign are given in paper i where we describe the observatory , the telescope , the spectrograph and the observing procedure in great detail . in short , targets were selected from the list of seambh candidates in w13 based on their coordinates and the slope of their 210 kev continuum requiring @xmath64 . all targets are classified , spectroscopically , as nls1s i.e. , 1 ) fwhm(h@xmath65 ; 2 ) /h@xmath66 ; 3 ) strong emission lines . only radio - quiet sources are selected to avoid contamination by relativistic jet emission to the optical continuum and , perhaps , emission lines line relative to the 5100 continuum and a couple of simultaneous @xmath67-ray flares . these results will be reported separately ( f. wang , 2014 in preparation ) . ] . the campaign started in october 2012 and lasted until june 2013 . we used the lijiang 2.4-m telescope of yunnan observatory , in china . we obtained optical spectra of all 10 radio - quiet selected sources almost every night with suitable weather conditions . flux calibration is obtained through the use of a nearby comparison star that was observed in the same slit with the agn . the variability of three of the sources ( mrk 335 , mrk 142 and iras f12397 + 3333 ) , and their bh mass and accretion rate , was reported in paper i. details of five additional sources , four of which were found to be seambhs , are given in table 1 . two additional radio quiet sources , that do not show significant time lags , will not be further discussed . rrrccccccc mrk 486 & w0mt010 & wfpc , p6 , f785lp & @xmath68 & & & & & psf & 1.278 + & & & @xmath69 & 0.11 & [ 2.0 ] & @xmath70 & @xmath71 & bulge & + & & & 16.75 & 2.41 & [ 1.0 ] & 0.41 & @xmath72 & disk & + & & & 0.032 & & & & & sky & + mrk 382 & u2e62i01 t & wfpc2 , pc1 , f606w & 17.16 & & & & & psf & 1.068 + & & & 16.77 & 0.51 & @xmath73 & 0.83 & @xmath74 & bulge & + & & & @xmath75 & @xmath76 & @xmath77 & 0.41 & @xmath78 & bar & + & & & @xmath79 & @xmath80 & [ 1.0 ] & 0.96 & @xmath81 & disk & + & & & 0.023 & & & & & sky & + mcg @xmath82 & u2e61l01 t & wfpc2 , pc1 , f606w & 18.22 & & & & & psf & 0.761 + & & & 18.43 & 0.07 & @xmath83 & 0.41 & @xmath84 & addl psf & + & & & 18.09 & 1.02 & 0.81 & 0.82 & @xmath85 & bulge & + & & & 17.05 & @xmath86 & 0.41 & 0.27 & @xmath87 & bar & + & & & 14.69 & @xmath88 & [ 1.0 ] & 0.85 & @xmath89 & disk & + & & & 0.019 & & & & & sky & + mrk 493 & u2e62o01 t & wfpc2 , pc1 , f606w & 17.34 & & & & & psf & 1.191 + & & & 17.26 & 0.59 & 0.11 & 0.83 & @xmath90 & n. spiral / ring & + & & & 17.42 & 1.52 & @xmath91 & 0.85 & @xmath92 & bulge & + & & & 16.23 & @xmath93 & 0.57 & 0.21 & @xmath94 & bar & + & & & 14.72 & @xmath95 & [ 1.0 ] & 0.57 & @xmath96 & disk & + & & & 0.015 & & & & & sky & + mrk 1044 & ibgu10 & wfc3 , uvis1 , f547 m & 15.35 & & & & & psf & 1.160 + & & & 15.71 & 0.85 & 1.86 & 0.98 & @xmath97 & bulge & + & & & 15.38 & 6.33 & 0.41 & 0.65 & @xmath98 & bar & + & & & 14.70 & @xmath99 & [ 1.0 ] & 0.87 & @xmath100 & disk & + & & & 0.011 & & & & & sky & all five targets were observed by the _ hubble space telescope _ ( hst ) prior to our campaign . we use the hst images to remove the host galaxy contaminations as described in paper i. for objects with only one exposure , we use l. a. cosmic ( van dokkum 2001 ) to remove cosmic rays in their images . unlike the three other sources , mrk 486 has only wf / pc-1 exposure with a resolution which is too low to constrain the srsic index of its bulge . since a large fraction of nls1s contain pseudo - bulges ( ryan 2007 ; orban de xivry 2011 ; mathur 2012 ) , we fix the srsic index of the bulge component to 2.0 . we also experimented with srsic indexes of 1.0 and 4.0 which changed the host flux by @xmath101 , well within the uncertainty of this measurement . the results of the fitting process are listed in table 2 and the images are shown in figure 1 . in most cases , the host contamination , given the slit size used in the observations ( 2.5@xmath102 ) , does not amount to more than @xmath103 at @xmath104 . the uncertainty on the bh distance , @xmath105 ( see below ) , due to this component is @xmath106 , where @xmath107 is the fractional contribution of the host at @xmath104 . the various panels of figure 2 show the light curves of the five new targets and table 4 presents the relevant measurements . we make use of the cross - correlation function ( ccf ) to measure the time lag of the h@xmath7 line relative to the observed continuum . the interpolated cross - correlation function ( iccf ; gaskell & sparke 1986 ; gaskell & peterson 1987 ) and the z - transformed discrete correlation function ( zdcf ; alexander 1997 ) methods were employed . the uncertainties of the time lags are determined using the flux randomization ( fr)/random subset sampling ( rss ) " method ( peterson 1998a , 2004 ) . more details are provided in paper i. in general , we prefer the use of the 5100 continuum for calculating the ccf . however , in one source , mrk 382 , we choose to use the better sampled @xmath8-band light curve despite the additional uncertainty due to the inclusion of several emission lines in this band . in general , the variability we find in our hard x - ray selected nls1 sample is different from what was found in previous monitoring of such objects where small amplitude variations in the optical bands were reported ( klimek 2004 ) . the variations are consistent with several earlier suggestions that show the variability amplitude decreases with the eddington ratio ( e.g. , zuo 2012 ; ai 2013 ) . these properties are important for the understanding of slim accretion disks and will be discussed in detail in future publications . cccccccccccccccccccc 29.3 & @xmath108 & @xmath109 & & 24.4 & @xmath110 & 24.4 & @xmath111 & & 115.3 & @xmath112 & @xmath113 & & 179.4 & @xmath114 & @xmath115 & & 201.4 & @xmath116 & @xmath117 + 31.3 & @xmath118 & @xmath119 & & 25.4 & @xmath120 & 25.4 & @xmath121 & & 117.3 & @xmath112 & @xmath122 & & 181.4 & @xmath123 & @xmath124 & & 202.4 & @xmath125 & @xmath126 + 36.2 & @xmath127 & @xmath128 & & 26.4 & @xmath129 & 26.4 & @xmath121 & & 152.4 & @xmath130 & @xmath131 & & 185.4 & @xmath132 & @xmath133 & & 203.4 & @xmath134 & @xmath126 + 39.2 & @xmath135 & @xmath136 & & 27.4 & @xmath137 & 27.4 & @xmath113 & & 153.4 & @xmath130 & @xmath138 & & 186.4 & @xmath139 & @xmath140 & & 204.4 & @xmath141 & @xmath142 + 41.2 & @xmath143 & @xmath144 & & 28.3 & @xmath145 & 28.3 & @xmath122 & & 172.3 & @xmath146 & @xmath113 & & 190.4 & @xmath147 & @xmath148 & & 206.4 & @xmath149 & @xmath126 + _ mrk 1044 _ : _ rosat _ observations show evidence for soft x - ray variability of this source ( boller 1996 ) . our ccf analysis shows a peak at @xmath150 days with a maximum cross - correlation coefficient of @xmath151 . the measured @xmath10 combined with the conservative assumption on the spin give @xmath152 , the largest in our sample . _ mrk 382 _ : this object shows the strongest host contamination in our sample . because of this , the 5100 flux calibration is uncertain and we chose to use the @xmath8-band light curve , with their improved precision , in the ccf analysis ( the only source in our sample ) . using the entire light curve we obtain a peak correlation of @xmath153 and a time lag of @xmath154 days . since the first part of the observations of this object suffers from bad - weather and problems with continuum subtraction from the h@xmath7 line , this resulted in a noisier light curve in the first part of the campaign . thus , we also analyzed only the second half of the light curve , with its improved sampling ( blue points in figure 2 ) . this gives a stronger correlation with @xmath155 corresponding to a time lag of @xmath156 days in the rest frame of the source . in the following analysis we use this time lag . _ mrk 486 _ : the source shows significant flux and sed variations in the x - ray band ( ballo 2008 ) . the gap in the data between jd260 - 290 ( see figure 2 ) is due to bad weather . the h@xmath7 light curve shows a monotonic decreases with time and much of the signal in the ccf is due to the flux increase at the end of the campaign , on july 14 and 15 , 2013 . the observations on those night were taken under very good conditions and we have no reason to suspect these measurements . _ mrk 493 _ : we obtained only 27 observations from april to june 2013 . this makes the quality of the ccf poorer than in other objects and results in a larger uncertainty on the time lag which is consistent with zero ( although the peak of the ccf is very clear with @xmath157 ) . we report on the various results obtained for this source but do not include it in the distance analysis . _ mcg + 06 - 26 - 012 : _ this nls1 is situated in a sb galaxy ( see _ hst _ image in figure 1 ) . we find fwhm(h@xmath7)=1685 @xmath158 ( see also grupe 1999 ; veron - cetty 2001 ) . the ccf shows a statistically significant time lag of @xmath159 days with @xmath160 . from eqn . ( [ eq : virial ] ) , we get @xmath161 . the host - galaxy subtracted luminosity is @xmath162 . using our method to determine accretion rate [ eqn . ( [ eq : dotm_min ] ) below ] we get @xmath163 , indicating that this object is not a seambh . finally we comment on the general method of measuring line intensities in this and earlier rm experiments . the h@xmath7 light curves reported here and in paper i were obtained by simple integration over the line profile using a locally determined continuum . these are not necessarily the most accurate light curves and spectral fitting methods that include other components , such as the host galaxy sed and the lines ( e.g. , barth et al . 2013 ) can lead to significant improvements and reduced uncertainties . in some rare cases , this can also be used to recover points in the light curve that were discarded due to poor weather conditions that prevented us from using the local comparison star as our flux calibrator . we are working on the improvement of such methods and will report the results in a future publication ( hu et al . in preparation ) . measuring bh mass through rm is a well established method . it is based on the idea that the velocities of the clouds emitting the broad emission lines are virialized in the bh gravitational potential and the emitted line intensities echo the variable ionizing continuum . the time - lag , @xmath164 , reflects the size and geometry of the variable part of the blr and the emissivity weighted radius is given by @xmath165 , where @xmath166 is the speed of light . rm experiments provide the necessary measurements of @xmath4 which , combined with the virial assumption , result in a simple expression for the bh mass , @xmath167 where @xmath168 is the gravitational constant , @xmath169 is a measure of the gas velocity , and @xmath170 is a constant which combines all the unknowns about the geometry and kinematics of the gas in the blr . the best value of @xmath170 is obtained by using bh mass estimated determined by the @xmath171 relationship in agn hosts where stellar absorption features can be observed ( woo et al . 2013 ) . we have used our observations to measure the mass of the bhs in our sample through eqn . ( [ eq : virial ] ) . in principal there are four possible choices to define @xmath169 ( e.g. , collin 2006 ) : 1 ) the full - width - at - half - maximum ( fwhm ) of the variable component of the line obtained from the rms spectrum ( peterson 1998b ; wandel 1999 ) , 2 ) the fwhm of the line obtained from the mean spectrum ( e.g. , kaspi 2000 ) , 3 ) the line dispersion ( so called @xmath172 ) obtained from the rms spectrum ( e.g. , fromerth & melia 2000 ; peterson 2004 ; collin 2006 ; denney 2013 ) and 4 ) the line dispersion obtained from the mean spectrum ( e.g. , collin 2006 ; bian 2008 ) . for example , the recent woo et al . ( 2013 ) work suggests that for @xmath172 measured from the rms spectrum , @xmath173 . despite many years of study , there is no empirical evidence , or a realistic model , to show that any of the methods is preferred over the others . following paper i , we have chosen to use the fwhm(h@xmath7 ) from the mean spectrum as our choice of @xmath169 . applying this to the woo et al . ( 2013 ) sample of 25 agns with measured @xmath174 gives @xmath175 ( woo , private communication ) with no dependence on bh mass or accretion rate . the scatter in the derived @xmath10 obtained in this way is very similar to the scatter observed by assuming @xmath176 with its corresponding @xmath170 . our measurements of @xmath10 based on this method are listed in table 4 and more references regarding this choice are provided in paper i. the seambhs discussed in this work belong to the category of slim accretion disks . they can be recognized by their @xmath177 that approaches and even exceeds unity . however , the uncertainties on present slim disk models , and the uncertainties on the observations , are too large to identify a specific value of @xmath177 beyond which photon trapping is significant . various models ( laor & netzer 1989 ; beloborodov 1998 ; sadowski 2011 ) suggests that this occurs at @xmath178 but even this range is uncertain . returning to seambhs with higher accretion rate , we note that the value of @xmath177 which is required to identify such objects can be determined from the measured @xmath10 and the global sed of the disk . however , we do not have access to the entire sed because of the galactic and intergalactic absorption . we can only estimate the mass accretion rates , @xmath15 , from the observed optical continuum and the results of the existing slim disk models that suggest this to be a good approximation since the wavelengths used correspond to the parts of the disk that are not affected by photon trapping . the uncertainty on bh spin , and @xmath20 and @xmath14 still remains . to overcome the uncertainty related to the bh spin , we adopted a conservative approach that uses the smallest possible @xmath177 based on the smallest possible @xmath14 . for a maximally rotating pro - grade disk ( @xmath179 ) , the radiative efficiency reaches its maximum of @xmath180 . from eqn . ( [ eq : ss_2 ] ) , the maximum accretion rate assuming @xmath181 is @xmath182 . if @xmath183 , this object has too low accretion rate to maintain a slim disk . in the same way we define a minimum accretion rate , @xmath184 by using the lowest possible @xmath185 corresponding to a retrograde spinning bh with @xmath186 , @xmath187 if @xmath188 , the disk must be slim , and the object is considered to be a seambh . obviously there are objects with @xmath189 which we do not consider seambhs by our very conservative criterion , that may be powered by slim accretion disks if @xmath14 is large enough . we can introduce an even more conservative criterion by combining @xmath190 and @xmath191 . since @xmath192 , we get @xmath193 for seambhs with @xmath194 . in the present paper we use @xmath195 as the only criterion . to decide whether they are powered buy thin or slim disks.,scaledwidth=40.0% ] 0.5 cm lccccccrc + mrk 335 & @xmath196 & @xmath197 & @xmath198 & @xmath199 & @xmath200 & 0.030 & 117.6 & @xmath201 + mrk 1044 & @xmath202 & @xmath203 & @xmath204 & @xmath205 & @xmath206 & 0.031 & 74.5 & @xmath207 + mrk 382 & @xmath208 & @xmath209 & @xmath210 & @xmath211 & @xmath212 & 0.043 & 154.5 & @xmath213 + mrk 142 & @xmath214 & @xmath215 & @xmath216 & @xmath217 & @xmath218 & 0.015 & 207.9 & @xmath219 + iras f12397 & @xmath220 & @xmath221 & @xmath222 & @xmath223 & @xmath224 & 0.017 & 201.0 & @xmath225 + mrk 486 & @xmath226 & @xmath227 & @xmath228 & @xmath229 & @xmath230 & 0.012 & 179.3 & @xmath231 + mrk 493 & @xmath232 & @xmath233 & @xmath234 & @xmath235 & @xmath236 & 0.022 & 143.5 & @xmath237 + + mrk 110 & @xmath238 & @xmath239 & @xmath240 & @xmath241 & @xmath242 & 0.011 & 162.1 & @xmath243 + mrk 202 & @xmath244 & @xmath245 & @xmath246 & @xmath247 & @xmath248 & 0.018 & 95.5 & @xmath249 + ngc 4051 & @xmath250 & @xmath251 & @xmath252 & @xmath253 & @xmath254 & 0.011 & 17.1 & @xmath255 + ngc 7469 & @xmath256 & @xmath257 & @xmath258 & @xmath259 & @xmath260 & 0.061 & 73.9 & @xmath261 + pg 0026 + 129 & @xmath262 & @xmath263 & @xmath264 & @xmath265 & @xmath266 & 0.063 & 700.4 & @xmath267 + pg 0844 + 349 & @xmath268 & @xmath269 & @xmath270 & @xmath271 & @xmath272 & 0.033 & 300.0 & @xmath273 + pg 1211 + 143 & @xmath274 & @xmath275 & @xmath276 & @xmath277 & @xmath278 & 0.030 & 383.6 & @xmath279 + pg 1700 + 518 & @xmath280 & @xmath281 & @xmath282 & @xmath283 & @xmath284 & 0.030&1565.9 & @xmath285 + pg 2130 + 099 & @xmath286 & @xmath287 & @xmath288 & @xmath289 & @xmath290 & 0.039 & 295.0 & @xmath291 a flow chart diagram for identifying seambhs is given in figure 3 . applying eqn . ( [ eq : dotm_min ] ) to our sample we find 7 targets ( the three from paper i and four from this work ) to have slim accretion disks . as explained earlier , we decided not to use mrk493 because of the large uncertainty on the time lag and hence the bh mass . all together , six of the new objects are suitable for the distance analysis . it is interesting to note that the fraction of seambhs from our targets reaches @xmath292 suggesting that the @xmath293-based selection is very efficient for identifying seambh candidates . we used the same method to search for seambhs among the entire sample ( @xmath294 ) of radio - quiet agns with previous rm - based bh mass measurements ( peterson 1998a ; kaspi 2000 , 2005 ; bentz 2013 ) . we found 9 objects with @xmath295 . thus the total number of newly identified seambhs with directly measured bh mass and @xmath296 , is 15 under the most conservative assumption about the bh spin . we also found 21 objects with @xmath297 , which must be powered by thin disks . the rest of the sources have @xmath298 . obviously the real number of objects containing slim disks could be larger , but because of our conservative estimate of @xmath32 , we have no way to prove it . the uncertainty on @xmath299 is obtained from the uncertainty on @xmath300 that includes the measured flux , the bh mass and the inclination to the line - of - sight . the uncertainty due to inclination is discussed below . to measure the distance from @xmath63 using the new method we have to know the bolometric correction factor defined by @xmath301 , where @xmath5 is the luminosity at 5100 . ideally , this could be determined by direct observations and integration over the entire seds . as explained , this is not possible in the uv and euv parts of the spectrum and hence this is normally replaced by an empirical estimate of @xmath302 such as the one calculated by marconi et al . ( 2004 ) . for geometrically thin disks , @xmath303 ( frank et al . as demonstrated by netzer & trakhtenbrot ( 2014 ) , the difference between these two approximations can be very large . the situation regarding slim disks is even more problematic since models of such objects are rather uncertain which reflects on the uncertainty in @xmath302 . the approach adopted here is to use the simple slim disk calculation ( abramowicz et al . 1988 ) to write a generic approximation for @xmath302 and use the observations to find the parameters in this equation . the expression is motivated by thin disk models and is given by @xmath304 where @xmath305 . for the simplest slim disks @xmath306 and @xmath307 ( mineshige et al 2000 ; see also wang et al . 1999b , shimura & manmoto 2003 ) . having defined the bolometric correction term , we define @xmath308 as the measured @xmath309 at @xmath29 , after correcting for foreground extinction . thus , @xmath310 . where @xmath311 is the radiation anisotropy factor which is a complex function of the disk geometry and inclination angle @xmath26 ( e.g madau 1988 ) . we approximate this by @xmath312 for a thin disk and by @xmath313 for a thick disk . for slim disks in type - i agns , @xmath314 and @xmath315 so @xmath316 is a good approximation . we can now combine @xmath317 , @xmath311 and @xmath318 into one parameter , @xmath319 where @xmath320 . this result in the final expression for the distance : @xmath321 where @xmath322 . 0.5 cm to understand the various uncertainties associated eqn . ( [ eq : dbh ] ) , we divide all the variables contributing to the error on the distance into three different groups . the first group contains three measurable quantities , the continuum flux @xmath308 , the velocity @xmath169 and time lag @xmath164 that are used to calculate the mass ( the uncertainty on the host galaxy flux is included in the uncertainty on @xmath308 ) . these uncertainties are listed in table 4 . the second group contains @xmath170 which is not measured in individual sources but derived from the known properties of rm agns with host that show absorption lines . as explained , we use the woo ( 2013 and private communication ) result of @xmath323 . in particular we use @xmath324 . the third group includes parameters that can in principle be obtained from the slim disk theory . for the simplified slim disk model used here these are @xmath317 , @xmath311 and @xmath302 that are all absorbed into @xmath325 ( eqn . [ eq : ell_kappa ] ) . the anisotropy factor , @xmath311 , is a complicated function of the disk geometry and include the effects of radiative transfer , inclination , reflection by the funnel walls , etc . as explained , the approximation adopted here is @xmath313 with an uncertainty similar to the uncertainty on @xmath326 . for type - i agns , the inclination angle @xmath26 can change over a small range of @xmath327 . with our chosen value of @xmath181 we get @xmath328 and @xmath329 . the total uncertainty on a given @xmath105 is obtained by combining the uncertainties associated with the bh mass measurements and the flux , with those estimated for @xmath311 and @xmath170 and are assumed to be the same for all objects . this should also be combined with the uncertainty on @xmath330 which can not be obtained directly from the simple disk model and hence required different considerations . as we show below , the remaining uncertainties and the missing constant of calibration , can be obtained by a comparison with the observations . we define @xmath331 and write eqn . [ eq : dbh ] for source @xmath26 as @xmath332 where @xmath333 , @xmath334 , @xmath335 and @xmath336 ( same for all sources ) . the new term , @xmath337 , is the intrinsic scatter associated with the new method . it represents the uncertainty related to additional physical parameters that were not included in our simple slim disk model ( e.g. , bh spin ) . if such terms are important , they will introduce a large scatter which will dominate the uncertainty in @xmath338 . we assume that @xmath337 has a gaussian distribution with a 1-@xmath339 width of @xmath340 . the calibration of @xmath341 and @xmath7 is achieved by requiring that @xmath342 , where @xmath343 is the luminosity distance derived from the standard cosmological model . for low redshift sources , the uncertainty on @xmath343 is equivalent to the uncertainty on @xmath344 which is less than 2% ( freedman & madore 2013 ) . this is significantly smaller than the uncertainties on @xmath105 ( see table 4 ) and we do not include this in our bayesian analysis . the values of cosmological parameters chosen here are based on the recent planck measurements ( ade et al . 2013 ) : @xmath345 , @xmath346 and @xmath347 ( note that the use of @xmath344 only is not enough since the redshift range is @xmath348 ) . using bayes theorem , the posterior probability is @xmath349 where @xmath350 is the likelihood function and @xmath351 is the prior probability . since the error distributions of @xmath352 are not gaussian , we employ asymmetric gaussians to approximate their distributions . the function @xmath353 is the prior probability , which is assumed to be uniformly distributed since we have no prior information about the three parameters . in appendix a , we derive the likelihood function . the normalised posterior function is given by @xmath354 the most likely values of @xmath355 and @xmath7 with @xmath340 are located where the posterior function is maximized . the joint confidence region for @xmath355 and @xmath7 is derived from the three - dimensional likelihood space via reducing the three - dimensional probability distribution function to two dimensions @xmath356 this is plotted in figure 4 left panel . the other two joint functions of confidence regions are given by @xmath357 and @xmath358 , which can be obtained from eq . ( [ eqn : likelihood ] ) by integrating @xmath341 and @xmath7 , respectively . employing the observational data given in table 4 , the joint functions of confidence regions are plotted as contours in figure 4 . integrating two of the three parameters in the joint confidence functions , we have the probability distributions of the three parameters in figure 5 . the most likelihood values and their uncertainties are given by @xmath359 the results obtained here show that the calibrated value of @xmath7 is consistent with @xmath360 , which is the value derived for ss73-disks . this is related to the fact that in slim disks , most of the 5100 flux originates at a few @xmath361 , where photon trapping is not important . therefore the bolometric correction factor follows a relation that is similar to the one obtained for thin accretion disks . the normalization factor @xmath330 is also consistent with the simple model of slim disks . most importantly , the value of @xmath340 which reflects several physical unknowns in the slim disk theory ( spin , exact geometry , etc . ) is very small ( 0.04 ) , well below the combined uncertainties of @xmath311 and @xmath170 . these physical unknowns can not influence much the accuracy of the new method confirming the suggestions that seambhs can be used as cosmological distances indicators provided the measurement errors can be substantially reduced . using the values of @xmath362 and @xmath7 found here , we can calculate the distances to all seambhs listed in table 4 and obtain the averaged residual distance ( @xmath363 ) and its scatter in our sample . this number is @xmath364 with a standard deviation of @xmath365 . figure 6 shows the comparison of the new distance with @xmath343 for all the sources in our sample . the correlation in the upper panel is very close to a line of 1:1 and the plotted @xmath363 , in the ( bottom panel , shows no dependence on the on the standard luminosity distance . in figure 7 we examine the dependence of @xmath363 on bh mass . here , again , there is no systematic deviation confirming the usefulness of this technique over a large mass range . the error bars on the points in both diagrams are the combination of all the uncertainties and are listed in the right column of table 4 . vs. bh mass . the dashed line is the mean value in our sample ( @xmath364).,scaledwidth=47.0% ] at @xmath366 , the distance modular difference between an accelerating universe and a constant expanding universe is about 0.3mag ( e.g. , riess et al . 1998 ) . using the scatter found here ( @xmath367 ) , we estimate that some 60 - 100 seambhs will be required , for the same redshift interval , to achieve this accuracy . obviously , a much larger number of sources will be required to constrain @xmath368 or its cosmological evolution . it is too early to estimate this number given the limited number of sources ( 15 ) used here and the potential improvements of the method once better slim disk models are available . how likely is it to find a large number of seambhs at high redshift and measure their bh mass to the same level of accuracy achieved here ? several recent papers describe searches for super - eddington accreting massive black holes in large agn samples like the sdss ( nobuta 2011 ; trakhtenbrot & netzer 2012 ; kelly & shen 2012 ; netzer & trakhtenbrot 2014 ) . such studies use bh mass estimates based on the known @xmath369 relation in combination with an empirical bolometric correction factor like the one derived by marconi ( 2004 ) . all these studies show a large fraction of sources with @xmath370 , i.e. , likely to be powered by slim accretion disks . the recent work by netzer & trakhtenbrot ( 2014 ) adopt an approach similar to the one used here . the paper assumes that all agns are powered by accretion disks and used the estimated @xmath10 and @xmath15 to derive the value of @xmath190 . this is then used to investigate the fraction of slim accretion disks in the agn population . they found that at redshifts of about 0.7 and beyond , most of the sdss detected agns contain slim accretion disks . thus , there is no lack of suitable sources and the main obstacle is observational , to measure the mass of so many bhs by rm . several major improvements of the methodology used here are around the corner promising to achieve higher accuracy in estimating @xmath170 , @xmath311 and @xmath362 ; the main contributers to the uncertainty . estimates of @xmath170 and @xmath311 can be improved by better modelling of the blr using methods like maximum entropy reconstruction space ( horne 2004 ) and/or the markov chain monte carlo simulations ( mcmc ) ( pancoast 2011 , 2013 ; li 2013 ) . the uncertainties on @xmath362 are due to theoretical limitations and incomplete understanding of slim accretion disks . the main ingredient that requires improvement is the treatment of the two - dimensional radiative transfer in the disk ( ohsuga 2002 , 2005 ; pountanen 2007 ; dotan & shaviv 2012 ; yang 2014 ) . better calculations will improve the estimates of @xmath14 in slim accretion disks and through it , the estimates of @xmath16 . as explained , the method does not depend much of such theoretical developments since @xmath362 can be calibrated experimentally . the first year of observations with the lijiang 2.4-m telescope resulted in successful time lag and bh mass measurements of 8 radio - quiet narrow line seyfert 1 galaxies . a conservative method to estimate the normalized accretion rate @xmath177 show that 7 of those are powered by slim accretion disks and hence are seambhs . the time lag in one of the 7 sources , mrk493 , is consistent with zero and hence removed from the rest of the analysis . literature search for sources with rm - based bh mass measurements revealed 9 more sources that are seambhs by our definition . we presented a novel method that can be used to infer cosmological distances using seambhs . we tested the method on our sample of 15 seambhs with @xmath296 and showed that it can be used to obtain cosmological distances . the key to the new method is the empirical calibration of the mean bolometric correction factor , and through it the saturated luminosity , by means of a baysian analysis . the resulting bolometric correction is in agreement with the simple version of slim disk models and the intrinsic scatter of the distance estimate is only 0.04 dex indicating that the unknown seambh physics is small and does not prevent us from using the method . for seambhs of @xmath371 , the intrinsic bolometric luminosities are few @xmath372 , which is much higher than sne - ia ( leibundgut 2001 ) . furthermore , sn - based cosmological distances at @xmath373 are rather uncertain due to small number statistics and such objects are hard to find at high-@xmath374 ( hook 2012 ) because of their faintness and/or the slow evolution of their progenitors ( kobayashi 2009 ) . in contrast , the number density of massive , very luminous fast accreting bhs is increasing with redshift ( trakhtenbrot & netzer 2012 ; nobuta 2012 ; kelly & shen 2013 ; netzer & trakhtenbrot 2014 ) and the saturated luminosity of slim disks is nearly independent of redshift - related factors such as the chemical composition of the accreted gas . future rm experiments based on the @xmath375 line , and perhaps other lines that will be proven to be good indicators for the motion of the blr gas , can be used to measure bh mass in a large number of high-@xmath374 seambhs . such a campaign will require large ground - based telescopes over a period of several years . the results can be used to measure bh mass and new cosmological distances and to explore the dynamics of the universe in the era when gravity was the dominant , but not the only factor affecting its expansion . the authors are very grateful to an anonymous referee for useful report that helped to improve the paper . we thank the staff of the lijiang station of the yunnan observatory for their great help in making this project successful in the first year of observations . we thank l. c. ho and y .- y . zhou for useful discussions . the research is supported by the strategic priority research program @xmath35 the emergence of cosmological structures of the chinese academy of sciences , grant no . this research is supported by the nsfc through nsfc-11173023 , -11133006 , -11233003 , and by the israel - china isf - nsfc grant 83/13 . 99 abramowicz , m. a. , 1988 , , 332 , 646 ade , p. a. r. 2013 , , arxiv:1303.5076 ai , y. et al . 2013 , aj , 145 , 90 ballo , l. 2008 , , 483 , 137 barth , a. et al . 2013 , , 769 , 128 beloborodov , a. m. 1998 , , 297 , 739 bentz , m. 2013 , , 767 , 149 boller , t. , brandt , w. n. & fink , h. 1996 , , 305 , 53 collin , s. 2002 , 388 , 771 collin , s. , 2006 , , 456 , 75 czerny , b. , hryniewicz , k. , maity , i. , schwarzenberg - czerny , a. , ycki , p. t. & bilicki , m. 2013 , , 556 , 97 denney , k. d. 2013 , apj , 775 , 60 dotan , c. & shaviv , n. j. 2012 , , 427 , 3071 du , p. ( seambh collaboration ) 2014 , , 782 , 45 ( paper i ) elvis , m. & karovska , m. 2002 , , 581 , l67 frank , j. , king , a. & raine , d. j. accretion power in astrophysics , cambridge univ . press ( 2002 ) freedman , w. l. & madore , b. f. 2010 , , 48 , 673 freedman , w. l .. & madore , b. f. 2013 , iaus , 289 , 3 fromerth , r. & melia , f. 2000 , apj , 533 , 172 fukue , j. 2000 , , 52 , 829 gaskell , c. m. & peterson , b. m. 1987 , , 65 , 1 gaskell , c. m. & sparke , l. s. 1986 , , 305 , 175 grupe , d. , beuermann , k. mannheim , k. & thomas , h .- c . 1999 , , 350 , 805 ho , l. c. 2008 , , 47 , 475 hook , i. m. 2012 , iau sym . , 285 , 63 hnig , s. f. 2014 , , 784 , l4 horne , k. , korista , k . t. & goad , m. r. 2003 , , 339 , 367 horne , k. , peterson , b. m. , collier , s. j. & netzer , h. 2004 , , 116 , 465 humphreys , e. m. l. et al . 2013 , , 775 , 13 kaspi , s. 2000 , , 533 , 631 kaspi , s. 2005 , , 629 , 61 kato , s. , mineshige , s. & fukue , j. , accretion disks kyoto university press ( 2007 ) kelly , b. c. & shen , y. 2013 , , 764 , 45 kim , a. 2013 , arxiv:1309.5382 klimek , e. s. , gaskell , c. m. , hedrick , c. h. 2004 , , 627 , 1072 kobayashi , c. & nomoto , k. 2009 , , 707 , 1466 la franca , f. et al . 2014 , , 787 , l12 laor , a. & netzer , h. 1989 , , 238 , 897 leibundgut , b. 2001 , , 39 , 67 li , y .- r . , wang , j .- m . , ho , l. c. , du , p. & bai , j .- m . 2013 , , 779 , 110 madau , p. 1988 , 327 , 116 marconi , a. 2004 , , 351 , 169 marziani , p. & sulentic , j. w. 2013 , arxiv:1310.3143 mathur , s. , fields , d. , peterson , b. & grupe , d. 2012 , apj , 754 , 146 matsumoto , r. , kato , s. , fukue , j. , okazaki , a. t. 1984 , , 36 , 71 mckinney , j. c. , tchekhovskoy , a. , sadowski , a. & narayan , r. 2013 , arxiv:1312.6127 melia , f. 2014 , jcap , 01 , 027 mineshige , s. , 2000 , , 52 , 499 mortlock , d. j. 2011 , , 474 , 616 muchotrzeb , b. & paczynski , b. 1982 , aca , 32 , 1 muchotrzeb - czerny , b. 1986 , aca , 36 , 1 narayan , r. & yi , i. 1994 , , 428 , l13 netzer , h. the physics and evolution of active galactic nuclei , cambridge univ . press ( 2013 ) netzer , h. & trakhtenbrot , b. 2014 , , 438 , 672 nobuta , k. , 2012 , , 761 , 143 ohsuga , k. , mineshige , s. , mori , m. & umemura , m. 2002 , , 574 , 315 ohsuga , k. , mori , m. , nakamoto , t. & mineshige , s. 2005 , , 628 , 368 osterbrock , d. e. & pogge , r. w. 1985 , , 297 , 166 orban de xivry , g. et al . 2011 , mnras , 417 , 2721 pancoast , a. , brewer , b. j. & treu , t. 2011 , , 730 , 139 pancoast , a. 2013 , , arxiv:1311.6475 paczynski , b. & bisnovatyi - kogan , g. 1981 , aca , 31 , 283 perlmutter , s. , 1999 , , 517 , 565 peterson , b. m. 2013 , sp . tmp 60.p peterson , b. m. , 1998a , , 501 , 82 peterson , b. m. , 1998b , , 110 , 660 peterson , b. m. , 2004 , , 613 , 682 poutanen , j. , lipunova , g. , fabrika , s. , 2007 , , 377 , 1187 ryan , c. j. 2007 , apj , 654 , 799 richards , g. 2006 , , 166 , 470 riess , a. g. 1998 , , 116 , 1009 riess , a. g. 2011 , , 730 , 119 sadowski , a. , abramowicz , a. 2011 , , 52 , a17 sadowski , a. , narayan , r. 2013 , , 436 , 3856 sirianni , m. , jee , m. j. , benitez , n. et al . 2005 , , 117 , 1049 shakura , n. i. & sunyaev , r. 1973 , , 24 , 337 shimura , t. & manmoto , t. 2003 , , 338 , 1013 szuszkiewicz , e. , malkan , m. a. & abramowicz , m. a. 1996 , , 458 , 474 teerikorpi , p. 2011 , , 531 , a10 trakhtenbrot , b. & netzer , h. 2012 , , 427 , 3081 van dokkum , p. g. 2001 , , 113 , 1420 veron - cetty , m .- p . , veron , p. & goncalves , a. c. 2001 , , 372 , 730 wandel , a. et al . 1999 , apj , 526 , 579 wang , j .- m . & netzer , h. 2003 , , 398 , 927 wang , j .- m . , du , p. , valls - gabaud , d. , hu , c. & netzer , h. 2013 , , 110 , 081301 wang , j .- m . & zhou , y .- y . 1999a , , 516 , 420 wang , j .- m . , szuszkiewicz , e. , lu , f .- j . & zhou , y .- y . 1999b , , 522 , 839 wang , j .- m . , watarai , k .- y . & mineshige , s. 2004 , , 607 , l107 watarai , k. & mineshige , s. 2001 , , 53 , 915 watson , d. , denney , k. d. , vestergaard , m. & davis , t. m. 2011 , , 740 , l49 weinberg , d. 2013 , phys . , 530 , 87 woo , j .- h . 2013 , , 772 , 49 yang , x .- h . , yuan , f. , ohsuga , k. & bu , d .- f . 2014 , , 780 , 79 yoshii , t. et al . 2014 , , 784 , l11 zuo , w .- w . , considering the independent measurements of @xmath352 , we have the likelihood function @xmath376 where the probability is given by @xmath377 here @xmath378 is the actual distance of source @xmath26 which is not known . thus we use the probability formulation of @xmath379 . the first term in the integral reads @xmath380 where @xmath381 is the @xmath382-function , @xmath383 is a uniform distribution , @xmath384 and @xmath385 are the actual black hole mass and flux , @xmath386 . so @xmath387 the observational uncertainties on @xmath388 and @xmath389 are independent hence , @xmath390 the likelihood function is @xmath391 where @xmath392 and @xmath393 are given by the asymmetric gaussian functions . we approximate the errors of the measured properties by asymmetric gaussian distributions , @xmath396 & ( { \rm if}~ m_i > m^{\prime}),\\ & \\ \displaystyle \frac{1}{\sqrt{2\pi}\delta_{m_i}}\exp\left [ -\frac{(m_i - m^{\prime})^2}{2\delta_{m_i}^2 } \right ] & ( { \rm if}~m_i\le m^{\prime } ) , \end{array } \right.\ ] ] and @xmath397 & ( { \rm if}~ f_i > f),\\ & \\ \displaystyle \frac{1}{\sqrt{2\pi}\delta_{f_i}}\exp\left [ -\frac{(f_i - f)^2}{2\delta_{f_i}^2 } \right ] & ( { \rm if}~f_i\le f ) , \end{array } \right.\ ] ] where @xmath398 and @xmath382 are the upper and lower error bars of @xmath388 and @xmath389 , respectively .

the advection of a scalar observable @xmath0 by a velocity field @xmath1 is described in classical hydrodynamics by the linear pde @xmath2 if @xmath1 is assumed to be solution of the navier - stokes equations in a turbulent rgime and the pclet number @xmath3 , which measures the ratio between the strength of the advective effects and the molecular diffusion @xmath4 in ( [ pde ] ) , is large @xmath5 ( @xmath6 and @xmath7 are the characteristic length and advection velocity in the problem ) and if a steady state is reached , an inertial range sets in where both the effects of the forcing @xmath8 limited to the large scales and those of the molecular diffusion acting mainly on the small scales can be neglected . in the inertial range no typical scale is supposed to characterise the flow . as a consequence , the structure functions of the scalar field @xmath9^p\right > \label{structurefunc}\ ] ] display a power law behaviour in the inertial range with anomalous scaling exponents @xmath10 @xcite . the word anomalous means that the exponents @xmath10 deviates from the linear behaviour predicted by a direct scaling analysis of ( [ pde ] ) . it was first realised by kraichnan @xcite that anomalous scaling can be observed in the mathematically more tractable case of the advection by a random homogeneous and isotropic gaussian velocity field , which is delta correlated ( white noise ) in time and has zero average and covariance in @xmath11 dimensions given by : @xmath12 . \nonumber\end{aligned}\ ] ] the power law behaviour of the covariance mimics an infinite inertial range for the velocity field . the scaling exponent @xmath13 is a free parameter characterising the degree of turbulence of the advecting field . the physically meaningful values range from zero to two . in the first limit the effect of the random advection is just to define an effective diffusion constant @xcite . in the latter case the velocity increments are smooth as it is expected for a laminar flow . the choice @xmath13 equal to four thirds represents the scaling of the velocity field conjectured by kolmogorov for the solution of the navier - stokes equation in turbulent rgime . the hypothesis of delta correlation in time is of great mathematical advantage for it allows to write the equation of motions of the scalar correlations in a linear closed form . the evolution of each correlation in the inertial range is specified by a linear differential operator , the inertial operator , plus matching conditions at the boundary of the inertial range . the occurrence of anomalous scaling has been related to the existence of zero modes of the inertial operators dominating the scaling properties of higher order correlations ( @xcite and for a recent review and more complete bibliography @xcite ) . the behaviour of the anomaly has also been numerically measured for the fourth order structure function versus the turbulence parameter @xmath13 @xcite . however , to implement accurate numerical experiments still remains a difficult task . therefore it turns out to be useful to use shell model as laboratories to test ideas and results related to the full pde model ( see @xcite for a general introduction to the shell model concept ) . in @xcite two different shell models advected by a delta correlated velocity field mimicking the kraichnan model were constructed . anomalous scaling was observed numerically and in the simpler case @xcite it was proven analytically that the anomaly of the fourth order structure function is related to the anomalous scaling of the dominant zero mode of the inertial operator . the passive scalar advection by a white noise velocity field is a useful mathematical model , but it still very far from a physical realistic velocity field which displays both time correlations and non gaussian fluctuations . a first small step in this direction is made by investigating how the introduction of a time correlation in a gaussian velocity field affects the statistical properties of the scalar field . in the present paper we introduce a time correlated velocity field in a shell model . this is done by substituting the white noise with the ornstein - uhlenbeck process which provides exponentially decaying time correlations ( section [ s : model ] ) . we investigate the model both analytically and numerically . by means of stochastic variational calculus , which we shortly review in appendices [ ap : one ] and [ ap : two ] , we show how to rewrite the equations of motion for the scalar correlations in integral non closed form . such an operation allows the evaluation of the correction to the white noise inertial operator stemming from the time correlated velocity field . this procedure has the further advantage that it creates a non - ambiguous relation between the coupling terms for the scaling exponent @xmath13 of white noise advection to the scaling exponent @xmath14 of the equal time correlation of the time - correlated velocity field ( section [ s : closure ] ) . the inertial operators can be expanded around the white noise limit in powers of an a - dimensional parameter which is interpreted as proportional to the ratio @xmath15 between the time correlation and the turn - over time of the advecting field . we focus on the features of the steady state . there we assume that the averages over the ornstein - uhlenbeck process of all the observables are time translational invariant . as a consequence the inertial operators become linear up to any finite order in @xmath15 . in the white noise case , when @xmath15 is equal zero , we generalise the procedure first introduced in @xcite and we show that the scaling of the zero modes of the inertial operator of any order is captured by focusing on nearest - shell interactions . the equations are closed with a scaling ansatz ( section [ s : wnclosure ] ) by postulating that the scalar field is `` close '' to a multiplicative process . furthermore we perturb the closure scheme in order to extract the first order corrections in @xmath15 to the anomalous exponents for different values of @xmath14 ranging from zero to two . the prediction of perturbation theory is an @xmath15 dependence ( non - universality ) of the exponents except for the second order @xmath16 ( section [ s : perturbative ] ) . the overall result is analogous to the one obtained in @xcite where a gaussian time correlated velocity field is considered for the advection of the scalar field in ( [ pde ] ) : the introduction of time correlation is seen to enhance intermittency . the anomalies vanish smoothly in the laminar limit @xmath14 equal two . to examine the validity of the results from the analytical calculations and explore the regime with long time - correlations ( @xmath17 ) we turned to numerical experiments . the occurrence of corrections to the anomalies predicted by the perturbation theory for small values of @xmath15 was confirmed . however , strong non - perturbative effects sets in and are dominating when the expansion parameter becomes of the order of unity . the model is defined by the equations @xmath18 @xmath19\theta_{m}(t)-\delta_{1m}f(t)=i[k_{m+1 } \theta_{m+1}^{*}(t)u_{m}^{*}(t)-k_{m}\theta_{m-1}^{*}(t)u_{m-1}^{*}(t)]\\ \label{ousde } & & u_{m}(t)=\frac{v_{m}}{\epsilon \sqrt{\tau_{m}}}\int_{0}^{t}ds\ , e^{-\frac{t - s}{\epsilon \tau_{m}}}\eta_{m}(s)\\ \label{force } & & f(t)=\frac{\tilde{f}}{\epsilon\sqrt{\tau } } \int_{0}^{t}ds\,e^{-\frac{t - s}{\epsilon \tau}}\eta(s)\end{aligned}\ ] ] where the star denotes complex conjugation , the @xmath20 s and @xmath21 are independent white noises with zero mean value and correlation : @xmath22 the boundary conditions are @xmath23 . the model might be regarded as a severe truncation of the equation of the passive scalar ( [ pde ] ) in fourier space . the field component @xmath24 is the representative of all the fourier modes in the shell with a wavenumber ranging between @xmath25 and @xmath26 . the parameter @xmath27 is the ratio between two adjacent scales and it is usually taken equal to two in order to identify each shell with an octave of wave numbers . the energy transfer in a turbulent flow is conjectured to occur mainly through the interactions of eddies of the same size . as a consequence the interactions in fourier space are assumed to be local . the `` localness '' conjecture @xcite is the motivation for the restriction to nearest neighbours of the couplings among the shells . in the absence of external forcing and dissipation the total `` energy '' of the passive field is conserved : @xmath28 far from the infra - red and the ultra - violet boundaries ( i.e. for @xmath29 ) the conservation of energy is expected to hold approximately giving rise to an inertial range . equations ( [ ousde ] ) and ( [ force ] ) describe the random evolution according to ornstein - uhlenbeck ( o - u ) processes respectively of the advecting and external force field . the o - u process has differentiable realisations thus making the random differential equations with multiplicative noise which specify the dynamics of the scalar @xmath30 independent of the discretisation prescription . the velocity correlations are for @xmath31 @xmath32 in the limit of large @xmath33 only the stationary part survive . the a - dimensional parameter @xmath15 appearing in the definition of the o - u processes ( [ ousde ] ) and ( [ force ] ) defines the strength of the time correlation in units of the typical times @xmath34 . in the white noise limit one has @xmath35 the factor two is reintroduced here with the proviso that the delta must be understood according to the midpoint prescription when its argument coincides with the upper boundary of integration as it occurs in practical computations in perturbation theory around the white noise limit . for any finite @xmath15 ordinary differential calculus holds true : the consistency conditions yields a stratonovich discretisation prescription when @xmath15 is set to zero and the recovery of the white noise advection model of @xcite . the information about the scaling of the correlations of the velocity field at equal times is stored in the constants @xmath36 . we assume the power law behaviour @xmath37 kolmogorov scaling is specified by @xmath38 while @xmath39 corresponds to a laminar regime . the @xmath40 s in ( [ ousde ] ) describe the typical correlation times for the random velocity field . a simple physical interpretation is to identify them with the turn - over times , i.e. , with the typical time rates of variation through non linearity of the advection field on each shell @xcite : @xmath41 the scaling of the correlation times is then fully specified in terms of the parameter @xmath14 . it is worth to note that for any @xmath14 less than two the @xmath40 s are always decreasing functions of the wave number . the evolution of the scalar @xmath30 is determined in the inertial range by its complex conjugate . it is useful to introduce an unified notation for the @xmath42 degrees of freedom . calling @xmath43 and @xmath44 one has for the n shells : @xmath45\,\theta_{\beta } + f\delta_{\alpha,1}+f^*\delta_{\alpha , n+1 } \label{passcomplete}\ ] ] with @xmath46\nonumber\\ & & b_{\alpha,\beta}^{n+m}=ik_{m+1}[\delta_{\beta , n+m+1}\,\delta_{\alpha , m}- \delta_{\beta , n+m}\,\delta_{\alpha , m}]\end{aligned}\ ] ] where latin and greek indices range respectively from @xmath47 to @xmath48 and from @xmath47 to @xmath42 . the set of matrices with constant entries @xmath49 do not commute within each other and with the @xmath50 matrix . the known sufficient condition ( see for example @xcite ) to have a solution of ( [ passcomplete ] ) in an analytic exponential form is therefore not satisfied . from the geometrical point of view non - commutativity means that the dynamics is confined on a manifold which turns into an hyper - sphere in @xmath51 in the inertial limit ( [ energy ] ) . the complex equations ( [ passcomplete ] ) are invariant under phase transformations . given two diagonal hermitian @xmath52 matrices with time independent random entries @xmath53 if @xmath54 is a realisation of the solution of the equations of motion then @xmath55 is still a solution . the phase symmetry is the remnant of the translational invariance of the original hydrodynamical equations in real space @xcite . from the phase symmetry ( [ identity ] ) it follows that at stationarity the only analytic non zero moments of the correlation are of the form @xmath56 in the inertial range such quantities display a power law behaviour . the diagonal sector of the moments whose scaling properties are specified by the exponents @xmath57 @xmath58 is in the shell model context the analogue of the structure functions ( [ structurefunc ] ) of the original pde model ( [ pde ] ) . the exponents @xmath57 s are said to be normal if they can be derived from dimensional analysis . under the assumption that a steady state is reached one matches the scaling of the inertial terms in ( [ passive ] ) with a power law decay of the solution @xmath59 the resulting prediction is a linear behaviour of the exponents versus the order @xmath60 of the diagonal correlation : @xmath61 the scaling argument ( [ matching ] ) neglects completely the random fluctuations of the passive scalar field . normal scaling holds if the statistics of the @xmath30-field is gaussian . deviations from normal scaling are then related to the occurrence of intermittency corrections to the gaussian statistics . a systematic account of the fluctuations is provided by the study of the equations of motion satisfied by the moments of the scalar field . in the white noise limit , @xmath15 equal zero , the furutsu - donsker - novikov formula @xcite and the delta correlation in time of the velocity ensure that the moments @xmath62 are specified by the solutions of closed linear systems @xcite . in the presence of finite time correlations stochastic calculus of variations @xcite allows to write non closed integro - differential equations for the correlations . a typical functional integration by parts relation is : @xmath63 where einstein convention holds for repeated _ greek _ indices . the matrix @xmath64 is the fundamental solution of the homogeneous system associated with ( [ passcomplete ] ) . a heuristic proof of the stochastic integration by parts formula and of ( [ typical ] ) is provided in appendices [ ap : one ] and [ ap : two ] . let us start with the second moment of the scalar field @xmath65 > from the equations of motion ( [ passcomplete ] ) one has @xmath66 c_m^{(2)}(t ) -2 \re\left\{\left<\theta_{n+m}(t)f(t)\right>\right\}\delta_{m,1}=}\nonumber\\ & & + 2 \re\left\{i\,k_{m+1}\left < u_{n+m}(t)\theta_{n+m+1}(t)\theta_{n+m}(t ) \right>\right\}+\nonumber\\ & & -2\,k_{m}\re\left\{i\,\left < u_{n+m-1}(t)\theta_{n+m-1}(t)\theta_{n+m-1}(t ) \right>\right\}. \label{2pointsdiff}\end{aligned}\ ] ] the integration by parts formula ( [ typical ] ) gives @xmath67 c_m^{(2)}(t)-2 \delta_{m,1}\re\ , \int_{0}^{t } \,ds \,\left < f(t)f(s)^*\right>\left < r_{n+m , n+1}(t , s)\right>= } \nonumber\\ & & 2\,k_{m+1}^2 \tau_{m } \int_{0}^{t}\,ds \ , \frac{\left < u_{m}(t)u_{n+m}(s)\right>}{\tau_m } \re{\cal f}_m^{(2)}(t , s)+\nonumber\\ & & -2\,k_{m}^2 \tau_{m-1 } \int_{0}^{t}\,ds \ , \frac{\left < u_{m-1}(t)u_{n+m-1}(s)\right>}{\tau_{m-1 } } \re{\cal f}_{m-1}^{(2)}(t , s ) \label{2points}\end{aligned}\ ] ] where @xmath68 , @xmath69 is the real part and : @xmath70 when a steady state is reached the lhs of ( [ 2points ] ) can be neglected through the whole inertial range . the rhs specifies the inertial operator of the theory . a further simplification is attained in the limit of very large shell number . for any @xmath14 less than two one has @xmath71 independently of @xmath15 . at equal times the resolvent matrix @xmath64 reduces to the identity . > from ( [ scalvel ] ) and ( [ tauscale ] ) it stems that @xmath72 hence for @xmath73 going to infinity the inertial operator is linearised in the form @xmath74 the slowest decay scaling solution compatible with a zero lhs is @xmath75 in other words we have proven that the scaling of the second moment is normal as it coincides with the dimensional prediction ( [ normalscaling ] ) . moreover since the result does not depend on @xmath15 it is universal versus the time correlation . it is worth stressing that the derivation of ( [ 2pointnorm ] ) requires that each of the terms appearing in ( [ largem ] ) has separately a finite non zero limit for @xmath73 going to infinity . the condition turns out to be not self - consistent when the same reasoning is applied to moments higher than the second . an important consequence of normal scaling of the @xmath76 s is the obukhov - corrsin @xcite law for the decay of the power spectrum @xmath77 of the passive scalar if the kolmogorov scaling is assumed for the advecting field : @xmath78 a second interesting limit is when @xmath15 tends to zero . neglecting all non stationary contributions to the velocity correlations the rhs of ( [ 2points ] ) becomes @xmath79 if @xmath15 is set exactly to zero the integral terms disappear and the white noise equations of @xcite are recovered . the information about the scaling of the velocity field is absorbed in the @xmath80 s . in a pure white noise theory it is convenient to redefine the turbulence parameter as @xmath81 a kolmogorov scaling of the velocity field corresponds to @xmath13 equal to four thirds which is also the value giving the obukhov - corrsin scaling in ( [ oc ] ) . the two definitions of the degree of turbulence coincide for @xmath14 equal to two ( batchelor limit ) . it is natural to identify @xmath13 with the turbulence parameter of the kraichnan model . the correspondence fixes the physical range of @xmath14 between @xmath82 $ ] . in the general case of the @xmath83-th even moment of the scalar @xmath84 one has : @xmath85 the multidimensional matrix @xmath86 is the linear inertial operator of the white noise theory . the integrand functions @xmath87 are given by the straightforward generalisation of ( [ ffunc ] ) . the lhs , as above , is set to zero as far as the steady state features of the inertial range are concerned . repeated integrations by parts in the large @xmath33 limit generate a laplace asymptotic expansion @xcite of integral terms in the rhs , the coefficient of which are the derivatives with respect to @xmath88 of the functions @xmath89 evaluated at equal times . when the steady state sets in we assume the latter quantities to be invariant under time translations for large @xmath33 . under such an assumption it will be proven in section [ s : perturbative ] that the equal time derivatives are specified at equilibrium by linear combinations of the @xmath62 s . the effect of a small time correlation is therefore to generate new couplings of order @xmath15 in the inertial operators . the observables we focus on are the scaling exponents . as discussed in the introduction , anomalies occur in the presence of non trivial scaling zero modes of the white noise inertial operators . it makes sense to relate the @xmath15 dependence of the anomalous exponents to a perturbation of the scaling zero modes derived for @xmath15 equal zero . a straightforward approach to the problem calls for the solution of @xmath90 linear equations . a further source of difficulty is that the exact determination of the zero eigenvectors of the inertial operators of any given order requires the matching of infra - red and ultra - violet boundary conditions . in the absence of an exact diagonalisation any analytical approach must rely on closure anstze to solve first the white noise problem and then to yield the corrections to the zero modes by linear perturbation theory . in the present section we present a closure strategy to compute the @xmath91 s in the case of white noise advection . as shown in the previous section the second diagonal moment is normal and universal versus the time correlation . the first nontrivial zero mode problem is provided by the fourth order inertial operator @xmath92 . in @xcite it has been shown that the anomalous exponent @xmath93 @xmath94 can be extracted up to a very good accuracy from the solution of only two non - linear algebraic equations . the stationary equations for @xmath95 in the inertial range far from the infra - red and ultra - violet boundaries are given by @xmath96 one recognises two kinds of couplings in @xmath97 . * `` global '' or `` unconstrained '' interactions : the indices @xmath98 and @xmath99 range respectively from @xmath100 to @xmath101 and from @xmath102 to @xmath103 . the coupling are independent upon the relative values of @xmath73 and @xmath104 : in this sense they are referred as global . * `` purely local '' interactions : they occur only for @xmath105 and correspond to the terms proportional to the kroenecker s @xmath106 in ( [ irwn4 ] ) . anomalous scaling in the inertial range is strictly related to the presence of such purely local interactions . were these latter neglected the fourth order moment would have a normal scaling solution @xmath107 the idea is to capture the anomalous scaling by looking at the `` renormalisation '' of global couplings by pure short range ones . disregarding the boundaries , the system is invariant under a simultaneous shift of the indices . hence assuming a perfect index - shift invariance there are , for the @xmath73-th shell , only two independent equations where @xmath106-like terms occur : @xmath108 the third equation involving a purely local interaction of the @xmath73-th shell with its nearest neighbours @xmath109 is generated from the second of the ( [ 4closed ] ) s by a simple index shift . therefore it is not regarded as independent . the pair ( [ 4closed ] ) contains all the relevant information to extract the scaling of the fourth moment . it forms a closed system of equations independently on the shell number @xmath73 as one imposes scaling relations to hold within the set of `` independent '' moments of fourth order : @xmath110 where the integer @xmath104 is taken larger than zero . as in the analysis of the interactions the concept of independence stems from the assumption of index shift invariance : the moments of the form @xmath111 are immediately reconstructed once ( [ 4diagonal ] ) and ( [ 4nondiagonal ] ) are given : @xmath112 let us analyse the closure ansatz in more detail . the first equation ( [ 4diagonal ] ) is a global scaling assumption of the `` diagonal '' sector of the fourth moment . its justification lies in the very definition of an inertial range . the second scaling assumption relates the diagonal sector to the non diagonal one via a marginal scaling . it is the analogous in the present context of an operator product expansion ( ope ) in statistical field theory @xcite . there renormalisation group ( rg ) techniques are able to describe the scaling behaviour of correlations of fields sampled at large real space distances one from the other . if an observable requires the evaluation of a correlation including the products of one field in two points at short distance i.e. @xmath113 the rg procedure can not be directly applied . the problem is overcome by an ope or short distance expansion . the prescription is to rewrite the product via a taylor expansion in terms of local composite operators sampled just in one point . such a point is now well separated from all the other appearing in the correlation function . the original correlation is substituted by a set correlations such that rg applies provided an extra renormalisation , renormalisation of composite operators ( rco ) , is introduced . the latter is understood by observing that in our example the first term in the taylor expansion gives @xmath114 the mathematical meaning of a field is that one of an operator - valued distribution . the product of two distributions at equal points i.e. @xmath115 requires a regularisation before the cut - off is removed in order to make itself sense as a distribution . this is the content of the rco . finally at leading order the relation between the renormalised quantities reads for the above example @xmath116_{r} ... \right > \sim c(dx)\left < \left[\phi(x)^{2}\right]_{r} ... \right>. \label{ope}\end{aligned}\ ] ] roughly speaking the small real space separations are associated with the uv behaviour of the fourier conjugated variable . in the shell model context the @xmath117 are representative of the scalar field variation over one octave . the moments @xmath118 correspond to the fourier transform of the fourth order structure functions of a homogeneous and isotropic real space solution of the passive scalar equation ( [ pde ] ) . the meaning of the non - diagonal closure ( [ 4nondiagonal ] ) is to assume the long range ( many shells ) behaviour of the fourier transform of the ope coefficient @xmath119 to scale inside the inertial range for large @xmath104 independently on @xmath73 . the constant @xmath120 renormalises the value of the first shell where deviation from scaling occur . the analogy with the ope is then summarised by writing @xmath121^{2}_{r } \nonumber\\ & & \lim_{|k|\uparrow \infty } \int d^{d}x e^{i k \cdot dx } c(dx ) \sim x\,k_{n-1}^{-h(2)}\end{aligned}\ ] ] the insertion of the scaling ansatz in ( [ 4closed ] ) leaves a non - linear system in the unknown variables @xmath122 and @xmath120 . by applying the definition @xmath123 one gets @xmath124 which after a straightforward manipulation provides @xmath122 as the physical root of a second order polynomial @xmath125 in terms of @xmath122 the anomaly is @xmath126 and it proves proves to be in fair agreement with the values obtained from the numerical solution of the exact equations ( [ irwn4 ] ) @xcite and from the numerical integration of ( [ passive ] ) for all the values of the turbulent exponent @xmath14 in the physical range @xmath127 $ ] ( see also fig . [ fig : wn ] ) . the sign of @xmath93 is always positive : the effect of the anomaly is to decrease the diagonal scaling exponent . the procedure presented in detail for the computation of the fourth order exponent is straightforwardly extended to any higher order moment when one recognises that two crucial observation hold in general . * in the absence of pure short range couplings the normal scaling prediction holds true far from the boundaries for the zero modes of the inertial operators of any order @xmath128 . * for any fixed shell @xmath73 there is a one to one correspondence between the number of independent equations and moments of order @xmath128 . in the case of @xmath62 there are @xmath129 equations : for any fixed reference shell @xmath130 the interaction with the second index @xmath131 is affected by a pure short range coupling if the latter is equal or one unit different from @xmath130 i.e. only two possible choices and so on until the @xmath60-th index is reached . on the other hand @xmath129 is the number of exponents which characterise the scaling of the @xmath128-th moment . the ansatz is that the marginal scaling of the non - diagonal sector is fully specified in terms of the diagonal scaling exponents of order less than @xmath83 . by means of the `` ope s '' one is able to close the zero mode equations in terms of @xmath129 unknown renormalisation constants and @xmath91 . the analogy with a field theoretical ope goes to show that the need for an infinite set of constant does not necessarily imply the non - renormalisability of the real space theory mimicked by the shell model @xcite . more concretely the diagonal scaling exponent of the sixth moment ( @xmath132 ) of the scalar field @xmath133 according to the above criterion requires four independent equations ( see appendix [ ap : sixthorder ] ) @xmath134 the ope inspired closure yields @xmath135 inserting the `` ope '' in ( [ 6closed ] ) one gets into the algebraic system for the unknown renormalisation constants @xmath136 and the diagonal scaling factor @xmath122 . @xmath137 after some algebra ( [ 6nonlin ] ) reduces to a single third order polynomial specifying the physical root of @xmath122 . it is worth to remark that from the functional dependence of the coefficient of ( [ 6nonlin ] ) the exponent @xmath138 depends upon the anomaly of @xmath139 . once again the anomaly evaluated from @xmath140 is in fair agreement with numerics ( see fig . [ fig : wn ] ) for different values of @xmath14 . in appendix [ ap : eightorder ] the same steps are performed in the case of the eight moment @xmath141 . the analytical predictions for the anomalous exponents are summarised in fig . [ fig : wnanomal ] . in all cases the anomalies are decreasing functions of the turbulence parameter @xmath14 vanishing smoothly when the laminar limit ( @xmath14 equal two ) is approached . the anomaly of the fourth order moment can be compared with the results of numerical experiments for the fourth order structure function of the kraichnan model @xcite . there the adopted turbulence parameter is @xmath13 . for values of @xmath13 of order one , i.e. , from the kolmogorov scaling up to the batchelor limit one indeed observes the same monotonically decreasing behaviour with values of the anomaly of the same order of those found in the shell model . for lower values of @xmath13 the anomaly in the kraichnan model display a maximum before decreasing to zero for @xmath13 equal to zero , i.e. , when @xmath14 tends to minus two . no sign of such behaviour is observed in shell model . the discrepancy might be an artifact of the shell model , which was originally designed to mimic the supposed local in scale character of the nonlinear interactions in a turbulent flow @xcite , fails to describe a rgime where strong non local effects become important . on a phenomenological level the energy transfer in the inertial range of turbulent field is related to the occurrence of a cascade mechanism as firstly conjectured by richardson @xcite . the conservation of energy in the inertial range imposes that the forcing occurring on large real space scales is transferred to small scales ( i.e. large wave numbers ) before being dissipated . a mathematical description of the cascade is provided by multiplicative stochastic processes @xcite . multiplicative modelling has been shown to account for most of the features observed in real and synthetic turbulence @xcite . in the present case the idea of a multiplicative structure is incorporated in the hypothesis that the scaling of the non diagonal sector of a given moment of order @xmath83 is reconstructed once the scaling of the lower moments is known . such an assumption together with the analysis of the couplings in the inertial operator of order @xmath83 yields with fair accuracy the scaling exponents of the model without the resort to an exact diagonalisation of inertial operator . let us now turn to the time correlated case . the idea is to evaluate the scaling behaviour of the dominant zero modes of the inertial operators ( [ 2omegabyparts ] ) linearised up to first order in @xmath15 by perturbing the white noise closure ansatz . the first order corrections in @xmath15 to the inertial operators are obtained by truncating the integration by parts to the terms linear in @xmath15 @xmath142 the use of ( [ coeff ] ) in the rhs stresses that the effective a - dimensional expansion parameter is @xmath143 : the range of reliability of first order perturbation theory is compressed to @xmath144 . as mentioned in section [ s : closure ] , in the limit @xmath33 going to infinity , one expects two time quantities to be stationary . in such a case the derivative with respect to the variable @xmath88 can be interchanged with the derivative with respect to @xmath33 and one can use the equations of motion to evaluate ( [ correctiongen ] ) . a direct differentiation with respect to @xmath88 is consistently taken with respect to the system of stochastic differential equations conjugated by time reversal of equations ( [ passive ] ) to ( [ force ] ) . the latter operation in general requires the knowledge of the probability density of the forward problem . in the stationary limit the time reversal operation for the o - u process reduces to the inversion of the sign of the drift term as in the deterministic case . after a slightly more lengthy algebra the result is equal to the differentiation with respect to @xmath33 with opposite sign . the computations in the general case are very cumbersome ( see appendices [ ap : fourthorder ] , [ ap : sixthorder ] and [ ap : eightorder ] ) . it is convenient to exemplify the procedure in the simpler case of the second order correlation . there are four contributions to @xmath145 : @xmath146 by definition @xmath147 the time derivative of @xmath148 is derived by a simple index shift . as the terms non - diagonal in the resolvent @xmath64 indices are zero , the second moment inertial operator is not by affected by first order perturbation theory . the result is not surprising . the second moment has only one free index . hence at any order of perturbation theory only global coupling can be generated which are forced by the symmetries of the model to be consistent with a normal scaling of the zero mode . moreover the @xmath149 s components of the scalar evolve only through the coupling with their complex conjugated @xmath150 s : their variation is a second order effect in @xmath15 . the only possible non zero corrections are viscous and can be consistently neglected . let us turn to draw the general picture when @xmath60 is larger than one . once again the phase symmetries ( [ identity ] ) and the fact that when @xmath151 is known all other terms are yielded by index shift or exchange operations , prevent the corrections to the global couplings from affecting the scaling properties : the resulting `` global '' sectors of inertial operators have a normal scaling zero mode . this is in agreement with the observation made in @xcite , where the dependence of the scaling exponents on the time correlation for generalised models of passive scalar advection is predicted to appear only through anomalies . the corrections to the purely short range couplings are therefore the relevant ones . they occur in two ways . on one hand new terms of order @xmath15 show up in the purely self and nearest - neighbours interactions . on the other hand , terms proportional to @xmath152 the latter ones are the most dangerous for they in principle perturb the logic of the white noise closure by introducing new independent equations and henceforth the need for more renormalisation constants in the non diagonal sector of the moments . nevertheless one can argue a priori in the spirit of the renormalisation group @xcite , only the nearest neighbours interactions are relevant for scaling . hence first order corrections can be obtained allowing an @xmath15 dependence in the renormalisation constant of the white noise closure and determining the first order coefficient of their taylor expansion . moreover for @xmath60 larger than two such a strategy is already able to take into account the corrections due to the purely second neighbours interactions . let us analyse in more detail the case of @xmath153 . the white noise closure is perturbed by introducing an @xmath15 dependence in the renormalisation constants @xmath154 the marginal scaling in the non - diagonal sector in ( [ 4perturbednondiag ] ) is assumed to stay universal as it is for @xmath155 while the @xmath15 dependence is stored in the pre - factor . the diagonal exponent is then determined up to first order as @xmath156 where @xmath157 is the derivative of @xmath122 at @xmath15 equal zero yielded by the perturbative solution of the system @xmath158\ , c_{p , q}^{(4)}(\epsilon ) = 0\nonumber\\ \sum_{p , q}[\,i_{m , m-1;p , q}^{(4;0)}+\epsilon\,i_{m , m-1;p , q}^{(4;1)}\ , c_{p , q}^{(4)}(\epsilon)=0 . \label{4closedpert } \end{array } \right.\ ] ] the correction to @xmath93 due to time correlation increases the anomaly leading to a slower decay of the diagonal moment . for @xmath157 is negative ( see fig . [ fig : ouanomalfour ] ) the overall anomaly is @xmath159 in the range of reliability of first order perturbation theory the effect is very small : for @xmath160 the prediction is a correction amounting to the three percent of the white noise exponent @xmath139 . the perturbative scheme just proposed does not take into account the emergence of pure second neighbours interactions . in order to weight their relevance for the diagonal scaling and simultaneously to check the hypothesis of normal scaling for the marginal scaling in ( [ 4perturbednondiag ] ) one can relax the closure in order to encompass the equation @xmath161\ , c_{p , q}^{(4)}(\epsilon)=0 \label{secondneighbours}\ ] ] which describe the independent , in the sense stated above , second neighbours interaction . consistency with the white noise theory imposes the latter equation to decouple when @xmath15 is set to zero . the requirement is satisfied if the closure is chosen in the form @xmath162 the pre - factor @xmath163 enforces @xmath164 to be a function of the white noise renormalisation constants . were the white noise closure exact it would fix the value of @xmath164 to one . in fig . [ fig : qzero ] the @xmath164 is plotted versus @xmath14 : through all the physical range it stays close to one with a maximal deviation on the order of four percent for @xmath14 equal to minus two . moreover as shown in fig . [ fig : ouanomalfour ] the time correlation induced correction to @xmath165 when ( [ secondneighbours ] ) is included has the same qualitative behaviour and is quantitatively very close to the nearest - neighbours prediction . the result is an a - posteriori check of the robustness of the closure approach . it confirms that first order corrections can be extracted within the logical scheme of the zero order one . it follows that the equations specifying the zero modes of the inertial operator acting on the sixth moment ( see appendix [ ap : sixthorder ] ) can be closed by assuming : @xmath166 the exponent @xmath165 is known perturbatively from ( [ fourperturbed ] ) while @xmath16 is universal . with the same rationale ( appendix [ ap : sixthorder ] ) one can evaluate @xmath167 . in fig . [ fig : ouanomal ] the analytic predictions for the corrections to the scaling exponents are summarised . in all cases the corrections are negative i.e. they carry a positive contribution to @xmath168 . the corrections increase with @xmath60 , the rate of the growth being slightly slower than the @xmath169 predicted in @xcite for time - correlation generalised pde kraichnan models . within the range of first order perturbation theory the overall effect of time correlation is seen to enhance intermittency . an intuitive understanding of the phenomenon might be obtained by interpreting the time correlation as a mechanism to increase the probability of coherent fluctuations of the scalar field . the latter are rare events felt in the tail of the probability density of the scalar field as extreme deviations from the gaussian behaviour of the typical events . the resort to numerical experiments has a double motivation . on one hand they can be used to test the predictions from the first order perturbation theory . on the other hand they provide a broader scenario of the features of the model beyond the grasp of perturbative approaches . the first task is far from being easy because a quantitative check of perturbation theory requires measurements of the scaling exponents within an accuracy smaller than two percent . the main feature of the inertial range is the conservation of the scalar `` energy '' . from the analytical point of view this is seen in the non - commutativity of the terms associated with the multiplicative noise , @xmath170 in ( [ passcomplete ] ) . this property rules out the use of a simple euler scheme , which can be applied in the case of delta correlated noise . in the case of white noise advection the multiplicative structure of the noise ( [ passcomplete ] ) , which is interpreted in the stratonovich sense , can be mapped into the corresponding it equations . the advantage is that the diagonal non zero average part of the noise is explicitly turned into an effective drift term @xcite . the non - diagonal terms in the taylor expansion of the scalar field @xmath54 are of the order three halves in @xmath171 , which are neglected in the euler scheme . this procedure becomes meaningless for a time - correlated noise . there ordinary calculus holds and in the taylor expansion of @xmath54 both diagonal and non - diagonal products of the noise are of the same order in @xmath171 . moreover the algorithm to be used must tend smoothly to a white noise limit , so that the same relative error is preserved independently on the value of @xmath15 . following burrage & burrage @xcite a reliable way out from the mentioned difficulties is to adopt the trotter - lie - magnus formula to integrate the equations of motion to first order . for each time increment @xmath171 ( [ passcomplete ] ) is solved in exponential form . fast matrix exponentiation algorithms are provided by the package expokit @xcite . to generate the correlated noise , the exact method described by miguel & toral @xcite is used . this method ensures that the noise is accurate down to the limit @xmath172 . the relevant time scale to measure the convergence of the solution is the slowest time scale in the system , namely the eddy - turn - over time of the first shell estimated as the maximum between @xmath173 and @xmath174 . as shown in fig . [ fig : contexample2 ] more than @xmath175 eddy - turn - over times are needed to achieve a converged solution for the sixth order structure function . the time step is set by the fastest time scale in the system , which is the one of the largest shell @xmath176 . the number of iterations needed to achieve convergence is then for @xmath15 less than one : @xmath177 which shows that the number of iterations needed grows like @xmath178 , making it difficult to get close to the white noise limit using the same algorithm . the scaling of the diagonal moments of higher order has been extracted by means of extended self - similarity @xcite , where the @xmath98-th order structure function is plotted versus the second order one , which is assumed to be normal . the scaling is found as the average slope of the logarithmic derivatives in the inertial range . we considered a system with twenty - five shells with wave numbers increasing as power of @xmath179 , with viscosity @xmath180 . this choice ensures that there are several shells in the dissipative range . we focused on the results for @xmath14 equal two thirds ( kolmogorov scaling ) . in fig . [ fig : structexample2 ] the `` normalised '' structure functions @xmath181 are shown . the quality of the scaling is demonstrated by the fact that the moments show scaling over a wide range of scales . a summary of the numerical experiments is given in fig . [ fig : compare ] where the scaling exponents are plotted versus the order of the moments of the scalar field for different values of @xmath15 . it is evident that the anomaly grows as the time correlation increases . when turning to the interpretation of the results in more detail , the uncertainty in the extraction of the scaling has to be kept in mind . for the sixth moment this uncertainty turned out to be on the order of a four percent . the changes in the scaling between different values of @xmath15 is also on the order of a few percent . this seems to exclude a proper resolution in the numerics to compare the results with the analytical predictions from the perturbation analysis . however , the results for different values of @xmath15 can still be compared with some confidence , as the relative uncertainty between the different runs is much smaller than the absolute uncertainty . this means that the slope of , i.e. , the sixth order structure functions vs. @xmath15 will be well resolved , while the absolute values can be shifted up and down a few percent . in fig . [ fig : epsilon ] the analytical prediction of the exponents is compared with the result of the numerics . the theoretical points are systematically below the numerical ones which is due to the absolute uncertainty as explained above . the slope is the same for the analytical calculation and the numerics , giving credibility to the results of the perturbation analysis . it should be noted that the effect of time correlation on the anomaly is quantitatively quite small even in the non perturbative range when @xmath15 is equal to one ( @xmath143 equal four ) . the global picture provided by the numerical experiments is that @xmath182 is seen to be a non linear function of @xmath15 which , after rapid initial decrease in the perturbative range , displays a much slower rate of variation . an interesting question is whether there is a limiting value of the scaling of the structure function as @xmath183 or not . however the quality of the numerics does not allow us to answer this question . we have presented a shell model for the advection of a passive scalar by a velocity field which is exponentially correlated in time . we developed a systematic procedure to calculate the exponents of the correlation of the diagonal moments ( the structure functions ) . for the delta correlated velocity we find good agreement between analytical and numerical calculations up to the eight order . we presented an analytical perturbative theory to compute the correction to the scaling exponents due to the exponentially correlated velocity field . the occurrence of anomalies in the exponents of the diagonal moments of the scalar and their non universality versus the intensity @xmath184 of the time correlation , is related to the presence of pure short range couplings in the corresponding inertial operator which provide for non trivial scaling of the zero modes . in the absence of such short range couplings , as is the case for the second moment , normal scaling would take place independently on the value of @xmath143 . the behaviour of the anomalous exponents in the non - perturbative regime was studied numerically . this was found to be a non linear monotonic function of @xmath143 , decreasing with a rate much slower than in the perturbative regime . it is thus clear that the addition of the time correlation to the advecting velocity field enhances the anomalous scaling . the anomaly found in the present study is still much smaller that what is found when the passive scalar is driven by a turbulent velocity field driven by navier - stokes turbulence or by a shell model for the velocity field @xcite . this indicates that the non - gaussian nature of the real turbulent velocity field plays a significant rle for the strong anomalous scaling observed for real passive scalars . the authors wish to thank a. vulpiani for drawing their attention to the problem . discussions with e. aurell , a. celani , p. dimon and e. henry are gratefully acknowledged . a particular thank to m. h. jensen for his interest and encouragement during our work and to m. van hecke for many physically insightful comments . pmg is supported by the tmr grant erb4001gt962476 from the european commission . a heuristic proof of the stochastic integration by parts formula is provided . for a rigorous treatment see @xcite , @xcite . let @xmath185 be a stochastic process whose realisations are defined as the solution of the it sde @xmath186 where @xmath187 is white noise . let @xmath188 the stochastic process specified by @xmath189 for @xmath15 equal the two sde s coalesce : ( [ a1sde2 ] ) can be derived from ( [ a1sde1 ] ) under the variation of the white noise @xmath190 . the integration by parts formula states that for any smooth functional @xmath8 the identity holds : @xmath191 where @xmath192 denotes the expectation values with respect to the measure induced by @xmath185 in order to prove it let us observe that the transition probability density for ( [ a1sde2 ] ) can be written formally as a path integral ( it discretisation ) : @xmath193}\nonumber\\ s_{\zeta}(x , t|x_0,0)&=&\int_{0}^{t } \,\,\frac{dt'}{2 } [ \frac{\dot{x_{t'}}-b(x_{t'})}{\sigma(x_{t}',t')}]^2\end{aligned}\ ] ] if one introduces the functional @xmath194 } \label{functional}\ ] ] one has by construction @xmath195 to each realisation of the solutions of ( [ a1sde2 ] ) corresponds a mapping @xmath196 . hence the last equality can be rewritten as the white noise average : @xmath197 which implies ( [ a1intbyparts ] ) when @xmath15 is set to zero . the derivative @xmath198 is a frchet derivative . the dynamics of the stochastic process @xmath199 is linear once the realisations @xmath200 of @xmath185 are known : @xmath201+h(x_t , t ) \sigma(x_t , t ) \label{variation}\end{aligned}\ ] ] it is worth to note that for @xmath202 the integration by parts formula ( [ a1intbyparts ] ) reduces to @xmath203 which is the gaussian integration by parts formula ( see e.g. @xcite ) applied to the wiener process @xmath204 . the generalisation to a multidimensional complex case proceeds straightforwardly by introducing @xmath205 variational parameters @xmath206 and applying the definitions @xmath207 for the white noise correlations . as in the above appendix we limit ourselves to the real case the generalisation to the complex case being straightforward . functional differentiation is formally derived from a frchet derivative with @xmath208 where @xmath88 is a parameter specifying the time when the white noise is perturbed . the variation is assumed to be non - anticipating ( causal ) : @xmath209 let us consider the system of sde s @xmath210 where @xmath211 is the coloured noise : @xmath212 functional differentiation gives @xmath213+\frac{e^{-\frac{t - s } { \epsilon \tau_l}}}{\epsilon \sqrt{\tau_l}}\sigma_{m , l}(x ) \label{a2var}\ ] ] the functional derivative is fully specified when it is known its form at the time @xmath88 when the variation of the white noise occurs . the latter is determined by the causality requirement @xmath214\ , \delta_{n , l } \label{varnoise}\ ] ] which implies the variation of the coloured noise to be nonzero _ only _ immediately after the instantaneous kick @xmath215 by differentiating ( [ a2var ] ) one finds @xmath216 > from the last equation it stems that for @xmath217 @xmath218 consistency with ( [ a2var ] ) then requires that the variation of the @xmath120 s associated with a non anticipating variation of the white noise at time @xmath88 fulfils the initial condition : @xmath219 the integration by parts formula ( [ a1intbyparts ] ) for a smooth functional @xmath220 @xmath221 the variation is the solution of the linear problem ( [ a2var ] ) of which we define @xmath64 to be the fundamental solution . it follows @xmath222\,r_{l , m}(t , s')\sigma_{m , n } ( x_s')\right>\ ] ] finally inverting the order of integration one obtains @xmath223\,r_{l , m}(t , s')\sigma_{m , n } ( x_s')\right > \label{result}\ ] ] this proves the real version of formula ( [ typical ] ) . the inertial operator acting on the fourth moment @xmath224 is in the large time limit @xmath225 the bi - dimensional matrix @xmath226 is the white noise linear inertial operator . the corrections to the white noise theory are generated by the time derivative at equal times of the integrand function @xmath227 @xmath228\right>+\nonumber\\ & & -\left < \theta_{n+m}(t ) \theta_{n+n}(t)\theta_{n}(t ) r_{n+m+1,n+m+1}(t , s)\theta_{m}(s)]\right > + \nonumber\\ & & + \left < \theta_{n+m+1}(t ) \theta_{n+n}(t)\theta_{n}(t ) r_{n+m , n+m}(t , s)\theta_{m+1}(s)]\right > + \nonumber\\ & & -\left < \theta_{n+m+1}(t ) \theta_{n+n}(t)\theta_{n}(t ) r_{n+m , n+m+1}(t , s)\theta_{m}(s)]\right > + \nonumber\\ & & + \left < \theta_{n+m}(t ) \theta_{n+m+1}(t)\theta_{n+n}(t ) r_{n , n+m}(t , s)\theta_{m+1}(s)]\right > + \nonumber\\ & & - \left < \theta_{n+m}(t ) \theta_{n+m+1}(t)\theta_{n+n}(t ) r_{n , n+m+1}(t , s)\theta_{m}(s)]\right > + \nonumber\\ & & + \left < \theta_{n+m}(t ) \theta_{n+m+1}(t)\theta_{n}(t ) r_{n+n , n+m}(t , s)\theta_{m+1}(s)]\right > + \nonumber\\ & & - \left < \theta_{n+m}(t ) \theta_{n+m+1}(t)\theta_{n}(t ) r_{n+n , n+m+1}(t , s)\theta_{m}(s)]\right > \label{4tobeint}\end{aligned}\ ] ] after a double integration by parts neglecting viscous contributions one gets into @xmath229 + \nonumber\\ & & + \delta_{n , m+1 } [ ( - { \frac{2\,{{\lambda } ^2}}{{{\tau } _ m}}}+ { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } + { \frac{34\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(4)}_{m+1,m}- { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } \,c^{(4)}_{m , m } + \nonumber\\ & & + ( { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(4)}_{m , m-1}\,- ( { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m}}}}+ { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(4)}_{m+1,m-1}- { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } \,c^{(4)}_{m+1,m+1 } + \nonumber\\ & & + ( { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } + { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) \,c^{(4)}_{m+2,m+1}\ , - ( { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ m}}}+ { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) \,c^{(4)}_{m+2,m}]+ \nonumber\\ & & + \delta_{n , m-1 } [ ( - { \frac{2\,{{\lambda } ^2}}{{{\tau } _ { -1 + m}}}}+ { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } + { \frac{34\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(4)}_{m , m-1 } - { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } \,c^{(4)}_{m , m}+ \nonumber\\ & & + ( { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(4)}_{m+1,m } - ( { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(4)}_{m+1,m-1}- { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } \,c^{(4)}_{m-1,m-1 } + \nonumber\\ & & + ( { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(4)}_{m-1,m-2}\,- ( { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(4)}_{m , m-2}]+ \nonumber\\ & & + \delta_{n , m+2 } [ ( -{\frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) c^{(4)}_{m+1,m}\ , ( -{\frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) c^{(4)}_{m+2,m+1 } + \nonumber\\ & & + ( { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } + { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) c^{(4)}_{m+2,m}\,]+ \nonumber\\ & & + \delta_{n , m-2 } [ ( -{\frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) c^{(4)}_{m , m-1}\ , + ( -{\frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) c^{(4)}_{m-1,m-2 } + \nonumber\\ & & + ( { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) c^{(4)}_{m , m-2}\ , ] \label{fourthorder}\end{aligned}\ ] ] the diagonal scaling exponent is derived up to first order is @xmath15 resorting to linear perturbation theory . if pure second neighbours interactions are taken into account the constant @xmath164 is specified by @xmath230 the result is approximately equal to one for all @xmath14 ranging between @xmath127 $ ] . the first order correction @xmath231 is extracted from the solution of the linear system : @xmath232 @xmath233 \,x'(0 ) + \nonumber\\ & & + [ { { \lambda } ^2 } + ( -3 - { { \lambda } ^{-h(2 ) } } - { { \lambda } ^{h(2 ) } } ) \,x(0){{\lambda } ^2 } + 2\,{{\lambda } ^{2 - 2\,h(2)}}\,x(0)\,z(0 ) ] \,z'(0)=\nonumber\\ & & = -13\,{{\lambda } ^4 } - 3\,{{\lambda } ^{4 + h(2 ) } } + 3\,{{\lambda } ^{4 + h(2)}}\ , ( 2 + { { \lambda } ^{-h(2 ) } } ) \,x(0 ) - 13\,{{\lambda } ^4}\,z(0 ) - 3\,{{\lambda } ^{4 - h(2)}}\,z(0)+ \nonumber\\ & & + { { \lambda } ^4}\ , ( 42 - 2\lambda^{-2\,h(2 ) } + 7\lambda^{-h(2 ) } + 9\,{{\lambda } ^{h(2 ) } } + q(0 ) ) \,x(0)\,z(0)+ \nonumber\\ & & + { { \lambda } ^4}\,x(0)\,{{z(0)}^2 } [ 1 - 10\lambda^{-2\,h(2 ) } + ( -1 + 2\,q(0))\lambda^{-3\,h(2 ) } + ( 3 + 2\,q(0))\lambda^{-h(2 ) } ] + \nonumber\\ & & + { { \lambda } ^{4 - 4\,h(2)}}\,q(0)\,x(0)\,{{z(0)}^3 } \nonumber\end{aligned}\ ] ] @xmath234 \,z'(0 ) + \nonumber\\ & & + { { \lambda } ^{2 - h(2)}}\,z(0)\ , [ 1 - z(0)\ , ( 1 + { { \lambda } ^{-2\,h(2 ) } } + { { \lambda } ^{-h(2 ) } } - q(0 ) + { { \lambda } ^{- 2\,h(2)}}\,q(0)\,z(0)\ , ) ] \,x'(0)+ \nonumber\\ & & + { { \lambda } ^{2 - h(2)}}\,{{z(0)}^2\ , [ x(0 ) } + { { \lambda } ^{- 2\,h(2)}}\,x(0)\,z(0 ) ] \,q'(0 ) = \nonumber\\ & & = 2\,{{\lambda } ^{4 + h(2)}}\ , ( -1 + { { \lambda } ^{-3\,h(2 ) } } - 4\lambda^{-2\,h(2 ) } - 2\lambda^{-h(2 ) } ) \,x(0)\,z(0 ) + \nonumber\\ & & + { { \lambda } ^{4 - h(2)}}\,{{z(0)}^2 } + 6\,{{\lambda } ^{2 + 2\ , ( 1 - h(2 ) ) } } \,x(0)\ , { { z(0)}^2}- 2\,{{\lambda } ^{4 - 4\,h(2)}}\,q(0)\,x(0)\ , { { z(0)}^2}+ \nonumber\\ & & - { { \lambda } ^4}\ , ( 1 + q(0 ) ) \,x(0)\ , { { z(0)}^2 } - { { \lambda } ^{4 - 3\,h(2)}}\ , ( -7 + 2\,q(0 ) ) \,x(0)\,{{z(0)}^2}+ \nonumber\\ & & + { { \lambda } ^{4 - h(2)}}\ , ( 2 - 7\,q(0 ) + { { q(0)}^3 } ) \,x(0)\,{{z(0)}^2 } + 2\,{{\lambda } ^{4 - h(2)}}\ , ( 1 + { { \lambda } ^{-2\,h(2 ) } } ) \,x(0)\,{{z(0)}^3}+ \nonumber\\ & & - { { \lambda } ^{4 - h(2)}}\ , ( 2 + { { \lambda } ^{-4\,h(2 ) } } + 7\lambda^{3\,h(2)}+{{\lambda } ^4 } + 2\,{{\lambda } ^4}\ , ( 1 + { { \lambda } ^{-h(2 ) } } ) \,z(0)+ \nonumber\\ & & + 2\lambda^{h(2 ) } ) \,q(0)\,x(0)\ , { { z(0)}^3 } + 2\,{{\lambda } ^{4 - 2\,h(2)}}\ , ( 1 + { { \lambda } ^{-3\,h(2 ) } } ) \,{{q(0)}^3}\ , x(0)\,{{z(0)}^3}+ \nonumber\\ & & + { { \lambda } ^{4 - 6\,h(2)}}\,{{q(0)}^3}\ , x(0)\,{{z(0)}^4}\end{aligned}\ ] ] under the hypothesis that pure second neighbours interactions do not require new equations to specify the diagonal scaling for small values of @xmath15 , there are only four equations describing how global coupling are `` renormalised '' by relevant pure short range interactions . given the @xmath73-th shell one has @xmath235\ , c_{p , q , r}^{(6);1}=\nonumber\\ & & = - ( { \frac{3\,{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } + { \frac{3\,{{\lambda } ^2}}{{{\tau } _ m } } } - { \frac{45\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{45\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{m , m , m } + \nonumber\\ & & + ( { \frac{9\,{{\lambda } ^2}}{{{\tau } _ m } } } - { \frac{9\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } - { \frac{36\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{207\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{m , m,1 + m } + \nonumber\\ & & + ( { \frac{9\,{{\lambda } ^2}}{{{\tau } _ { -1 + m}}}}- { \frac{9\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{36\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{207\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{m , m,-1 + m}+\nonumber\\ & & + ( { \frac{72\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{72\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{-1 + m , m,1 + m } + { \frac{72\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } \,c^{(6)}_{1 + m,1 + m , m}+\nonumber\\ & & + { \frac{72\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } \,c^{(6)}_{-1 + m,-1 + m , m } + { \frac{9\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } \,c^{(6)}_{m , m,2 + m } + { \frac{9\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } \,c^{(6)}_{m , m,-2 + m } \nonumber\end{aligned}\ ] ] @xmath236\,c_{p , q , r}^{(6)}= \nonumber\\ & & = ( { \frac{{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } - { \frac{5\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{23\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{m , m , m } + ( { \frac{10\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{15\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{m , m,1 + m } + \nonumber\\ & & - ( { \frac{{{\lambda } ^2}}{{{\tau } _ { -2 + m } } } } + { \frac{2\,{{\lambda } ^2}}{{{\tau } _ m } } } + { \frac{7\,{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } - { \frac{17\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{40\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{169\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{m , m,-1 + m } + \nonumber\\ & & + ( { \frac{4\,{{\lambda } ^2}}{{{\tau } _ m } } } - { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } - { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{36\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{96\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{-1 + m , m,1 + m } + \nonumber\\ & & + ( { \frac{4\,{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } - { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{12\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{124\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{-1 + m,-1 + m , m } + \nonumber\\ & & + ( { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{-1 + m,-1 + m,1 + m } + { \frac{8\,\epsilon \,{{\lambda } ^4}\,}{{{\tau } _ m } } } c^{(6)}_{1 + m,1 + m,-1 + m } + \nonumber\\ & & + { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } \,c^{(6)}_{-1 + m,-1 + m,-1 + m } + ( { \frac{12\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{24\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } ) \,c^{(6)}_{-2 + m,-1 + m , m } + \nonumber\\ & & + ( { \frac{{{\lambda } ^2}}{{{\tau } _ { -2 + m}}}}- { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -3 + m } } } } - { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{6\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{17\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } ) \,c^{(6)}_{m , m,-2 + m } + \nonumber\\ & & + ( { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{-2 + m , m,1 + m } + { \frac{4\,\epsilon \lambda^4}{\tau_{1 + m } } } \,c^{(6)}_{-1 + m , m,2 + m } + { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -3 + m } } } } \,c^{(6)}_{m , m,-3 + m } \nonumber \end{aligned}\ ] ] @xmath237\,c_{p , q , r}^{(6)}= \nonumber\\ & & = { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } \,c^{(6)}_{m , m , m } + ( { \frac{4\,{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } - { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{12\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{124\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{m , m,-1 + m } + \nonumber\\ & & + ( { \frac{12\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{24\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{-1 + m , m,1 + m } + \nonumber\\ & & - ( { \frac{{{\lambda } ^2}}{{{\tau } _ m } } } + { \frac{2\,{{\lambda } ^2}}{{{\tau } _ { -2 + m } } } } + { \frac{7\,{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } - { \frac{17\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{40\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{169\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{-1 + m,-1 + m , m } + \nonumber\\ & & + ( { \frac{{{\lambda } ^2}}{{{\tau } _ m } } } - { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{6\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{17\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) \,c^{(6)}_{-1 + m,-1 + m,1 + m } + \nonumber\\ & & + ( { \frac{{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } - { \frac{5\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{23\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{-1 + m,-1 + m,-1 + m } + \nonumber\\ & & + ( { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{m , m,-2 + m } + ( { \frac{15\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{10\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{-1 + m,-1 + m,-2 + m } + \nonumber\\ & & + ( { \frac{4\,{{\lambda } ^2}}{{{\tau } _ { -2 + m } } } } - { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -3 + m } } } } - { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{36\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{96\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } ) \,c^{(6)}_{-2 + m,-1 + m , m } + \nonumber\\ & & + { \frac{\epsilon \lambda^4}{\tau_{1 + m } } } \,c^{(6)}_{-1 + m,-1 + m,2 + m } + ( { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{-2 + m,-1 + m,1 + m } + \nonumber\\ & & + { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } \,c^{(6)}_{-2 + m,-2 + m , m } + { \frac{4\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -3 + m } } } } \,c^{(6)}_{-3 + m,-1 + m , m } \nonumber\end{aligned}\ ] ] @xmath238c_{p , q , r}^{(6)}=\nonumber\\ & & = ( { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{m , m , m } + ( { \frac{{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } - { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } - { \frac{12\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{22\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{m , m,1 + m } + \nonumber\\ & & + ( { \frac{{{\lambda } ^2}}{{{\tau } _ m } } } - { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{12\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{22\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{m , m,-1 + m } + \nonumber\\ & & + ( { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{6\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{1 + m,1 + m , m } + ( { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } + { \frac{6\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{-1 + m,-1 + m , m } + \nonumber\\ & & - ( { \frac{{{\lambda } ^2}}{{{\tau } _ { -2 + m } } } } + { \frac{{{\lambda } ^2}}{{{\tau } _ { 1 + m } } } } + { \frac{4\,{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } + { \frac{4\,{{\lambda } ^2}}{{{\tau } _ m } } } - { \frac{18\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m}}}}- { \frac{18\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } - { \frac{84\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{84\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{-1 + m , m,1 + m } + \nonumber\\ & & + ( { \frac{{{\lambda } ^2}}{{{\tau } _ m } } } - { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } - { \frac{20\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{1 + m,1 + m,-1 + m } + \nonumber\\ & & + ( { \frac{{{\lambda } ^2}}{{{\tau } _ { -1 + m } } } } - { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } - { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{20\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } ) \,c^{(6)}_{-1 + m,-1 + m,1 + m } + \nonumber\\ & & + ( { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } + { \frac{6\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) \,c^{(6)}_{-1 + m,1 + m,2 + m } + ( { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{6\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } ) \,c^{(6)}_{-2 + m,-1 + m,1 + m } + \nonumber\\ & & + ( { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) \,c^{(6)}_{-1 + m,-1 + m,2 + m } + ( { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } + { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) \,c^{(6)}_{m , m,2 + m } + \nonumber\\ & & + ( { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{m , m,-2 + m } + ( { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } ) \,c^{(6)}_{1 + m,1 + m,-2 + m } + \nonumber\\ & & + ( { \frac{{{\lambda } ^2}}{{{\tau } _ { -2 + m } } } } - { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -3 + m } } } } - { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } - { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{18\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } ) \,c^{(6)}_{-2 + m , m,1 + m } + \nonumber\\ & & + ( { \frac{{{\lambda } ^2}}{{{\tau } _ { 1 + m } } } } - { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 2 + m } } } } - { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } - { \frac{3\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ m } } } - { \frac{8\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -1 + m } } } } - { \frac{18\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) \,c^{(6)}_{-1 + m , m,2 + m } + \nonumber\\ & & + ( { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -2 + m } } } } + { \frac{2\,\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 1 + m } } } } ) \,c^{(6)}_{-2 + m , m,2 + m } + { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { 2 + m } } } } \,c^{(6)}_{-1 + m , m,3 + m } + { \frac{\epsilon \,{{\lambda } ^4}}{{{\tau } _ { -3 + m } } } } \,c^{(6)}_{-3 + m , m,1 + m } \label{6closedexplicit}\end{aligned}\ ] ] the set of independent equations is finally given as : @xmath239 c_{p , q , r , s}^{(8)}=0\nonumber\\ \sum_{p , q , r , s}[i_{m , m , m , m-1;p , q , r , s}^{(8,0)}+\epsilon i_{m , m , m , m-1;p , q , r , s}^{(8;1 ) } ] c_{p , q , r , s}^{(8)}=0\nonumber\\ \sum_{p , q , r , s}[i_{m , m , m-1,m-1;p , q , r , s}^{(8,0 ) } + \epsilon i_{m , m , m-1,m-1;p , q , r , s}^{(8;1)}]c_{p , q , r , s}^{(8)}=0 \nonumber\\ \sum_{p , q , r , s}[i_{m , m , m+1,m-1;p , q , r , s}^{(8,0 ) } + \epsilon i_{m , m , m+1,m-1;p , q , r , s}^{(8;1 ) } ] c_{p , q , r , s}^{(8)}=0\nonumber\\ \sum_{p , q , r , s}[i_{m , m-1,m-1,m-1;p , q , r , s}^{(8,0)}+\epsilon i_{m , m-1,m-1,m-1;p , q , r , s}^{(8;1 ) } ] c_{p , q , r , s}^{(8)}=0 \nonumber\\ \sum_{p , q , r , s}[i_{m , m+1,m-1,m-1;p , q , r , s}^{(8,0)}+\epsilon i_{m , m+1,m-1,m-1;p , q , r , s}^{(8;1)}]c_{p , q , r , s}^{(8)}=0 \nonumber\\ \sum_{p , q , r , s}[i_{m , m+1,m+1,m-1;p , q , r , s}^{(8,0)}+\epsilon i_{m , m+1,m+1,m-1;p , q , r , s}^{(8;1)}]c_{p , q , r , s}^{(8)}=0 \nonumber\\ \sum_{p , q , r , s}[i_{m , m+1,m-1,m-2;p , q , r , s}^{(8,0)}+\epsilon i_{m , m+1,m-1,m-2;p , q , r , s}^{(8;1 ) } ] c_{p , q , r , s}^{(8)}=0 \label{8closed } \end{array } \right.\ ] ] the closure is provided again assuming scaling for all the possible conditioned expectation values with respect to a given shell . it follows @xmath240 99 u.frisch , _ turbulence : the legacy of a.n . kolmogorov _ ( cambridge university press 1995 ) r.h . kraichnan , phys . lett . * 72 * , 1016 ( 1994 ) d.bernard , k.gawedzki and a.kupiainen phys . rev . * e 54 * 2564 ( 1996 ) k. gawedzki and a. kupiainen , phys . lett . * 75 * 3834 ( 1995 ) m. chertkov , g. falkovich , i. kolokov and v. lebedev , phys . rev . * e 52 * 4924 ( 1995 ) m.chertkov and g.falkovich , phys . lett . * 76 * 2706 ( 1996 ) k.gawedzki , chao - dyn/9803027 u.frisch , a.mazzino and m.vergassola , phys . lett . * 80 * , 5532 ( 1998 ) and cond - mat/9802192 m.h . jensen , g. paladin and a. vulpiani , phys . rev.*a 45 * , n. 10 , 7214 ( 1992 ) t.bohr , m.h . jensen , g.paladin and a.vulpiani _ dynamical systems approach to turbulence _ cambridge university press ( 1998 ) a.m. obukhov , ivz . sssr , serv geogr . geofiz . * 13 * , 58 ( 1949 ) s. corrsin , j.appl . phys . * 22 * , 469 ( 1951 ) a.erdlyi _ asymptotic expansions _ ( dover publicatins inc . new york 1956 ) m.chertkov , g.falkovich and v.lebedev phys.rev.lett . * 76 * , 3707 ( 1996 ) l.arnold : _ stochastic differential equations : theory and applications _ ( wiley , new york 1974 ) a.wirth and l.biferale , phys.rev . e * 54 * , 4982 ( 1996 ) r. benzi , l.biferale and a.wirth , phys . lett . * 78 * , p. 4926 r.f.bass , _ diffusions and elliptic operators _ new york springer - verlag 1998 d.nualart , _ the malliavin calculus and related topics _ new york london springer - verlag 1995 j.cardy _ scaling and renormalization in statistical physics _ cambridge lecture notes in physics , cambridge university press ( 1996 ) j zinn - justin _ quantum field theory and critical phenomena _ clarendon press oxford ( 1989 ) l.f . weather prediction by numerical process _ cambridge university press , cambridge ( 1922 ) g. parisi and u.frisch _ turbulence and predictability in geophysical fluid dynamics _ , proceed . intern . school of physics `` e.fermi '' varenna 1983 eds m.ghil , r.benzi and g.parisi . north holland , amsterdam . l.biferale , g.boffetta , a. celani and f.toschi chao - dyn/9804035 m.le bellac _ quantum and statistical field theory _ oxford university press 1991 p.e.kloeden and e.platen _ numerical solution of stochastic differential equations _ springer - verlag 1995 k.burrage and p.burrage preprint _ high strong order methods for non commutative ordinary stochastic differential equations and the magnus formula _ http://www.maths.uq.edu.au/ kb / papers.html r.b.sidje , to appear in acm - transactions on mathematical softwere m.san miguel and r.toral in _ instabilities and nonequilibrium structures vi _ e.tirapegui and w.zeller editors kluwer acad . pub ( 1997 ) and cond - mat/9707147 r.benzi , s.ciliberto , r.tripiccione , c baudet and s.succi phys.rev e. * 48 * r29 ( 1993 )

a puzzling question in modern cosmology has been the nature and origin of the dark energy responsible for the apparent acceleration @xcite of the universe in the present epoch . the simplest explanation is , perhaps , that of a cosmological constant , or a vacuum energy in the form of a `` quintessence '' scalar field slowly evolving along an effective potential . the existence of a cosmological constant or quintessence , however , leads inevitably to the well known fine tuning problem as to why the present dark energy is so small compared to the natural scales of high - energy physics . there is also a cosmic coincidence problem as to why the universe has conspired to produce nearly equivalent energy contents in matter and dark energy at the present time . moreover , most quintessence models are now ruled out @xcite by the fact that the equation of state parameter , @xmath2 is so close to -1 . this implies that there is little evidence for evolution in a quintessence field so that a cosmological constant remains the most likely interpretation . accounting for such dark energy , however , poses an unsolved theoretical challenge . ultimately , one would hope that the existence of this dark energy could be accounted for in the context of string theory or some other unified theory . this problem is exacerbated , however , by the fact that it appears difficult to accommodate a cosmological constant in string theory @xcite , although some progress has been made @xcite . recently , it has been proposed @xcite that the observed cosmic acceleration may be an artifact of inhomogeneities in the distribution of mass - energy . that possibility , though compelling for its simplicity , has not yet been established @xcite . hence , in this paper , we consider it worthwhile to analyze an alternative mechanism by which the observed cosmic acceleration could be produced even without the need to invoke dark energy and its associated complexities . specifically , we explore models in which the cosmic acceleration is driven @xcite by the flow of dark matter from a higher dimension ( the bulk ) into our three - space ( the brane ) . this study is also in part motivated by the currently popular view that our universe could be a sub - manifold embedded in a higher - dimensional space - time . this paradigm derives from the low - energy limit of the heterotic m - theory @xcite which becomes an 11-dimensional supergravity compactified on a line segment to two 10-dimensional @xmath3 gauge theories . hence , the universe can be envisioned as two smooth 10-dimensional manifolds ( 9-branes ) embedded in a bulk dimension . a thin three - brane ( randall - sundrum brane ) embedded in a five - dimensional anti - de sitter space @xmath4 has been proposed as a practical phenomenological model @xcite in which to explore such higher dimensional physics . this approach poses an alternative to the standard kaluza - klein ( kk ) compactification of extra dimensions , through the localization of the graviton zero mode on the brane . this brane approach also provides a new way of understanding the hierarchy between the four - dimensional planck scale @xmath5 and the electro - weak scale @xcite . of relevance to the present study is that in this phenomenological approach , one can relax the requirement that the gauge groups be rigorously confined to the brane . for example , particles might be localized on a defect in the higher dimensional space @xcite . the simplest example of this would be a domain wall ( brane ) in ( 4 + 1 ) dimensions . in such a picture , the extra ( bulk ) dimension can be infinite and the observed 3-dimensional fields are represented by zero modes of bulk fields in the domain - wall background . these modes are localized and thus behave like 3-dimensional mass - less fields . in this domain - wall scenario , physical matter fields are dynamically confined to this sub - manifold , while gravity can reside in the extra bulk dimension . the stability of massive matter fields has been analyzed @xcite in this scenario where it was shown that such massive particles are metastable on the brane . that is , for both scalar and fermion fields , the quasi - normal modes are metastable states that decay into continuum states . from the viewpoint of an observer in the 4-dimensional space time , these massive particles appear to propagate for some time in three spatial dimensions and then disappear into the fifth dimension . this disappearance of massive particles from the 3-brane constitutes an energy flow from the brane to the bulk . the cosmological constraints on such disappearing matter have been studied in ref . @xcite . moreover , if the massive particles can also exist in the bulk , it becomes possible to consider the inverse flow from the bulk into our three - brane . in this paper , we build a model with such mass - energy exchange , in which the flow from the bulk to the brane provides the present observed acceleration @xcite of the universe . as noted in ref . @xcite a heavy ( @xmath6tev ) dark - matter candidate such as the lightest supersymmetric partner is likely to have the largest tunneling rate between the brane and bulk . moreover , a recent study @xcite of energy transfer from the bulk through parametric resonance indicates that massive particles are most efficient for depositing energy on the brane . hence , in the present work we mainly consider the exchange between the bulk and brane involving a growth of the cold dark - matter component on the brane . this model [ which we refer to as growing cold dark matter ( gcdm ) ] is , thus , an alternative to the standard @xmath1 plus cold dark matter ( s@xmath1cdm ) cosmology to an observer on the 3-brane . we test this model by analyzing the observations of type ia supernovae at high redshift , the temperature fluctuation spectrum of the cosmic microwave background ( cmb ) , and the matter power spectrum . surprisingly , all of these constraints can be satisfied in this model even without introducing a cosmological constant on the brane . a shocking feature of the best fit model , however , is that the present dark - matter content becomes as much as an order of magnitude larger than that of a @xmath7cdm cosmology . the influence of this excess dark matter , however , is suppressed by the large dark - radiation component associated with the flow of matter from the bulk dimension . in fact , this model provides a natural explanation for the observed suppression of the cmb power spectrum for the lowest multipoles due to the effect of the late arriving dark matter on the integrated sachs - wolf effect . we show here that the matter power spectrum exhibits the correct power on large and small scales , however , a somewhat large bias factor is required . we begin with the five - dimensional einstein equations @xmath8 where @xmath9 is the 5-dimensional gravitation constant with @xmath10 the 5-dimensional planck mass . @xmath11 is the five - dimensional total energy - momentum tensor . we use the index notation @xmath12 to clearly distinguish the bulk fifth dimension . the metric is assumed to be of the randall - sundrum ( rs ) @xcite form . for simplicity , however , we consider a coordinate system with a static bulk dimension and an expanding brane or moving @xcite brane . hence , we choose a metric in gauss - normal coordinates @xmath13 where @xmath14 is a maximally symmetric 3-metric . as an illustrative model , we allow the hubble expansion of the brane relative to a static bulk to produce a nonzero 5-component of the bulk fluid velocity as discussed below . when matter is present in the bulk , the evolution of the brane is not autonomous . it becomes necessary to explicitly define the bulk energy - momentum tensor . the general bulk energy - momentum tensor for a perfect fluid in this 5-dimensional space - time is , @xmath15 where @xmath16 and @xmath17 denote density and pressure in 4-dimensional space . in gaussian - normal coordinates , however , the bulk 5th dimension is fixed so that the four density and pressure are proportional to the three - volume like ordinary three - space quantities . for normal matter on the brane we denote the density and pressure as @xmath18 and @xmath19 . in our model in which dark matter can exist both on the brane and in the bulk we denote the dark - matter three density and pressure on the brane as @xmath20 and @xmath21 . since the density , pressure and equation of state can be different in the bulk we denote the four density of dark matter in the bulk as @xmath22 and @xmath23 . in a frame for which the three - brane is at rest we can decompose the energy - momentum tensor into three parts . @xmath24 where the @xmath25-function identifies the location of the brane at @xmath26 to which the standard - model particles are assumed to be confined . @xmath27 , corresponds to the usual three - density and pressure , @xmath18 and @xmath19 , of ordinary relativistic and non - relativistic particles plus the tension @xmath28 on the brane . @xmath29 the bulk is taken to be @xmath4 , so that the five - dimensional cosmological constant @xmath30 is negative in the bulk . hence , we write the vacuum energy - momentum for the bulk , @xmath31 , as @xmath32 in our model dark matter is allowed to exist both on the brane and in the bulk . hence , we decompose the dark - matter energy - momentum tensor , @xmath33 , into the usual three - density @xmath34 and pressure @xmath35 of dark matter on the brane , plus the bulk components @xmath36 . @xmath37 where , @xmath38 represents the matter - energy flow from the bulk to the brane , while @xmath39 represents a bulk pressure in the limit of vanishing @xmath40 . as we shall see , the cosmology for an observer living on the three - brane will depend upon the equation of state ( eos ) properties for matter in the bulk dimension . it is not guaranteed , however , that matter will have the same eos properties in the bulk as on the brane . inevitably , assumptions must be made @xcite about the form of the energy - momentum tensor for matter in the bulk . for example , a bulk field @xcite , or radiation emitted from the brane to the bulk @xcite have been considered along with possible thermal effects @xcite . in a string theory sense , particles in the bulk may appear as a topological defect between two branes . for the purposes of this paper , we adopt @xcite the usual parametrization of the eos for matter in the bulk . scaling the bulk dark - matter density to the present critical density @xmath41 on the brane , we write : @xmath42 where @xmath43 is the scale factor on the brane , defined as @xmath44 , and for a fixed bulk dimension the scaling parameter @xmath45 can be written @xmath46 , where @xmath47 is the usual equation of state parameter with , @xmath48 for normal cold dark matter , while @xmath49 for relativistic matter . a value of @xmath50 corresponds to a vacuum energy , while @xmath51 for a string - like topological defect . to illustrate this cosmology we consider two models for the motion of matter in the bulk relative to the brane . in the gaussian - normal coordinates , the five velocity of matter in the static bulk with respect to the expanding 3-brane is just @xcite @xmath52 where @xmath53^{1/2}~~,\ ] ] is the bulk curvature radius @xcite . as a second illustration we also consider the case of a constant @xmath40 component . such a model might correspond , for example , to gravitational accretion of matter from the bulk toward the brane at a constant rate . for the purposes of scaling , we parameterize the 0 - 5 component of the bulk dark - matter energy - momentum tensor as @xmath54 where we have assumed a constant transition rate of matter between the bulk and the brane absorbed into the dimensionless parameter @xmath55 which can be either positive or negative . for the case of constant @xmath40 we replace @xmath56 in the above scaling , where @xmath57 is the present hubble parameter . the cosmological equations of motion with brane - bulk energy exchange have previously been formulated in refs . the covariant derivative of the energy - momentum tensor leads to the usual energy conservation condition for various components @xmath58 of normal matter on the brane ; @xmath59 plus a new condition for the dark matter which takes into account the flow of matter from the brane world , @xmath60 where @xmath61 is the discontinuity of the ( 0,5 ) component of @xmath36 at @xmath26 . the ( 0,0 ) component of the einstein equation produces a modified friedmann cosmology to an observer on the brane : + @xmath62 while the ( 5,5 ) component leads to an equation for the dark radiation term @xmath63 , @xmath64 in eqs . ( [ a ] ) and ( [ chi ] ) the usual change variables has been introduced : @xmath65 @xmath66~~,\ ] ] and @xmath67 introducing these identities into eq . ( [ a ] ) leads to the usual friedmann equation plus extra terms with @xmath63 and @xmath68 . this modified friedmann equation can be normalized in the usual way so that the spatial curvature takes on values @xmath69 . in the limit of an empty bulk and no exchange between the bulk and the brane , @xmath70 , the quantity @xmath63 varies with scale factor on the brane as @xmath71 . thus , @xmath63 reduces to the standard dark - radiation term when all matter fields are confined to the brane . now we assume that the dark matter is cold on the brane : @xmath72 . first , we consider the special case @xmath73 . a solution of eq . ( [ rho - dm ] ) is @xmath74 . this is precisely the disappearing dark matter cosmology introduced in our previous paper @xcite . thus , @xmath73 describes matter - energy flow from the brane into the bulk . in this cosmology dark - matter particles can exist in both the brane and the bulk . hence , negative @xmath75 describes a reverse flow from the bulk to the brane . however , in the present paper , we only consider the inflow from the bulk into the brane . to see how an accelerating cosmology arises , consider the simple example of a constant flow rate from the bulk into the brane , i.e. @xmath76 for a positive constant @xmath77 . in the case of @xmath78 and @xmath79 , ( nearly realized in the present universe ) equations ( [ rho - dm])-([chi ] ) have a fixed point @xmath80 governed by , @xmath81 the constant r.h.s . of equation ( [ h * ] ) then is manifestly equivalent to a cosmological - constant dominated universe . in general , the fixed point remains even in the case of @xmath82 . thus , the present accelerating universe need not be attributed to @xmath83 at all , but could also arise from constant bulk components of the five dimensional stress - energy tensor , @xmath84 and @xmath85 . when a fluid is static , @xmath85 represents a pressure , it is then natural to set the pressure to be zero . but @xmath85 is in general non - vanishing for a moving flow . another possibility for non - vanishing @xmath85 is that the dark matter in the bulk is in a vacuum state . in this case , @xmath85 is a vacuum pressure , and it is likely that @xmath86 . we consider two growing dark matter models of interest to the present work . for the case of vanishing @xmath88 and constant @xmath40 , the evolution equations of energy densities are rewritten as , @xmath89 where @xmath90 is the dark - radiation term which plays the role of dark energy . for the case of @xmath91 , the evolution equations for the energy densities are @xmath92 = 0~,\\ & & \dot\rho + 3h(\rho+p ) = 0 ~,\\ & & \dot\rho_\chi + h\biggl[4\rho_\chi + \alpha / a^q \times \rho_{cr } \biggr ] = 0~. \label{rhochidot}\end{aligned}\ ] ] these equations satisfy the conservation s law on the brane . note , that in this latter case , the fact that @xmath93 means that eq . ( [ rhochidot ] ) quickly evolves to @xmath94 for the case when @xmath95 . indeed , as long as @xmath96 an accelerating cosmology eventually emerges @xcite . hence , the dark radiation contribution becomes constant and indistinguishable from a cosmological constant . for the @xmath97 accelerating cosmology of interest here , the modified friedmann equation for cosmic expansion can be written @xmath98 here we have neglected the @xmath99 term . this term decays rapidly as @xmath100 in the early radiation dominated epoch and hence is insignificant for the present studies . note , that eq . ( [ friedmann ] ) corresponds to the limit in which @xmath101 , and @xmath102 . figure [ fig : energy ] illustrates the evolution of various components on the brane in a simple @xmath103 cosmology . this cosmology separates into three characteristic epochs on this figure . first , the usual early radiation dominated epoch ( @xmath104 ) . second , a dark - matter dominated epoch ( @xmath105 ) . third , the dark matter and dark radiation dominated accelerating epoch ( @xmath106 ) . both of our two models for the growth rate of dark matter are illustrated on figure [ fig : energy ] for the case of @xmath107 , @xmath108 . during the first and second epochs , the dark radiation component for the case of constant @xmath40 evolves as @xmath109 and @xmath110 , respectively . for the case of brane expansion into the bulk , @xmath111 , the dark radiation term is initially much larger than for the case of constant @xmath40 and it remains at a nearly constant value . nevertheless , during the radiation- and matter - dominated epochs , the dark radiation term for both models is insignificant until recent history near @xmath112 . hence , this cosmology is not constrained by primordial nucleosynthesis . @xcite this model thus explains the smallness problem of the apparent cosmological constant as simply due to the slow tunneling of matter from the bulk onto the brane . the cosmic acceleration only occurs now , because this is the epoch for which the fixed - point solution could be obtained . moreover , in what follows we show that the fits to observations based upon these two models are indistinguishable from each other and from the best standard @xmath7cdm model . eventually , in the third region the cosmology of interest to this paper emerges . the dark radiation component dominates . in the case of @xmath113 it becomes a constant dark - energy density . hence , the dark radiation associated with in - flowing cold dark matter leads the cosmic acceleration without the need for a cosmological constant on the right hand side of eq . ( [ friedmann ] ) as long as matter in the bulk has the right equation of state . having defined the cosmology of interest we now analyze the various cosmological constraints as a test of this hypothesis by solving equations ( 24)-(30 ) numerically . in particular , we examine the magnitude - redshift relation for type ia supernovae ( snia ) , the cosmic microwave background ( cmb ) and the matter power spectrum @xmath114 . parameters summarizing the best fits to these constraints are given in table 1 for models with and without growing cold dark matter . the s@xmath1cdm fits are consistent with the usually inferred cosmological parameters ( e.g. @xcite . an important constraint on the gcdm model is that the sum of @xmath115 be approximately equivalent to the sum of @xmath116 in the standard cosmology . this is evidenced by the fifth column of table 1 . note , however , that the magnitude of @xmath117 and @xmath118 individually can be quite large in the gcdm model , as long as @xmath119 . the apparent brightness of the type ia supernova standard candle with redshift is given @xcite by a simple relation which we slightly modify to incorporate the brane - world cosmology given in eq . ( [ friedmann ] ) . the luminosity distance becomes , @xmath120^{-1/2 } \biggr\}~~,\end{aligned}\ ] ] where @xmath57 is the present value of the hubble constant , and @xmath121 for @xmath122 , @xmath123 , for @xmath124 and @xmath125 for @xmath126 . the @xmath127 are the energy densities normalized by the current critical density , i.e. @xmath128 . note the @xmath117 has a nontrivial redshift dependence in the present cosmology via eq . ( [ rho - dm ] ) . the look - back time @xmath129 is a function of redshift becomes , @xmath130^{-1/2 } dz ' \biggr\}~~.\end{aligned}\ ] ] we have found the best fits to the supernova magnitude - redshift relation by maximizing the likelihood functions in an effective @xmath131 analysis , @xmath132 where the second term corresponds to and analytic marginalization over the absolute magnitude of the snia data with a flat prior @xcite . figure [ fig : mag ] compares various cosmological models with some of the recent combined data from the high - z supernova search team and the supernova cosmology project @xcite . the lower figure highlights the crucial data points at the highest redshift which constrain the redshift evolution during the dark - matter dominated decelerating phase . shown are the k - corrected magnitudes @xmath133 vs. redshift . curves are plotted relative to an open @xmath134 , @xmath135 cosmology . the best - fit growing cold dark matter ( gcdm ) models shown on this figure correspond to either the constant @xmath40 or @xmath136 models . they are indistinguishable from each other on this plot . of particular interest on figure [ fig : mag ] is the fact that our best fit @xmath0 growing cold dark matter ( gcdm ) models are nearly indistinguishable from the best fit standard @xmath137cold dark matter ( s@xmath1cdm ) model . thus , our model realizes the present cosmic acceleration without a cosmological constant . this figure also illustrates an important point regarding the eos parameter for matter in the bulk dimension . an accelerating cosmology requires that @xmath138 be nearly constant . this means that no matter what the eos for matter when it is on the brane , a constant @xmath138 requires that @xmath139 be small in eqs . ( [ rho1 ] ) to ( [ rhochidot ] ) for matter in the bulk . this is illustrated by the @xmath140 curve of figure [ fig : mag ] which is unable to reproduce the required cosmic acceleration . for the snia constraint alone , the best fit curve from table 1 is for @xmath141 . the high - resolution measurement of the spectrum of temperature fluctuations in the cosmic microwave background ( cmb ) by the wilkinson microwave anisotropy probe ( wmap ) has become one of the most stringent tests for cosmology . the standard procedure for obtaining cosmological parameters @xcite is based upon bayesian statistics . this approach gives the posterior probability distribution from which the optimal set of the cosmological parameters and their confidence levels can be deduced . in our analysis we followed the markov chain monte carlo ( mcmc ) approach @xcite and explored the likelihood in an eight dimensional parameter space consisting of six wmap standard parameters ( @xmath142 , @xmath143 , @xmath144 , @xmath145 , @xmath146 , @xmath147 ) plus the two brane - world parameters , @xmath55 and @xmath45 . figure [ fig : cls ] shows best - fit theoretical cmb power spectra along with the combined data set used to constrain our models . this figure illustrates how the flow of mass - energy exchange can modify the spectrum . there are essentially two ways in which growing cold dark matter models alter the cmb power spectrum . first , gcdm means that there is less dark matter at earlier times . this leads to a smaller amplitude of the third acoustic peak @xcite . second , the decay of the gravitational potential at late times is diminished due to the inflow of dark matter . this leads to a smaller late integrated sachs - wolfe effect ( lisw ) and correspondingly less power for the smallest multipoles . the integrated sachs - wolf effect accounts for changes on the cmb anisotropy due the time evolution of the gravitational potentials as the photons travel from the last scattering surface to us . writing the comoving metric perturbations as @xmath148\ ] ] for a photon in the absence of anisotropic stress @xmath149 and the isw effect on the temperature power spectrum depends upon a simple integral over @xmath150 @xcite . include anisotropic stress we have : @xmath151 d\tau ~~,\ ] ] where @xmath152 along the photon trajectory , and @xmath153 . figure [ swolf ] illustrates the evolution of the @xmath154 and @xmath155 potential derivatives with scale factor . in the @xmath7cdm model the gravitational potentials are rapidly decaying at late times which causes enhanced power on the largest scales ( smallest multipoles ) . in the gcdm model , however , the potentials on the largest scale are changing less . this accounts for the relative suppression of cmb power for the lowest multipoles . for the combined snia and cmb data we again used the mcmc approach in a seven dimensional parameter space ( @xmath142 , @xmath156 , @xmath157 , @xmath146 , @xmath147 , @xmath55 , @xmath45 ) . it should be noted that in the gcdm cosmology of interest here we must treat @xmath158 and @xmath117 as independent parameters in the snia analysis contrary to the usual cosmologies . this is because the evolution of the energy densities @xmath159 and @xmath160 obey different functions of the cosmic scale factor @xmath161 in the gcdm cosmology . these data imply a minimum in @xmath131 of 1659.8 for the gcdm model without a cosmological constant , compared to a value of 1661.7 obtained in the s@xmath1cdm model . note , however , that in our model there are seven parameters ( @xmath142 , @xmath156 , @xmath157 , @xmath146 , @xmath147 , @xmath55 , @xmath45 ) , while in the s@xmath1cdm model there are only six ( @xmath142 , @xmath156 , @xmath157 , @xmath146 , @xmath147 , @xmath162 ) . hence , there is no significant improvement in the reduced @xmath131 per degree of freedom as evidenced in table 1 . although the gcdm model explains the apparent suppression of the cmb at low multipoles , the optical depth is rather large for the optimum fit to the cmb alone ( @xmath163 ) . this is due to a degeneracy among parameters so that a slightly better fit is obtained for the combination of a large @xmath28 offset by a smaller @xmath144 and @xmath117 . in the combined fit with the snia data , however , @xmath144 is better constrained so that a smaller value of @xmath164 results . note also , that in all of these fits a large value of @xmath165 is obtained . this value is offset , however , by the negative dark radiation component as discussed below . the key constraint is that @xmath166 as evidenced in the fifth column of table 1 . .2 in it is straight forward to determine the galactic matter power spectrum @xmath114 to compare with that deduced from large - scale structure surveys @xcite . for the usual power - law spectrum of primordial fluctuations @xmath167 a transfer function , @xmath168 @xcite is required to convert from the amplitude of the perturbation as wave number @xmath169 enters the horizon , to the present - day power spectrum , @xmath114 . transfer functions are easily computed using the code cmbfast @xcite for various sets of cosmological parameters . an adequate approximate expression for the structure power spectrum is @xmath170 this expression is only valid in the linear regime , which in comoving wave number is up to approximately @xmath171 mpc@xmath172 and therefore adequate for our purposes . however , we also correct for the nonlinear evolution of the power spectrum @xcite . here it is worth noting the way in which the power spectrum is computed in our gcdm model . it is necessary to specify the distribution of the dark matter as it enters from the bulk dimension . it would be most general to suppose that there are fluctuations in the bulk mass - energy distribution just as there are in the brane . material entering from the bulk would then appear with such fluctuations . however , we make the simplifying assumption that the dark matter and dark radiation enter with uniform distributions . after that , they are allowed to evolve as non - relativistic and relativistic particles , respectively , along with the already evolved structures on the brane . hence , we define the matter over density as @xmath173 where @xmath174 is the effective dark matter density as would be deduced , for example , by the gravitational potential of a galaxy cluster . for the gcdm model the effective over density is @xmath175 and the effective gravitation mass energy density is @xmath176 as usual , a bias parameter is introduced to account for the difference between the galaxy and matter power spectrum , @xmath177 we have made a mcmc simultaneous fit to the cmb+snia+@xmath114 data . we get the best fit parameters , @xmath178 and @xmath179 . figure [ fig : pk ] shows a comparison of the observed power spectrum with that of a s@xmath1cdm model and also the gcdm model which best fits the cmb+snia+@xmath114 data . the power spectrum derived in the best fit growing dark matter model is almost indistinguishable from a s@xmath1cdm model until one gets to the very largest structures for which there is no data . the parameters associated with this model are summarized in table 1 . in these fits the bias @xmath180 is a marginalized parameter . it is perhaps of note that the bias parameter deduced in this way is somewhat larger @xmath181 than that deduced in the usual s@xmath1cdm models , @xmath182 . this derives from the fact that the dark matter potentials are not as deep at early times , and hence , galaxy formation must be more efficient to produce the observed amplitude . we have considered models in which the apparent cosmological constant derives from the energy exchange of cold dark matter from the bulk dimension to the brane . the energy momentum tensor in this extra dimensional brane cosmology leads naturally to terms resembling a cosmological constant at the present time . if such energy exchange occurs our universe accelerates without the need to invoke a cosmological constant on the 3-brane . we find that such gcdm exchanges are consistent with observations including the supernova magnitude - redshift relation , temperature fluctuations in the cmb , and the matter power - spectrum data . this cosmology is even slightly preferred as it fits better the suppression of the cmb power spectrum at low multipoles . we have thus demonstrated that this cosmology represents an alternative model to the s@xmath1cdm cosmology for an observer on the 3-brane . a consistent fit to the observational constraints , however , requires that the eos parameter for matter in the bulk be small @xmath183 . this eos is consistent with a need for an ads@xmath184 geometry in the bulk . a peculiar feature of the present best fit models , is the fact that the true value of @xmath117 is much larger than in the standard cosmology , though its gravitational effect is canceled by the dark - radiation contribution . indeed , the key constraint is that @xmath166 as evidenced in the fifth column of table 1 . one consequence of such a large and growing dark - matter contribution is the need for a somewhat large bias parameter . on the other hand , such a large dark matter content suggests new observational tests of this cosmology . for example , if the dark - matter content at the present time is much larger than in a s@xmath1cdm model , then direct terrestrial measurements of the total density of cold dark - matter particles could indicate a much higher density than expected based upon their mass and gravitation effect . another test of this cosmology is that there should be a suppression of the matter power spectrum on the scale of the horizon compared to a s@xmath1cdm cosmology . other tests of this paradigm are that a suppression of the third acoustic peak in the cmb power spectrum occurs so that better data near @xmath185 could help distinguish this cosmology . we also note that this cosmology produces large oscillations in the cmb polarization power spectrum so that when better polarization data are available it should help to eliminate of confirm this cosmology . an amusing feature of this model stems from the requirement that the present dark radiation from in - flowing matter must cancel the effects of excess existing dark matter . hence , if the flow were to cease , the universe would change to a matter - dominated @xmath186 cosmology which would begin to collapse in about a hubble time . in that sense , this is a new kind of doomsday cosmology @xcite . this work has been supported in part by the mitsubishi foundation , the grants - in - aid for scientific research ( 13640313 , 14540271 ) and for specially promoted research ( 13002001 ) of the ministry of education , science , sports and culture of japan . k.i.s work has been supported by a grant - in - aid for jsps . work at the university of notre dame supported by the u.s . department of energy under nuclear theory grant de - fg02 - 95-er40934 . t. banks , m. dine and l. motl , jhep , * 1 * , 31 ( 2001 ) ; s. hellerman , n. kaloper , and l. susskind , jhep , 06 , 003 ( 2001 ) ; w. fischler et al . , jhep , 07 , 003 ( 2001 ) ; t. banks and m. dine , jhep , * 10 * , 012 ( 2001 ) . federico and t. shinji , j. cosm . phys , * 07 * , 004 ( 2004 ) ; v. sahni and y. shtanoc , j. cosm . astrop . phys , * 11 * , 014 ( 2003 ) ; m. bastero - gil and l. mersini , phys . rev . d 67 , 103519 ( 2003 ) ; a. chamblin and n. d. lambert , phys . rev . * d65 * , 066002 ( 2002 ) ; s. nojiri and s. d. odintsov , phys . lett . * b576 * , 5 ( 2003 ) . f. a. brito , f. f. cruz , j.f.n . oliveira , phys . rev . * d71 * , 083516 ( 2005 ) . e. w. kolb , s. matarrese , n. alessio , and a. riotto , hep - th/0503117 e. w. kolb , s. matarrese , and a. riotto , astro - ph/0506534 g. geshnizjani and d. j.h . chung , astro - ph/0503553 c. m. hirata and u. seljak , astro - ph/0503582 e. kiritsis , g. kofinas , n. tetradis , t. n. tomaras and v. zarikas , jhep , * 02 * , 035 ( 2003 ) . n. tetradis , phys . lett . , * b569 * , 1 ( 2003 ) . p. s. apostolopoulos and n. tetradis , class . * 21 * , 4781 ( 2004 ) . y.s . myung , j.y . kim , , class . , * 20 * , l169 ( 2003 ) . n. tetradis , class . grav . , * 21 * , 5221 ( 2004 ) . p. s. apostolopoulos and n. tetradis , phys . , * d71 * , 043506 ( 2005 ) . horava & e. witten , npb , 460 , 506 ( 1996 ) . l. randall and r. sundrum , phys . lett . * 83 * , 3370 ( 1999 ) ; * 83 * , 4690 ( 1999 ) . n. arkani - hamed , s. dimopoulos , n. kaloper and r. sundrum , phys . b480 * 193 , ( 2000 ) . s. kachru , m. schulz and e. silverstein , phys . * d62 * , 045021 , ( 2000 ) . v. a. rubakov and m. e. shaposhnikov , phys . , * b125 * , 136 ( 1983 ) . k. akama , lect . notes phys . * 176 * , 267 ( 1982 ) [ arxiv : hep - th/0001113 ] . s. l. dubovsky , v. a. rubakov and p. g. tinyakov , phys . d 62 , 105011 ( 2000 ) . k. ichiki , p. m. garnavich , t. kajino , g. j. mathews , and m. yahiro , phys . rev . d 68 , 083518 ( 2003 ) . diakonos , e.n . saridakis and n. tetradis , class . lett . , * b605 * , 1 ( 2005 ) . p. krauss , jhep , * 12 * , 011 ( 1999 ) . c. van se bruck , m. dorca , c.j.a . martins and m. parry , phys . b495 * , 183 ( 2000 ) . a. chamblin and h.s . reall , nucl . phys . , * b562 * , 133 ( 1999 ) ; k. maeda and d. wands , phys . , * d62 * , 124009 ( 2000 ) ; c. barcelo and m. visser , phys . , * d63 * , 024004 ( 2001 ) ; a. menim and r.a . battye , class . quantum grav . , * 18 * , 2171 ( 2001 ) . d. langlois , l. sorbo , and m. rodriguez - martinez , phys . lett . , * 89 * , 171301 ( 2002 ) ; e. leeper , r. maartens and c.f . sopuerta , class . quantum grav . , * 21 * , 1125 ( 2004 ) . a. chamblin , a. karch , and a. nayeri , phys . * b509 * , 163 ( 2001 ) ; i.z . rothstein , phys . , * d64 * , 084024 ( 2001 ) s. mukohyama , t. shiromizu and k. i. maeda , phys . d * 62 * , 024028 ( 2000 ) k. ichiki , m. yahiro , t. kajino , m. orito and g. j. mathews , phys . d * 66 * , 043521 ( 2002 ) [ arxiv : astro - ph/0203272 ] . s. m. carroll , w. h. press , and e. l. turner , ann . , * 30 * , 499 ( 1992 ) . s. l. bridle , r. crittenden , a. melchiorri , m. p. hobson , r. kneissl and a. n. lasenby , mon . not . soc . * 335 * , 1193 ( 2002 ) [ arxiv : astro - ph/0112114 ] . a. g. riess _ et al . _ [ supernova search team collaboration ] , astron . j. * 116 * , 1009 ( 1998 ) [ arxiv : astro - ph/9805201 ] ; a. g. riess _ [ supernova search team collaboration ] , astrophys . j. * 607 * , 665 ( 2004 ) [ arxiv : astro - ph/0402512 ] ; l. verde , et al . ( _ wmap collaboration _ , astrophys . j. suppl . , * 148 * , 195 ( 2003 ) . n. christensen , r. meyer , l. knox and b. luey , class . and quantum grav . , * 18 * , 2677(2001 ) ; a. lewis and s. bridle , phys . rev . * d66 * , 103511 ( 2002 ) w. hu , d. scott , n. sugiyama , and m. white , phys . , * d52 * , 5498 ( 1995 ) . a. r. liddle and d. h. lyth in _ cosmological inflation and large - scale structure _ , ( cambridge u. press , cambridge ) ( 2000 ) . s. dodelson , et al . ( sdss collaboration ) , astrophys . j. , * 572 * , 140 ( 2002 ) ; m. tegmark , a. j. s. hamilton , and y. xu , mnras , * 335 * , 887 ( 2002 ) . w. percival , et al . ( 2df collaboration ) , mnras , * 328 * , 1039 ( 2001 ) . g. p. efstathiou , in _ physics of the early universe _ , ( sussp publications , edinburgh , 1990 ) , eds . a. t. davies , a. heavens , and j. peacock . u. seljak , and m. zaldarriago , astrophys . j. , * 469 * 437 ( 1996 ) . j. a. peacock and s. j. dodds , mon . not . * 280 * , l19 ( 1996 ) [ arxiv : astro - ph/9603031 ] . r. kallosh , j. kratochvil , a. linde , e. v. linder , and m. shmakova , j. cosmology and astroparticle physics , * 10 * , 015 ( 2003 ) 155mmlccccccccccccc fit gcdm & @xmath55 & q & @xmath142 & @xmath115 & @xmath117 & @xmath118 & h & @xmath187 & @xmath146 & @xmath28 & @xmath180 & @xmath188 + fit s@xmath1cdm & & & & @xmath116 & & @xmath189 + snia only & & & & & & & & & & + best fit gcdm & 11.0 & 0.006 & 0.022 & 0.93 & 3.31 & -2.38 & 0.58 & - & - & - & - & 1.23 + best fit s@xmath1cdm & - & - & 0.022 & 0.97 & 0.26 & 0.71 & 0.71 & - & - & - & - & 1.24 + cmb only + best fit gcdm & 2.14 & 2.92 & 0.029 & 0.93 & 1.91 & -0.98 & 0.64 & 29.1 & 1.18 & 0.533 & - & 1.02 + best fit s@xmath1cdm & - & - & 0.023 & 0.94 & 0.23 & 0.71 & 0.71 & 14.9 & 0.97 & 0.13 & - & 1.01 + snia + cmb + best fit gcdm & 8.45 & 0.023 & 0.023 & 0.95 & 3.14 & -2.19 & 0.71 & 15.0 & 0.97 & 0.133 & - & 1.04 + best fit s@xmath1cdm & - & - & 0.023 & 0.96 & 0.25 & 0.71 & 0.70 & 13.3 & 0.96 & 0.111 & - & 1.04 + snia + cmb + p(k ) + best fit gcdm & 8.33 & 0.037 & 0.024 & 0.95 & 3.05 & -2.44 & 0.71 & 15.3 & 0.98 & 0.140 & 2.1 & 1.03 + best fit s@xmath1cdm & - & - & 0.023 & 0.95 & 0.24 & 0.71 & 0.70 & 13.7 & 0.97 & 0.117 & 1.05 & 1.03 +

since its discovery in 1974 @xcite , the charmonium system has become the prototypical ` hydrogen atom ' of meson spectroscopy @xcite . the experimentally clear spectrum of relatively narrow states below the open - charm dd threshold of 3.73 gev can be identified with the 1s , 1p and 2s levels predicted by potential models , which incorporate a color coulomb term at short distances and a linear scalar confining term at large distances . spin - dependent interquark forces are evident in the splittings of states within these multiplets , and the observed splittings are consistent with the predictions of a one gluon exchange ( oge ) breit - fermi hamiltonian , combined with a linear scalar confining interaction . discussions of the theoretical importance and experimental status of heavy quarkonium , including recent experimental results for charmonium , have been given by quigg @xcite , galik @xcite , the cern quarkonium working group @xcite , seth @xcite and swarnicki @xcite . recently there has been a resurgence of interest in charmonium , due to the realization that b factories can contribute to the study of the missing states @xcite , and to high - statistics experiments at bes @xcite and cleo @xcite and the planned gsi @xmath3 facility @xcite . the possibility of contributions from b factories was dramatically illustrated by the recent discovery of the long missing @xmath4 @xmath5 state by the belle collaboration @xcite , which has since been confirmed by babar @xcite , and has also been observed by cleo in @xmath6 collisions @xcite . additional interest in spectroscopy has followed the discovery of the remarkable x@xmath7 by belle @xcite and cdf @xcite in b decays to @xmath8 ; assuming that this is a real resonance rather than a threshold effect , the x@xmath7 is presumably either a dd@xmath9 charmed meson molecule @xcite or a narrow @xmath10 d - wave state @xcite . very recent observations of the x@xmath7 in @xmath11 and @xmath12 by belle support a @xmath13 dd@xmath9 molecule assignment @xcite . there has also been experimental activity in the spin - singlet p - wave sector , with recent reports of the observation of the elusive @xmath14 @xmath15 state by cleo @xcite . finally , the surprisingly large cross sections for double charmonium production in @xmath16 reported by belle @xcite suggest that it may be possible to study c @xmath17 @xmath0 states in @xmath16 without using the higher - order @xmath18 two - photon annihilation process . one open topic of great current interest in spectroscopy is the search for the @xmath19 and @xmath20 states , which are expected to be quite narrow due to the absence of open - charm decay modes . a second topic is the lorentz nature of confinement ; in pure models this is tested by the multiplet splittings of orbitally - excited states . for example , with pure scalar confinement as is normally assumed there is no spin - spin hyperfine interaction at o@xmath21 , so the masses of spin - singlets ( such as the @xmath22 @xmath15 ) are degenerate with the corresponding triplet c.o.g . ( center of gravity ) ; here this is the @xmath1p@xmath23 c.o.g . , at 3525 mev . in the original cornell model @xcite it was assumed that confinement acts as the time component of a lorentz vector , which lifts the degeneracy of the @xmath15 and the @xmath1p@xmath23 c.o.g . another possibility is that confinement may be a more complicated mix of scalar and timelike vector @xcite . of course these simple potential model considerations may be complicated by mass shifts due to other effects , such as couplings to open - flavor channels @xcite . a third topic is the search for exotica such as hybrids ; the level of mixing between conventional quarkonium and hybrid basis states falls rapidly with increasing quark mass , which suggests that nonexotic hybrids may be more easily distinguished from conventional quarkonia in charmonium than in the light quark sectors . since lgt ( lattice gauge theory ) predicts that the lightest hybrids lie near 4.4 gev @xcite , there is a strong incentive to establish the background " spectrum of conventional states up to and somewhat beyond this mass . a final topic of current interest is the importance of mixing between quark model @xmath24 basis states and two - meson continua , which has been cited as a possible reason for the low masses of the recently discovered d@xmath25 states @xcite . the effects of unquenching the quark model " by including meson loops can presumably be studied effectively in the system , in which the experimental spectrum of states is relatively unambiguous . the success of the @xmath24 quark model is surprising , in view of the probable importance of corrections to the valence approximation ; the range of validity of the naive quenched " @xmath24 quark model is an interesting and open question @xcite . motivated by this revived interest in spectroscopy , we have carried out a theoretical study of the expected properties of charmonium states , notably the poorly understood higher - mass levels above dd threshold . two variants of potential models are used in this study , a conventional nonrelativistic model based on the schrdinger equation with a coulomb plus linear potential , and the godfrey - isgur relativized potential model . we give results for all states in the multiplets 1 - 4s , 1 - 3p , 1 - 2d , 1 - 2f and 1 g , comprising 40 resonances in total . predictions are given for quantities which are likely to be of greatest experimental interest , which are the spectrum of states , e1 ( and some m1 ) electromagnetic transition rates , and strong partial and total widths of states above open - charm threshold . similar results for many of the electromagnetic transition rates have recently been reported by ebert _ the @xmath26 leptonic and two - photon widths are not discussed in detail here , as they have been considered extensively elsewhere ; see for example @xcite and @xcite and references cited therein . as a minimal model of the charmonium system we use a nonrelativistic potential model , with wavefunctions determined by the schrdinger equation with a conventional quarkonium potential . we use the standard color coulomb plus linear scalar form , and also include a gaussian - smeared contact hyperfine interaction in the zeroth - order potential . the central potential is v_0^(c|c)(r ) = - + br + _ ( r ) _ c _ |c where @xmath27 . the four parameters @xmath28 are determined by fitting the spectrum . the spin - spin contact hyperfine interaction is one of the spin - dependent terms predicted by one gluon exchange ( oge ) forces . the contact form @xmath29 is actually an artifact of an o@xmath30 expansion of the t - matrix @xcite , so replacing it by an interaction with a range @xmath31 comparable to @xmath32 is not an unwarranted modification . we treat the remaining spin - dependent terms as mass shifts using leading - order perturbation theory . these are the oge spin - orbit and tensor and a longer - ranged inverted spin - orbit term , which arises from the assumed lorentz scalar confinement . these are explicitly v_spin - dep = . the spin - orbit operator is diagonal in a @xmath33 basis , with the matrix elements @xmath34/2 $ ] . the tensor operator t has nonvanishing diagonal matrix elements only between @xmath35 spin - triplet states , which are ^3l _ | | ^3l _ = . for experimental input we use the masses of the 11 reasonably well established @xmath0 states , which are given in table [ table_spectrum ] ( rounded to 1 mev ) . the parameters that follow from fitting these masses are @xmath36 . given these values , we can predict the masses and matrix elements of the currently unknown states ; table [ table_spectrum ] and fig . [ spectrum ] show the predicted spectrum . the godfrey - isgur model is a `` relativized '' extension of the nonrelativistic model of the previous section . this model assumes a relativistic dispersion relation for the quark kinetic energy , a qcd - motivated running coupling @xmath37 , a flavor - dependent potential smearing parameter @xmath38 , and replaces factors of quark mass with quark kinetic energy . details of the model and the method of solution may be found in ref.@xcite . the hamiltonian consists of a relativistic kinetic term and a generalized quark - antiquark potential , = h_0 + _ q|q ( , ) where _ 0 = + . just as in the nonrelativistic model , the quark - antiquark potential v@xmath39 assumed here incorporates the lorentz vector one gluon exchange interaction at short distances and a lorentz scalar linear confining interaction . to first order in @xmath40 , v@xmath41 reduces to the standard non - relativistic result given by eqns.(1 ) and ( 2 ) ( with @xmath42 replaced by a running coupling constant , @xmath37 ) . the full set of model parameters is given in ref.@xcite . note that the string tension and quark mass ( @xmath43 gev@xmath44 and @xmath45 gev ) are significantly larger than the values used in our nonrelativistic model . one important aspect of this model is that it gives reasonably accurate results for the spectrum and matrix elements of quarkonia of all @xmath46 quark flavors , whereas the nonrelativistic model of the previous section is only fitted to the system . the spectra predicted by the nr and gi models ( table [ table_spectrum ] and fig . [ spectrum ] ) are quite similar for s- and p - wave states , largely because of the constraints provided by the experimental candidates for these multiplets . we note in passing that these potential model results are very similar to the most recent predictions of the charmonium spectrum from lgt @xcite . at higher l we have only the l=2 @xmath47d@xmath48 and @xmath49d@xmath48 states @xmath50 and @xmath51 to constrain the models , and the predicted mean d - wave multiplet masses differ by _ ca . _ 50 mev . for @xmath52 the absence of experimental states allows a relatively large scatter of predicted mean masses , which differ by as much as @xmath53 mev in the 1 g multiplet . ( the splittings within higher l multiplets in contrast are rather similar . ) the mean multiplet masses predicted by the two models differ largely because of the values assumed for the string tension @xmath54 , which is @xmath55 gev@xmath44 in the gi model but is a rather smaller @xmath56 gev@xmath44 in the nr model . identification of any l=3 or l=4 state would be very useful as a constraint on the spectrum of higher - l states generally . this is unfortunately a difficult problem , since these states are not easily accessible experimentally . as we shall see , one possibility is to produce the @xmath1f@xmath57 @xmath58 state , which is formed in e1 radiative transitions from the @xmath51 and decays dominantly to dd and dd@xmath9 . ( for simplicity in this paper we abbreviate the final state d@xmath59 as dd " , the state d@xmath60d@xmath9 as dd@xmath9 " , and so forth . ) 1.0 cm ) . the solid lines are experiment , and the broken lines are theory ( nr model left , gi right ) . spin triplet levels are dashed , and spin singlets are dotted . the dd open - charm threshold at 3.73 gev is also shown.,width=321 ] radiative transitions of higher - mass charmonium states are of interest largely because they provide one of the few pathways between states with different quantum numbers . since typical e1 radiative partial widths of charmonia are 10s to 100s of kev , corresponding to significant branching fractions of @xmath61 to @xmath62 , large event samples of radially excited s - wave states produced in @xmath16 annihilation could be used to identify radially excited p - wave states , which are not otherwise easily produced . similarly , the e1 radiative transition of the nominally 2@xmath1d@xmath48 @xmath51 can be used to produce an f - wave state ; this multiplet would likely be very difficult to reach using other mechanisms . we evaluate these e1 radiative partial widths using _ e1 ( n ^2+1l _ ^2+1l_ + ) = c_fi _ e_c^2 |_f | r | _ i |^2 e_^3 , where @xmath63 is the @xmath64-quark charge in units of @xmath65 , @xmath66 is the fine - structure constant , e@xmath67 is the final photon energy , @xmath68 is the total energy of the final state , @xmath69 is the mass of the initial state , the spatial matrix element @xmath70 involves the initial and final radial wavefunctions , and the angular matrix element @xmath71 is c_fi=(, ) ( 2 + 1 ) \ { 1 } ^2 . this is the result of ref.@xcite , except for our inclusion of the relativistic phase space factor of @xmath72 , which is usually not far from unity . ( the gi results do not include this phase space factor . ) we evaluate these radiative partial widths in both the nr and gi models . for the nr model the matrix elements latexmath:[$\langle { n'}^{2{\s}'+1}{\l}'_{{\j } ' } |\ ; r \ ; linear plus smeared hyperfine wavefunctions of the potential model described in sec.iia , and for the gi model they were evaluated using the wavefunctions of ref.@xcite . since the masses predicted for unknown states differ in the two models , the assumed photon energy e@xmath67 differs as well ; these photon energies are given in the e1 and m1 transition tables ( tables [ e1rad_stop]-[m1rad ] ) together with the radiative partial widths . some e1 transitions that are of special importance for the study of higher charmonium states are discussed in the text . transitions from initial @xmath74 states are of greatest interest in this regard , since these can be studied with high statistics at @xmath16 machines . these can provide access to the spin - triplet members of the 2p and 3p multiplets in particular , starting from the @xmath75 and @xmath76 . e1 radiative transitions may also be useful in identifying the narrow 1d @xmath77 and @xmath78 states , since they are all predicted to have large partial widths ( _ ca . _ 300 kev ) to the 1p @xmath79 and @xmath15 states . although m1 rates are typically rather weaker than e1 rates , they are nonetheless interesting because they may allow access to spin - singlet states that are very difficult to produce otherwise . it is also interesting that the known m1 rates show serious disagreement between theory and experiment . this is in part due to the fact that m1 transitions between different spatial multiplets , such as @xmath80 ( 2s @xmath81 1s ) , are nonzero only due to small relativistic corrections to a vanishing lowest - order m1 matrix element . the m1 radiative partial widths are evaluated using _ m1 ( n ^2+1l _ ^2+1l_ + ) = 2j+12l+1 _ _ , 1 e_c^2 |_f | _ i ( see the previous e1 formula for definitions . ) the gi m1 radiative rates do not incorporate the @xmath72 phase space factor , but do include a @xmath82 recoil factor . we quote nr results both with and without this recoil factor . the photon energies e@xmath67 depend on the model in most cases , since we have assumed theoretical masses for unknown states . as with the e1 rates , we quote m1 radiative widths in both the nr potential model of sec.iia and the gi model @xcite of sec.iib . although the m1 decay rates involve a relatively simple off - diagonal matrix element of the magnetic moment operator , we note that our predicted rates show considerable variation with the model assumptions , and agreement with the two known m1 rates is not good . in the off - diagonal ( 2s @xmath81 1s ) transition @xmath80 this is because the zeroth - order m1 matrix element vanishes due to orthogonality of the spatial wavefunctions , and the nonzero predicted rate results from rather model - dependent corrections to the wavefunctions and recoil effects . evidently the corrections we have included do not accurately predict the observed partial width ; for this reason the other predicted m1 rates between different spatial multiplets are suspect , and in any case are evidently strongly dependent on recoil factors . better experimental data will be very important for improving our description of these apparently simple but evidently poorly understood m1 radiative transitions . -0.2 cm -0.2 cm the dominant strong decays ( when allowed by phase space ) are transitions to open - flavor final states . in these open - flavor decays the initial meson decays through production of a light @xmath24 quark - antiquark pair @xmath83 , followed by separation into two open - charm mesons . remarkably , the qcd mechanism underlying this dominant decay process is still poorly understood . in quark model calculations this decay process is modelled by a simple phenomenological pair production amplitude . the pair is usually assumed to be produced with vacuum @xmath84 quantum numbers , and variants of the decay model make different assumptions regarding the spatial dependence of the pair production amplitude relative to the initial pair . the simplest of these models is the @xmath1p@xmath2 model , originally introduced by micu @xcite , which assumes that the new @xmath24 pair is produced with vacuum quantum numbers ( @xmath1p@xmath2 ) by a spatially constant pair - production amplitude @xmath85 . leyaouanc _ et al . _ subsequently applied the @xmath1p@xmath2 model to meson @xcite and baryon @xcite open flavor strong decays in a series of publications in the 1970s . they also evaluated strong decay partial widths of the three states @xmath50 , @xmath75 and @xmath76 in the @xmath1p@xmath2 model @xcite ; the relation between this early work and our contribution will be discussed . this decay model , which has since been applied extensively to the decays of light mesons and baryons , was originally adopted largely because of the success in predicting the d / s amplitude ratio in the decay @xmath86 @xcite . another stringent test of strong decay models is provided by the very tight limit on the spin singlet to spin singlets transition @xmath87 from the ves collaboration @xcite , @xmath88 this branching fraction is predicted to be zero in the @xmath1p@xmath2 model , but would not necessarily be negligible in a different decay model or if final state interactions are important . ( final state interactions combined with the @xmath1p@xmath2 model allow two - stage transitions such as @xmath89 , although the direct decay @xmath90 is forbidden in the @xmath1p@xmath2 model . ) recent variants of the @xmath1p@xmath2 model modify the pair production vertex @xcite or modulate the spatial dependence of the pair production amplitude to simulate a gluonic flux tube @xcite . ( the latter is the flux - tube decay model " , which gives very similar predictions to the @xmath1p@xmath2 model in practice . ) another class of decay models assumes that the pair production amplitude transforms as the time component of a lorentz vector . the cornell group used a decay model of this type for charmonium @xcite . this model appears to describe the partial and total widths of some charmonium states reasonably well , although it is known to disagree with experimental amplitude ratios in the light quark sector @xcite . in this work we employ a formalism that is equivalent to the original constant - amplitude version of the @xmath1p@xmath2 decay model , although we have simplified the calculations by deriving momentum - space feynman rules @xcite instead of using the real - space convolution integrals of leyaouanc _ et al . _ @xcite . in our formalism the @xmath1p@xmath2 model describes decay matrix elements using a pair production hamiltonian which is the nonrelativistic limit of h_decay = _ q 2 m_q d^ 3 x |_q _ q . here @xmath91 is a dirac quark field , @xmath92 is the constituent quark mass , and @xmath85 is the dimensionless pair production amplitude , which is fitted to data . the decay arises from the matrix element of this model hamiltonian between an initial meson state @xmath93 and a final meson pair @xmath94 , which is nonzero due to the @xmath95 term . here we will use sho ( simple harmonic oscillator ) wavefunctions for the mesons , with a universal width parameter @xmath96 . the pair - production strength parameter @xmath85 is found to be roughly flavor - independent in light meson decays involving pair production of @xmath97 , @xmath98 and @xmath99 pairs . in our recent extensive studies of light ( @xmath100 ) meson decays @xcite we found that the value = 0.4 gives a reasonably accurate description of the overall scale of decay widths . the open - flavor decay amplitudes of the three states @xmath50 , @xmath75 and @xmath76 were evaluated in the @xmath1p@xmath2 model by leyaouanc _ et al . _ in the late 1970s @xcite . their calculations are generally quite similar to the approach used here , although we note three important differences : 0.2 cm 1 ) leyaouanc _ et al . _ fitted the pair production strength @xmath85 and wavefunction length scale @xmath101 to the charmonium decay data , whereas we used the same pair production amplitude @xmath85 as in light meson decays , and our sho wavefunction length scale @xmath102 is taken from potential models . 0.2 cm 2 ) we use the correct reduced mass coordinates for the unsymmetric open - charm meson wavefunctions , whereas leyaouanc _ et al . _ assumed symmetric sho wavefunctions for all mesons . this approximation proves to be quite important numerically , since @xmath103 . 0.2 cm 3 ) we consider all energetically allowed open - flavor decay modes . leyaouanc _ et al . _ did not consider some s+p decays of the @xmath76 , nor did they consider the important j@xmath104 d meson singlet - triplet mixing angle . ( the masses of the p - wave charmed mesons were unknown at that time . ) 0.2 cm although we have improved on the earlier calculations of leyaouanc _ et al . _ , and evaluate a much more extensive set of decay amplitudes , we do concur with their important conclusion that the dd mode of the @xmath75 is naturally suppressed given a @xmath105s@xmath48 assignment , and also that s+p modes are important for the @xmath76 . other more recent applications of the @xmath1p@xmath2 decay model to charmonium include a study of the strong decays of various @xmath0 x(3872 ) candidates to dd @xcite , and an examination of the hypothesis that the @xmath75 and @xmath106 are linear superpositions of a hybrid and a conventional @xmath107 state @xcite . the only other detailed studies of the open - flavor strong decays of charmonia of which we are aware are the early work by the cornell group @xcite , and recent more detailed applications of this model to other charmonium states @xcite . the decay model of the cornell group assumes a pair creation operator which is the time component of a lorentz vector ( @xmath108 ) , rather than the lorentz scalar assumed by the @xmath1p@xmath2 model . as we noted previously , this @xmath108 model does not agree well with some light meson decay amplitude ratios . in the original cornell model decay calculations only the combined channel contributions to @xmath101 were derived , and only channels containing pairs of s - wave open - charm mesons were included . the very interesting relative amplitudes in the d@xmath9d@xmath9 channel have not been evaluated in this model . tables [ strong_3s4s]-[strong_1 g ] present the partial widths and strong decay amplitudes we find for all kinematically allowed open - flavor decay modes of all the charmonium states listed in table [ table_spectrum ] . we have evaluated these in the @xmath1p@xmath2 model , assuming sho wavefunctions with a width parameter of @xmath109 gev and a pair production amplitude of @xmath110 . sho wavefunctions have the advantage that the decay amplitudes can be determined analytically , and we have found that the numerical results are usually not strongly dependent on the details of the spatial wavefunctions @xcite , unless they are near a node . this width parameter was chosen by comparing the overlap of nr and gi potential model wavefunctions with sho wavefunctions , which was found to be largest for both charmonium and open - charm mesons for a @xmath96 near 0.5 gev . as a test of the accuracy of these parameters , especially the value assumed for the pair production amplitude @xmath85 ( which is taken from light meson decays ) , in fig . [ gamma ] we compare the predicted total widths for @xmath109 gev and variable @xmath85 to the values observed for the four known states above open - charm threshold , @xmath50 , @xmath75 , @xmath51 and @xmath76 . although there is some scatter in the value of @xmath85 specified by these widths , a choice of @xmath111 is evidently near optimum , and yields an average error of only 29% , which is very reasonable for this phenomenological model . in the decay tables we will quote numerical results for our standard light - meson value of @xmath110 , which is rather close to this value . a subsequent change in @xmath85 will simply scale the widths as @xmath112 . for this paper the strong decay amplitudes were first evaluated analytically , following which numerical values were determined given our parameters @xmath109 gev , @xmath110 , and our constituent quark masses @xmath113 gev , @xmath114 gev and @xmath115 gev . ( quark mass _ ratios _ are required to specify the final open - charm meson wavefunctions . ) we determine the decay kinematics using external meson masses taken from the particle data group @xcite ( charge averaged where appropriate ) and from recent belle results @xcite . these masses are m@xmath116 gev , m@xmath117 gev , m@xmath118 gev @xcite , m@xmath119(narrow)@xmath120 gev , m@xmath121(broad)@xmath122 gev @xcite , m@xmath123 gev , m@xmath124 gev , m@xmath125 gev , m@xmath126 gev , m@xmath127 gev . the focus collaboration @xcite estimate a somewhat higher mass for the scalar @xmath128 , however their result is complicated by the presence of both scalar and broad axial vector ( @xmath129 ) contributions @xcite . the @xmath130 axial vector @xmath131 and @xmath132 mesons @xmath133 and @xmath129 are assumed to be coherent superpositions of quark model spin - singlet and spin - triplet states , @xmath134 we define this singlet - triplet mixing angle in the ls coupling scheme , in accord with the conventions of ref.@xcite . this reference also discusses other conventions for this mixing angle that have appeared in the literature . in the heavy - quark limit we expect to find a magic " mixing angle , due to the quark mass dependence of the spin - orbit and tensor terms , which is @xmath135 if the expectation of the heavy - quark spin - orbit interaction is negative ( positive ) @xcite . in the following we assume the first case , which is supported by the reported widths of the @xmath136 p - wave @xmath131 mesons . we note however that finite quark mass effects and mixing induced by higher - order fock states can substantially modify this mixing angle , so it should more generally be treated as a free parameter . we will discuss the dependence of our strong decay amplitudes on this mixing angle for a few sensitive cases . the four known states above dd threshold , @xmath50 , @xmath75 , @xmath51 and @xmath76 are of special interest because they are easily produced at @xmath16 machines . accordingly we will first discuss our predicted strong decay amplitudes and widths for these states . this will be followed by a more general discussion of strong decays by multiplet . the @xmath50 is generally assumed to be the @xmath1d@xmath48 state , perhaps with a significant @xmath49s@xmath48 component @xcite . ( this additional component can explain the leptonic width , which is much larger than expected for a pure @xmath1d@xmath48 state @xcite . ) for a pure @xmath1d@xmath48 state at the @xmath50 mass we predict a dd width of 43 mev with our parameters , which is rather larger than the experimental value of @xmath137 mev . the partial width of a mixed 2s - d state function of the mixing angle @xmath138 . in this simple model , fitting the experimental @xmath50 width requires a mixing angle of @xmath139 @xmath140 @xmath141 . assuming that the leptonic widths scale as the s - wave component squared , this mixing angle predicts an @xmath16 partial width ratio of |_thy . = 0.10 0.03 . it is interesting that this is consistent with experiment , |_expt . = 0.12 0.02 , so the small @xmath50 total width may indeed be an effect of @xmath1d@xmath48 - 2@xmath1s@xmath48 mixing . as a function of the @xmath1d@xmath48 - 2@xmath1s@xmath48 mixing angle @xmath138 . the experimental @xmath142 limits are shown as horizontal dashed lines.,width=302 ] next in mass is the @xmath75 , which is a very interesting case for the study of strong decays . four open - charm modes are energetically allowed , dd , dd@xmath9 , d@xmath9d@xmath9 and d@xmath143d@xmath143 , although d@xmath9d@xmath9 has little phase space . there is some experimental evidence for the three nonstrange modes from mark i at slac @xcite . remarkably , the reported relative branching fractions ( scaled by @xmath144 ) show a very strong preference for d@xmath9 final states , @xmath145 . this motivated suggestions that the @xmath75 might be a @xmath146 molecule @xcite . the mark i results were ( ^*0^*0 : ^0^*0 : ^0 ^ 0 ) = -0.8 cm 12840 : 4.0 0.8 : 0.20.1 . 1.0 cm this is much less striking once the @xmath144 factors are restored , which gives ( ^*0^*0 : ^0^*0 : ^0 ^ 0 ) = -0.8 cm 10.31 : 0.95 0.19 : 0.120.06 , 0.8 cm where we have used the @xmath147 mode to set the scale . these relative branching fractions should be compared to our theoretical @xmath75 branching fractions , not the more often quoted @xmath148 ratios . comparison with table [ strong_3s4s ] shows that the predicted @xmath1p@xmath2 model partial and total widths for a 3@xmath1s@xmath48 @xmath75 state are actually in accord with the mark i results . the total width is predicted to be 74 mev , which is reasonably close to the pdg experimental average of @xmath149 mev . seth @xcite estimates a rather larger width of @xmath150 mev for this state from crystal ball and new bes data , which is closer to our predicted width . the @xmath151 and @xmath146 branching fractions are predicted to be approximately equal , consistent with mark i. the dd mode is predicted to be much smaller , despite its larger phase space , as it is accidentally near a node in the @xmath1p@xmath2 model decay amplitude . ( this was previously noted by leyaouanc _ _ @xcite . ) this set of predictions constitutes a very nontrivial success of the @xmath1p@xmath2 decay model . the actual width to the suppressed dd mode is probably a few mev ; the 0.1 mev quoted in the table depends strongly on the location of the node , and varying the sho width parameter @xmath96 by @xmath152 suggests that a dd partial width of a few mev is plausible . this is again consistent with the mark i result . the @xmath146 mode is especially interesting for strong decay studies , since there are three independent decay amplitudes for @xmath153 , @xmath154p@xmath48 , @xmath155p@xmath48 and @xmath155f@xmath48 . if the @xmath75 is a pure s - wave state , we predict a zero @xmath155f@xmath48 @xmath146 decay amplitude , and the ratio of the nonzero amplitudes is @xmath155p@xmath156p@xmath157 ( independent of the radial wavefunction ) . as we shall see when we discuss the @xmath51 , very different @xmath146 amplitude ratios are predicted for a @xmath1d@xmath48 initial state . the unobserved mode d@xmath143d@xmath143 is predicted to have a branching fraction of about @xmath158 ; this prediction is fortunately not especially sensitive to the location of the decay amplitude node for @xmath75 to two pseudoscalars . this branching fraction is of special interest because it determines the event rates available for studies of @xmath159 weak decays at @xmath16 colliders such as cleo . finally , we note that a large sample of @xmath75 events might be used to access the 2p multiplet , since the e1 radiative branching fractions for @xmath160 are expected to be _ ca . _ @xmath161 ( table [ e1rad_stop ] ) . comparison of the mass of the @xmath74 @xmath51 with potential model predictions ( fig . [ spectrum ] ) immediately suggests a 2@xmath1d@xmath48 assignment . there may also be a significant s - wave component , since the @xmath51 has a much larger @xmath16 width than one would expect for a pure d - wave state @xcite . ( this was also the case for the @xmath50 . ) there are five open - flavor decay modes available to a @xmath74 vector at this mass , dd , dd@xmath9 , d@xmath9d@xmath9 , d@xmath143d@xmath143 and d@xmath162d@xmath163 . the predicted decay amplitudes and partial widths for a 2@xmath1d@xmath48 @xmath51 are given in table [ strong_1d2d ] . the theoretical total width of 74 mev is in good agreement with the experimental value of @xmath164 mev . seth @xcite quotes a similar width of @xmath165 mev for this state from crystal ball and recent bes data . the leading mode is predicted to be d@xmath9d@xmath9 with a branching fraction of @xmath166 , followed by comparable dd and d@xmath162d@xmath163 modes ( both @xmath167 ) , and a somewhat weaker d@xmath143d@xmath143 ( @xmath168 ) . the dd@xmath9 branching fraction is predicted to be very small , since it is suppressed by a decay amplitude node near the physical point . the mode d@xmath9d@xmath9 is again especially interesting , due to the three decay amplitudes allowed for this final state . ( the other modes have only single amplitudes . ) for a pure d - wave assignment the ratio of the two @xmath146 p - wave amplitudes is independent of the radial wavefunction , and is @xmath155p@xmath156p@xmath169 . the @xmath155f@xmath48 amplitude is predicted to be the largest given this 2@xmath1d@xmath48 assignment , whereas it is zero for an s - wave state . clearly , a determination of these d@xmath9d@xmath9 decay amplitude ratios would be an extremely interesting test of decay models . experimentally , to date nothing has been reported regarding the exclusive hadronic decay modes of the @xmath51 . the final resonance known above dd threshold is the @xmath74 @xmath76 . again , the mass of this resonance relative to potential model predictions suggests a assignment , in this case 4@xmath1s@xmath48 . of course this identification requires independent confirmation , since hybrids are predicted to first appear near this mass by lgt simulations @xcite , and the lightest hybrid multiplet in the flux tube model includes a @xmath74 state @xcite . ten open - charm strong decay modes are allowed for the @xmath76 , seven with @xmath131 meson final states ( @xmath170 ) , and three with @xmath132 . the predicted branching fractions of a 4@xmath1s@xmath48 state at this mass are quite characteristic , and ( if the @xmath1p@xmath2 decay model is accurate ) may be useful in confirming this assignment . as a note of caution , a 4s state has three radial nodes , and some of the smaller predicted branching fractions are sensitive to the locations of the nodes . the predicted total width of a 4@xmath1s@xmath48 meson at this mass with our parameters is 77 mev , somewhat larger than the experimental pdg average width of @xmath171 mev . seth @xcite notes however that the width of this state in crystal ball and recent bes data is rather larger , and quotes an average of @xmath172 mev , on the opposite side of the @xmath1p@xmath2 decay model prediction . the largest exclusive mode is predicted to be the s+p combination dd@xmath48 , where d@xmath48 is the narrower of the two @xmath136 @xmath131 axial mesons near 2.42 - 2.43 gev . since the d@xmath48 is rather narrow ( @xmath173-@xmath174 mev ) and decays dominantly to d@xmath175 , there should be a strong @xmath76 signal in dd@xmath175 final states . although both s - wave and d - wave dd@xmath48 final states are allowed , in the @xmath1p@xmath2 model the hqet ( heavy quark effective theory ) d@xmath48 mixing angle @xmath138 is just the value needed to give a zero s - wave @xmath176 amplitude . thus we have the striking prediction that the dominant @xmath76 decay mode is dd@xmath48 , in d - wave rather than s - wave . the second - largest @xmath76 branching fraction is predicted to be another s+p mode , dd@xmath177 . the d@xmath177 is also moderately narrow , so this isobar can also be isolated from the observed final state . ( the pdg quotes a neutral d@xmath177 total width of @xmath178 mev , but belle @xcite and focus @xcite find somewhat broader values of @xmath179 mev and @xmath180 mev respectively ; see @xcite for a recent discussion . the d@xmath177 has significant branching fractions to both d@xmath175 and d@xmath181 , so the dd@xmath177 mode of the @xmath76 should be observable in both dd@xmath181 and dd@xmath175 . a final important @xmath76 mode is predicted to be d@xmath9d@xmath9 , which should be comparable in strength to dd@xmath177 . as noted in the previous discussions of @xmath75 and @xmath51 decays , d@xmath9d@xmath9 is an especially interesting decay mode because it has three amplitudes , and the @xmath155p@xmath48/@xmath154p@xmath48 amplitude ratio is independent of the radial wavefunction for pure s - wave or d - wave states . if the @xmath76 is indeed an s - wave ( 4@xmath1s@xmath48 ) state to a good approximation , we expect this ratio to be @xmath155p@xmath48/@xmath154p@xmath157 , and the @xmath155f@xmath48 amplitude should be zero . it is interesting to note that @xmath76 decays may provide access to the recently discovered d@xmath182 . although the channel d@xmath183d@xmath184 has a threshold of 4429 mev , 14 mev above the nominal mass of the @xmath76 , the width of the @xmath76 and the fact that the decay @xmath185d@xmath183d@xmath184 is purely s - wave implies that one may observe significant d@xmath184 production just above threshold , near e@xmath186 mev . this is illustrated in fig . [ 4415_scan ] , in which we show theoretical @xmath187 decay model partial widths to d@xmath188d@xmath189 and d@xmath143d@xmath190 as functions of the @xmath191s@xmath48 mass . ( this calculation assumes pure @xmath132 d@xmath189 and d@xmath190 states , and should be modified accordingly if they have significant non-@xmath132 components . ) unfortunately , @xmath76 decays are not expected to be similarly effective in producing the d@xmath190 , because the assumed magic - mixed " hqet state @xmath192 predicts a vanishing @xmath193 s - wave decay amplitude . and @xmath194 as a function of the assumed @xmath195 mass . , width=321 ] the two unknown states in the 3s and 4s multiplets are the 3@xmath154s@xmath2 and 4@xmath154s@xmath2 pseudoscalars . to evaluate their strong decays we have assigned potential model masses to these states , specifically 4043 mev and 4384 mev . the resulting total widths are predicted to be moderate , 80 mev and 61 mev respectively , and both states should be observable in dd@xmath9 and d@xmath9d@xmath9 ( see table [ strong_3s4s ] ) . no other open - charm strong decay modes are allowed for a 3@xmath154s@xmath2 @xmath196 . although these modes are also important for the 4@xmath154s@xmath2 @xmath197 , its largest branching fraction is predicted to be to the s+p combination dd@xmath177 ( @xmath198 ) . these states may be observable in @xmath6 and hadronic production ( through @xmath199 fusion ) , analogous to the @xmath200 and @xmath5 . m1 transitions from the higher vectors to these states are unfortunately predicted to be rather weak , since they are either hindered or have little phase space . ( see table [ m1rad ] . ) the mean 2p multiplet mass is predicted to be near 3.95 gev . there is an interesting disagreement between the nr and gi models regarding the scale of mass splittings within this multiplet , as well as in the 3p states . ( see table [ table_spectrum ] and fig . [ spectrum ] . ) there are few open - flavor strong modes available to the 2p states ( table [ strong_2p3p ] ) . one result of this restricted phase space is the prediction of a fairly narrow 2@xmath1p@xmath2 @xmath201 ( assuming the nr model mass ) that decays only to dd , with a total width of just 30 mev . this small width is due in part to a decay amplitude node , and a variation of the total width by a factor of two could easily be accommodated . the two axial states 2@xmath1p@xmath48 @xmath202 and 2@xmath154p@xmath48 @xmath203 can only decay to dd@xmath9 , and both have nonzero s- and d - wave decay amplitudes ; their d / s ratios are @xmath204 and @xmath205 respectively . observation of the d / s ratio in a dd@xmath9 enhancement could be a useful check of the assumption of a resonant @xmath136 contribution . the relatively narrow 2@xmath1p@xmath2 state @xmath201 can be produced in @xmath6 , as can its somewhat wider 2@xmath1p@xmath57 @xmath206 partner . additional possible production mechanisms include @xmath199 fusion ( for the 2@xmath1p@xmath2 and 2@xmath1p@xmath57 ) , e1 radiative transitions from higher @xmath74 states ( to all 2p @xmath79 states ) , and b decays . the e1 branching fractions of the @xmath75 and @xmath51 to 2p @xmath79 states are predicted to be sufficiently large ( _ ca . _ @xmath161 , see tables [ e1rad_stop],[e1rad_1d2d ] ) to allow identification of these states at bes and cleo . although no 2p @xmath0 state has been clearly established experimentally , there are two recent reports of enhancements which may be due to states in this multiplet . the belle collaboration has reported evidence of an enhancement in @xmath12 with a mass and width of @xmath207 mev and @xmath208 mev @xcite ; this is compatible with expectations for 2p @xmath79 @xmath0 states . ( although 2p @xmath0 states should strongly favor open - charm strong decay modes , inelastic fsis ( final state interactions ) will allow weaker transitions to closed - charm final states such as @xmath12 . these fsis should be most important in s - wave , as in the virtual second - order processes @xmath209 . ) a search for evidence of this state in open - charm final states , with much larger branching fractions than to @xmath12 , will be a crucial test of a charmonium assignment . assuming that the enhancement is indeed due to a 2p @xmath79 state , if it is the @xmath210 state it will populate only dd , if @xmath13 only dd@xmath9 , and if @xmath211 comparable branching fractions to dd and dd@xmath9 are expected . there is a second report of a possible 2p state , also from the belle collaboration @xcite , in double charmonium production . an enhancement is seen in @xmath212 at a mass of @xmath213 mev , with a limit on the total width of @xmath214 mev ( @xmath215 ) . as this enhancement is not seen in @xmath12 , it has been suggested @xcite that it is distinct from the 3940 mev signal discussed above . consideration of possible 2p candidates ( table [ strong_2p3p ] ) suggests that this enhancement may be due to the @xmath216p@xmath217 , which is predicted to have a similar mass of 3925 mev . ( the predicted total width however is a somewhat larger 165 mev . ) the 3p states have an expected mean multiplet mass of about 4.3 gev . as with the 2p states there is a significant difference between the nr and gi models in the scale of mass splittings predicted within the multiplet . this leads for example to a 90 mev difference between the models for the predicted mass of the 3@xmath1p@xmath2 scalar . many open - charm strong decay channels are open at the 3p mass scale . the individual decay amplitudes however are typically somewhat smaller than for the 2p states , due in part to nodes in the decay amplitudes . this has the interesting result that the predicted mean total width of 3p states is smaller than for 2p states ( 58 mev for 3p versus 91 mev for 2p ) . the predicted branching fraction to the mode d@xmath9d@xmath9 is large for all four 3p states , and is the largest for all except the spin - singlet 3@xmath154p@xmath48 @xmath218 ; in this case the leading mode is dd@xmath219 . for all 3p states except the 3@xmath1p@xmath48 the d@xmath9d@xmath9 decay mode has multiple amplitudes , which are predicted to be comparable in strength . an experimental determination of any of these amplitude ratios would provide an interesting test of decay models . the decays of the tensor 3@xmath1p@xmath57 @xmath220 to dd@xmath48 and dd@xmath221 are also interesting , in that the prediction that the broad d@xmath221 mode dominates is a test of the axial - d mixing angle . as with the 2p states , the most important experimental problem is to find an effective production mechanism . again , @xmath6 and @xmath199 fusion can produce the @xmath211 and @xmath210 states ( here 3@xmath1p@xmath57 and 3@xmath1p@xmath2 ) , although there should be moderate suppression of the production amplitudes relative to the 1p ground states due to the higher masses and smaller short - distance wavefunctions . e1 transitions from the @xmath76 could lead to the identification of the 3p states , since radiative partial widths of 50 - 100 kev are expected to the 3@xmath1p@xmath57 and 3@xmath1p@xmath48 ; these correspond to branching fractions of _ ca . _ @xmath161 . production of the 3@xmath1p@xmath2 state however is expected to be suppressed , due to a radiative decay amplitude node . since the @xmath1d@xmath57 and @xmath154d@xmath57 states do not have allowed open - flavor decay modes , and we have already considered @xmath1d@xmath48 decays in the section on the @xmath50 , we restrict our discussion to the @xmath1d@xmath222 . this currently unknown state is predicted to have a mass of 3806 mev in the nr potential model . it is expected to be quite narrow simply because it has little phase space to the only open - flavor mode dd , which has an f - wave centrifugal barrier ; with our parameters the predicted total strong width is just 0.5 mev . the radiative width for the e1 transition @xmath223 is predicted to be about 0.3 mev ( table [ e1rad_1d2d ] ) , so the total width of this @xmath1d@xmath222 state should be near 1 mev , with comparable strong and radiative branching fractions . finding an effective production mechanism for this interesting state is again a crucial problem . as it has c @xmath224 , it can not be made in @xmath225 or @xmath199 fusion . one possibility is that the @xmath1d@xmath222 may be observable in e1 transitions from the radial @xmath49p@xmath57 tensor , since the radiative partial width is predicted to be a relatively large @xmath53 kev . similarly , the narrow @xmath1d@xmath57 and @xmath154d@xmath57 states may be observable in e1 transitions from 2p states , assuming that a production mechanism can be found for 2p states . all three of these states can also be produced in b - decays @xcite . the states in the 2d multiplet are predicted to be essentially degenerate in both the nr and gi potential models . the @xmath49d@xmath48 candidate @xmath51 suggests a 2d multiplet mass near 4.16 gev , which is consistent with potential model expectations . since the @xmath51 was discussed previously , here we will consider only the unknown states @xmath49d@xmath222 , @xmath49d@xmath57 and @xmath226d@xmath57 . these states are predicted to be rather broad , with theoretical total widths ranging from 78 mev for the @xmath49d@xmath48 ( consistent with the @xmath51 ) to 148 mev for the @xmath49d@xmath222 . the energetically allowed open - flavor modes are all of s+s type , since the first s+p mode is dd@xmath227 , at a nominal mass of 4175 mev . the decays of 2d states are dominated by dd , dd@xmath9 and d@xmath9d@xmath9 , with smaller contributions from the charm - strange final states d@xmath143d@xmath143 and d@xmath162d@xmath163 . if these states ( besides the @xmath51 ) can be produced , there are several interesting tests of their strong decay amplitudes . the @xmath49d@xmath222 @xmath228 , for example , is predicted to have d@xmath9d@xmath9 as the dominant mode , with both @xmath154f@xmath222 and @xmath155f@xmath222 amplitudes present , in the ratio @xmath155f@xmath222/@xmath154f@xmath229 . the remaining 2d states have many multiamplitude decays , in which the higher - l partial waves have large or dominant amplitudes . ( compare the p- and f - waves in the decays in table [ strong_1d2d ] . ) the four states in the 1f multiplet are predicted to be almost degenerate in both the nr and gi models , although the models disagree somewhat regarding this mass . in the nr potential model the states are expected near 4025 mev , whereas in the gi model the expected mass is near 4095 mev . this discrepancy is mainly due to the different values assumed for the string tension . it is unfortunate that production amplitudes for these higher - l states are expected to be quite weak , since the 1f multiplet is predicted to contain a rather narrow state , the @xmath1f@xmath230 @xmath231 ( table [ strong_1f ] ) . assuming our nr model @xmath1f@xmath230 mass of 4021 mev , we predict a strong width for this state of just 8.3 mev . this is partly a result of the centrifugal barrier ( both open modes , dd and dd@xmath9 , are g - waves ) , and is also because there is essentially no phase space for d@xmath9d@xmath9 . if the gi masses are more accurate we would no longer expect a very narrow @xmath1f@xmath230 state , since the d@xmath9d@xmath9 partial width increases rapidly above threshold . at the gi mass of 4095 mev the @xmath1f@xmath230 is predicted to have a partial width to d@xmath9d@xmath9 of 38 mev , and a total width of 60 mev . the single well - established mechanism that might be exploited to produce a 1f state is an e1 radiative transition from the @xmath51 to the @xmath1f@xmath57 @xmath232 . this transition however has a rather weak , model - dependent partial width of 20 - 50 kev ( table [ e1rad_1d2d ] ) , and the @xmath1f@xmath57 is predicted to be the broadest 1f state , with a total width of about 160 mev . if an effective production mechanism for 1f states is identified , the e1 decays of these states can be used to populate the narrow 1d states ; the e1 transitions @xmath1f@xmath233d@xmath222 , @xmath1f@xmath234d@xmath57 and @xmath154f@xmath235d@xmath57 are predicted to have rather large radiative widths of _ ca . _ 300 - 400 kev ( see table [ e1rad_1f ] ) . to complete our study we have evaluated the spectrum and decays of all states in the 2f and 1 g multiplets ( tables [ strong_2fa ] and [ strong_1 g ] ) , although the lack of a clear experimental route to these states suggests that it may be quite difficult to test these predictions . the multiplets are again predicted to be nearly degenerate in both the nr and gi potential models , with somewhat lower masses predicted by the nr model . the mean 2f and 1 g masses are approximately 4350 mev and 4225 mev respectively in the nr model , and 4425 mev and 4320 mev in the gi model . assuming nr model masses , these states are predicted to have total widths that range from a minimum of 58 mev for the @xmath1g@xmath236 @xmath237 to a maximum of 180 mev for the @xmath49f@xmath57 @xmath238 . the individual decay modes do show some interesting features ; although the dominant modes are usually of s+s type , some s+p modes are important in 2f decays . for example , the largest decay modes of the @xmath49f@xmath222 @xmath239 and @xmath49f@xmath57 @xmath238 are both s+p , with branching fractions of b(@xmath240 dd@xmath241 and b(@xmath242 dd@xmath243 respectively . we have computed the spectrum , all allowed e1 and some m1 electromagnetic partial widths , and all allowed open - charm strong decay amplitudes of the 40 charmonium states expected to _ ca . _ 4.4 gev . these were evaluated using two potential models , the quark model formulation of electromagnetic transition amplitudes , and the @xmath187 strong decay model . the predictions of the spectrum should be useful in the identification of new states , and in tests of the lorentz structure of confinement and the nature of spin - dependent forces . the transition amplitudes will be useful in searches for new higher - mass states , as they suggest which final states should be populated preferentially by the decays of a given initial state , as well as predicting the often characteristic strong decay amplitude ratios within a given final state . these predictions may also be useful in distinguishing between conventional charmonia and exotica such as charmed meson molecules ( perhaps including the x(3872 ) ) and hybrids . these results all implicitly test the accuracy of the pure - assumption made in these models , and can serve as benchmarks for calculations that relax this assumption . future experimental measurements of strong partial widths and decay amplitude ratios can provide very important tests of strong decay models , the @xmath1p@xmath2 model in particular . we have discussed many specific examples of these tests in the previous sections ; here we will summarize some of our most important observations . the @xmath50 is the lightest known state above open - charm threshold . we noted that the relatively large @xmath16 width of the @xmath50 is difficult to understand if it is assumed to be a pure @xmath1d@xmath48 state , but this problem is solved if the @xmath50 has a significant admixture of the @xmath244 @xmath0 basis state . this mixing also solves the @xmath50 total width problem , since the mixing angle required to fit the @xmath16 width is consistent with the value required to fit the @xmath1p@xmath2 model dd strong width to experiment . this mixing angle has been discussed in earlier references ( see for example refs.@xcite ) , although the observation that the same mixing angle resolves both the @xmath16 and total width discrepancies has not been noted previously . the relative branching fractions of the @xmath75 to dd , dd@xmath9 and d@xmath9d@xmath9 , although often cited as anomalous , are naturally explained in the @xmath187 model by a node in the dd decay amplitude . a very important test of this strong decay model and the nature of the @xmath75 follows from a simple measurement of the ratios of the three independent d@xmath9d@xmath9 decay amplitudes . the @xmath187 model predicts that the @xmath245 amplitude is zero , and the ratio of the two p - wave amplitudes is @xmath246 , independent of the radial wavefunction , provided that the @xmath75 is a pure s - wave state . similarly , the d@xmath9d@xmath9 amplitude ratios in @xmath51 decays test the decay model and the assumed @xmath247 assignment for this state . the @xmath245 d@xmath9d@xmath9 amplitude is predicted to be the largest for a pure @xmath247 @xmath51 , and the p - wave amplitude ratio is predicted to be @xmath248 , again independent of the radial wavefunction . finally , we noted that it may be possible to identify a higher - l state ( the @xmath249 @xmath232 member of the f - wave multiplet ) in e1 radiative @xmath51 decays . the @xmath76 has ten open - flavor modes , and their branching fractions have never been determined . we predict that the largest is the unusual s+p combination dd@xmath48 . it is especially notable that strong decays of the @xmath76 may provide a novel production mechanism for the enigmatic @xmath250 , through the above - threshold decay @xmath251 , as shown in fig . [ 4415_scan ] . one may also produce higher - mass charmonium states from the @xmath76 through electromagnetic transitions . for example , @xmath76 e1 decays are predicted to produce 3p states with branching fractions at the per mil level . these states could then be identified through their subsequent open - charm strong decays . as discussed in section v.c , two new states near 3940 mev have recently been reported by belle . 2p @xmath79 states are natural assignments , as they have theoretical quark model masses ranging from 3852 to 3972 mev and widths of 50 - 100 mev , similar to the reported values . a measurement of the relative branching fractions to @xmath252 , @xmath151 and @xmath146 and comparisons to our @xmath1p@xmath2 decay model predictions should allow the determination of the quantum numbers of these states , and will show whether they are indeed consistent with 2p @xmath0 assignments . a better determination of the properties of conventional charmonium states at higher masses is important not only because of the improved understanding of qcd that will follow ( especially aspects of confinement and strong decays ) , but also because qualitatively different types of resonances are expected at these masses . these new states include charmed meson molecules and charmonium hybrids , and the identification of these novel excitations will obviously be easier if the conventional charmonium spectrum is well established . the recent results from b factories and new programs at bes , cleo and gsi have led to a resurgence of interest in the physics of charmonium . we argue that a detailed experimental investigation of the spectrum of excited charmonium states and their decay properties will considerably improve our understanding of the nonperturbative aspects of qcd . this research was supported in part by the natural sciences and engineering research council of canada , the u.s . national science foundation through grant nsf - phy-0244786 at the university of tennessee , and the u.s . department of energy under contracts de - ac05 - 00or22725 at oak ridge national laboratory and de - fg02 - 00er41135 at the university of pittsburgh , and by pparc grant pp / b500607 . j. j. aubert _ et al . _ [ e598 collaboration ] , phys . rev . * 33 * , 1404 ( 1974 ) . j. e. augustin _ et al . _ [ slac - sp-017 collaboration ] , phys . * 33 * , 1406 ( 1974 ) . t. appelquist and h. d. politzer , phys . lett . * 34 * , 43 ( 1975 ) . a. de rujula and s. l. glashow , phys . * 34 * , 46 ( 1975 ) . t. appelquist , a. de rujula , h. d. politzer and s. l. glashow , phys . lett . * 34 * , 365 ( 1975 ) . e. eichten , k. gottfried , t. kinoshita , j. b. kogut , k. d. lane and t. m. yan , phys . lett . * 34 * , 369 ( 1975 ) [ erratum - ibid . * 36 * , 1276 ( 1976 ) ] . c. quigg , arxiv : hep - ph/0403187 . r. s. galik , arxiv : hep - ph/0408190 . n. brambilla _ et al . _ , arxiv : hep - ph/0412158 . k. k. seth , arxiv : hep - ex/0501022 . this reference in effect quotes a combined average mass for the @xmath15 of @xmath253 mev from new cleo measurements . k. k. seth , arxiv : hep - ex/0504050 . k. k. seth , arxiv : hep - ex/0504052 . t. skwarnicki , arxiv : hep - ex/0505050 . e.j.eichten , k.lane and c.quigg , phys . 89 , 162002 ( 2002 ) [ arxiv : hep - ph/0206018 ] . the besiii detector . preliminary design report , ihep - bepch - sb-13 ( jan . 2004 ) . s .- k.choi _ et al . _ ( belle collaboration ) , phys . 89 , 102001 ( 2002 ) , err . 89 , 129901 ( 2002 ) . [ arxiv : hep - ex/0206002 ] . b. aubert _ et al . _ [ babar collaboration ] , phys . 92 , 142002 ( 2004 ) [ arxiv : hep - ex/0311038 ] . ( cleo collaboration ) , phys . lett.92 , 142001 ( 2004 ) [ arxiv : hep - ex/0312058 ] . s. k. choi _ et al . _ [ belle collaboration ] , phys . 91 , 262001 ( 2003 ) [ arxiv : hep - ex/0309032 ] . d. acosta _ et al . _ [ cdf ii collaboration ] , phys . rev . 93 , 072001 ( 2004 ) [ arxiv : hep - ex/0312021 ] . f. e. close and p. r. page , phys . lett . b578 , 119 ( 2004 ) [ arxiv : hep - ph/0309253 ] . e. s. swanson , phys . b588 , 189 ( 2004 ) [ arxiv : hep - ph/0311229 ] . n. a. tornqvist , phys . b * 590 * , 209 ( 2004 ) [ arxiv : hep - ph/0402237 ] . t. barnes and s. godfrey , phys . d69 , 054008 ( 2004 ) [ arxiv : hep - ph/0311162 ] . e. j. eichten , k. lane and c. quigg , phys . d69 , 094019 ( 2004 ) [ arxiv : hep - ph/0401210 ] . k. abe , arxiv : hep - ex/0505037 . k. abe , arxiv : hep - ex/0505038 . a. tomaradze , j. phys . * 9 * , 119 ( 2005 ) [ arxiv : hep - ex/0410090 ] . the reported @xmath15 signal is in the decay chain @xmath254 , @xmath255 . the masses found in two different inclusive analyses were @xmath256 mev and @xmath256 mev ( with an estimated systematic error of @xmath257 mev , and @xmath258 mev in exclusive decays ( with six different identified @xmath200 final states ) . k. abe _ et al . _ [ belle collaboration ] , phys . 89 , 142001 ( 2002 ) [ arxiv : hep - ex/0205104 ] . k. abe _ et al . _ [ belle collaboration ] , phys . d70 , 071102 ( 2004 ) [ arxiv : hep - ex/0407009 ] . p. pakhlov , [ arxiv : hep - ex/0412041 ] . p. pakhlov ( private communication ) . e. eichten , k. gottfried , t. kinoshita , k. d. lane and t. m. yan , phys . d * 17 * , 3090 ( 1978 ) [ erratum - ibid . d * 21 * , 313 ( 1980 ) ] . d. ebert , r. n. faustov and v. o. galkin , arxiv : hep - ph/0503012 . c. w. bernard _ et al . _ [ milc collaboration ] , phys . d * 56 * , 7039 ( 1997 ) [ arxiv : hep - lat/9707008 ] . x.liao and t.manke , [ arxiv : hep - lat/0210030 ] . z. h. mei and x. q. luo , int . j. mod . phys . a * 18 * , 5713 ( 2003 ) [ arxiv : hep - lat/0206012 ] . g. s. bali , eur . j. a * 19 * , 1 ( 2004 ) [ arxiv : hep - lat/0308015 ] . b. aubert _ et al . _ [ babar collaboration ] , phys . 90 , 242001 ( 2003 ) [ arxiv : hep - ex/0304021 ] . d. besson _ et al . _ [ cleo collaboration ] , aip conf . 698 , 497 ( 2004 ) [ arxiv : hep - ex/0305017 ] . e. s. swanson , phys . b * 582 * , 167 ( 2004 ) [ arxiv : hep - ph/0309296 ] . d. ebert , r. n. faustov and v. o. galkin , phys . d * 67 * , 014027 ( 2003 ) [ arxiv : hep - ph/0210381 ] . t. barnes , [ arxiv : hep - ph/0406327 ] . li , f. e. close and t. barnes , phys . d * 43 * , 2161 ( 1991 ) . e. s. ackleh and t. barnes , phys . d * 45 * , 232 ( 1992 ) . e. s. ackleh , t. barnes and f. e. close , phys . rev . d * 46 * , 2257 ( 1992 ) . t. barnes and g. i. ghandour , phys . lett . b118 , 411 ( 1982 ) . s.godfrey and n.isgur , phys . d32 , 189 ( 1985 ) . m. okamoto _ et al . _ [ cp - pacs collaboration ] , phys . d * 65 * , 094508 ( 2002 ) [ arxiv : hep - lat/0112020 ] . s. hashimoto and t. onogi , ann . rev . nucl . part . sci . * 54 * , 451 ( 2004 ) [ arxiv : hep - ph/0407221 ] . w.kwong and j.l.rosner , phys . d38 , 279 ( 1988 ) . l. micu , nucl . b10 , 521 ( 1969 ) . a. le yaouanc , l. oliver , o. pne and j .- c . raynal , phys . d8 , ( 1973 ) 2223 . a. le yaouanc , l. oliver , o. pne and j .- c . raynal , phys . d9 , 1415 ( 1974 ) . a. le yaouanc , o. pne , j .- c . raynal and l. oliver , nucl . b149 , 321 ( 1979 ) . a. le yaouanc , l. oliver , o. pne and j .- c . raynal , phys . b71 , 397 ( 1977 ) . le yaouanc , l. oliver , o. pne and j .- c . raynal , phys . b72 , 57 ( 1977 ) . r. kokoski and n. isgur , phys . d35 , 907 ( 1987 ) . e.s.ackleh , t.barnes and e.s . swanson , phys . d54 , 6811 ( 1996 ) . [ arxiv : hep - ph/9604355 ] . m. nozar _ et al . _ [ e852 collaboration ] , phys . lett . b541 , 35 ( 2002 ) [ arxiv : hep - ex/0206026 ] . d. v. amelin _ et al . _ [ ves collaboration ] , [ arxiv : hep - ex/9810013 ] . ( particle data group ) , phys . b592 , 1 ( 2004 ) . w.roberts and b.silvestre-brac , phys . d57 , 1694 ( 1998 ) . [ arxiv : hep - ph/9708235 ] . t.barnes , f.e.close , p.r.page and e.s.swanson , phys . d55 , 4157 ( 1997 ) . [ arxiv : hep - ph/9609339 ] . t.barnes , n.black and p.r.page , phys . d68 , 054014 ( 2003 ) [ arxiv : nucl - th/0208072 ] . f. e. close and p. r. page , phys . b * 366 * , 323 ( 1996 ) [ arxiv : hep - ph/9507407 ] . p. geiger and e. s. swanson , phys . d50 , 6855 ( 1994 ) [ arxiv : hep - ph/9405238 ] . h. g. blundell and s. godfrey , phys . d53 , 3700 ( 1996 ) [ arxiv : hep - ph/9508264 ] . et al . _ [ belle collaboration ] , phys . d69 , 112002 ( 2004 ) [ arxiv : hep - ex/0307021 ] . e. w. vaandering [ focus collaboration ] , [ arxiv : hep - ex/0406044 ] . r.kutschke , private communication . s.godfrey and r.kokoski , phys . d43 , 1679 ( 1991 ) . j. l. rosner , [ arxiv : hep - ph/0411003 ] . j. l. rosner , [ arxiv : hep - ph/0405196 ] . et al . _ , b69 , 503 ( 1977 ) . m.b.voloshin and l.b.okun , jetp lett . 23 , 333 ( 1976 ) [ pisma zh . fiz . 3 , 369 ( 1976 ) ] . a.derjula , h.georgi and s.l.glashow , phys . 38 , 317 ( 1977 ) . s.iwao , lett . nuovo cim . 28 , 305 ( 1980 ) . k. k. seth , [ arxiv : hep - ex/0405007 ] . n.isgur , r.kokoski and j.paton , phys . 54 , 869 ( 1985 ) . j. m. link _ et al . _ [ focus collaboration ] , phys . b586 , 11 ( 2004 ) [ arxiv : hep - ex/0312060 ] . p.s.cooper , new results in charm meson spectroscopy from focus and selex " , presented at ghp2004 ( fermilab , 24 - 26 oct 2004 ) . k. abe _ et al . _ [ belle collaboration ] , [ arxiv : hep - ex/0408126 ] . t. skwarnicki , int j. mod . a * 19 * , 1030 ( 2004 ) . y.zhu , ph.d . thesis , california institute of technology , 1988 , caltech report calt-68 - 1513 ( unpublished ) .

the ground state of a weak - disorder quantum hall bilayer has an exciton condensate when the total landau level filling factor @xmath0 and the semiconductor quantum wells are close together . the condensate state is characterized by spontaneous inter - layer phase coherence.@xcite condensate properties are often probed most effectively by transport experiments , both because the two - dimensional electron system is buried beneath the surface of a three - dimensional host and because these measurements probe@xcite condensate superfluidity . indeed , the presence of a condensate has been inferred on the basis of wide variety of transport anomalies.@xcite although these experiments do not directly measure the condensate order parameter , it is generally acknowledged that they do not allow plausible alternate explanations . on the other hand the transport properties of systems in which the condensate is believed to have formed have largely defied attempts to achieve fully quantitative interpretations . this conundrum is related in part to the importance of edge states in quantum hall transport@xcite and , as we explain in this paper , to the sensitivity of transport measurements to disorder and to contact properties that are not normally characterized . the description that we develop for transport in the deeply coherent regime is based on a microscopic mean - field theory of the exciton condensate state , on a scattering formulation of transport theory,@xcite and on a two - channel contact model which accounts explicitly for the two contact layers . we show that transport properties can be non - trivial even when the current is small enough that the coherence order parameter is time - independent.@xcite a simplified version of this picture with a single parameter , which we refer to as _ p - ology _ , allows for virtual occupation of a pair - breaking edge channel and can be used to provide an approximate description of the crossover between strong and weak interlayer coupling regimes . our paper is organized as follows . in section ii we explain the general philosophy of our approach , which has similarities to theories of the non - equilibrium properties of both superconductors @xcite and magnetic conductors.@xcite the edge state physics of quantum hall systems is essential to the way in which the theory is developed . indeed , the presence of edge states at the fermi level , even when the bulk is gapped , is responsible for many of the complications which make the transport anomalies difficult to describe quantitatively . in section iii we detail the landauer - buttiker theory that we use to describe transport of the mean - field - theory quasiparticles , and illustrate it by evaluating two probe conductances for systems with well developed coherence . in section iv we discuss transport in the case of a non stationary condensate , focusing on the simplest case in which both coherent and incoherent inter - layer tunneling are absent and illustrating our picture by calculating the interlayer voltage for a drag experiment with a grounded drag layer . in section v we develop _ p - ology _ , which is intended to describe the influence of weaker coherence but assumes that quasiparticle transport is still edge dominated . this phenomenology makes it clear that large longitudinal drags are always present when the hall drag voltage drops below its quantized value , as seen in experiment.@xcite sec . vi contains a summary of our main results and discusses their relationship to other ideas for explaining the transport properties of coherent bilayers . in this paper we adopt a hartree - fock - like mean - field description of the exciton condensate which is similar to the mean - field - theories used , generally successfully , to describe transport in metallic ferromagnets and superconductors . this type of theory has an @xmath1 vector order parameter constructed in our case from the pseudospin-@xmath2 layer degree of freedom , and an associated pseudospin effective magnetic field which arises from microscopic exchange interactions . for a typical bilayer with weak single - particle tunneling , the exchange field exceeds the single - particle symmetric - antisymmetric level splitting , @xmath3 , by many orders of magnitude . @xmath3 is therefore normally negligible in the description of the quasiparticle properties . the dynamics of the order parameter , on the other hand , is described by a landau - lifshitz - like equation which depends only on an external field term proportional to @xmath3 , on other terms proportional to order parameter derivatives and on transport currents.@xcite ( characteristic length scales then emerge from a comparison of @xmath3 and the coefficients of the order - parameter derivative terms . ) the @xmath3 term must be present@xcite to allow consistently for the net interlayer charge transfer which occurs in many transport experiments . typical electrical measurements performed on a coherent quantum hall bilayer involve establishing a source to drain voltage bias which injects transport quasiparticles , and measuring local electrochemical potentials at positions along the sample perimeter . when the quantum hall effect is well developed the transport quasiparticles are localized at the sample edges . electrochemical potential measurements therefore often reflect properties of the bulk condensate somewhat indirectly . contacts to the coherent bilayer resemble normal metal - superconductor junctions , at which normal currents are converted into supercurrents . this conversion between normal current and supercurrent is accomplished via andreev scattering . a similar conversion process occurs in a coherent bilayer ; even for vanishingly small single - particle tunneling a quasiparticle that has a definite layer quantum number in the leads becomes coherently mixed between the layers far away from the leads . the order parameter exchange field provides a broken symmetry contribution to the quasiparticle hamiltonian with a large inter - layer tunneling amplitude . this process does not violate charge conservation since an excitonic supercurrent is always launched to compensate@xcite for charge transfer between layers . when the condensate is described collectively , the full system of equations consists of a landau - lifshitz - like equation for the condensate with a _ pseudospin - transfer _ torque@xcite term to account for normal current to exciton super current conversion , combined with a transport theory for quasiparticles influenced by the condensate . the two sets of equations need to be solved self - consistently . at low temperatures , inelastic scattering is weak and quasiparticle transport can be described using the landauer - buttiker formalism . application of this picture is complicated in coherent quantum hall bilayer systems by the role of chiral edge states , which can carry nearly dissipationless currents by a mechanism completely different from superfluidity . two signature properties of coherent bilayers are i ) the quantization of hall drag resistance at @xmath4 ( the hall resistance is quantized whether measured in the layer in which current is flowing or in the layer in which no current flows ) , and ii ) the absence of longitudinal drag at filling factor @xmath5 per layer . together the two properties suggest that a quantum hall effect is established at a filling factor related to the combined densities of the two layers , and that the quasiparticle current flows coherently through both layers ; there is no layer dependence in chemical potential measurements . because of the quantum hall effect , and its associated bulk gap , the bilayer coherent quasiparticles must flow at the sample edges . since these properties are normally measured in samples in which @xmath3 is extremely small , the quasiparticle current in the drag layer must be exactly compensated by a condensate counterflow supercurrent . we will be interested in analyzing transport both in four - probe geometries , which include voltage - probe contacts in addition to the current source and drain contacts , and in two - probe geometries in which the additional voltage probe contacts are absent . as we explain , some interesting information about the system can be obtained from two - probe measurements . one interesting issue that arises in the two - probe case concerns a theory of the maximum conductance possible when source and drain contacts are connected to opposite layers . our analysis relates two - probe conductances to the influence of the condensate on electron scattering near the contacts , information which will prove helpful in analyzing four - probe experiments . an appropriate microscopic model for the _ normal _ contact region in which interlayer coherence is absent plays an essential role in our analysis . we assume that the active area of the bilayer in which coherence is established is incompressible and surrounded by a single current - carrying edge state channel shared between the two layers.@xcite as a convenience we imagine the leads as consisting of narrow hall bars with separate @xmath6 states in the two layers and associated ingoing and outgoing states which can be viewed as being located to some degree on opposite edges of these leads . we recognize that this is a somewhat idealized model of the incoherent bilayer contacts , but believe that it is general enough to capture all the essential physics . the contacts are indeed realistically viewed as narrow hall bars , but the filling factors in each layer depends on the device layout , especially on the arrangement of the main top and bottom gates.@xcite because there is a single channel in the active region , however , we can in principle always identify a linear combination of lead channels into which electron waves incident from the sample are transmitted and also a linear combination of lead channels which is responsible for all transmission . the incoming and outgoing channels in the leads in our model should be understood as represented these sample - dependent channels . with this understanding , even if other channels are present they do not contribute to transport . when the temperature @xmath7 , this description should apply on the bilayer @xmath8 hall plateau , which can in principle extend over the filling factor range @xmath9 . at finite temperatures there will always be corrections due to thermally excited mobile bulk carriers , and these corrections will be small over a narrower filling factor range . the single edge state in the active area is a certain coherent superposition of the top and bottom edge states , to be specified more completely below . under most circumstances the quasiparticles have equal weight in the two layers corresponding to the pseudospin order parameter lying in the @xmath10-plane . the phase relationship between quasiparticle components in the two layers is dictated by the local direction of the pseudospin effective zeeman field . occupied edge states therefore add to the condensate strength . this picture is summarized schematically in fig.[fig : edgestates ] . occupation of the orthogonal quasiparticle state with the opposite interlayer phase would be pair breaking , _ i.e. _ it would weaken the condensate . this state has higher energy and does not participate in low temperature , low bias - voltage transport . later we consider the possibility of virtual occupation of the pair - breaking edge - state channel when the condensate is weak . coherent quantum hall bilayer . the normal region has two layers , each with a left - going and right - going channel that can transmit electrons toward or away from the coherent layer . on a @xmath6 plateau the coherent region supports a single edge channel with an inter - layer phase relationship that is dictated by the local condensate phase . shaded regions schematically represent places where scattering in the contacts takes place . like a ferromagnet , and unlike a @xmath11 superconductor , the ordered region has conducting quasiparticles . unlike typical ferromagnets , the conducting pseudospins are localized at the sample edge and share the same pseudospin state , in the jargon of magnetism the system is half - metallic . ] disorder gives rise to a random potential difference between edges which mixes edge state channels . this can give rise to memory of the layer from which an electron has been injected that persists along the edge . when layer memory is important for the electrical measurements we need to employ a different formulation of the scattering transport theory which has a larger number of free parameters . the description of scattering at a junction separating two incompressible states , one representing the contact arms taken to have @xmath12 , and one representing the active area having @xmath8 , represents a new and interesting example of the application of landauer - buttiker transport theory to the qhe@xcite regime . in this case the two arms can be alternatively thought of as a combined system with two edge states , or as two separate leads attached to the bulk . in coherent bilayer transport experiments current is often injected into or drained from only one layer . we model separate - contacting to a single layer simply by setting the current to zero for one of the layers . in this way one arm of the contact acts like a voltage probe and the other like a current probe . in experiment the voltage on the open layer can normally not be read externally , although our theory does predict its value . the bulk of the bilayer contains a single edge state , as is appropriate for @xmath8 . the scattering problem near a contact , depicted in fig . [ fig : edgestates ] , is then defined via transmission probabilities from the upper `` coherent '' edge state to top and bottom edge states in the arms , @xmath13 , and from top and bottom edge states into the lower coherent edge state , @xmath14 , as well as the layer - diagonal reflection probabilities ( @xmath15 ) , the cross - layer ( andreev ) reflection probabilities ( @xmath16 ) , and finally the probability for a coherent edge state to bypass a lead , @xmath17 . note that we adopt the convention that the first index of a transmission or reflection probability denotes the outgoing scattering state channel , while the second index specifies the incoming channel . using the above scattering probabilities , one can relate the electrochemical potentials of incoming and outgoing states for the problem illustrated in fig . [ fig : edgestates ] : @xmath18 unitarity and microreversibility@xcite impose the following conditions on the scattering probabilities : @xmath19 eqs . ( [ eq : unitarity ] ) demonstrate that there are only three independent scattering probabilities which fully characterize a lead / coherent sample junction . we choose these to be @xmath20 , and @xmath21 . as a convenience we will assume in our discussion that all contacts are identical ; differences between contacts could readily be accounted for in interpreting experiments if they were adequately characterized . the way a particular contact connects to the outer world specifies additional relations between ` in ' and ` out ' electrochemical potentials . for example , when the top arm is open , one has @xmath22 . this applies equally well to voltage probes . when a voltage probe is connected to the the top layer only , for example , we set @xmath22 as usual . however we also set @xmath23 . even though the bottom layer chemical potential is not _ read _ externally it is still true that no current is flowing to the outside through this arm . contacts that are not layer selective are electrically connected to the same external reservoir and therefore can be represented by setting @xmath24 , whether used as simply as voltage probes or as a current source or drain . when current is allowed to flow between contact arms through an external resistor , the difference between their chemical potentials is proportional to this current . the unitarity and microreversibility relations , eqs . ( [ eq : unitarity ] ) ensure that the voltage probe readings made by probes that contact both layers or either layer individually agree . . there are a large variety of possible measurements in which two of the four contacts are used as source and drain and two as voltage probes . contacts are normally connected to only one of the two layers , as described in the text . ] having described the scattering problem near the contacts , we can proceed to calculate various experimentally relevant quantities . in general the effective single - particle scattering problem for the quasiparticles must be solved together with equations for the condensate . ( the quasiparticle mean - field hamiltonian and hence the transmission probabilities depend on the condensate configuration , which is in turn influenced by the transport currents . ) at small voltages and currents , however , we can neglect this dependence , and treat the transmissions and reflections as constants dependent on the equilibrium condensate configuration . the typical experimental setup is depicted in fig . [ fig : setup ] . in the two - probe case the additional voltage probes are absent and it does not matter which pair of contacts we take as a source and a drain . the three reflection probabilities that characterize the contacts , @xmath25 , and @xmath26 can be determined by three independent measurements which take advantage of the freedom that the current at both source and drain can flow through only one layer or through both . we take these experiments to be _ parallel flow _ , _ drag _ , and _ tunneling _ measurements . in _ parallel flow _ the layers are shorted at both source and drain . in _ drag _ , current flows in from and out through the drive layer ; no current flows through the drag layer arms at either source or drain , or obviously at any voltage probes which might be present . finally , in the _ tunneling _ measurement current is injected into one layer , and extracted from the other one . in most of this paper we assume that the condensate is time independent . for tunneling experiments this requires that currents be smaller than a critical tunneling current.@xcite we find that the parallel , drag , and tunneling linear conductances are given respectively by @xmath27 in the above expression for @xmath28 the skipping probability , @xmath17 , is given by @xmath29 , as dictated by eqs . ( [ eq : unitarity ] ) . the drag conductance was evaluated for the case of a floating drag layer , _ i.e. _ an uncontacted drag layer . the same expression holds in the case of a grounded drag layer with a large ( compared to the critical tunneling current ) total current flowing through the bilayer . when the drag layer is grounded , the condensate phase will be time - dependent as discussed in the next section . one might view as ideal the situation in which the two - layers behave identically , the skipping probability is small , and there is no memory of incoming layer when reflection by the active coherent region occurs . these conditions together imply that @xmath30 , @xmath31 and give ideal values for @xmath32 , and for @xmath33 . the parallel flow two - contact conductance is the usual ideal quantum hall effect value , while the drag and tunnel conductances are half as large . the maximum possible tunnel conductance is actually somewhat larger and reached at @xmath34 , when it becomes @xmath35 . for these values @xmath36 . vanishing normal reflection and maximum andreev reflection seem unlikely in semiconductor bilayers , since reflection in the same layer should be , if anything , more likely than reflection to the other layer . a more plausible extreme set of values might be @xmath30 , @xmath37 and @xmath38 which yields @xmath39 , and @xmath40 . experimental two - probe tunnel resistances in the neighorhood of @xmath41 are indeed commonly observed in coherent bilayers . even for a strong static condensate the maximum tunneling conductance is always associated with a large resistance , which can be viewed as a contact resistance . in this sense there is no ideal dissipationless josephson effect in semiconductor bilayers . the finite resistances do not necessarily , as it is often assumed,@xcite imply that the condensate is lost because of disorder and/or quantum fluctuations . we will return to this point in the the discussion section . measurements of two - probe linear conductances for several different contacting geometries can determine the numerical values of @xmath42 , @xmath43 and @xmath44 , and in this way shed light on the nature of the lead / active region junction . so far we have been concerned with problems with a time - independent condensate . there is a class of experiments , however , in which the contacting geometry and bias voltages lead to a net current flow between layers which exceeds the critical value . the interlayer current flow can either be forced by the contacting geometry , as in counterflow and tunneling experiments , or implied by quasiparticle transport theory . an example of the second case is provided by a drag experiment in which the drag layer is grounded , _ i.e. _ it is connected to the drain contact rather than floating as in the classic drag experiment . the description of the two types of experiments is slightly different : in counterflow and tunneling , there _ must _ be interlayer current , and the full system of equations for the condensate and quasiparticles needs to be solved to find it , as well as the corresponding distribution of interlayer voltages . in the drag experiment with a grounded drag layer , on the other hand , there is always a current path connecting source and drain that lies in the drive layer only . however if a naive landauer - buttiker description with a time - independent condensate predicts an interlayer current larger than the condensate can deliver , there must be an interlayer voltage to ensure that almost all the current flows through the drive layer . we leave calculating the full i - v curve for tunneling experiments to future work,@xcite and concentrate here on a simpler case of a drag experiment with the drag layer grounded at the drain . we will show that the measurement of the interlayer voltage generated in this type of measurement can provide a test of the strong - coherence theory developed in the previous section . for definiteness , let as assume that electrons are injected into the west bottom layer in fig . [ fig : setup ] , and extracted from both north top and bottom layers , going counter - clockwise along the sample edge . application of the landauer - buttiker formalism predicts that the current that exits through the drag ( top ) layer , @xmath45 , is @xmath46 where @xmath47 is the total current through the system . when this current exceeds the critical tunneling current the condensate can not be stationary@xcite and a finite interlayer voltage , developed through interlayer polarization , should be measurable . this additional voltage will reduce the current flowing to ground through the drag layer to a value below the critical current . experimentally , the total current is quite often much larger than the critical tunneling current . in this case interlayer tunneling can be neglected altogether and the interlayer voltage can then be found from the condition that the entire current flows through the drive layer . the latter condition follows from the fact@xcite that for a non stationary condensate collective tunneling becomes an ineffective sink for condensate currents , and the interlayer resistance due to incoherent tunneling is much larger than the resistance associated with the carrying the current in the drive layer only . since we neglect real interlayer electron transfer , we can remove the time dependence of the condensate by switching to a frame rotating with the order parameter . it must be noted that the distribution of interlayer electrostatic potentials has to be calculated taking full account of a bilayer electrostatics , modified by the presence of the condensate.@xcite however , for a well developed qhe the top and bottom edges running from west to north , and from north to west must be equipotentials . moreover , the interlayer voltage must be equal at the above edges , otherwise there would be an ever - increasing counterflow from one edge to the other . in principle , this can be counteracted by counterflow decay due to real tunneling , which we take to be negligible at present , or phase slippage events,@xcite but these should not occur at small currents , since vortices are pinned . hence we conclude that the rate at which the order parameter rotates , @xmath48 , where @xmath49 is its phase must be constant at the system edge , and it is this value of @xmath50 at the edge that is measured by a voltage probe as an interlayer electrochemical potential difference . the latter statement assumes that the quasiparticle dynamics is fast on the condensate time scales , thus the quasiparticles are in equilibrium with the local instant value of the condensate . now let us calculate the value of the interlayer electrochemical potential difference that is necessary for the absence of the current through the drag ( top ) layer . switching to a rotating frame corresponds to a gauge transformation @xmath51 , where @xmath52 is the annihilation operator in the top and bottom layers , respectively . this transformation removes the time dependence of the order parameter by shifting the electrochemical potentials in the top and bottom layers by @xmath53 . since the quasiparticles now see a time - independent spontaneous tunneling amplitude , we can apply landauer - buttiker formulae to the above shifted electrochemical potentials substituting @xmath54 in eq . ( [ eq : chempots ] ) . note that @xmath55 remains unchanged . the electrochemical potential difference is found from the condition that there is no current flowing out of the top layer at north , and the _ unshifted _ electrochemical potential in north top arm coincides with the drain electrochemical potential . the result is @xmath56 where , as before , @xmath47 is the current flowing through the bilayer . we note that the conductance in this case is still given by the corresponding expression in eq . ( [ eq : conductances ] ) . the interlayer voltage , eq . ( [ eq : interlayerv ] ) , depends on the same set of reflection probabilities that can be , in principle , extracted from the two - probe conductances measurements . ( see eqs . ( [ eq : conductances ] ) ) . confirmation of this prediction would provide strong confirmation that the strong coherence regime modeled in the previous section can be achieved experimentally . when the charge gap responsible for the quantum hall effect is reduced , disorder enables quasiparticles to deviate from perfect chiral routes around the sample perimeter and travel across the bulk . processes of this type , although always present at finite temperatures , are extremely weak when the quantum hall effect is well developed . the bulk scattering events which lead to deviations from a perfect quantum hall effect are normally well characterized by the value of the longitudinal resistance measured in a four - probe geometry . in this section we show that in coherent bilayers , longitudinal voltages can develop in both drag and drive layers when the current source and drain contacts are layer selective , even when transport is perfectly localized at the edge and the quantum hall effect is still perfect when the source and drain contacts are not layer selective . the longitudinal voltages emerge when interlayer coherence is weak enough to permit evanescent conduction in the _ pair - breaking _ bilayer edge channel , leading to a non - negligible layer - memory length . below we explain a model intended to describe electrical measurements in this regime , which we refer to the weak coherence regime . we start this discussion of deviations from the strong coherence case by briefly reviewing the main assumptions made in the previous section . there it was assumed that an electron arriving at a voltage probe travelled in a single chiral channel localized along the edge of the active coherent region . under such an assumption , unitarity of scattering near the probe ensures that selective voltage probes connected to either layer measure the same electrochemical potentials . on the other hand , electrons propagating along the aforementioned edge channel could have come from either top or bottom layers of , say , an upstream current probe , and these upstream probes could have different electrochemical potentials in top and bottom layers . in order to allow for the possibility of layer memory being retained between probes , we have to weaken our assumptions about electron propagation between probes . losing _ which layer _ memory is , of course , a process of quantum mechanical scattering , which happens on a certain length scale around a contact . up to now , we have been assuming that this length scale is small , which corresponds to what we are referring to as the strong coherence regime . as coherence is weakened , say , by increasing temperature or changing the total filling fraction , one can imagine a situation when this _ which layer _ memory length becomes considerable and eventually comparable to the system size . indeed , as we explain below , there are good reasons to expect that disorder can cause this length to become comparable to sample dimensions even when the normal quantum hall effect is still reasonably well established . voltage probe readings will then depend on which layer is contacted . quasiparticle transport in this regime is most generally described by a phenomenology which allows all possible chiral transmission coefficients between neighboring two - layer contacts . the most general model consistent with a perfect @xmath6 integer quantum hall effect in the active region , when normally contacted , is one in which outgoing electrons in a lead are either reflected within the same lead or transmitted to the nearest downstream contact but satisfy the constraint @xmath57 below we will construct a one - parameter landauer - buttiker - type parameterization to describe the weak coherence regime , which is less general but physically transparent . in order to capture layer memory we need to relax the microscopic single - channel edge model for the active region edge . in order to explain the main ideas we apply this approach mostly to the case of drag experiments , although the same type of analysis could ( and has been ) applied to tunnel , counterflow , and a wide variety of other contacting arrangements . longitudinal drag always accompanies weakening coherence in bilayers , and has previously been analyzed using an effective medium theory due to stern and halperin which associates longitudinal drag with loss of coherence in the bulk of the system.@xcite we show here that longitudinal drag should still be expected even when transport currents still flow perfectly chirally at the edge of the system and the normal quantum hall effect is still accurately quantized . we expect our model to be most relevant at low temperature and small @xmath58 when the bulk effective medium theory appears to fail.@xcite our approach provides a complementary understanding of the large longitudinal drag resistances that can appear in coherent bilayers . in what follows we assume that current is injected into and drained from the bottom layer so that the top layer is the drag layer . we assume that electrons flow counter - clockwise around the sample as in fig . [ fig : setup ] . quantities related to the edge running from the electron source to the electron drain ( simply source and drain from now on ) are labeled with the superscript `` + '' , and those for the edge running from drain to source - with the superscript `` - '' . for the `` + ' ' ( `` - '' ) edge , the coordinate along the edge , @xmath59 , is counted from the source ( drain ) . correspondingly , the distances from source to drain ( along `` + '' edge ) , and from drain to source ( along `` - '' edge ) are denoted @xmath60 . for simplicity , we take all contacts to be identical . finally , we neglect the `` skipping '' probability @xmath17 , and the `` andreev '' reflection probability , @xmath26 ( see eqs . ( [ eq : unitarity ] ) ) . complications which arise when these assumptions are relaxed will be discussed below . to describe the loss of _ which layer _ memory with a minimal set of parameters , we assume that electrons reflect in the same layer in the leads with probability @xmath61 . these values are appropriate ( dictated by eqs . ( [ eq : unitarity ] ) ) near the strong coherence limit if andreev and skipping probabilities are neglected , and there is a symmetry between layers . to specify position - dependent transmissions , we assume that if a voltage probe is physically close to a lead , electrons retain full memory of the lead from which they originate , _ i.e. _ @xmath62 , while @xmath63 . for a voltage probe located far from a lead , on the other hand , the layer memory is completely lost , @xmath64 . the simplest single - length - scale and layer - symmetric@xcite _ ansatz _ for the loss of _ which layer _ memory is as follows : @xmath65 where @xmath66 is the characteristic memory - loss length scale . a similar _ ansatz _ was used previously@xcite to describe bilayer transport at total filling factor @xmath67 . having defined the transmission probabilities , we can now write down the equations for the voltmeter readings . given top and bottom layer incoming electrochemical potentials @xmath68 and @xmath69 , the equations for voltmeters readings along the edges are @xmath70 when these equations are written in the form @xmath71 the matrix @xmath72 satisfies the property @xmath73 . thus , the presence of a voltmeter does not alter readings of any voltmeters placed downstream . we can therefore characterize a given edge by a position dependent chemical potential . ( this property can be traced to our assumption in this section that andreev scattering in the contact regions is not important . ) to close eqs . ( [ eq : chempot_xdep ] ) we need to relate @xmath74 to @xmath75 and the current flowing through the system . taking into account that @xmath74 correspond to readings of voltmeters situated just before the west ( source ) and east ( drain ) leads , respectively , and that the top arm is open at east and west , we can apply eqs . ( [ eq : chempot_xdep ] ) for @xmath76 to get the boundary conditions , @xmath77 the expression for the current , defined as the one flowing through the bottom arm of west or east contacts is @xmath78 eqs . ( [ eq : chempot_xdep ] ) , ( [ eq : boundarycond ] ) , and ( [ eq : totalcurrent ] ) allow for a full characterization of electrical measurements on a bilayer under the conditions of weak coherence , and imply a well defined @xmath8 qhe when the layers are regularly contacted . the considerations above are simplified in several ways , but have the advantage that they leads to a theory with a single parameter , which can be fixed by one measurement and compared with the wide variety of alternate measurements on the same sample that differ only by switching contacts between top and bottom layers or using a probe to contact both layers . more complete theories depend on aspects of the contacts that are not characterized in typical experiments . in what follows we present examples of quantities that can be calculated in the present simple framework for the typical experimental geometry of fig . [ fig : setup ] . we then compare our predictions to the experimental results of refs . [ ] , [ ] , and [ ] , where longitudinal and hall drags are measured as a function of total filling factor@xcite or effective layer distance.@xcite the longitudinal drag resistance is normally defined in terms of the difference between probe voltage values measured in the drag layer and along the same edge . the hall drag is defined similarly , except that the voltage probes are placed on opposite edges . the geometry of kellogg _ et al . _ @xcite corresponds to that described in fig . [ fig : setup ] . the longitudinal drag resistance is measured with source and drain being at the west and north contacts , while voltage probes situated at the south and east contacts , fig . [ fig : setup ] . if we denote the length of the square mesa edge as @xmath79 , in this setup @xmath80 and @xmath81 . the longitudinal drag resistance is then @xmath82 . the hall resistance is measured between the south and north contacts with the source and drain being at the west and east , respectively . thus in the hall drag measurement @xmath83 , and the hall drag resistance is given by @xmath84 . it is natural to express the resistances in terms of the parameter @xmath85 , the probability of retaining layer memory while transiting one edge of the sample , hence _ p - ology_. by solving eqs . ( [ eq : chempot_xdep ] ) , ( [ eq : boundarycond ] ) , and ( [ eq : totalcurrent ] ) one obtains the following expressions : @xmath86 the values of these drag resistances do not depend on whether the drag layer is grounded or not ( as in section [ sec : timedepcond ] we assume a current much larger than critical tunneling one flowing through the bilayer ) . longitudinal drag resistance also does not depend on a particular chirality of electron motion around the sample ( assumed counter - clockwise in fig . [ fig : setup ] ) , while the hall drag resistance changes its sign with a change in the chirality . the configuration used for the measurement of the longitudinal drag resistance described above is also convenient to measure the four - probe interlayer resistance . this resistance is defined in the same way as the longitudinal drag resistance , except the voltage at the east contact is measured in the drive layer : @xmath87 . this resistance does depend on whether the drag layer is grounded or not , and direction of electron motion around the sample . to proceed we combine the present considerations with those from section [ sec : timedepcond ] . then in the case of a counter - clockwise propagation ( as shown in fig . [ fig : setup ] ) we obtain @xmath88 while for clockwise propagation ( _ i.e. _ the opposite to the one shown in fig . [ fig : setup ] ) we get @xmath89 here the superscripts specify the drag configuration that the interlayer resistances pertain to . we observe that in the case of a floating drag layer the interlayer resistance is an odd function of the magnetic field , while in the grounded drag layer case it does not have a definite parity for @xmath90 . we note once again that the expression for the interlayer resistance for the grounded drag configuration assumes that the total current floating through the bilayer is much larger than the critical tunneling one , which typically holds in drag experiments . finally , we can also make predictions for two - probe quantities , in particular , for how the values of conductances described in section [ sec : lb ] change with deviation from the strong coherence limit . the most interesting cases measure tunneling conductances . we assume a tunneling geometry that is identical to that of the hall drag measurements described above , except that source and drain contacts are attached to opposite layers . this leads to @xmath91 this value applies to currents small compared to the critical one . we emphasize that all expressions in eqs . ( [ eq : resistances ] ) , ( [ eq : resis_interlayer ] ) , and ( [ eq : condpology ] ) are expected to be more reliable close to strong coherence , @xmath92 , where the quantum hall effect in the parallel channel is well - developed , and transport is edge - dominated . for instance , the formal limit @xmath93 in the expression for @xmath94 is meaningless . it is derived under the assumption that the condensate must be dynamical in the grounded drag case when the total current is much larger then the critical tunneling current , in order to prevent current from floating through the drag layer , as would be predicted by naive close - to - strong - coherence landauer - buttiker theory . ( see section[sec : timedepcond ] for the discussion of this point . ) we observe that @xmath95 in eq . ( [ eq : resistances ] ) is a positive non - monotonic function of parameter @xmath96 , vanishing both for @xmath97 ( strong coherence ) and @xmath93 ( weak coherence ) . it reaches a maximum of @xmath98 at @xmath99 , or @xmath100 . the positiveness of @xmath95 as defined above implies that the sign of the voltage drop in the drag layer is opposite to the sign of the longitudinal voltage in the drive layer , as found experimentally . since @xmath95 and @xmath101 depend on a single parameter , the relationship between @xmath95 and @xmath101 is universal in this simplified theory . one quantity of interest characterizing the rate at which longitudinal drags grow as the hall drag deviates from perfect quantization : @xmath102 in refs . [ ] and [ ] a maximum drag resistivity of @xmath103 was observed at the lowest temperatures ( @xmath104mk ) both by varying the total filling and by varying the effective layer distance , and the extracted slope of the parametric @xmath105 dependence was @xmath106 . so we see the in the case of aforementioned experiments the present theory predicts a longitudinal drag about twice as large as observed . the discrepancy is most likely due to the fact that in experimental systems deviations from perfect parallel transport hall conductance and loss of layer memory appear together . ( this point is discussed at greater length in the following section . ) when the ordinary hall conductance is not perfectly quantized some of the transport is able to travel from source to drain without passing the longitudinal hall voltage probes , reducing the measured voltage difference . although eqs . ( [ eq : resistances ] ) are not in quantitative agreement with experiment , they do clearly demonstrate that longitudinal drag _ must _ accompany deviations from perfect hall drag quantization . large longitudinal drags do not imply phase separation into coherent and incoherent regions , and , in our view , this macroscopic phase separation is unlikely to be present in samples studied experimentally . we note that the strong coherence limit eq . ( [ eq : condpology ] ) gives a value of tunneling conductance close to the one observed in experiment at low temperature and small @xmath58(e.g . in refs . [ ] and [ ] ) . however , the conductance deteriorates much faster with decreasing coherence than predicted by eq . ( [ eq : condpology]),when the value of @xmath96 is taken from hall drag measurements . ( this dependence is reported in ref . this may indicate that the critical current has decreased to a value smaller than the smallest measurement current levels used in these experiments . in ref . [ ] a rectangular sample was used , and the longitudinal and drag resistivities per square , @xmath107 and @xmath108 ( @xmath109 and @xmath79 being the width and length of the active area ) were reported . for the sample studied @xmath110 . the explicit expressions for resistances for this case are not very instructive so we present our findings graphically , comparing the parameter - free implicit dependence of @xmath111 on @xmath112 predicted here , with the experimental results reported in ref . [ ] and with the theoretical results in ref . [ ] . we note that close to the strong coherence limit ( @xmath113 ) and at low temperatures our curve appears to be close to the one observed in experiment , whereas at higher temperatures the one predicted by stern and halperin@xcite is a good match , as illustrated in fig . [ fig : resistances ] ( note a change in the sign convention for the hall drag resistance ) . in the experiment of ref . [ ] the spontaneous coherence state is destroyed by changing filling factor , rather than effective layer separation . we still apply our contact model to predict the outcome of four - probe measurements though , as discussed in section [ sec : lb ] . on @xmath112 . solid lines represent the experimental data of ref . [ ] , where the filling fraction was varied from @xmath114 to @xmath115 . the maximum in the longitudinal resistivity at low temperatures is reached around @xmath116 the dotted line corresponds to the semicircle law of ref . [ ] , the dashed line is the prediction of this paper . ] we can conclude from this analysis that the presence of a non - zero longitudinal drag and an imperfectly quantized hall drag alone do not necessarily imply that inter - layer coherence is lost over a large fraction of the sample area . in this section we present a theoretical estimate of the value of layer - memory - loss length , @xmath66 , introduced in the previous section , that is based on a highly simplified model of electron propagation along the edge and intended to capture the essential roles of layer asymmetric disorder and interlayer coherence . we map the problem of transport electrons propagating along an edge onto that of bilayer electrons in a single chiral charge - carrying channel which propagates at the magnetoplasmon velocity , @xmath117 . layer antisymmetric disorder and interlayer coherence are taken into account by fields that act on the @xmath118 and @xmath119 components of the _ which layer _ pseudospin degree - of - freedom . the schrdinger equation for the problem reads @xmath120 where , as before , @xmath59 is the coordinate along the edge . without loss of generality , we can assume that the part of the exchange field due to the interlayer coherence has only a @xmath121 component and for simplicity we take it to be spatially independent . the @xmath122 term in eq . ( [ eq:1dse ] ) accounts for the difference between the disorder potentials in the two layers ; we assume that @xmath123 changes sign on the disorder correlation length scale @xmath124 , which is usually assumed to be comparable to the distance to the dopant layer and therefore much longer than a magnetic length . because it is first order in derivatives , eq . ( [ eq:1dse ] ) maps to the time dependent schrdinger equation for a local spin in a time dependent magnetic field with components @xmath125 . for weak coherence , the regime of present interest , @xmath126 , so that the _ local _ eigenstates of eq . ( [ eq:1dse ] ) correspond to a definite layer index almost everywhere in space ( or at almost all times in the local spin mapping ) , except for branching points at which the disorder potential changes its sign . near such points , we can map eq . ( [ eq:1dse ] ) to the well - known landau - zener problem,@xcite in which an electron has a probability to follow the instantaneous trajectory ( _ i.e. _ retain its layer index ) and a probability of following the adiabatic path under the influence of @xmath127 , changing its layer index . the large value of @xmath117 favors the former process can be dominant and prevent an electron from losing its layer memory . at each landau - zener branching point , _ i.e. _ each instantaneous level crossing , the probability to follow the nonadiabatic path and stay in the same layer is @xmath128 correspondingly , the probability to stay on the adiabatic path is @xmath129 , which would mean retaining the layer index . we introduce probabilities @xmath130 for an electron to be in the locally high and low energy states in a region where @xmath131 after @xmath132 landau - zener branching events . using the branching rules we find that these probabilities satisfy @xmath133 it follows that the transport quasiparticle layer polarization @xmath134 satisfies @xmath135 in the continuum limit . we identify the layer - memory loss length with the characteristic length which emerges from this analysis of polarization decay : @xmath136 since the edge magnetoplasmon velocity in large hall bar samples@xcite has a logarithmic enhancement related to the long - range of the coulomb interaction @xmath137 . furthermore @xmath72 can be @xmath138 due to weakened screening on hall plateaus , whereas @xmath127 has a maximum value @xmath138 when coherence is complete . from these estimates it is clear that the layer memory length on the right hand side of eq . ( [ eq : memorylength ] ) must become large as interlayer coherence is weakened , leading to longitudinal and hall voltages whenever current source or drain contacts are layer selective as explained in the previous section . these layer memory effects can in principle be separated experimentally from normal deviations from perfect quantum hall effects , also expected to become stronger when coherence weakens , by comparing measurements made with layer selective and non - selective source and drain contacts . in this article we have developed a landauer - buttiker description of quasiparticle transport in separately - contacted bilayer quantum hall superfluids . for well developed condensates we first assume that current is carried between leads by a conducting channel consisting of quasiparticles that occupy both layers simultaneously and have inter - layer phase coherence . when this assumption applies , voltages measured along the sample edge have the same value for probes that contact both layers and probes that contact either of the individual layers . we refer to this regime as the _ strong coherence _ regime . we show that two - probe conductances are then dependent on the three independent landauer - buttiker scattering probabilities that characterize the link between two - layer leads and the condensate edge . assuming that all contacts to the system are identical , these three parameters can be determined by separate two - probe conductance measurements for parallel , drag , and tunnel contacting geometries . the two - probe linear conductance values predicted in this approach tend to range from @xmath139 to @xmath140 depending on the contacting geometry and the contact scattering properties . indeed these values appear to be generally consistent with coherent bilayer experiments performed at low temperature and small @xmath58 values.@xcite at higher temperatures and larger @xmath58 conductances tend to decrease . a part of the decrease is readily accounted for by deviations from a perfect quantum hall effect due to scattering across the bulk of the sample , which is always present at finite temperature and beyond the scope of the present theory . tunnel conductances decrease more rapidly than parallel transport or two - probe drive - layer conductances in drag experiments however . this observation points to the development of layer memory in quasiparticle transport . we have therefore elaborated our strong coherence scattering formalism to allow for layer memory , which we relate microscopically to disorder at the edge and resulting evanescent transport in the pair - breaking bilayer edge - state channel . this picture leads to a layer memory length which is proportional to @xmath141 , where @xmath127 is the bulk quasiparticle gap of the coherent state . in addition to accounting for two - point tunnel geometry conductances that are much smaller than @xmath140 , this analysis shows that longitudinal drag voltages and deviations from quantized hall drag voltages develop when layer memory lengths become comparable to distances between contacts . we refer to this circumstance as the _ weak coherence _ regime . in this way we are able to establish a relationship between hall drag voltages and the longitudinal drag voltages which appear at weak coherence which is qualitatively consistent with experiment.@xcite this picture also accounts for the sign of the longitudinal drag voltage which is opposite to the sign of the longitudinal voltages in the drive layer . we have considered the transport properties of coherent bilayers in both time - independent condensate , and time - dependent condensate ( interlayer current above the critical tunneling current ) cases . in either regime measured voltages reflect the properties of fermionic quasiparticle waves at the interface between the sample and bilayer leads . when the condensate is time - dependent , differences between chemical potentials measured by leads attached to different layers can reflect mainly this time dependence , and therefore the physics which controls its value in a particular geometry . generally , the phase variation rate assumes a value sufficient to pump the required amount of current between layers via external electrical connections between layers , via inelastic quasiparticle tunneling between layers ( the mechanism which appears to dominate experimentally at large bias voltage ) , or via local dynamic fluctuations of the condensate ( as in the model used by hyart and rosenow,@xcite discussed further below ) . in some cases a finite phase variation rate is necessary to make sure that current _ does not _ flow between layers , as in the grounded drag configuration discussed in section [ sec : timedepcond ] of this paper . before comparing the present considerations to related theoretical work , we would like to point out that a substantial part of the literature on the transport anomalies of coherent quantum hall bilayers addresses tunneling@xcite and drag@xcite experiments separately . in our view , an adequate understanding of the superfluid transport anomalies can be achieved only by describing all possible transport measurements within the same framework . this point is emphasized further below . one approach that is usually applied only to tunneling experiments accounts for collective fluctuations of bosonic particle - hole pairs.@xcite these theories assume that tunneling can be treated as a weak perturbative correction to a condensate with an interlayer phase precessing at an average rate proportional to the inter - layer potential @xmath142 , _ i.e. _ @xmath143 . this picture is close to the one we employ when the tunneling current is much larger than its critical value , except that the fermionic quasiparticle degrees of freedom on which we focus do not appear explicitly . we return to this point below . a perturbative treatment of tunneling can be valid only when the system does not possess _ spontaneous _ long - range order . indeed the absence of spontaneous long range order , even at @xmath11 , appears likely in this system because of the presence of meron condensate textures that are known to be induced by potential fluctuations . a recent careful examination of a large body of experimental data by hyart and rosenow@xcite has shown that tunneling and closely related loop counterflow tunneling experiments can be accounted for by a weak - tunneling bosonic picture with phenomenological phase correlation lengths and times , presumed to be related respectively to the typical distance between merons and the time scale of their fluctuations . this picture leads to a non - linear bilayer tunneling i(v ) curve with a characteristic current scale @xmath144 proportional to the bare tunneling amplitude @xmath145 , a maximum differential conductance at zero bias that is proportional to @xmath146 , where @xmath147 is the phase fluctuation correlation time , and a region of sharply negative differential resistance . in comparison , the picture which underlies the present work is one in which the characteristic current is a critical current below which the condensate is time - independent , and the differential conductance at zero bias is limited by quasiparticle scattering properties to values below around @xmath148 in the strong coherence regime and to smaller values in the weak coherence regime . we will argue below that the characteristic current scale should be the same size in both pictures . nevertheless it should be possible to distinguish the two pictures experimentally . as we have previously emphasized,@xcite we have assumed that the condensate can be time - independent if the net current carried between layers inside the active region of the coherent bilayer is sufficiently small . when the condensate is time - independent , voltage measurements reflect only elastic quasiparticle scattering rates . for example all drag transport measurements ( including drag - counterflow@xcite ) in which there is no net current flow between layers can be consistently understood in terms of the scattering properties of quasiparticles alone . for tunneling , series counterflow,@xcite and other measurements in which there is net current - flow between layers , the condensate can be static only if the supercurrent injected into the sample by andreev quasiparticle scattering can be sunk by time - independent coherent condensate tunneling . the critical current in essence@xcite is the maximum possible value of @xmath149 where @xmath150 is the full sample area . here @xmath132 is the condensate density which is equal to @xmath151 in the absence of disorder and quantum fluctuations . it is quite clear that the experimental critical currents , if non - zero , are much smaller than the value which would be obtained using this value of @xmath132 and setting @xmath152 everywhere . it is natural to ascribe the large reduction in critical currents to the presence of the merons,@xcite which are known ( see refs . [ ] and [ ] and references therein ) to be induced by disorder in coherent bilayers and assumed to be present in all experimental samples . since the phase of the condensate winds by @xmath153 around a meron core , setting the spatially average value of @xmath154 , it is clear that the presence of merons will drastically influence the critical current . it seems clear that the presence of disorder in bilayers implies that long - range spontaneous phase order is not present even at @xmath11 . as first emphasized in ref . [ ] , and later detailed in ref . [ ] , the evolution of transport anomalies with in - plane fields suggests a correlation length @xmath155 , possibly related to the distance between merons , which is comparable to the spatial separation distance between the bilayer and the set - back dopant layer . long - range phase order will , however , certainly be induced in response to the non - zero inter - layer tunneling amplitude present in every sample . assuming that the degree of long - range order is determined by a balance between the coupling of the condensate to inter - layer tunneling , and the exchange fields associated with meron - induced condensate phase variation , we conclude that the effective condensate density is reduced from its uniform system value by a factor of @xmath156 , where @xmath157 is the condensate stiffness . with this modification @xmath158 in eq . ( [ eq : isink ] ) gives as a critical current , the same characteristic interlayer current value that emerges from the perturbative condensate tunneling theory . we remark that this critical current estimate is not sensitive to the detailed meron - core configuration , so thermal or quantum fluctuations in meron positions will produce small relative fluctuations in critical current values . in summary , the transport anomalies of coherent quantum hall bilayers are similar to those of superconductors connected to normal metal leads , but are enriched by the possibility of contacting the two - layers separately , and by the presence of processes including transport through external conducting links , which violate separate particle - number conservation in the two layers . this paper describes a conceptual framework for describing these anomalies which combines a landauer - buttiker picture of quasiparticle transport with a description of pseudospin - pumping by time - dependent condensates . coherent bilayers are very sensitive to disorder because it tends to induce merons ( vortices ) in the order - parameter field and limit the maximum counterflow supercurrents which can be sunk by coherent interlayer tunneling . we have argued that critical currents remain finite even when merons are present because of the distortion of meron configurations by inter - layer tunneling . the authors acknowledge helpful interactions with w. dietsche , p. r. eastham , j. p. eisenstein , a. d. k. finck , b. i. halperin , k. von klitzing , b. rozenow , l. tiemann , e. tutuc , and y. yoon . we are grateful to j.p . eisenstein , a. d. k. finck , and e. tutuc for providing their experimental data . p - ology _ discussion in section [ sec : pology ] elaborates on ideas developed in collaboration with j.p . this work was supported by welch foundation grant tbf1473 , and the nsf - nri swan project .

in the successful standard cosmological model where structure in the universe originates from gaussian random density fluctuations in the initial conditions , all statistical properties of cosmological structure observables depend on a single quantity : the linear power spectrum of mass fluctuations . the evolution of this spectrum depends on the properties of the dark energy in a precisely calculable way . the task of extracting dark energy information from cosmological structures reduces to determining the relationship between observables and the underlying linear mass power spectrum . deep , multi - color photometric galaxy surveys measure two sets of cosmological observables : the angular distribution of the galaxies and the weak lensing shear induced on their shapes . from these observables , three types of two - point correlations can be constructed : the angular correlations between the positions of the foreground galaxies , the shear - shear correlations between background galaxies , and the galaxy - shear cross correlations induced by the association of dark matter with foreground galaxies . whereas these correlations have so far been analyzed separately and/or with data from different surveys , the joint analysis of these measurements will be feasible from forthcoming surveys . in this paper , we consider what can be learned from the combined two point correlations . we shall see that with the multitude of observables available , prospects for the joint determination of the cosmology and the relationship between galaxies and mass are bright . galaxy - shear cross correlations , also known as galaxy - galaxy lensing correlations , were first detected by brainerd _ @xcite , following earlier upper limits @xcite . the red - sequence cluster and virmos - descart surveys ( @xcite , also @xcite ) are examples of the current state of the art in deep surveys . though shallower , the sloan digital sky survey ( sdss ) complements these given the larger population of foreground and background galaxies @xcite . these observations have been interpreted in terms of the dark matter distribution associated with the lensing galaxies and their environment under the halo model for galaxies @xcite . ongoing and future surveys such as the canada france hawaii legacy survey , pan - starrs , lsst , and snap @xcite will produce multi - color catalogs of galaxies . photometric redshifts can then be estimated for these galaxies , allowing for measurements with foreground galaxies extending to @xmath6 and background galaxies up to twice as far in multiple redshift bins . photometric redshifts significantly augment the prospects for joint galaxy - lensing studies beyond the current state - of - the - art . the number of observable cross - correlation functions between the galaxies and the shear scales as the product of the number of lens and source redshift bins and can easily exceed the number of galaxy - galaxy clustering observables and independent shear - shear correlation observables . we consider prospects for constraining the evolution of the dark energy through such surveys . the combination of galaxy - shear and galaxy - galaxy correlations is particularly fruitful in that they allow for a joint solution of the evolution of the matter and galaxy distributions . these determinations can be cross checked against those from the shear - shear correlation . the latter depend only on the mass power spectrum but are typically subject to more severe systematic uncertainties . we begin in [ sec : statistics ] with a description of the statistical methods employed . these methods may be applied to any model of the two - point correlations . in the appendices , we describe the parameterization of the linear mass spectrum in the standard cosmological model , including convenient fitting functions for the dark energy effects , and a generalization of the halo model for the association of galaxies with the mass . our model for the latter is based on recent developments in galaxy simulations which rely on matching the observed number density of galaxies to the predicted number density of halos @xcite . utilizing these parameterized statistics , we study prospects for constraining the properties of the dark energy in a spatially flat universe in [ sec : darkenergy ] . we conclude in [ sec : discussion ] . in this section , we describe the basic statistical approach to the joint study of galaxy and lensing power spectra . these methods are independent of the specific parameterization of the power spectra described in the appendix and employed in the following section . we begin with a brief review of the relationship between angular and three dimensional power spectra in [ sec : power ] . we relate these to the traditional galaxy - galaxy lensing observables in [ sec : cross ] . in [ sec : fisher ] we describe how the sensitivity of power spectra to underlying parameters may be quantified . under the assumptions of statistical isotropy and small angles , the two - point observables of a set of two dimensional scalar fields @xmath7 , where @xmath8 represents the direction on the sky , is given by their angular power spectra @xmath9 where @xmath10 is the fourier wavevector or multipole @xmath11 here and throughout we use the same variable to represent the field in angular and fourier space . suppose these angular fields are related to three dimensional source fields @xmath12 by a weighted projection @xmath13 where @xmath14 is the angular diameter distance in comoving coordinates . we will use semicolons to denote arguments that will be suppressed where no confusion will arise . the limber approximation @xcite then relates the two dimensional power spectra to the three dimensional power spectra as @xmath15 where @xmath16 is the hubble parameter and @xmath17 defines the three dimensional source power spectrum . for our purposes , the two dimensional fields will be the `` lens '' galaxy number density fluctuations and the electric or @xmath18 component of the weak lensing shear field measured with `` source '' galaxies . the lens and source galaxies of a given survey may be divided into bins according to redshift , luminosity , color or other criteria . we will here concentrate on redshift binning . however for notational convenience we will suppress the subscripts @xmath19 denoting the bins for the rest of this section . for the galaxy fluctuations , the source field is the three dimensional number density @xmath20 or rather its fluctuations @xmath21 and the weight for the angular fluctuation field @xmath22 is the normalized redshift distribution function @xmath23 where the normalization factor @xmath24 is the angular number density in @xmath25 . note that the weights are normalized so that @xmath26 . for the weak lensing shear , the underlying scalar field is the electric component of the shear field @xmath27 which manifests itself as a shearing of background galaxy images according to the complex shear @xmath28 where @xmath29 is the azimuthal angle of the fourier vector with respect to the @xmath30 axis . the @xmath18 field itself is a projection of the mass density fluctuation @xmath31 and hence is equal to the convergence @xmath32 . the weights are given by the efficiency for lensing a population of source galaxies @xmath33 where the angular diameter distance to the lens is @xmath34 , to the source @xmath35 and between the lens and the source @xmath36 . here @xmath37 is the comoving coordinate distance and note that @xmath38 in a flat universe . the distribution of source galaxies @xmath39 need not be the same as for the ( lens ) galaxies above . furthermore @xmath40 , the normalized redshift distribution , is the direct observable so that the efficiency @xmath41 for a known @xmath40 may be used to probe cosmology . the complete two point statistics of the shear and galaxy correlations are thus specified by a choice of cosmology and a description of the underlying three dimensional power spectra @xmath42 , @xmath43 and @xmath44 as a function of wavenumber @xmath45 and redshift @xmath46 . the latter two will depend not only on cosmology but also on galaxy properties . the galaxy - shear cross power spectra are related to the more familiar cross correlation functions through the fourier transform relations eqn . ( [ eqn : fouriertransform ] ) , ( [ eqn : shear ] ) @xmath47 where @xmath48 is the angular separation vector with magnitude @xmath49 and azimuthal angle @xmath50 with respect to the @xmath51 axis . we have used the identity @xmath52\,.\ ] ] note that the correlation functions depend on both the magnitude of the separation vector @xmath49 and the azimuthal angle @xmath50 . despite this complication , ( [ eqn : pixelcorrelation ] ) can be straightforwardly used to generalize a maximum likelihood estimator for the shear - shear angular power spectrum ( e.g. @xcite ) . although the galaxy - shear cross power spectrum is thus a direct observable , observations to date have focused on the tangential shear component around galaxies due mainly to systematic effects in the shear measurement . it is therefore useful to relate the two approaches . we can express the tangential shear about a galaxy at the origin as @xmath53 the angular correlation function then becomes a function of @xmath49 alone and is given by @xmath54 under the ergodic assumption , this quantity can be reinterpreted as the average tangential shear around lens galaxies in the sample volume . for a narrow redshift distribution of source and lens galaxies , the tangential shear directly probes the galaxy - mass power spectrum at the lens redshift . substitution of the limber equation ( [ eqn : limber ] ) in eqn . ( [ eqn : tangential ] ) gives @xmath55 where @xmath56 pc@xmath57 and @xmath58 is the distance transverse to the line of sight . the critical surface density is given by @xmath59 for a single lens galaxy with an azimuthally symmetric density profile @xmath60 , these quantities are related to the projected or surface mass density @xmath61 through @xmath62 , where the average is over transverse distances interior to @xmath58 . note that all distances are in comoving coordinates . the tangential shear technique throws away information by combining the two components of the shear before averaging . it contains the majority of the signal when considering a spherically symmetric density distribution about a galaxy . at large radii , the usual approach is to attempt a reconstruction of the scalar @xmath63 ( @xmath64 , the convergence in the weak lensing limit ) out of both components ( e.g. @xcite ) . this reconstructed field then acts as a template density map for the correlation @xmath65 alternately , one can go the other way and define a template shear @xcite @xmath66 and construct the correlation @xmath67 formally the two techniques construct the same correlation function and preference for one versus the other is a matter of considering errors in the reconstruction . given angular power spectra that are defined by a set of cosmological and galaxy parameters @xmath68 , forecasts on how well such parameters can be extracted from the data is an exercise in error propagation . in general , the observed two points statistics of the angular fields @xmath69 will receive a contribution from noise sources which we will assume to be statistically isotropic @xmath70 we will further assume that the noise contributions for the galaxy and shear fields arise from uncorrelated shape and shot noise @xmath71 where @xmath72 is the rms shear in each component arising from the intrinsic ellipticity of the galaxies and measurement noise . the shear fields are expected to be nearly gaussian with respect to power spectrum errors for @xmath73 due to linearity and projection @xcite as borne out in simulations @xcite . for the galaxy fields , the transition scale is at somewhat lower @xmath74 depending on the width of the projection ( e.g. @xcite ) . however for both , the noise contributions at @xmath75 dominate the errors and mask the non - gaussianity of the underlying fields . under the assumption of gaussian statistics for the fields , the information contained in the power spectra can be quantified by the fisher matrix @xmath76\ , , \label{eqn : fisher}\ ] ] where the sum is over bands of width @xmath77 in the power spectrum and @xmath78 is the amount of sky covered by the survey . the rough @xmath78 scaling is valid for contiguous regions with comparable extent in each of the two angular directions . here we have suppressed the @xmath79 indices in a matrix notation and @xmath80^{ij } \equiv d_{l\alpha}^{x_i x_j } = { \partial c_l^{x_i x_j } \over \partial p_\alpha}\ , . \label{eqn : derivs}\ ] ] the inverse fisher matrix approximates the covariance matrix of the parameters @xmath81 . one can also break this information down into subsets @xmath82^{-1 } { d}^{x_m x_n}_{l \beta } \label{eqn : fishersub}\ ] ] where sum over @xmath83 and @xmath84 pairs can run over a restricted subset of the observables . the covariance matrix of the subsetted power spectra is given by @xmath85^{ij , mn}= \tilde c_l^{x_i x_m } \tilde c_l^{x_j x_n } + \tilde c_l^{x_i x_n } \tilde c_l^{x_j x_m}. \label{eqn : crosscov}\ ] ] in the limit that the sum is over all combinations , eqn . ( [ eqn : fishersub ] ) returns eqn . ( [ eqn : fisher ] ) . this subsetting also clarifies the role of the gaussian assumption . gaussianity implies a diagonal covariance matrix in @xmath74 and reduces its form to the product of power spectra in eqn . ( [ eqn : crosscov ] ) . non - gaussianity from non - linear structure formation correlates the power between high @xmath74 bands but in a fairly simple way : all bands share a common normalization whose variance is determined not by gaussian statistics but by the sample variance near the non - linear scale . under the halo model of appendix b , this behavior arises because the shape reflects the shape of halo profiles whereas the amplitude reflects their abundance . this abundance fluctuates with the local mean density . under the limber approximation of eqn . ( [ eqn : limber ] ) , given @xmath86 lens galaxy samples in disjoint redshift bins and @xmath87 source galaxies samples , there are @xmath86 distinct galaxy spectra , @xmath88 shear spectra , and @xmath89 galaxy - shear cross spectra . the potentially large number of cross spectra offers great opportunities for studies of galaxy evolution and cosmology . consider then the fisher matrix of cross spectra alone @xmath90^{-1 } { \partial c_l^{g_b \epsilon_j } \over \partial p_\alpha } \label{eqn : fishercross}\ ] ] where @xmath91^{ai , bj}= \tilde c_l^{g_a g_b } \tilde c_l^{\epsilon_i \epsilon_j } + \tilde c_l^{g_a \epsilon_j } \tilde c_l^{g_b \epsilon_i}\,.\ ] ] note that the total variance of the shear and galaxy fields contributes to the noise of the cross correlation as a type of sample variance . furthermore in the low signal - to - noise regime , the covariance is dominated by the product of power spectra of the galaxy and shear fields not the sample variance of the signal . in this case , the @xmath74-diagonal form of the fisher matrix depends on the assumption of statistical isotropy and remains valid even when the shear and galaxy fields are strongly non - gaussian . the framework described here is general and may be applied to any parameterized model for the underlying three dimensional power spectra @xmath92 , @xmath93 and @xmath94 and any selection criteria for the galaxies . in appendix a we describe the well - tested standard model for the linear mass spectrum and how it depends on the dark energy . in appendix b we develop the halo model for the galaxy and cross spectra . motivated by recent simulations which associate galaxies with substructure in dark matter halos , we utilize 5 free functions of redshift to describe the occupation of galaxies in dark matter halos . by discretizing these functions into observed redshift bins we obtain a halo model parameterization with @xmath95 parameters because even this multi - dimensional halo model may not be sufficiently realistic , we will attempt in the next section to separate cosmological information that does and does not depend on the details of the halo model . for instance in the large scale regime where galaxies clustering is nearly fully correlated with the mass , a measurement of the galaxy auto and cross power spectra is essentially a measurement of the mass power spectrum . furthermore , the angular cross spectra as a function of source galaxy redshift scale with distance in a known way through the lensing efficiency for any choice of the underlying three dimension power spectrum @xmath93 . in this section , we study the prospects for dark energy constraints with galaxy - lensing power spectra . we begin in [ sec : fiducial ] by defining the fiducial survey and the galaxy selection . in [ sec : jt ] we study the distance related or halo model independent information in the galaxy - shear cross correlation and in [ sec : full ] the joint constraints from all power spectra . for illustrative purposes , let us define a fiducial survey that is loosely based on a next generation ground based lensing survey . we take a source redshift distribution of the form @xmath96 with @xmath97 corresponding to a median redshift @xmath98 and an angular number density of @xmath99 arcmin@xmath57 . this corresponds roughly to a magnitude limit of @xmath100 . for the shape noise we take a shear rms per component of @xmath101 which reflects the intrinsic ellipticity and measurement errors of ground based observations ( see e.g. @xcite ) . note that the noise variance scales as @xmath102 in eqn . ( [ eqn : noise ] ) . we take a survey area of 4000 deg@xmath0 . from the survey galaxies we choose a galaxy lens population from the high luminosity tail . to balance signal strength and halo model robustness per lens galaxy against lens abundance , we choose lens galaxies with an abundance corresponding to a mass threshold of @xmath103 in the fiducial model . due to this tradeoff , the net signal - to - noise ratio is only weakly dependent on the threshold . note that the mass threshold is not held fixed as halo and cosmological parameters are varied ( see appendix b ) the fiducial mass threshold simply defines the redshift distribution and angular density of the lens galaxies . these are the quantities held fixed under the variations . for the redshift binning , we typically choose between 2 - 5 photometric redshift bins in the source distribution . further divisions do not substantially enhance constraints on the dark energy given the broad efficiency function @xcite . these are taken to have equal extent in redshift out to @xmath104 but with the last bin containing the remaining high redshift galaxies . for the lens galaxies , we limit the populations to @xmath105 since they will require more accurate photometric redshifts . we typically take @xmath106 lens galaxy bins reflecting a photometric redshift accuracy of @xmath107 . the total lens population then has an angular number density of @xmath108 gal arcmin@xmath57 with a median redshift of @xmath109 . finally , we allow for uncertainties in the shape of the mass power spectrum and initial normalization by taking priors of @xmath110 corresponding to a conservative interpretation of current constraints ( see appendix a and e.g. @xcite ) . with the fiducial lens survey , dark energy results depend only weakly on these prior assumptions . note that we take no prior constraints on the dark energy parameters so that the projected constraints reflect only the potential of galaxy - lensing power spectra . = 3.0 in comparing lensing observables for different source redshifts has been proposed as a way of measuring distances with galaxy clusters @xcite . recently , wide field weak lensing statistics have also been developed to exploit this test @xcite . in the limit of infinitesimal lens galaxy redshift bins , i.e. in eqn . ( [ eqn : efficiency ] ) @xmath111 the ratio of the galaxy - shear cross correlation in multiple shear source bins depends only on the ratio of efficiencies . this in turn depends only on angular diameter distances and redshifts . in terms of the cross power spectra @xmath112 hence the galaxy - shear cross correlation provides information on the dark energy which is immune to uncertainties in the modeling of the underlying galaxy - mass correlation . the ratio is also immune to sample variance in the _ signal _ , i.e. galaxy to galaxy variations in the underlying mass correlation and hence the non - gaussianity of the signal . in the fisher matrix of eqn . ( [ eqn : fishercross ] ) , if the noise terms @xmath113 and @xmath114 vanish , the covariance matrix becomes singular implying infinitesimal errors in distance parameters . in fact the technique works for individual galaxies in principle . unfortunately with realistic noise estimates , the tradeoff between model independence and sensitivity is severe . for example , a 10% change in the equation of state parameter @xmath115 typically yields an @xmath116 change in the efficiency ratio and so measurement of the effect is only possible with large galaxy and shear surveys . contrast this with the several percent change in the absolute growth or shear amplitude implied by eqn . ( [ eqn : growthfunction ] ) and illustrated in fig . [ fig : param ] . even for this statistic the halo model enters in three ways : by defining the strength of the signal , the sample variance of the _ noise _ and the accuracy to which the redshifts of the galaxy lenses must be known . the sample variance arises from contributions to the shear from structure along the line of sight not associated with the galaxy population and from the intrinsic clustering of galaxies . furthermore , with finite - width redshift bins in the lens galaxy distributions , the efficiency ratios depend on the model for the evolution of the underlying power spectra across the bins @xcite . even with our 5 halo parameter model for the evolution in each bin , the constraints are compromised . no constraints are possible with photometric redshift bins of @xmath117 if we marginalize all 5 parameters for a given bandpower in @xmath74 . by combining multiple bandpowers , one can recover dark energy information , but that amounts to utilizing information from the shape of the underlying correlation and again degrades the model independence of the effect . we will return to this type of constraint in the next section . to study the efficiency ratio test , we instead marginalize a _ single _ halo parameter per redshift and bandpower bin . this model is equivalent to marginalizing a constant amplitude or bias per bandpower in the underlying power spectrum . hence it eliminates information from the shape but retains an assumption on the evolution of the galaxy - mass power spectrum . results do not depend on the number of bands employed so long as the signal and noise dominated regimes are separated since the additional parameters are employed simply to remove shape information from measurements of different bands . we take the parameter to be the satellite normalization @xmath118 for convenience and have verified that the results do not depend on this choice . in fig . [ fig : sample ] we show the constraints in the constant equation of state , dark energy density ( @xmath115-@xmath119 ) plane for @xmath106 , @xmath120 ( see appendix a for dark energy parameter definitions ) . note that when moving to a varying @xmath121 , the errors on equation of state at the best constrained redshift @xmath122 remain the same @xmath123 , whereas those on @xmath119 increase . we can also assess the effect of sample variance by artificially removing the terms in the covariance ( [ eqn : crosscov ] ) that are proportional to @xmath124 . this causes an improvement in the errors by @xmath125 indicating that the sample variance is comparable to shot variance . this near equality reflects the fact that these halos have a projected scale radius of order @xmath126 ( see fig . [ fig : gamt ] ) . on these scales , the assumed shape noise @xmath127 is comparable to the cosmic shear . we include the information from all scales ; in practice the signal to noise has converged well before our numerical maximum @xmath128 . the discrepancy of our results with those of @xcite is due to a factor of two error in the calculation of the signal ( since corrected , astro - ph/0306046 v3 ) , to a more realistic model for halo profiles , and the inclusion of sampling errors . comparison of the results also requires implementing their prior on @xmath119 and their shape noise specifications . these results agree with an independent and concurrent study which also extended the model - independent techniques to shear - shear correlations @xcite . raising the number of source ( @xmath87 ) and lens ( @xmath86 ) divisions or choosing less rare galaxy tracers only slightly improves these constraints . note that the latter entails employing less massive objects with higher number density but weaker shear signal . finally note that the template technique of @xcite can be reexpressed as a measurement of the zero lag correlation of eqn . ( [ eqn : sheartemplate ] ) or a single bandpower @xmath129 although the efficiency ratio constraints are relatively weak , they are fairly robust to both the halo model and non - gaussianity of the correlation on small scales . moreover , they become more powerful when combined with complementary probes of shape and amplitude of the various power spectra on large scales as we shall now see . = 3.0 in = 3.0 in given the underlying halo model parameterization , cosmological and halo model parameters can be jointly fit to the observable power spectra . let us first consider constraints that are based on the galaxy - mass cross - correlation alone . in fig . [ fig : cross ] ( dashed lines ) we show the cross power constraints for constant @xmath115-@xmath119 and @xmath106 , @xmath120 . we marginalize @xmath130 halo parameters and vary the maximum @xmath74 employed in eqn . ( [ eqn : fishercross ] ) [ see appendix b ] . let us focus on the opposite region to the previous section , that of large scales @xmath131 where non - gaussianity in the fields and inadequacies in the halo model are minimized . under the halo model prescription , there is sufficient information in the power spectra to constrain a degenerate combination of @xmath115 and @xmath119 . note that the errors in the best constrained direction are insensitive to the maximum @xmath132 and hence to uncertainties in the non - gaussianity of power spectrum errors . they are also then less sensitive to inadequacies in the halo modeling that appear on small scales . the degeneracy line follows a line of constant linear power spectrum amplitude and distance at the typical redshift of the lenses . breaking this degeneracy then depends both on the internal or external determination of parameters in the halo model and its overall validity . fortunately , the efficiency ratio information from small scales is complementary in direction . furthermore , information from the other spectra can included . in particular the galaxy - galaxy correlations are more sensitive to the assumptions of the halo model than the galaxy - mass correlation ( see fig . [ fig : param ] ) . it may be used to cross check and calibrate the halo model parameters and potentially extend the parameter space as needed . in fig . [ fig : cross ] , we illustrate the utility of a joint analysis . adding in galaxy - galaxy power spectra information out to the same @xmath132 substantially assists dark energy parameter constraints by effectively acting as a prior and consistency check on the halo model parameters . the small interior ellipse illustrates the potential for simultaneous determinations of a constant @xmath115 and @xmath119 by combining this information with that from the efficiency ratio . = 3.0 in just as galaxy - galaxy correlations provide a powerful cross check on halo model parameter determinations , shear - shear correlations provide a powerful cross check on cosmological parameter determinations . it is well known that in principle shear - shear correlations are an extremely powerful probes of the dark energy parameters ( e.g. @xcite ) . in fig . [ fig : summary ] we show that with shear - shear correlations alone all three dark energy parameters @xmath122 , @xmath133 and @xmath119 can be simultaneously measured . note that we do not place external prior constraints on @xmath119 unlike the convention commonly found in the literature but that the effect of a prior can be readily read off these figures since the @xmath119 dimension is shown . we have chosen here @xmath106 and @xmath134 . however shear - shear correlations are more susceptible to systematic errors since the cross power would null out systematics that are not common to both . they also require a modeling of the statistics and their covariance in the non - linear regime . here we have taken @xmath135 . the combined galaxy - shear and galaxy - galaxy power spectra potentially have constraining power that is comparable to the shear - shear spectra . the combination of all three reduces the errors by of order @xmath136 or more versus shear alone , leading to marginalized errors of @xmath1 , @xmath137 [ @xmath138 , and @xmath3 . more importantly they provide important cross checks against shear systematics on the one hand and inadequate halo modeling on the other . for comparison with other probes , it is useful to note that the dark energy pivot point of the combined power spectra information is @xmath139 , the epoch of dark energy domination . equation ( [ eqn : rotation ] ) then maps the errors back into a more conventional description such as @xmath140 ( @xmath141 ) . for example if we add in projected cmb constraints for the planck satellite @xcite , the errors improve to @xmath142 , @xmath143 [ @xmath144 , @xmath145 as shown in fig . [ fig : w0 ] . note that the combined pivot point shifts to higher redshift . here we have employed the sensitivities in @xcite amounting to e.g. @xmath146 where @xmath147 is the angular diameter distance to last scattering ; note that current constraints are at the 0.04 level . finally , with all three power spectra one can probe the evolution of the halo parameters and hence aspects of the evolution of the underlying galaxy population . we show in fig . [ fig : halo ] the resulting errors on the 5 halo parameters as a function of redshift employing all @xmath148-point information out to @xmath135 . the errors for different redshifts are nearly independent whereas the errors between the 5 halo parameters at a given redshift are strongly correlated . the correlation indicates degeneracies between the parameters as can also be seen in fig . [ fig : param ] . thus there are combinations of halo parameters that are better determined than implied by fig . [ fig : halo ] . these combinations control the shape and amplitude of the spectra . for example , the linear combination that controls the bias and correlation coefficient across the @xmath149 bin and at scales central to the constraint , e.g. @xmath150 mpc@xmath151 are separately constrained at the level of @xmath152 and @xmath153 ( fiducial values : @xmath154 and @xmath155 ) . constraints on the correlation coefficient mainly reflect its insensitivity to halo parameters on large scales ; they weaken in the deeply non - linear regime . constraints on the bias as a function of scale remain at the percent level from the linear to well into the non - linear regime . the joint analysis of galaxy clustering and lensing data from next - generation surveys offers unique opportunities to simultaneously determine the evolution of clustering in the dark matter and galaxies . these surveys are expected to provide multi - color catalogs of galaxies with well characterized photometric redshifts well beyond @xmath5 . we have shown here that with redshift information , even at the 2-point or power spectrum level , there is enough information in the galaxy - shear correlation , especially when combined with galaxy - galaxy correlations , to jointly solve for a model with @xmath156 parameters to describe the galaxy evolution and @xmath157 dark energy parameters . we have conservatively allowed the galaxy parameters to vary independently as a function of lens redshift and marginalized the associated parameters when quoting dark energy constraints . the dark energy parameter determinations are statistically competitive with the shear - shear correlations and should be more robust to systematic errors in the shear determinations . they furthermore can provide a better redshift localization of dark energy effects given the broad lensing kernel of the shear . these determinations are assisted by two relatively robust features in the halo model : on large scales , the combination of galaxy - mass correlations and galaxy - galaxy correlations may be used to determine the underlying mass - mass correlations in a manner that is only weakly sensitive to halo model assumptions ; on small scales the ratio of galaxy - shear correlations at different source redshifts yields distance ratio information that is only weakly dependent on the evolution of the galaxy - mass correlation . equally important , by combining galaxy - galaxy and galaxy - shear clustering one can determine whether the halo model employed here suffices as a description of the relationship between galaxies and dark matter halos . consistency of the halo determinations can also be checked by selecting lens galaxies with multiple cuts on luminosity or rarity . consistency of the dark energy determinations can be checked against the shear - shear correlations which depend only on the mass spectrum . a further extension to include contributions from the many galaxy - shear bispectra would also improve parameter accuracies and provide cross - checks @xcite . halo model parameter determinations will be valuable in testing models of galaxy formation ( e.g. @xcite ) . galaxy parameters measured from surveys as a function of galaxy type , luminosity and redshift can be compared with simulations and semi - analytic model predictions of the same parameters . although these studies may show that the our halo model requires further modification and extension , we believe that the prospects are bright for a joint solution of galaxy and dark matter clustering . _ acknowledgments : _ we thank g. bernstein , d. holz , a. kravtsov , m. takada , a. tasitsiomi and especially e. sheldon for many useful discussions . is supported by nasa nag5 - 10840 , the doe , and the packard foundation . b.j . is supported by nasa nag5 - 10923 and the keck foundation . the linear mass power spectrum @xmath158 which underlies all observable power spectra depends only on well - motivated cosmological parameters in the context of the successful @xmath159cdm cosmology . in this section , we describe its parameterization in terms of the initial conditions [ sec : initial ] and evolution [ sec : evolution ] . we begin by assuming that massive neutrinos make a negligible contribution to the matter density . the shape of the mass power spectrum is then specified by the baryon density @xmath160 , the dark matter density @xmath161 , and the spectrum of initial curvature fluctuations @xmath162 @xmath163 where @xmath164 mpc@xmath151 is the normalization scale . we take fiducial values for these parameters that are consistent with the wmap determinations : @xmath165 , @xmath166 , @xmath167 and @xmath168 @xcite . the current uncertainties in these parameters are at the 10% level or better . note that @xmath169 is related to the wmap normalization parameter by @xmath170 and current and future uncertainties in this parameter are expected to be dominated by uncertainties in the thomson optical depth to reionization @xmath171 , i.e. @xmath172 thus the power spectrum as a function of @xmath45 in mpc@xmath151 ( not @xmath173 mpc@xmath151 ) in the matter dominated regime can be considered as largely known . the shape of this initial power spectrum does not change during dark energy domination on scales below the sound horizon of the dark energy . the amplitude of the linear power spectrum depends on the initial normalization @xmath174 and the growth function " , here the decay rate of potential fluctuations @xmath175 @xmath176 ^ 2 p(k,0 ) \,,\ ] ] where @xmath177 and we assume that all relevant scales are sufficiently below the maximal sound horizon of the dark energy . @xmath178 can be evaluated from any one of a number of einstein - boltzmann codes . the normalization of the linear power spectrum today is conventionally given at a scale of @xmath179mpc @xmath180 where @xmath181 is the fourier transform of a top hat window . the approximation in eqn . ( [ eqn : sigma8 ] ) is valid to the 1% level for individual variations of the parameters in the regime @xmath182 , @xmath183 , @xmath184 , @xmath185 which more than span the current observational errors . note that because the normalization is given in @xmath186 mpc , there is a strong scaling with the hubble constant . this scaling actually assists dark energy determinations since in a flat universe @xmath187 and precise measurements of @xmath188 makes @xmath173 depend on @xmath119 only . a measurement of the hubble constant is a measurement of the dark energy density . likewise a measurement of @xmath189 is a measurement of a specific combination of dark energy parameters . we will hereafter limit ourselves to flat universes but correspondingly _ neglect _ the dark energy information from the cmb unless otherwise specified . the angular diameter distance to recombination @xmath147 has been measured to @xmath190 @xcite and this constraint will continue to improve with better measurements of @xmath161 from the peaks as ( see e.g. @xcite ) @xmath191 the rationale behind dropping this constraint is that this measurement will be used in conjunction with galaxy and lensing constraints to eliminate any small curvature contribution that might exist . in a flat universe , @xmath192 closely follows @xmath193 in its dark energy dependence and so may be used as a powerful consistency test for the absence of spatial curvature . we discuss this issue further in [ sec : full ] . the growth function @xmath175 then depends only on the dark energy density @xmath194 and equation of state @xmath195 through the equation ( e.g. @xcite ) @xmath196 \frac{d g}{d \ln a } \nonumber\\ & + & \frac{3}{2}[1-w(a)]\omega_{{{\rm de}}}(a ) g = 0\ , , \label{eqn : growth}\end{aligned}\ ] ] where the initial conditions are @xmath197 and @xmath198 at an epoch well before dark energy domination @xmath199 , @xmath200 where @xmath201 and @xmath202 . given a constant equation of state , @xmath193 follows the approximate form @xmath203\ , , \nonumber\\ f(x ) & = & ( 1 + 0.498 x + 4.88 x^3)^{-1}\ , , \label{eqn : growthfunction}\end{aligned}\ ] ] where the approximation holds to @xmath204 for separate variations of @xmath205 and @xmath206 . here and throughout we denote a dark energy parameterization for which @xmath121 is constant with @xmath207 . note the fairly strong scaling of @xmath193 with @xmath115 around the fiducial model @xmath208 . a dynamical form of dark energy is unlikely to possess a strictly constant equation of state . since it only has observable consequences at @xmath209 , it is convenient to describe the function with its first order taylor expansion . we therefore choose an equation of state parameterization @xmath210 with the expansion around some epoch @xmath211 ; this generalizes the form employed in @xcite where @xmath212 . we instead choose @xmath213 such that the errors in @xmath122 and @xmath133 for a given observable are uncorrelated @xmath214 . the pivot point @xmath215 can be derived from a general representation via the transformation @xmath216 for the transformation to the pivot representation @xcite @xmath217 from which it follows that the errors decorrelate for a shift in the pivot of @xmath218 moreover the resulting errors on @xmath122 are then equal to those on @xmath115 , a constant @xmath121 . the pivot redshift is therefore also the redshift at which @xmath121 is best constrained . the drawback to choosing the pivot redshift is that it is specific to the observable , survey and dark energy model . as the redshift where the dark energy evolution is best constrained , the pivot point for growth measurements tends to coincide roughly with @xmath219 or @xmath220 in the fiducial @xmath159 model . in fact , further taking @xmath221 in eqn . ( [ eqn : growthfunction ] ) yields an approximation to the growth function that is good to several percent across a wide range of @xmath133 . choosing this standard normalization epoch then provides the benefits of the pivot point without the drawbacks . although we will employ the pivot redshift for the galaxy - lensing study below , it is sufficiently close to @xmath220 that it may be interpreted in this manner note that the pivot point for distance - based dark energy measurements can be at even higher redshift @xmath222 so that it becomes even more important to choose a @xmath223 for the characterization of constraints . in summary , our linear matter power spectrum is specified by @xmath224 parameters : 4 that are already well constrained by the cmb @xmath225 , @xmath226 , @xmath227 , @xmath228 , and 3 dark energy parameters @xmath229 , @xmath230 ( or @xmath140 ) , @xmath231 which galaxy - lensing power spectra can help constrain . parameter values of the fiducial cosmology are given in parentheses . given a linear power spectrum and cosmology , cosmological simulations can accurately predict the fully non - linear mass power spectrum and hence the shear - shear 2-point correlations . on the other hand , the power spectra involving galaxies will require some semi - analytic modeling for the foreseeable future . we now turn to a halo model for the parameterization of the underlying relationship between the observable power spectra and the linear theory predictions . to study the information contained in the lensing and galaxy two point observables , we require a model for the underlying three dimensional galaxy and mass density power spectra . recent work on comparing simulations to galaxy clustering data have shown that to first approximation , galaxies selected on luminosity are assigned to a mass - based selection of dark matter ( sub)halos of the same spatial number density @xcite . this mapping avoids the traditional problem of defining explicit halo mass - luminosity relationships . in [ sec : satellite ] , we build this underlying ansatz into a halo description of the galaxies and mass ( see @xcite for a review ) and obtain a parameterized model for their joint power spectra in [ sec : halopower ] . we then describe the phenomenology of the fiducial halo model in [ sec : phenomenology ] and study the sensitivity of the power spectra to variations in the halo model and cosmological parameters in [ sec : sensitivity ] . under the halo model , the power spectra are described by the abundance , clustering , density profile , substructure and galaxy occupation of dark matter halos . for the comoving abundance of host halos , we take the mass function @xcite @xmath232 where @xmath233 is the matter density today @xmath234 , @xmath235 @xmath236 \exp[-a\nu^2/2]\,.\end{aligned}\ ] ] here @xmath237 is the rms of the linear density field smoothed with a top hat of a radius that encloses the mass @xmath238 . we choose @xmath239 , @xmath240 , @xmath241 , and @xmath242 such that @xmath243 . the mass function of the fiducial model is shown in fig . [ fig : massfn ] bracketing the redshifts of interest for lensing ( @xmath244 ) . the halo clustering is then given by the peak - background split as @xcite @xmath245}\ , . \label{eqn : bias}\ ] ] for the halo density profile , we take the nfw form @xcite @xmath246 where @xmath247 with the virial radius @xmath248 we take the concentration @xcite @xmath249 where @xmath250 is defined by @xmath251 . we convert between the halo mass @xmath238 , assumed to be defined at an overdensity of 180 times the mean density , and the virial mass @xmath252 defined at an overdensity @xcite @xmath253 ^ 2 \over 1 + \omega_{\rm de}(z)}\,,\ ] ] using the nfw profile ( see e.g. @xcite ) . the profile enters into the power spectra through its normalized fourier transform @xmath254 although there are current uncertainties in these descriptions of the halo mass function , bias and concentration , we choose _ not _ to associate free parameters with these _ dark matter _ clustering properties . the characterization of these properties and more importantly the mass power spectrum itself can in principle be fixed by better simulations . similarly for the association with galaxies , since we will be matching number densities of objects , the halo mass definition employed here may be replaced with any variable that defines the selection of objects in the simulations and is a monotonic function of galaxy luminosity on average . the halo model predicts the galaxy and galaxy - mass power spectra under an assumption for the statistics of the occupation of the host halo by galaxies @xcite . since each galaxy also carries its own dark matter halo we will hereafter distinguish between the host halo of mass @xmath255 and its satellite halos of mass @xmath256 . for simplicity we take the satellites to also have nfw profiles but with an adjustable concentration ; a more sophisticated model would account for the change in the functional form due to truncation from tidal stripping and trends in mass . we begin by taking a simulation - motivated universal form for the number of dark matter satellites in the host halo as a function of their mass ratio @xcite @xmath257 for @xmath258 where simulations suggest that @xmath259 and @xmath260 @xcite . the cutoff to low masses prevents a mild logarithmic divergence in the total mass ; this arbitrary cutoff only affects the mass power spectrum at unobservably small scales . we next associate galaxies with these satellites . the total number of satellite galaxies above a given threshold mass @xmath261 then becomes @xmath262 to this population of satellite galaxies we add the central galaxy associated with the host halo itself to obtain the total number of galaxies above the threshold ( see fig . [ fig : massfn ] ) @xmath263 these two pieces have different statistical properties . the central galaxy may be either above or below threshold and hence either occupied or unoccupied leading to @xmath264 . we model the average value as a step function at some limiting mass @xmath261 smoothed by a gaussian in @xmath265 to reflect scatter in the conversion of a magnitude limit to a mass limit @xmath266 the dark matter satellites follow a poisson distribution with @xmath267 @xcite . given a fixed shape for @xmath268 , the threshold mass @xmath261 is not a free parameter . rather it is fixed by matching the space density of galaxies to the space density inferred by the observed number counts @xmath269 note that by choosing rare objects , the implied @xmath270 and so the population will be dominated by host galaxies rather than satellite galaxies ( see fig . [ fig : massfn ] ) . this fact will be useful in minimizing the uncertainties and systematic errors associated with the model for @xmath271 and the satellite profiles , e.g. the mass scaling of the concentration parameters . the satellite contribution can be quantified by considering the satellite mass function @xmath272 and comparing it to the host halo mass function itself @xmath273^{-1}\,.\ ] ] this ratio depends mainly on the peak height threshold @xmath274 for a given form for the shape of @xmath275 ( see fig . [ fig : satellite ] ) . with our fiducial choice , it saturates at the low mass end at @xmath276 . with a selection of rare objects , e.g. corresponding to @xmath277 in the fiducial model , the satellite fraction can be less than @xmath278 at all redshifts . the power spectra are defined by the host and satellite distributions as @xmath279 the indices @xmath280 and @xmath148 refer to the number of points in a single halo . the first piece then represents two points in separate host halos correlated by the linear power spectrum @xmath158 @xmath281{dn_h \over d\ln m_h } b(m_h)\ , , \nonumber \end{aligned}\ ] ] and the second piece is the contribution from two points within a parent halo including its satellite contributions @xmath282\ , , \nonumber\\ i_{2 g } & = & { 1 \over \bar n_v^2 } \int { d m \over m } [ \bar n_s^2 y_g^2 + 2 \bar n_h \bar n_s y_g ] { d n_h\over d\ln m } \ , , \nonumber\\ i_{2c } & = & { 1 \over \bar n_v } \int { d m \over m } \left ( { m \over \rho_{0 } } \right ) \big [ { d n_h\over d\ln m } ( \bar n_h y_h + \bar n_s y_h y_g)\nonumber\\ & & + { d n_s\over d\ln m } h(m - m_{\rm th } ) y_s \big ] \ , , \label{eqn : singlehalo}\end{aligned}\ ] ] here we have employed the shorthand convention @xmath283 , @xmath284 and the step function @xmath285 for @xmath286 and @xmath287 for @xmath288 . the effect of satellite galaxies occupying satellite halos takes the same form as the central galaxy occupying the host halo @xcite except that we have neglected the small effect of smoothing around the threshold mass . in this description , the satellite - satellite mass correlation and satellite - halo mass correlation are implicitly included as the replacement of mass lost to subhalos in the term involving the parent profiles @xmath289 in eqn . ( [ eqn : singlehalo ] ) . five functions of redshift specify our halo model , the satellite - host normalization @xmath118 , the slope of the satellite mass function and occupation distribution @xmath290 , the scatter in the mass observable relation @xmath291 , the concentration of the satellites in the host halo @xmath292 and the concentration of the mass profiles of the satellites @xmath293 . we assume that the latter two follow the mass scaling of isolated halos in eqn . ( [ eqn : concentration ] ) but have an arbitrary normalization . we further represent the functions by a set of @xmath294 parameters that specify their values at the redshifts of the @xmath86 lens redshift bins . the values of these parameters will in the future be determined by fitting to the joint power spectra . however , to study the potential of these future data sets we must specify their values in the fiducial model . the sensitivity of observables to variations in their values around this model is then quantified through the fisher matrices of [ sec : fisher ] . we choose fiducial values that are roughly consistent with low redshift measurements from the sdss lensing data @xcite and @xmath295-body simulations : @xmath296 , @xmath297 , @xmath298 for all redshift bins . these three parameters define the shape of @xmath268 shown in fig . [ fig : massfn ] . in our fiducial model the concentration parameters @xmath299 . in fig . [ fig : bias ] , we show the galaxy - galaxy and galaxy - mass power spectra with the fiducial halo and cosmological parameters . they are shown relative to the mass - mass spectrum at @xmath300 for halo abundances corresponding to several different mass thresholds in the fiducial @xmath159cdm cosmology ( see [ sec : linear ] ) . in the linear regime @xmath301 mpc@xmath151 the ratios @xmath302 and @xmath303 both return the constant linear bias of the objects . in other words , in the linear regime the correlation coefficient @xmath304 independent of the population or halo model parameters . in this regime , a measurement of @xmath93 and @xmath94 is a measurement of the mass power spectrum @xmath305 and hence in combination can be used to study the dark - energy dependent growth rate . the linear bias increases with the mass threshold @xmath261 or more generally with the rarity of the objects . in the non - linear regime , the bias becomes strongly scale dependent reflecting the correlations within a parent halo . on the other hand , the ratio @xmath306 remains remarkably constant @xcite . formally the combination indicates that the galaxy - mass correlation @xmath307 . under the halo model , every galaxy has a dark matter halo around it and so the relevant power spectrum for computing the correlation coefficient is corrected for the excess shot noise power @xmath308 . the scale dependence of @xmath309 and @xmath306 can be used to pin down the halo parameters . for fixed halo parameters , it marks the non - linear scale which also depends on the dark energy through the linear growth rate . these properties remain qualitatively true for angular power spectra with the caveat that projection effects can broaden the non - linear transition regime for wide redshift shells for the lens galaxies and the broad efficiency for lensing . in fig . [ fig : ktol ] we show the correspondence of @xmath45 and @xmath74 in the fiducial cosmology and under the limber approximation . in fig . [ fig : cl ] , we show an example of the limber projection to the angular power spectra . note that the inflection caused by the transition between the one host halo and ( linear ) two host halo regimes occurs at wavenumbers of @xmath310 . note that there are @xmath311 galaxy spectra , @xmath312 cross spectra , and @xmath313 shear spectra . in fig . [ fig : gamt ] , we highlight the one halo regime by showing the predictions for the tangential shear from eqn . ( [ eqn : tangential ] ) for a given source and lens configuration . note that the shear turns over as the angular radius resolves the scale radius of the halo and shows another break on large scales marking the beginning of the two halo regime . the sensitivity of the power spectra to halo and cosmological parameters is quantified by their derivatives with respect to the parameters in eqn . ( [ eqn : derivs ] ) . in fig . [ fig : param ] we compare the halo parameter sensitivity of galaxy - galaxy and galaxy - mass spectra to the equation of state parameter @xmath115 at @xmath300 for a population with an abundance in the fiducial model corresponding to @xmath314 . since the initial normalization of the power spectrum is here fixed , sensitivity to @xmath115 is equivalent to sensitivity to the present - day normalization @xmath315 ( see eqn . ( [ eqn : growthfunction ] ) ) . the galaxy - galaxy spectrum is strikingly insensitive to @xmath115 compared with the halo parameters that control @xmath316 : @xmath317 , @xmath318 and @xmath319 . changes in the parameters that control its shape change the power spectrum on all scales since they enter into the calculation of the mass threshold @xmath261 and hence the bias of the host halos . a change in the cosmology that for example lowers the present amplitude of the matter power spectrum makes galaxies of a given number density rarer . since rarer objects are more highly biased tracers of the matter , the galaxy - galaxy power spectrum remains largely unchanged . the same is not true for the galaxy - mass power spectra . indeed the galaxy - mass power spectra are substantially more sensitive to cosmological parameters than halo parameters . the combination of galaxy - galaxy and galaxy - mass power spectra is then particularly powerful for simultaneously determining the halo and cosmological parameters . neither set of spectra are very sensitive to the concentration parameters @xmath293 and @xmath292 except at very @xmath45 values that correspond to @xmath75 for the redshifts in question ( see fig . [ fig : ktol ] ) . by definition , the concentration parameters only affect the power spectrum on small scales . furthermore since we have chosen a high mass threshold , the fraction of objects that are satellites and hence affected by these parameters is low . this insensitivity helps justify our crude treatment of the profiles . as for the mass - mass power spectrum , and hence the shear - shear power spectrum , it is of course the most sensitive to the dark energy parameters being directly related to the linear growth function of eqn . ( [ eqn : growth ] ) . it formally also depends on the satellite mass function and hence the halo model parameters that control it . however we here take the perspective that the form of the mass - mass power spectrum as a function of cosmology will in the future be fixed directly by simulations replacing the halo model description . hence we take the uncertainties in the halo parameters to only affect the galaxy - galaxy and galaxy - mass power spectra . operationally we assume that variations in @xmath92 due to the variations in the satellite mass function are localized to the galaxy mass regime and compensated by variations at other masses to leave the spectrum invariant . this corresponds to dropping the parameter derivative terms in the fisher matrix ( [ eqn : fisher ] ) .

it is well known that primordial black holes ( pbhs ) could have formed in the early universe @xcite . a comparison of the cosmological density at any time after the big bang with the density associated with a black hole shows that pbhs would have of order the particle horizon mass at their formation epoch : m(t ) 10 ^ 15(t10 ^ - 23 s ) g. pbhs could thus span an enormous mass range : those formed at the planck time ( @xmath1s ) would have the planck mass ( @xmath2 g ) , whereas those formed at 1 s would be as large as @xmath3 , comparable to the mass of the holes thought to reside in galactic nuclei . pbhs would could arise in various ways @xcite . since the early universe is unlikely to have been exactly friedmann , they would form most naturally from initial inhomogeneities but they might also form through other mechanisms at a cosmological phase transition . the realization that small pbhs might exist prompted hawking to study their quantum properties . this led to his famous discovery @xcite that black holes radiate thermally with a temperature t = c^38gmk 10 ^ 26 ( mg)^-1 k 10 ^ -7 ( m)^-1 k. this means that they evaporate on a timescale ( m ) 10 ^ 64 ( m)^3 . only black holes smaller than @xmath4 g would have evaporated by the present epoch , so eqn ( 1 ) implies that this effect could be important only for black holes which formed before @xmath5s . despite the conceptual importance of this result , it is bad news for pbh enthusiasts . for since pbhs with a mass of @xmath4 g , which evaporate at the present epoch , would have a temperature of order 100 mev , the observational limit on the @xmath6-ray background intensity at 100 mev immediately implies that their density could not exceed @xmath7 times the critical density @xcite . not only does this render pbhs unlikely dark matter candidates , it also implies that there is little chance of detecting black hole explosions at the present epoch @xcite . despite this conclusion , pbh evaporations could still have interesting cosmological consequences . in particular , they might generate the microwave background @xcite or modify the standard cosmological nucleosynthesis scenario @xcite or contribute to the cosmic baryon asymmetry @xcite . pbh evaporations might also account for the annihilation - line radiation coming from the galactic centre @xcite or the unexpectedly high fraction of antiprotons in cosmic rays @xcite . pbh explosions occurring in an interstellar magnetic field might also lead to radio bursts @xcite . even if pbhs had none of these consequences , studying such effects leads to strong upper limits on how many of them could ever have formed and thereby constrains models of the early universe . indeed pbhs serve as a probe of times much earlier than that associated with most other relicts " of the big bang . while photons decoupled at @xmath8y , neutrinos at @xmath9 s and wimps at @xmath10s , pbhs go all the way back to the planck time . therefore even if pbhs never formed , their non - existence gives interesting information . we review the formation mechanisms and evaporation constraints on pbhs in section 2 . much of this material is also contained in my contribution to the 1996 chalonge school @xcite . however , there have been several interesting developments since then and these are covered in section 3 . the remaining sections will examine how the pbh constraints are modified if the value of the gravitational constant " @xmath0 was different at early times . as reviewed in section 4 , this idea has a long history and should no longer be regarded as exotic . it arises in various scalar - tensor theories of gravity and these are a natural setting for many currently popular models of the early universe . black hole formation and evaporation could be greatly modified in variable-@xmath0 cosmologies , since many of their properties ( eg . their hawking temperature ) depend explicitly on @xmath0 . however , as emphasized by barrow @xcite and discussed in section 5 , the nature of the modification depends upon whether the pbh preserves the value of @xmath0 at its formation epoch ( corresponding to what is termed `` gravitational memory '' ) or always maintains the changing background value . there would be interesting modifications to the cosmological consequences of pbh evaporations in both cases but they would be more dramatic in the first . barrow & carr @xcite considered the implications of these two scenarios in detail and this work has been taken further with goymer @xcite . we will review the conclusions of these papers in section 6 and highlight a particularly interesting development in section 7 . one of the most important reasons for studying pbhs is that it enables one to place limits on the spectrum of density fluctuations in the early universe . this is because , if the pbhs form directly from density perturbations , the fraction of regions undergoing collapse at any epoch is determined by the root - mean - square amplitude @xmath11 of the fluctuations entering the horizon at that epoch and the equation of state @xmath12 . one usually expects a radiation equation of state ( @xmath13 ) in the early universe . in order to collapse against the pressure , an overdense region must be larger than the jeans length at maximum expansion and this is just @xmath14 times the horizon size . on the other hand , it can not be larger than the horizon size , else it would form a separate closed universe and not be part of our universe @xcite . this has two important implications . firstly , pbhs forming at time @xmath15 should have of order the horizon mass given by eqn ( 1 ) . secondly , for a region destined to collapse to a pbh , one requires the fractional overdensity at the horizon epoch , @xmath16 , to exceed @xmath6 . providing the density fluctuations have a gaussian distribution and are spherically symmetric , one can infer that the fraction of regions of mass @xmath17 which collapse is @xcite ( m ) ~(m ) where @xmath18 is the value of @xmath11 when the horizon mass is @xmath17 . the pbhs can have an extended mass spectrum only if the fluctuations are scale - invariant ( i.e. with @xmath11 independent of @xmath17 ) but this is expected in many scenarios . the fluctuations required to make the pbhs may either be primordial or they may arise spontaneously at some epoch . one natural source of fluctuations would be inflation @xcite and , in this context , @xmath19 depends implicitly on the inflationary potential . pbhs formed before inflation would be drastically diluted but new ones could form from the fluctuations generated after inflation . many people have studied pbh formation in this context @xcite as an important way of constraining the inflationary potential . this was the focus of my 1996 erice lecture , so it will not be covered again here . note that the gaussian assumption has been questioned in the inflationary context @xcite , so eqn ( 4 ) may not apply , but one still finds that @xmath20 depends very sensitively on @xmath11 . in some situations eqn ( 4 ) would fail qualitatively . for example , pbhs would form more easily if the equation of state of the universe were ever soft ( @xmath211 ) . this might apply if there was a phase transition which channelled the mass of the universe into non - relativistic particles or which temporally reduced the pressure . in this case , only those regions which are sufficiently spherically symmetric at maximum expansion can undergo collapse ; the dependence of @xmath20 on @xmath11 would then have the form @xcite = 0.02 ^ 13/2 , which is much weaker than indicated by eqn ( 4 ) , but there would still be a unique relationship between the two parameters . some formation mechanisms for pbhs do not depend on having primordial fluctuations at all . for example , at any spontaneously broken symmetry epoch , pbhs might form through the collisions of bubbles of broken symmetry @xcite . pbhs might also form spontaneously through the collapse of cosmic strings @xcite or domain walls @xcite . in these cases @xmath22 depends , not on @xmath19 , but on other cosmological parameters , such the bubble formation rate or the string mass - per - length . these mechanisms were discussed in more detail in my 1996 erice contribution @xcite . in all these scenarios , the current density parameter @xmath23 associated with pbhs which form at a redshift @xmath24 or time @xmath15 is related to @xmath20 by @xcite _ r(1+z ) 10 ^ 6 ( ts)^-1/2 10 ^ 18(m10 ^ 15g)^-1/2 where @xmath25 is the density parameter of the microwave background and we have used eqn ( 1 ) . the @xmath26 factor arises because the radiation density scales as @xmath27 , whereas the pbh density scales as @xmath28 . any limit on @xmath23 therefore places a constraint on @xmath22 and the constraints are summarized in fig . 1 . the constraint for non - evaporating mass ranges above @xmath4 g comes from requiring @xmath29 . stronger constraints are associated with pbhs smaller than this since they would have evaporated by now @xcite . the strongest one is the @xmath6-ray limit associated with the @xmath4 g pbhs evaporating at the present epoch @xcite . other ones are associated with the generation of entropy and modifications to the cosmological production of light elements . the constraints below @xmath30 g are based on the ( not necessarily secure ) assumption that evaporating pbhs leave stable planck mass relics , in which case these relics are required to have less than the critical density @xcite . the constraints on @xmath22 can be converted into constraints on @xmath19 using eqn ( 4 ) and these are shown in fig . 2 . also shown here are the ( non - pbh ) constraints associated with the spectral distortions in the cosmic microwave background induced by the dissipation of intermediate scale density perturbations and the cobe quadrupole measurement , as well as lines corresponding to various slopes in the @xmath18 relationship . this shows that one needs the fluctuation amplitude to decrease with increasing scale in order to produce pbhs . [ f1 ] [ f2 ] recent hydrodynamical calculations for the @xmath31 case have refined the criterion @xmath32 for pbh formation and this modifies the estimate for @xmath22 given by eqn ( 4 ) . niemeyer & jedamzik @xcite find that one needs @xmath33 rather than @xmath34 to ensure pbh formation , and shibata & sasaki @xcite reach similar conclusions . they also find that there is little accretion after pbh formation , as expected theoretically @xcite . another interesting development has been the application of critical phenomena " to pbh formation . studies of the collapse of various types of spherically symmetric matter fields have shown that there is always a critical solution which separates those configurations which form a black hole from those which disperse to an asymptotically flat state . the configurations are described by some index @xmath35 and , as the critical index @xmath36 is approached , the black hole mass is found to scale as @xmath37 for some exponent @xmath38 . this effect was first discovered for scalar fields @xcite but subsequently demonstrated for radiation @xcite and then more general fluids with equation of state @xmath39 @xcite . in all these studies the spacetime was assumed to be asymptotically flat . however , niemeyer & jedamzik @xcite have recently applied the same idea to study black hole formation in asymptotically friedmann models and have found similar results . for a variety of initial density perturbation profiles , they find that the relationship between the pbh mass and the the horizon - scale density perturbation has the form m = k m_h(- _ c)^ where @xmath40 is the horizon mass and the constants are in the range @xmath41 , @xmath42 and @xmath43 for the various configurations . since @xmath44 as @xmath45 , this suggests that pbhs may be much smaller than the particle horizon at formation ( although it is clear that a fluid description must break down if they are too small ) and it also modifies the mass spectrum @xcite . there has been particular interest recently in whether pbhs could have formed at the quark - hadron phase transition at @xmath2s . this is because the horizon mass is of order @xmath46 then , so such pbhs would naturally have the sort of mass required to explain the macho microlensing results @xcite . this is discussed in more detail in my other lecture at this meeting . one might expect pbhs to form more easily at that epoch because of a temporary softening of the equation of state . if the qcd phase transition is assumed to be of 1st order , then hydrodynamical calculations show that the value of @xmath16 required for pbh formation is indeed reduced below the value which pertains in the radiation case @xcite . this means that pbh formation will be strongly enhanced at the qcd epoch , with the mass distribution being peaked around the horizon mass . one of the interesting implications of the pbh macho scenario is the possible existence of a halo population of _ binary _ black holes @xcite . with a full halo of such objects , there could then be @xmath47 binaries inside 50 kpc and some of these could be coalescing due to gravitational radiation losses at the present epoch @xcite . current interferometers ( such as ligo ) could detect such coalescences within 50 mpc , corresponding to a few events per year . future space - borne interferometers ( such as lisa ) might detect 100 coalescences per year . if the associated gravitational waves were detected , it would provide a unique probe of the halo distribution ( eg . its density profile and core radius @xcite . kohri & yokoyama @xcite have recently improved the constraints on @xmath48 which come from cosmological nucleosynthesis considerations . constraints from neutrino background have also been presented by bugaev & konischev @xcite . the recent detection of a galactic @xmath6-ray background @xcite , measurements of the antiproton flux @xcite , and the discovery of very short period @xmath6-ray burts @xcite may even provide positive evidence for such pbhs . this is discussed in detail elsewhere @xcite . some people have emphasized the possibility of detecting very high energy cosmic rays from pbhs using air shower techniques @xcite . however , recently these efforts have been set back by the claim of heckler @xcite that qed interactions could produce an optically thick photosphere once the black hole temperature exceeds @xmath49 gev . in this case , the mean photon energy is reduced to @xmath50 , which is well below @xmath51 , so the number of high energy photons is much reduced . he has proposed that a similar effect may operate at even lower temperatures due to qcd effects @xcite . this is discussed further in the contribution of kapusta at this meeting @xcite . however , these arguments should not be regarded as definitive : macgibbon @xcite claim that heckler has not included lorentz factors correctly in going from the black hole frame to the centre - of - mass frame of the interacting particles ; in their calculation qed interactions are never important . most variable - g scenarios associate the gravitational `` constant '' with some form of scalar field @xmath52 . this notion has its roots in kaluza - klein theory , in which a scalar field appears in the metric component @xmath53 associated with the 5th dimension . einstein - maxwell theory then requires that this field be related to @xmath0 @xcite . although this was assumed constant in the original kaluza - klein theory , dirac @xcite noted the the ratio of the electric to gravitational force between to protons ( @xmath54 ) and the ratio of the age of the universe to the atomic timescale ( @xmath55 ) and the square - root of the number of particles in the universe ( @xmath56 ) are all comparable and of order @xmath57 . this unlikely coincidence led him to propose that these relationships must _ always _ apply , which requires g t^-1 , gm / r ~1 , where @xmath58 is the horizon scale . the first condition led jordan @xcite to propose a theory in which the scalar field in kaluza - klein theory is a function of _ both _ space and time , and this then implies that @xmath59 has the same property . the second condition implies the mach - type relationship @xmath60 , which suggests @xcite that @xmath52 is a solution of the wave equation @xmath61 . this motivated brans - dicke ( bd ) theory @xcite , in which the einstein - hilbert lagrangian is replaced by l = r - _ , _ , g^ + l_m , where @xmath62 is the matter langrangian and the constant @xmath63 is the bd parameter . the potential @xmath52 then satisfies = ( ) t , where @xmath64 is the trace of the matter stress - energy tensor , and this has the required machian form . since @xmath52 must have a contribution from _ local _ sources of the form @xmath65 , this entails a violation of the strong equivalence principle . in order to test this , the ppn formalism was introduced . applications of this test in a variety of astrophysical situations ( involving the solar system , the binary pulsar and white dwarf cooling ) currently require @xmath66 , which implies that the deviations from general relativity can only ever be small in bd theory @xcite . the introduction of generalized scalar - tensor theories @xcite , in which @xmath63 is itself a function of @xmath52 , led to a considerably broader range of variable-@xmath0 theories . in particular , it permitted the possibility that @xmath63 may have been small at early times ( allowing noticeable variations of @xmath0 then ) even if it is large today . in the last decade interest in such theories has been revitalized as a result of early universe studies . inflation theory @xcite has made the introduction of scalar fields almost mandatory and extended inflation specifically requires a model in which @xmath0 varies @xcite . in higher dimensional kaluza - klein - type cosmologies , the variation in the sizes of the extra dimensions also naturally leads to a variation in @xmath0 @xcite . the currently popular low energy string cosmologies necessarily involve a scalar ( dilaton ) field @xcite and bosonic superstring theory , in particular , leads @xcite to a lagrangian of the form ( 9 ) with @xmath67 . the intimate connection between dilatons , inflatons and scalar - tensor theory arises because one can always transform from the ( physical ) jordan frame to the einstein frame , in which the lagrangian has the standard einstein - hilbert form @xcite l = |r - 2_,_,|g^ + l_m . here the new scalar field @xmath68 is defined by d= ( ) ^1/2 and the barred ( einstein ) metric and gravitational constant are related to the original ( jordan ) ones by g _ = a()^2 |g_,g = [ 1+^2 ( ) ] a()^2 |g , a/a , where the function @xmath69 specifies a conformal transformation and a prime denotes @xmath70 . thus scalar - tensor theory can be related to general relativity plus a scalar field , although the theories are not identical because particles do not follow geodesics in the einstein frame . the behaviour of homogeneous cosmological models in bd theory is well understood @xcite . their crucial feature is that they are vacuum - dominated at early times but always tend towards the general relativistic solution during the radiation - dominated era . this is a consequence of the fact that the radiation energy - momentum tensor is trace - free [ i.e. @xmath71 in eqn ( 10 ) ] . this means that the full radiation solution can be approximated by joining a bd vacuum solution to a general relativistic radiation solution at some time @xmath72 , which may be regarded as a free parameter of the theory . however , when the matter density becomes greater than the radiation density at @xmath73s , the equation of state becomes that of dust @xmath74 and @xmath0 begins to vary again . for a @xmath75 model , one can show that in the three eras @xcite g = g_0 ( t_0/t_e)^n , a t^(2-n)/3 ( t > t_e ) g = g_e g_0 ( t_0/t_e)^n , a t^1/2 ( t_1<t < t_e ) g = g_e ( t / t_1)^-(n + ) /2 , a t^(2 - n - ) /6 ( t < t_1 ) where @xmath76 is the value of @xmath0 at the current time @xmath77 , @xmath78 and @xmath79 . since the bd coupling constant is constrained by @xmath66 , which implies @xmath80 , eqns ( 14 ) to ( 16 ) imply that the deviations from general relativity are never large if the value of @xmath81 is always the same . however , as we now explain , it is also interesting to consider bd models in which @xmath81 and @xmath63 can vary and thus violate the current constraints . the behaviour of cosmological models in more general scalar - tensor theories depends on the form of @xmath82 but they still retain the feature that the general relativistic solution is a late - time attractor during the radiation era . since one requires @xmath83 to @xmath84 at the epoch of primordial nucleosynthesis @xcite , one needs the vacuum - dominated phase to end at some time @xmath85 1 s. the theory approaches general relativity in the weak field limit only if @xmath86 and @xmath87 but @xmath82 is otherwise unconstrained . barrow & carr consider a toy model in which 2 + 3 = 2(1-/_c)^- where @xmath88 and @xmath20 are constants . this leads to 2 + 3 t^-/(2- ) , /^3 t^(1 - 2)/(1- ) ( t < t_v ) , so one requires @xmath89 in order to have @xmath86 and @xmath90 as @xmath91 . in the @xmath92 case , one finds g t^-2/(3- ) , a t^(1-)/(3- ) , ( t < t_v ) . during the vacuum - dominated era , such models can therefore be regarded as bd solutions in which @xmath63 is determined by the parameter @xmath20 and unconstrained by any limits on @xmath63 at the present epoch . after @xmath93 , @xmath0 is constant and one has the standard radiation - dominated or dust - dominated behaviour . the consequences of the cosmological variation of @xmath0 for pbh evaporation depend upon how the value of @xmath0 near the black hole evolves . barrow @xcite introduces two possibilities : in scenario a , @xmath0 everywhere maintains the background cosmological value ( so @xmath52 is homogeneous ) ; in scenario b , it preserves the value it had at the formation epoch near the black hole even though it evolves at large distances ( so @xmath52 becomes inhomogeneous ) . on the assumption that a pbh of mass @xmath17 has a temperature and mass - loss rate t = ( 8gm)^-1 , - ( gm)^-2 , with @xmath94 in scenario a and @xmath95 in scenario b , barrow & carr calculate the evaporation time @xmath96 for various values of the parameters @xmath81 and @xmath72 in bd theory @xcite . the results are shown in fig . 3(a ) for scenario a and fig . 3(b ) for scenario b. here @xmath97 is the mass of a pbh evaporating at the present epoch , @xmath98 is the mass of a pbh evaporating at time @xmath99 and @xmath100 is the mass of a pbh evaporating at the present epoch in the standard ( constant @xmath0 ) scenario . in scenario a with @xmath101 , there is a maximum mass of a pbh which can ever evaporate and this is denoted by @xmath102 . the results for the scalar - tensor with @xmath82 given by eqn ( 17 ) with @xmath103 are shown in fig . 3(c ) for scenario b with various values of the parameters @xmath104 and @xmath93 . the corresponding modifications to the constraints on @xmath22 in all three cases are shown in fig . 3(d ) , which should be compared to fig . 1 . barrow & carr considered both scenarios a and b but did not attempt decide which was more plausible . in this section we address this question more carefully . the main argument for scenario a comes from an important result of hawking @xcite . he showed that in bd theory , providing the weak energy condition holds , the gradient of @xmath52 must be zero everywhere for stationary , asymptotically flat black holes . this means that such black holes are identical to those in general relativity . this result can be generalized to all scalar - tensor theories and suggests that such theories are in agreement with the `` no - hair '' theorem . numerical calculations support this theorem @xcite . collapse is accompanied by outgoing scalar gravitational radiation , which radiates away the scalar mass until the black hole settles down to the schwarzschild form with a constant scalar field . in particular , harada et al . @xcite have investigated oppenheimer - snyder collapse in which a ball of dust described by a @xmath105 friedmann interior and a schwarzschild exterior collapses to a black hole . in these calculations the scalar field is taken to be constant before the collapse and its back - reaction on the metric is assumed to be always negligible . it is found that , as the collapse proceeds , a scalar gravitational wave propagates outwards before the scalar field settles down to being constant again . it should be stressed that the scalar no hair theorem has only been proved for asymptotically flat spacetimes , so it is not clear that it also applies in the asymptotically friedmann case . while the no hair theorem suggests that @xmath52 should tend to a _ locally _ constant value ( close to the black hole ) , it is not obvious that this needs to be the asymptotic cosmological value . indeed , since the homogeneizing of @xmath52 is only ensured by scalar wave emission , one might infer that this can only be achieved on scales less than the particle horizon . one way to determine what happens is to seek a precise mathematical model for a black hole in a cosmological background . for example , one can try to match a black hole and cosmological solution over some boundary @xmath106 . such a matching is provided in general relativity by the einstein - straus or swiss cheese " model @xcite . here a friedmann exterior is matched with a general spherically symmetric interior . if there is no scalar field , it turns out that the latter has to be the static schwarzschild solution but the situation may be more complicated in the present context due to the presence of scalar gravitational radiation . in general one can show that the following continuity conditions must apply at @xmath106 : = 0 , = [ g _ u^ n^ ] = 0 , = 0 , = 0 where @xmath107 and @xmath108 are 4-vectors normal and tangent to @xmath106 , respectively @xcite . unfortunately , it turns out that an einstein - straus type solution does not exist in bd theory . this is because the only way to satisfy the junction conditions ( 21 ) is if @xmath52 is spatially and temporally constant , which is just the general relativistic case . jacobsen @xcite has addressed the problem analytically by looking for a spherically symmetric solution which represents a perturbation of the schwarzschild solution near the origin but is asymptotically friedmann at large distances , with @xmath52 satisfying the appropriate cosmological conditions . he presupposes that the black hole event horizon is much smaller than the particle horizon , so that the cosmological timescale is much longer than the black hole timescale . in this case , he finds that there is little lag between the value of @xmath52 at the event horizon and particle horizon , which suggests that memory can only be weak . however , it must be emphasized that this conclusion need not follow if the black hole has a size _ comparable _ to the particle horizon at formation and , as indicated in section 2 , this is expected for a pbh . another way to investigate the problem is to study the collapse of dust in a tolman - bondi background using the same approximation employed by harada et al . @xcite , i.e. neglecting the back reaction of the scalar field , but requiring that @xmath52 have the required cosmological time - dependence at large distances . one puts in an initial density perturbation for the dust but assumes that @xmath52 is initially homogeneous . our preliminary numerical calculations @xcite use the characteristic method to determine the evolution of the scalar field perturbation along null and constant - time hypersurfaces . we find that @xmath52 does initially build up near the centre but it then gets smoothed out , tending eventually to homogeneity . although this suggests that there is no gravitational memory , it should be stressed that this conclusion only applies for dust and if one neglects the back reaction . since we lack definite knowledge about the evolution of the scalar field when a black hole forms in a cosmological background , it is useful to consider a range of scenarios which go beyond the two possibilities envisaged by barrow . in general , the background scalar field will have a present value @xmath109 and a value @xmath110 when the black hole first formed . however , it is likely to develop inhomogeneties for at least some intervening period . in the following discussion , we will characterize the degree of gravitational memory by comparing the value of the scalar field at the black hole event horizon ( @xmath111 ) and the cosmological particle horizon ( @xmath112 ) . we first consider the two extreme situations described by barrow @xcite : scenarioa : _ eh(t ) = _ ph(t ) for all t a schwarzschild black hole forms at time @xmath113 with its event horizon radius being @xmath114 . if @xmath115 evolves with time , then the black hole adjusts quasi - statically through a sequence of schwarzschild states approximated by @xmath116 , see fig . 4(a ) . in this scenario there is no gravitational memory . scenario b : _ eh(t ) = _ eh(t_f ) for all t a schwarzschild black hole of size @xmath117 forms at time @xmath113 and , while @xmath115 equals the evolving background value beyond some scale - length @xmath118 , it remains constant within @xmath119 , see fig . 4(b ) . in this case the black hole size is determined by @xmath120 even at the present epoch and this means that the region @xmath121 has a memory of the gravitational `` constant '' at the time of its formation . .2 in neither of these scenarios can be completely realistic since they both assume that @xmath52 is homogeneous almost everywhere . however , even if @xmath52 were homogeneous initially , one would expect it to become inhomogeneous as collapse proceeds . indeed , in the dust case , this is confirmed by the numerical calculations described above @xcite . therefore , if the background value is increasing ( as usually applies ) , one would expect @xmath52 in the collapsing region to become first _ larger _ than the background value on a local dynamical timescale and then _ smaller _ than it on a cosmological timescale . such behaviour would necessarily entail a variation of @xmath52 in space as well as time . we must also allow for the possibility that @xmath52 may vary interior to @xmath119 but on a slower or faster timescale than the background . we therefore propose two further scenarios : .2 in scenarioc : |_eh(t)| |_ph(t)| for all t where the dot represents a time derivative . this implies that the scalar field evolves faster at the event horizon than at the particle horizon until it eventually becomes homogeneous , see fig . we describe this as _ short - term _ gravitational memory and it reduces to scenario a as the timescale to become homogeneous tends to zero . this would apply , for example , if @xmath52 were to change on the dynamical timescales of the black hole since this is usually less than the cosmological timescale . scenario d : [ beta ] |_eh(t)| < |_ph(t)| for all t this implies that @xmath52 evolves faster at the particle horizon than the event horizon , see fig . we describe this is as _ weak _ gravitational memory and it reduces to scenario b when the left - hand - side of eqn ( [ beta ] ) is zero . in this case , the evolution of @xmath52 is again dominated by the black hole inside some length - scale @xmath119 . note that , in either this scenario or the last one , the length - scale @xmath119 need not be fixed , since it could either grow or shrink as scalar gravitational radiation propagates . a particular example of this , to which we return shortly , would be _ self - similar _ gravitational memory , in which the ratio of @xmath111 to @xmath112 always remains the same . .2 in [ f4 ] in general relativity there is an equivalence between a scalar field and a stiff fluid and this can be exploited in studying gravitational memory . in the einstein frame , the energy momentum tensor for a perfect fluid is |t _ = ( + p)u _ u _ + |g _ p where @xmath122 is the velocity of the fluid . if we define a velocity field by [ fvel ] u _ = , this gives |t _ = - + p |g_. by comparing this to the energy - momentum tensor for a scalar field , we find that p = = - |g^ | _ | _ , so we have a stiff fluid . this equivalence applies provided that the derivative of the scalar field is timelike . otherwise the velocity field defined in ( [ fvel ] ) would be imaginary . this is relevant to the gravitational memory problem because we can now interpret the various scenarios discussed in section 6 in terms of the _ accretion _ of a stiff fluid . if the black hole does not accrete at all or accretes very little , this will correspond to strong or weak gravitational memory ( scenarios b and d , respectively ) . however , if enough accretion occurs to homogenize @xmath52 , this will correspond to short - term gravitational memory ( scenario c ) . the faster the accretion , the shorter the memory , so scenario a corresponds to the idealization in which homogenization is instantaneous . a simple newtonian treatment @xcite for a general fluid suggests that the accretion rate in the einstein frame should be = 4r^2_a v_s , where @xmath123 is the accretion radius and @xmath124 is the sound - speed in the accreted fluid . for a stiff fluid , @xmath125 and @xmath126 , while @xmath127 in a friedmann universe at early times , so we have . this can be integrated to give [ mass ] m where @xmath128 is the black hole mass at the time @xmath113 when it formed . if we define a parameter @xmath129 , then eqn ( [ mass ] ) implies mm_f(1 - ) ^-1 as t. if @xmath130 , the black hole could not grow very much . however , if @xmath38 is close to 1 , which must be the case if @xmath131 , then the black hole could grow significantly . in particular , in the limit @xmath132 , eqn ( [ mass ] ) implies @xmath133 , so the black hole grows at the same rate as the universe . this simple calculation suggests that a black hole surrounded by a stiff fluid can accrete enough to grow at the same rate as the universe . since the above calculation neglects the effects of the cosmological expansion , one needs a relativistic calculation to check this . the newtonian result suggests that one should look for a spherically symmetric _ self - similar _ solution , in which every dimensionless variable is a function of @xmath134 , so that it is unchanged by the transformation @xmath135 for any constant @xmath136 . this problem has an interesting but rather convolved history . by looking for a black hole solution attached to an exact friedmann solution via a sonic point , carr & hawking first showed that there is no such solution for a radiation fluid @xcite and the argument can be extended to a general @xmath39 fluid with @xmath137 . lin et al . @xcite subsequently claimed that there is such a solution in the special case @xmath138 . however , bicknell & henriksen @xcite then showed that this solution is unphysical , in that the density gradient diverges at the event horizon . this suggests that the black hole must soon become much smaller than the particle horizon , after which eqn ( 32 ) implies there will be very little further accretion . therefore the stiff fluid analysis suggests that there should be at least _ weak _ gravitational memory . we have seen that studying the formation and evaporation of pbhs can place interesting constraints on models of the early universe even if they never existed . on the other hand , if they did exist , pbhs can provide unique information about times much earlier than those probed by any other relics of the big bang . in particular , they may provide information about the variation of @xmath0 at early times . the precise signature of such a variation depends upon the degree to which a black hole can remember " the value of @xmath0 at its formation epoch . this is still unclear but various methods are being pursued to resolve this issue . zeldovich and i.d . novikov , sov . 10 ( 1967 ) , 602 . s.w . hawking , mnras 152 ( 1971 ) , 75 . carr , in _ observational and theoretical aspects of relativistic astrophysics and cosmology _ , ed . j.l.sanz & l.j.goicoechea ( world scientific , 1985 ) p 1 . s.w . hawking , nature 248 ( 1974 ) , 30 . d.n.page and s.w . hawking , ap.j . 206 ( 1976 ) , 1 . n.a . porter and t.c . weekes , nature 277 ( 1979 ) 199 . zeldovich and a.a . starobinsky , jetp lett . 24 ( 1976 ) 571 . novikov , a.g . polnarev , a.a . starobinsky and ya.b . zeldovich , astron.astrophys . 80 ( 1979 ) 104 . barrow , mnras 192 ( 1980 ) 427 . okeke and m.j . rees , astron.astrophys . 81 ( 1980 ) 263 . p. kiraly et al . , nature 293 ( 1981 ) 120 . rees , nature 266 ( 1977 ) 333 . carr , in _ current topics in astrofundamental physics _ n.sanchez & a. zichichi ( world scientific , 1997 ) , p 306 . j. d. barrow , phys . d. 46 ( 1992 ) , r3227 . j. d. barrow and b. j. carr , phys . d. 54 ( 1996 ) , 3920 . c. goymer and b. j. carr , preprint ( 1999 ) . b. j. carr and s. w. hawking , mnras 168 ( 1974 ) , 399 . carr , ap.j . 201 ( 1975 ) , 1 . khlopov , b.e . malomed and ya.b . zeldovich , mnras 215 ( 1985 ) , 575 . naselsky and a.g . polnarev , sov.astron . 29 ( 1985 ) , 487 . b.j . carr and j.e . lidsey , phys.rev.d . 48 ( 1993 ) , 543 . carr , j.h.gilbert and j.e . lidsey , phys.rev.d . 50 ( 1994 ) , 4853 . p. ivanov , p. naselsky and i. novikov , phys.rev.d . 50 ( 1994 ) , 7173 . j. garcia - bellido , a. linde and d. wands , phys.rev.d . 54 ( 1997 ) , 6040 . a.m. green & a.r . liddle , phys.rev.d . 56 ( 1997 ) 6166 ; 60 ( 1999 ) , 0635616 . l. randall , m. soljacic and a.h . guth , nuc . b. 472 ( 1996 ) , 377 . j. yokoyama , astron . 318 ( 1997 ) , 673 . bassett and s. tsujikawa , preprint ( 2000 ) ; astro - ph/0008328 . a.m. green and k.a . malik , preprint ( 2000 ) ; hep - ph/0008113 . bullock and j.r . primack , phys.rev.d . 55 ( 1997 ) , 7423 . p. ivanov , phys . d. 57 ( 1998 ) , 7145 . m.yu . khlopov and a.g . polnarev , phys.lett.b . 97 ( 1980 ) , 383 . m. crawford and d.n . schramm , nature 298 ( 1982 ) , 538 . s.w . hawking , i. moss and j. stewart , phys.rev.d . , 26 ( 1982 ) , 2681 . d. la and p.j.steinhardt , phys.lett.b . 220 ( 1989 ) , 375 . polnarev and r. zemboricz , phys.rev.d . 43 ( 1988 ) , 1106 . s.w . hawking , b. 231 ( 1989 ) , 237 . j. garriga and m. sakellariadou , phys.rev.d . 48 ( 1993 ) , 2502 . r. caldwell and p. casper , phys . 53 ( 1996 ) , 3002 . macgibbon , r.h . brandenberger and u.f . wichoski , phys . d. 57 ( 1998 ) , 2158 . rubin , m.yu . khlopov and a.s . sakharov , grav . supp . 6 ( 2000 ) , 1 . carr , ap.j . 206 ( 1976 ) , 8 . macgibbon and b.j . carr , ap.j . 371 ( 1991 ) , 447 . macgibbon , phys . rev.d . 44 ( 1991 ) , 376 . macgibbon and b.r . webber , phys.rev.d . 41 ( 1990 ) , 3052 . macgibbon , nature 329 ( 1987 ) , 308 . barrow , e.j . copeland and a.r . liddle , phys.rev.d . 46 ( 1992 ) , 645 . dolgov , p.d . naselsky and i.d . novikov , preprint ( 2000 ) ; astro - ph/0009407 . j. niemeyer and k. jedamzik , phys.rev.d . 59 ( 1999 ) , 124013 . m. shibata and m. sasaki , phys.rev.d . ( 1999 ) ( in press ) ; gr - qc/9905064 . choptuik , phys . 70 ( 1993 ) , 9 . c. r. evans and j. s. coleman , phys . 72 ( 1994 ) , 1782 . d. maison , phys . lett b 366 ( 1996 ) , 82 . t. koike , t. hara and s. adachi , phys . d 59 ( 1999 ) , 104008 . j. niemeyer and k. jedamzik , phys.rev.lett . 80 ( 1998 ) , 5481 . h. kim , preprint ( 1999 ) . a.m. green and a.r . liddle , phys . d. 60 ( 1999 ) , 063509 . k. jedamzik , phys.rev.d . 55 ( 1997 ) , r5871 ; phys . 307 ( 1998 ) , 155 . k. jedamzik and j. niemeyer , phys.rev.d . 59 ( 1999 ) , 124014 . t. nakamura , m. sasaki , t. tanaka and k. thorn , astrophys.j . 487 ( 1997 ) , l139 . k. ioka , t. tanaka and t. nakamura et al . d. 60 ( 1999 ) , 083512 . k. ioka , astrophys.j . ( 1999 ) ( in press ) . k. kohri and j. yokoyama , preprint ( 1999 ) . bugaev and k.v . konischev , preprint ( 2000 ) ; astro - ph/0005295 . wright , ap.j . 459 ( 1996 ) , 487 . k. maki , t. mitsui and s. orito , phys.rev.lett . 76 ( 1996 ) , 3474 . d.b.cline , d.a . sanders and w. hong , ap.j . 486 ( 1997 ) , 169 . carr , phys . 307 ( 1998 ) 141 . coyne , c. sinnis & r. somerville , in _ proceedings of the houston advanced research center conference on black holes _ ( 1992 ) . f. halzen , e. zas , j. macgibbon , & t.c.weekes , nature 298 ( 1991 ) 538 . a. heckler , phys . 55 ( 1997 ) 840 . a. heckler , phys . b. 231 ( 1997 ) 3430 . j. kapusta , this volume ( 2001 ) . macgibbon , b.j . carr and d.n . page ( 2000 ) . t. kaluza , sitz.d.preuss.akad.d.wiss.physik-mat.klasse ( 1921 ) , 966 . p.a.m.dirac , proc . a165 ( 1938 ) , 199 . p. jordan , _ schwerkraft und weltall _ , vieweg ( braunschweig , 1955 ) . d. sciama , mnras 113 ( 1957 ) , 34 . c. brans and r. h. dicke , phys . 124 ( 1961 ) , 925 . will , _ theory and experiment in gravitational physics _ ( cambridge , 1993 ) . bergmann , int.j.theor.phys . 1 ( 1968 ) r.v.wagoner , phys.rev.d.1 ( 1970 ) , 3209 . k. nordtvedt , astrophys . j. 161 ( 1970 ) , 1059 , a. guth , phys . rev . d. 23 ( 1981 ) , 347 . p.g.o.freund , nuc . b. 209 ( 1982 ) , 146 . k. maeda , class . grav . 3 ( 1986 ) , 233 . kolb , m.j.perry and t.p.walker , phys.rev.d . 33 ( 1986 ) , 869 . callan et al . , nuc . b. 262 ( 1985 ) 597 . t. damour and a.m. polyakov , nuc.phys.b 423 ( 1994 ) 532 . p.j . steinhardt and f.s . accetta , phys . 64 ( 1990 ) 2740 . barrow and p. parsons , phys . d , 55 , 1906 ( 1997 ) . s. w. hawking , commun . phys . 25 ( 1972 ) , 167 . m. a. scheel , s. l. shapiro and s. a. teukolsky , phys . rev . d. 51 ( 1995 ) , 4208 . m. a. scheel , s. l. shapiro and s. a. teukolsky , phys . d. 51 ( 1995 ) , 4236 . j. kerimo and d. kalligas , phys . d. 58 ( 1998 ) 104002 . t. harada , t , chiba , k. nakao and t. nakamura , phys . d. 55 ( 1997 ) , 2024 . a. einstein and e. g. straus , rev . 17 ( 1945 ) , 120 . c. b. g. mcintosh , phys . 43a(1973 ) , 33 . t. jacobsen , phys . 83 ( 1999 ) , 2699 . t. harada , b.j . carr and c.a . goymer , preprint ( 2000 ) . d. n. c. lin , b. j. carr and s. m. fall , mnras 177 ( 1976 ) , 51 . g. v. bicknell and r. n. henriksen , ap . j. 219 ( 1978 ) , 1043 .

the markovian quantum master equation ( qme ) @xcite provides a key paradigm for the study of nonequilibrium statistical physics . the structure of this framework is transparent , the generality of its derivation is sufficient , and the approximations applied are well defined , especially when the system hamiltonian is time - independent . in the last decades , therefore , the markovian qme has allowed a broad range of applications for the study of open quantum systems that have _ time - independent _ system hamiltonians @xcite . in recent years , however , the quantum dynamics driven by _ time - dependent _ external field ( see also fig . [ fig : schematic ] ) has been an area of growing importance in various contexts , such as adiabatic quantum computation ( aqc ) @xcite , quantum annealing ( qa ) @xcite , quantum heat engines @xcite , bose - einstein condensates in optical lattices @xcite , and semiconductor quantum dots @xcite because the external driving is essential for the use of the quantum systems as nontrivial physical resources . nevertheless , no general way of rigorously constructing the markovian qme is currently known for the time - dependent system hamiltonian , @xmath1 @xcite . one therefore often comes across a fundamental problem of describing the open quantum dynamics simultaneously subject to the external driving . one possible way to circumvent this difficulty is the floquet formalism if the system hamiltonian is driven periodically in time @xcite . however , this approach is obviously inadequate in situations with no periodicity . another natural approach is then to assume the slow temporal change of @xmath1 . however , the following two questions immediately arise in turn : how can we quantify the temporal change timescale of @xmath1 , which we denote by @xmath0 , and what is the other timescale to be compared with @xmath0 when we say slow ? in relation to these questions , various different statements can be found in literature . in particular , an important assumption employed by many authors is the adiabatic regime that satisfies the ordinary adiabatic theorem @xcite , in which any non - adiabatic transitions between instantaneous eigenstates are suppressed @xcite . however , under this assumption , any non - adiabatic effect can not be discussed despite the recent experimental demand ; in the context of the qa , for example , it is essential to consider the non - adiabatic transitions with dissipation in practice @xcite . alternatively , another assumption sometimes encountered is the slow temporal change of @xmath1 compared to the decay timescale of the bath correlation function , @xmath2 . however , due to the lack of the general definition of @xmath0 , there is no consensus for this condition even if simple systems , e.g. the dissipative landau - zener ( dlz ) models , are considered @xcite . the conditions for the markovian qme are thus subject to long - standing debate when the system hamiltonian depends on time , originally started from davies and spohn @xcite . in this article , our purpose is two - fold . first , under the weak - coupling approximation ( wca ) , we settle this long - standing debate by studying the conditions to justify the markovian qme with @xmath1 in a general manner . we introduce an explicit definition of @xmath0 and approximations required to obtain the qme under the wca . as a result , we shall see that the markovian qme is naturally derived _ without the adiabatic theorem_. our route of the formulation successfully removes the ambiguities of the relevant conditions . second , in a broad range of situations , we further find that there is no need to even assume the slow temporal change of @xmath1 compared to @xmath2 . this is in contrast to the common belief that the markovian description breaks down when the system hamiltonian changes more rapidly than @xmath2 @xcite . in consequence , the non - adiabatic regime becomes definitely accessible by the markovian qme . our route of the formulation and the well - defined approximations allow us to clearly understand the structure of the framework with sufficient generality . hence , the approach is immediately applicable to a wide range of physical systems . as an example , the results are demonstrated by applying the framework to the dlz model . our scheme thus resolves the contentious issue , achieves an extension of the applicable range , and as a result opens up a new avenue for exploring the frontier in driven open quantum systems beyond the adiabatic regime . the paper is organized as follows . in section [ sec : adiabaticity ] , we introduce @xmath0 that quantifies the temporal change of @xmath1 by assuming the _ closed system_. we also discuss the similarities and differences between the adiabaticity of @xmath1 and the ordinary adiabatic theorem . next , in section [ sec : qme ] , we study the general derivation of the markovian qme under the wca . to help explain the framework clearly , we shall make a brief review of the _ time - independent _ hamiltonian case . we then introduce an approximation based on the instantaneous eigenbasis and illustrate the validity beyond the adiabatic regime . in section [ sec : dlz ] , we demonstrate the results by applying the framework to the dlz model . finally , in section [ sec : conclusions ] , we conclude and give an outlook . throughout the paper , we set @xmath3 for simplicity . we consider an open quantum system with the time - dependent system hamiltonian @xmath1 . although this topic has been studied by many authors in the past @xcite , there is no discussion on how to quantify the temporal change timescale @xmath0 of the system hamiltonian . therefore , in this section , we start with the definition of @xmath0 by considering the _ closed system_. for this purpose , it is convenient to consider a problem about in what condition @xmath4 remains unchanged from @xmath1 , @xmath5 where @xmath6 is an arbitrary time . we note that this problem is different from the traditional adiabatic theorem @xcite in the following sense . in the adiabatic theorem , the time evolution of the system s state @xmath7 is considered according to the schrdinger equation @xmath8 . then , one finds that the transitions between the instantaneous eigenstates are suppressed when @xmath1 changes sufficiently slowly if there is no degeneracy . in contrast , in eq . , we do _ not _ focus on the time evolution of @xmath7 but our attention is paid to the time dependence of @xmath1 itself . therefore , it is always possible to find @xmath9 that satisfies eq . , no matter how fast @xmath1 changes when @xmath1 is analytic . in this context , @xmath9 would be arbitrary as long as it is sufficiently small compared to a certain timescale . here , in fact , it is natural to introduce @xmath10 by this certain timescale because , if @xmath10 denotes the temporal change timescale of @xmath1 , eq . should be satisfied when @xmath9 is much shorter than @xmath10 , @xmath11 therefore , we can say that @xmath1 is adiabatic , or `` slow '' , within the time @xmath9 when @xmath9 satisfies the condition of eq . . in this sense , @xmath0 would give a clear measure for the adiabaticity of @xmath1 itself . however , eqs . and alone do not make much practical sense because we do not know how to estimate @xmath0 . as a convenient way to obtain @xmath10 , therefore , we use the instantaneous eigenbasis , defined by @xmath12 , where @xmath13 and @xmath14 denote the @xmath15-th instantaneous eigenstate and eigenenergy , respectively ; @xmath16 labels the ground state . the states @xmath13 are assumed to be normalized and nondegenerate . we then read eq . in this basis as @xmath17 for all @xmath15 . here , in a similar manner to eq . , we can separately introduce @xmath18 and @xmath19 that describe the temporal change timescales of the instantaneous eigenenergy and eigenstate . hence , the first and second approximations in eq . are appropriate , respectively , when @xmath20 as a result , @xmath10 is given by @xmath21 because @xmath22 and @xmath23 should both be satisfied for @xmath24 . since the taylor expansion of @xmath25 up to the first order yields @xmath26 , we obtain @xmath18 as @xmath27 then , one can indeed obtain @xmath22 if @xmath28 . on the other hand , for @xmath29 , @xmath30 should be satisfied when @xmath23 because @xmath31 . therefore , since @xmath32 , @xmath19 is given by @xmath33 where the maximum is taken over all @xmath34 and @xmath15 except for @xmath35 . thus , @xmath10 can be estimated from eqs . with the instantaneous eigenbasis . for the convenience of the reader , the timescales and the situations discussed in this paper are summarized in tables [ tbl : timescales ] and [ tbl : conditions ] . while there are a multitude of timescales , we should mention that @xmath0 , @xmath18 and @xmath19 are the only timescales essentially introduced in this paper . the other timescales are just straightforward extensions of previously known timescales . at this stage , let us discuss the time evolution of @xmath7 . this is actually equivalent to discuss the evolution operator defined by @xmath36 where @xmath37 ( @xmath38 ) denotes the chronological ( anti - chronological ) time ordering operator . the formal solution of the schrdinger equation is then given by @xmath39 with the initial time @xmath40 because @xmath41 satisfies [ eq : evprop2 ] @xmath42 for @xmath43 later than @xmath44 ( @xmath45 ) . we note , in addition , that the evolution operator also satisfies [ eq : evprop1 ] @xmath46 by definition . then , if @xmath6 satisfies eq . , the evolution operator from @xmath47 to @xmath48 , i.e. @xmath49 , can be approximated by @xmath50 where the integral variable is changed to @xmath51 [ see also eq . ] in the first line and eq . has been used in the second line . this equation means that the evolution of the eigenstates can be approximated just by the dynamical phase shift , @xmath52 when @xmath9 is much shorter than @xmath0 . these discussions based on the concept of the adiabaticity of @xmath1 will play an important role when we develop the markovian qme with @xmath1 in the following sections . here , it would be instructive to see the similarities and differences between the adiabaticity of @xmath1 and the ordinary adiabatic theorem . for this purpose , we introduce an _ intrinsic evolution timescale _ of @xmath1 by @xmath53 where @xmath54 denotes the instantaneous energy gap . hence , @xmath55 also corresponds to the longest timescale of the intrinsic oscillation period in the off - diagonal density matrix elements . one can then easily confirm that the adiabatic theorem is validated if @xmath56 is satisfied for all relevant time @xmath48 ( see ref . , for example ) by using @xmath57 . equation means that all the eigenstates @xmath13 should be adiabatic with respect to @xmath58 . in other words , the eigenstates @xmath13 should remain unchanged at least within the intrinsic oscillation period @xmath58 to satisfy the adiabatic theorem . in that case , the evolution operator @xmath41 can be approximated by @xmath59 with @xcite @xmath60 where @xmath61 and @xmath62 denotes the berry connection . notice that , according to eq . , the @xmath15-th eigenstate @xmath63 at time @xmath44 will evolve to the @xmath15-th eigenstate @xmath64 at time @xmath43 with the phase shift @xmath65 . therefore , eq . directly means that there are no transitions between the instantaneous eigenstates . we refer to this type of dynamics as the _ adiabatic evolution_. by comparing eq . with eq . [ eq . with eq . ] , the adiabaticity of @xmath1 is formally similar to the adiabatic theorem . however , they are conceptually different from each other , as described below eq . . according to eq . , we can indeed always find @xmath9 even if @xmath0 becomes short due to a rapid change of @xmath1 . as a result , eq . is validated , in which the eigenstates are left unchanged except for their phase factors . in contrast , eq . fails when @xmath19 becomes short due to the rapid change of @xmath1 . this is because @xmath55 is fixed once the structure of @xmath1 is determined . therefore , the adiabatic evolution [ eq . ] is valid only when @xmath1 varies slowly to satisfy eq . . with the help of this distinction , we shall see below that the markovian qme is naturally derived without the adiabatic theorem . [ cols="^ , < , < , < " , ] we now turn to the study of the markovian qme under the wca . here , in order to fix our notations as well as to give a self - contained presentation , let us shortly review the standard steps to obtain the markovian qme . in the schrdinger picture , the general hamiltonian we consider is @xmath66 where @xmath67 is the hamiltonian of the bath and @xmath68 is the system - bath interaction hamiltonian . without loss of generality , we assume that @xmath69 and @xmath70 are the hermitian operators acting only on the system and bath hilbert spaces , respectively @xcite . the total density operator @xmath71 evolves according to the von neumann equation ; @xmath72 $ ] . let @xmath73 describe the evolution operator for @xmath74 that is defined in the same way as eq . except for the replacement @xmath75 . we then transform into the interaction picture with respect to @xmath74 by introducing @xmath76 where @xmath77 is an arbitrary operator . in the interaction picture , therefore , the von neumann equation is written as @xmath78 $ ] . as a result , the standard time - convolutionless technique using projection superoperators @xcite yields the equation of motion for the reduced density operator @xmath79}$ ] , @xmath80 ] } , \label{eq : nmredfield1}\end{aligned}\ ] ] up to the second order in @xmath81 , or equivalently within the wca ; see also appendix [ tcl ] . in the derivation , we have assumed that the initial state is separable , @xmath82 , and that the odd moments of @xmath83 with respect to @xmath84 vanish ; @xmath85}=0 , \label{eq : oddmoments}\end{aligned}\ ] ] where @xmath86 . by inserting @xmath87 into eq . , with a change of the integration variable to @xmath88 , we obtain @xmath89 where @xmath90 denotes the elapsed time from the initial time @xmath40 and @xmath91}$ ] is the bath correlation function . the typical decay timescale of @xmath92 will be denoted by @xmath93 . in the derivation of eq . , we have assumed that @xmath84 is in the steady state , i.e. @xmath94 = 0 $ ] , to obtain @xmath95 } = { \text{tr}_{\text{b } } [ \vb_{\alpha}(\tau ) \vb_{\beta}(0 ) \orho_{\b}]}$ ] for simplicity . the thermal bath with the temperature @xmath96 is the most typical example ; @xmath97}$ ] . we remark that eqs . and are time - convolutionless in the sense that there is no time convolution in terms of @xmath98 . however , the time convolution between @xmath99 and @xmath100 is still required , which will cause a practical difficulty when we consider the time - dependent hamiltonian explicitly . _ the markovian approximation_here , in eq . , @xmath99 decays to zero well after the decay timescale of the bath correlation function @xmath101 . therefore , the upper limit of the integral may be approximated by infinity if we are only interested in the system dynamics over the time which is longer than @xmath2 ; @xmath102 under this approximation , known as the markovian approximation @xcite , we finally obtain @xmath103 we note however that , in the vicinity of the initial time , @xmath104 , the markovian approximation is not validated because @xmath99 does not yet sufficiently decay . as a result , errors accumulated in @xmath104 sometimes largely affect the subsequent evolution even for @xmath105 . this problem is nevertheless avoidable by using the renormalized or slipped " initial condition @xcite ( see also appendix [ slp ] ) . therefore , in the followings , we limit ourselves only to the markovian qme under the wca , which is equivalent to the qme under the born - markov approximation @xcite . we do not go into the non - markovian dynamics recently highlighted by several authors @xcite but the concept of the markovianity is used in the same spirit that there is no memory effect coming back from the bath @xcite . _ time - independent case_equation can be rewritten in a more tractable form when the system hamiltonian is time - independent ; @xmath106 , @xmath107 , and @xmath108 . to see this , let us define @xmath109 where the summation in eq . is taken over all @xmath34 and @xmath15 satisfying @xmath110 . then , the summation over all possible @xmath111 ( the bohr frequencies of the system ) yields the spectral decomposition , @xmath112 due to the completeness of the eigenbasis . since the evolution operator @xmath113 [ eq . ] is simply described by @xmath114 for @xmath106 , @xmath100 in eq is given by @xmath115 where @xmath116 is the interaction picture of @xmath117 ; @xmath118 . note that the key point in eq . is that the @xmath119-dependence of @xmath120 is separated from the operator part . therefore , by inserting eq . into eq . , the time convolution between @xmath99 and @xmath100 yields @xmath121 where @xmath122 denotes the one - sided fourier transform of the bath correlation function @xmath123 notice that this is an analogy to the ordinary convolution theorem , i.e. the convolution in time domain is equivalent to the multiplication in the fourier domain . as a result , one obtains the standard form of the qme in the interaction picture , @xmath124 where the dissipator @xmath125 is defined by @xmath126 however , the dissipator @xmath125 is still not described by the lindblad form and this means that the complete positivity is not guaranteed . therefore , the secular approximation ( sa ) , or equivalently the rotating wave approximation ( rwa ) , is often applied , where rapidly oscillating terms in eq . is neglected ( averaged out to zero ) . this is performed by putting @xmath127 [ cf . ] into eq . and by neglecting all terms with @xmath128 . we note that this approximation is based on the oscillating forms of @xmath129 and @xmath130 . then , we finally arrive at @xmath131 with @xmath132 which indeed results in the lindblad form . note however that the sa is justified only when all the differences of the bohr frequencies , i.e. @xmath133 for @xmath128 , are larger than the relaxation rate of the system @xcite . on the one hand , the relaxation rate of the system @xmath134 is characterized by @xmath135 where @xmath136 . on the other hand , we can define an intrinsic timscale @xmath137 by using the minimum of @xmath133 for @xmath128 ; @xmath138 therefore , the condition for the sa is described by @xmath139 , or equivalently , @xmath140 here , we refer to @xmath137 as the intrinsic beat timescale to distinguish from eq . , while @xmath137 is also called the intrinsic evolution timescale @xcite . this is because the differences of the bohr frequencies correspond to the beat frequencies of the intrinsic oscillation of the off - diagonal density matrix elements . if the condition is not satisfied , the sa sometimes leads to unphysical results and one often has to use @xmath125 or alternatively approximated lindblad qme @xcite . now , we return to the problem of the time - dependent system hamiltonian , @xmath1 . in this case , our starting point is eq . . although this equation can be solved numerically in principle , the numerical cost would not be low in practice because eq . requires the numerical @xmath119-integration at each time step . it is therefore desirable to analytically reduce eq . into a similar form to @xmath125 [ eq . ] or @xmath141 [ eq . ] . in this context , one often assumes that @xmath1 is slowly varying , and then , replaces the bohr frequencies @xmath111 , the eigenenergies @xmath142 , and the eigenstates @xmath143 in the dissipators by the corresponding time - dependent ones , @xmath144 , @xmath14 , and @xmath13 , respectively . indeed , such an approach has been employed by childs _ et al_. @xcite and its microscopic derivation has recently been shown by albash _ et al_. , based on the ` ideal ' adiabatic evolution operator @xmath145 [ eq . ] , or equivalently , the adiabatic theorem @xcite . in contrast , our purpose below is to derive such dissipators _ without using the adiabatic theorem_. to this end , we first introduce @xmath146 in the same manner as eqs . and . therefore , the summation over all possible @xmath144 yields the decomposition similar to eq . , @xmath147 again due to the completeness of the ( instantaneous ) eigenbasis . as illustrated in the previous section , the advantage to use such a decomposition was to separate the @xmath119-dependence from @xmath100 , based on the explicit calculation of the time dependence in the interaction picture [ see eq . ] . then , the time convolution between @xmath99 and @xmath100 becomes possible [ eq . ] . however , this approach is now _ not _ allowed for the time - dependent hamiltonian @xmath1 because the evolution operator @xmath113 includes the time integration of @xmath1 [ eq . ] . in order to avoid this difficulty , according to eq . , we rewrite @xmath100 in the following form , @xmath148 where the evolution operator @xmath149 has been separated into two parts ; @xmath150 . since @xmath119 is positive in eq . , the form of @xmath151 is reminiscent of eq . . by considering that the @xmath119-integration in eq . converges for @xmath152 due to the decay of @xmath99 , we can indeed approximate @xmath151 by @xmath153 when @xmath154 as a result , eq . yields @xmath155 where @xmath156 is the interaction picture of @xmath157 , @xmath158 , and we have used eqs . . it is important here to notice that the @xmath119-dependence is separated from the operator part and eq . corresponds to a straightforward extension of eq . . we refer to eq . as the instantaneous eigenbasis approximation ( iea ) . note that the iea becomes exact and recovers eq . when @xmath1 is time - independent . this is consistent with the condition for the iea [ eq . ] because @xmath0 goes to infinity based on eqs . when @xmath106 . by applying the iea [ eq . ] , we can then perform the same steps as eqs . to obtain the qme in the interaction picture . the result is @xmath159 where the dissipator @xmath160 is defined by @xmath161 these equations are formally the same as eq . except that @xmath162 , @xmath163 , and @xmath164 . it follows that eqs . and do not guarantee the complete positivity due to the non - lindblad form . therefore , the sa is again required to obtain the lindblad form in a similar manner to the time - independent case . however , we remark that @xmath165 in @xmath160 even though @xmath130 holds true in @xmath125 because the interaction picture can not be explicitly obtained when the system hamiltonian is time - dependent . actually , this difference between @xmath160 and @xmath125 makes the application of the sa more difficult than the time - independent case since the sa is based on the oscillating forms of @xmath166 and @xmath116 , as seen above eq . . nevertheless , we can show that the sa is still possible ( see also appendix [ sec : sa ] ) when @xmath167 are simultaneously satisfied . here , @xmath168 and @xmath169 are the straightforward extensions of eqs . and , @xmath170 where @xmath168 and @xmath169 denote the characteristic relaxation timescale of @xmath98 and the intrinsic beat timescale between the different bohr frequencies @xmath144 and @xmath171 , respectively . physically speaking , therefore , eq . means that @xmath98 and @xmath1 must remain unchanged until the beat oscillations of @xmath166 and @xmath116 are sufficiently developed . hence , eq . is a natural extension of eq . in the spirit of the sa . under this condition , we can obtain @xmath172 with @xmath173 which is again formally the same as @xmath141 [ eq . ] . thus , the dissipator can be described in the lindblad form . finally , for practical use , it is better to transform back into the schrdinger picture . with the help of eq . , eq . gives @xmath174 + \od^{\i\e\a}_t\orho_{\s}(t ) , \label{eq : sp::qmeiea}\end{aligned}\ ] ] where the dissipator @xmath175 is @xmath176 after the application of the sa , @xmath177 is justified with @xmath178 in the schrdinger picture . these markovian qmes are the direct generalizations of the standard markovian qmes for the time - independent hamiltonian . obviously , the obtained qmes recover the standard qmes if we assume @xmath106 . in the previous section [ subsec : iea ] , we have derived the markovian qmes within the iea . however , we have to mention that the forms of the dissipators [ eqs . and ] have already been known or expected from a long time ago @xcite . this is _ not _ surprising because the derived dissipators can be obtained just by intuitively replacing the time - independent variables by the corresponding time - dependent ones . however , it has been believed that the applicable range is limited to the adiabatic - evolution regime that satisfies the ordinary adiabatic theorem . following the recent arguments in ref . , for example , the evolution operator @xmath179 has been approximated by the ` ideal ' adiabatic - evolution operator @xmath180 [ eq . ] to avoid the difficulty originating from the time convolution between @xmath99 and @xmath181 . however , such an approximation is _ not _ needed at all in the formulation shown above . in this sense , the applicable range is already extended beyond the adiabatic regime . this is our first important result . however , based on the formulation above , the iea is still limited to the time regime of @xmath48 satisfying @xmath182 . this means that @xmath1 should remain unchanged at least within @xmath2 to justify @xmath175 . nevertheless , in a broad range of situations , our claim of the validity is _ not _ restricted to just the regime of @xmath182 for the iea ; the @xmath175 is still well justified even if @xmath182 fails . to show this result , we consider the time regime satisfying @xmath183 in this regime , @xmath1 is driven much more rapidly than the relaxation time scale . this means that @xmath175 [ the second term in eq . ] has only a minor effect on the dynamics , compared to the hamiltonian dynamics [ the first term in eq . ] . in such a situation , only a rough evaluation of the superoperator @xmath175 would be sufficient . in other words , the effect of the relaxation is negligible if we focus on the dynamics in the timescale of @xmath0 when @xmath184 . we can therefore state that the qme with @xmath175 [ eqs . and ] strongly breaks down only when the iea is invalid [ @xmath185 , and simultaneously , the effect of the relaxation is non - negligible [ @xmath186 , namely , @xmath187 however , this condition can never be satisfied as long as the sufficient condition for the wca holds , as shown below . for the wca , which is correct in the second order of @xmath81 , the fourth order contribution must sufficiently be small in comparison to the second order @xcite because the third order contribution vanishes due to the assumption of eq . . here , the second order contribution is roughly estimated by @xmath188 in the @xmath119-integral of eq . , whereas the corresponding forth order is similarly estimated by @xmath189 . the wca therefore remains valid when @xmath190 , or equivalently , @xmath191 for all @xmath192 and @xmath193 . here , due to the fourier - transform relation between @xmath194 and @xmath99 , the spectral bandwidth of @xmath194 , denoted by @xmath195 , satisfies @xmath196 . we can therefore estimate @xmath197 , where @xmath198 is the peak value of @xmath194 . putting this into eq . , we can find @xmath199 as the sufficient condition for the wca , where @xmath200 is defined by @xmath201 since @xmath202 holds by definition , the sufficient condition for the wca [ eq . ] results in @xmath203 which is , however , incompatible with eq . . therefore , if we limit ourselves to the wca regime , eq . can never be satisfied . thus , the strong breakdown of the qme with @xmath175 [ eqs . and ] is entirely avoidable as long as the wca condition is satisfied . although it is well - known that simple hamiltonian evolution is sufficient to characterize the system dynamics for @xmath204 , the important point here is that @xmath204 is _ always _ satisfied when @xmath205 because @xmath206 under the wca . as a result , we obtain our second important result ; @xmath175 is well justified even though @xmath0 is comparable to or even much shorter than @xmath2 , i.e. @xmath207 . since @xmath175 is well justified also for @xmath208 [ eq . ] as described in section [ subsec : iea ] , this means that the markovian qme can be justified regardless of the speed of @xmath209 . this is in contrast to the previous studies @xcite in which it is believed that the markovian description breaks down when @xmath209 changes more rapidly than @xmath210 . the applicable range of the iea can thus be significantly extended to the regime where the temporal change of @xmath1 is faster than @xmath2 . nevertheless , we have to take care the accumulated time of the iea being invalid , i.e. @xmath207 . this is because the markovian qme still weakly breaks down in this regime and the error would be accumulated . therefore , the presented approach of @xmath175 would not be applicable if the accumulated time becomes comparable to the relaxation time . in that case , we should return to eq . even though its numerical cost is not low in general . however , even in such a case , we stress that our discussions on the adiabaticity of @xmath1 ( section [ sec : adiabaticity ] ) give a clear guide to reduce the numerical effort in the following way . since @xmath0 is now well defined by eqs . , we can calculate @xmath0 simultaneously with the density operator @xmath211 . then , if the iea is valid at this moment , @xmath182 , the next time step can be obtained , based on @xmath175 . if not , @xmath212 , the next time step will be calculated by eq . . thus , it is only the time region @xmath212 that requires eq . and the numerical cost may be greatly reduced by this approach . finally , let us mention the applicability of the sa in a similar line of thought to the above discussion . to perform the sa , we showed that @xmath213 and @xmath214 [ eq . ] are further required . the first condition is not related to the adiabaticity of @xmath1 , and therefore , we assume that @xmath213 is indeed satisfied . in contrast , the second condition breaks down when @xmath1 rapidly varies within @xmath169 . however , the dissipator plays only a minor role if @xmath184 [ eq . ] as discussed above . therefore , the qme with @xmath215 strongly breaks down only when @xmath214 fails [ @xmath216 , and simultaneously , the effect of the relaxation is non - negligible [ @xmath186 , namely , @xmath217 however , this condition can never be satisfied as long as the first condition , @xmath213 , holds . as a result , @xmath215 is still well justified as long as @xmath213 if the accumulated time of @xmath214 being invalid is still much shorter than the relaxation timescale . to demonstrate the presented general ideas , in the following , we apply the markovian qme to the dlz model , in which the landau - zener ( lz ) transition in a dissipative environment is studied . this model has been studied in various contexts @xcite because it provides the simplest model to describe the adiabatic and non - adiabatic transitions at an avoided level crossing with dissipation . especially , in the aqc and the qa , the dlz model plays a key role in understanding their computational ability . this is because their practical performance is essentially determined by the non - adiabatic transitions at the minimum gap between the ground state and the first excited state in the presence of noise @xcite . furthermore , the model is appropriate for our purpose because the exact solution is known at @xmath218 @xcite . the system hamiltonian of the dlz model is described by @xmath219 where @xmath220 is the lz sweep velocity , @xmath221 is the constant tunneling amplitude , and @xmath222 describe the pauli operators . we denote the eigenstates of @xmath223 by @xmath224 , i.e. @xmath225 . a schematic illustration of the model is shown in fig . [ fig : dlzmodel ] . here , @xmath226 denotes the difference between the two eigenenergies and we can see the avoided level crossing at @xmath227 . the characteristic time for the eigenstates to pass through the minimum gap @xmath228 around @xmath227 is described by @xmath229 the quantities of our interest is then the adiabatic and non - adiabatic transition probabilities @xmath230 and @xmath231 with dissipation by assuming that the system is initially in @xmath232 at @xmath233 . note that @xmath234 by definition . in this situation , the timescales related to @xmath1 ( section [ sec : adiabaticity ] ) are analytically obtained as @xmath235 and @xmath236 according to the definitions listed in table [ tbl : timescales ] . then , @xmath0 is given by @xmath237 here , we note that @xmath19 reaches its minimum value @xmath238 at @xmath227 , while @xmath18 reaches the same minimum value @xmath238 at @xmath239 . therefore , the minimum value of @xmath0 also becomes @xmath238 , which is achieved not only at @xmath240 but also at @xmath239 [ cf . [ fig : evolution](d ) ] . we remark that the expression of @xmath0 can not be obtained without eqs . even in this simple model and this is why there has been no consensus in the previous studies when discussing the timescale of @xmath1 . at @xmath241 . then , the transition probability @xmath230 ( @xmath231 ) to find the system in the ground ( excited ) state at @xmath242 is discussed . the difference between the two eigenenergies is given by @xmath243 . @xmath244 denotes the characteristic time for the eigenstates to pass through the minimum gap @xmath228 around @xmath227 . when the bath is at zero temperature ( @xmath245 ) , there is no thermal excitation and relaxation but only the spontaneous emission can occur due to the vacuum fluctuation . ] ( blue ) and @xmath246 ( red ) at @xmath218 , as a function of the lz sweep velocity @xmath220 . symbols : the results of the markovian qme . we remark that the imaginary part of @xmath247 ( the lamb shift ) is not neglected in the calculation . solid lines : the exact probabilities of @xmath248 with @xmath249 @xcite . note that @xmath250 ( diagonal coupling ) gives @xmath251 , which results in @xmath252 . this is exactly the same as the lz transition probability _ without dissipation _ , as pointed out in ref . . parameters : @xmath253 and @xmath254 . the left arrows indicate the regimes by taking @xmath255 as @xmath256 for all @xmath48 ( @xmath257 , @xmath258 ) . one is the ordinary adiabatic - evolution regime and the other is the adiabatic regime with respect to @xmath2 . the green open squares correspond to the time evolutions in fig . [ fig : evolution ] . ] the system is further coupled to the bosonic bath @xmath259 with the interaction hamiltonian , @xmath260 where @xmath261 describes the coupling angle and @xmath262 is the system - bath coupling strength . for definiteness , in our analysis , the spectral density defined by @xmath263 is assumed to be the ohmic one , @xmath264 with a cutoff energy @xmath265 and a dimensionless coefficient @xmath266 . in comparison with the general form of @xmath81 [ eq . ] , eq . allows us to define @xmath267 where we have dropped the subscript ` @xmath192 ' . therefore , by assuming that the bath @xmath84 is in the thermal equilibrium , @xmath268 with @xmath269}$ ] [ see below eq . ] is given by @xmath270 j(\omega)\vartheta(\omega ) , \label{eq : dlz::gamma}\end{aligned}\ ] ] where @xmath271 is the bose distribution , @xmath272 is the step function . here , it is important to note that , if the sa is performed , one can find that @xmath273 and @xmath274 describe the excitation and relaxation of the system , respectively , while @xmath275 describes the pure dephasing . this means that the terms proportional to @xmath276 and @xmath277 denote the thermal excitation and relaxation of the system , respectively . in contrast , the ` @xmath278 ' term corresponds to the spontaneous emission due to the vacuum fluctuation of the bath . in particular , at @xmath245 , only the spontaneous emission term survives in eq . and @xmath279 holds . this situation is also depicted in fig . [ fig : dlzmodel ] . since the dissipative nature of the system is essentially determined by @xmath280 , let us briefly discuss the two parameters @xmath266 and @xmath265 before studying numerical results . in the case of zero temperature , @xmath279 gives the bath correlation function @xmath281 by @xmath282 where the fourier transform relation between @xmath283 and @xmath281 has been used . hence , we can estimate the decay time of the bath correlation function @xmath281 as @xmath284 we note that this is equivalent to the statement that the spectral band width of @xmath283 can be estimated by @xmath285 . as a result , if we set @xmath286 , there is no spectral density at the transition energy @xmath287 , i.e. @xmath288 , and the dissipative nature plays no significant role . for our purpose , therefore , we should set @xmath289 . moreover , by using eqs . and , we can estimate the condition for the wca , @xmath290 due to eq . . the wca becomes valid when this condition is satisfied . based on these considerations , we have performed numerical calculations by setting @xmath291 and @xmath292 . although the sa will not be applied in our calculation , we note that the sa ( @xmath177 ) does not show any significant difference in the parameter range presented below . figure [ fig : probabilities ] shows the transition probabilities @xmath293 and @xmath246 as a function of the lz sweep velocity @xmath220 . for comparison , the exact results are also shown by solid lines , where the transition probability for @xmath250 ( diagonal coupling ) is exactly the same as the lz transition probability _ without dissipation _ @xcite . in fig . [ fig : probabilities ] , when @xmath294 for all @xmath48 ( the pink shaded area ) , the lz sweep velocity @xmath220 is slow enough to satisfy the adiabatic theorem . in this regime , the spontaneous emission does not play any role because the state is always in the instantaneous ground state . as a result , @xmath295 holds for all @xmath261 . by increasing the velocity @xmath220 , the condition for the adiabatic theorem is violated and the non - adiabatic transition becomes discernible . in this regime that still satisfies @xmath182 for all @xmath48 ( the aqua shaded region ) , @xmath230 decreases from one , and instead , @xmath231 increases . here , we can notice that @xmath230 for @xmath296 and @xmath297 are greater than for @xmath250 . this means that , for @xmath296 and @xmath297 , the ground state is recovered by the spontaneous emission even though the non - adiabatic transition is possible . by increasing the velocity @xmath220 further ( @xmath298 ) , the decrease of @xmath230 becomes pronounced for all @xmath261 and @xmath231 approaches one . in this regime , the recovery of the ground state becomes incomplete because the transition energy @xmath287 rapidly goes through the spectral bandwidth of @xmath283 before the ground state is sufficiently recovered . now , it is important to point out that excellent agreements between the exact probabilities ( solid lines ) and the results of the markovian qme ( symbols ) are obtained , regardless of the speed of @xmath299 . this is notable in the following two points . first , it has been believed that the markovian description can not capture the exact probabilities of the dlz model due to the non - markovian effect @xcite . nevertheless , fig . [ fig : probabilities ] clearly shows that the markovian description is sufficient at least in the weak coupling regime . second , more importantly , the agreements are remarkable even if the temporal change of @xmath1 , i.e. @xmath0 , is comparable to or even faster than @xmath2 . this is in stark contrast to the conventional expectation that the markovian qme is applicable only when the system hamiltonian varies much more slowly than @xmath2 . figure [ fig : probabilities ] clearly shows that the markovian qme still works well not only beyond the ordinary adiabatic - evolution regime but also beyond the slowly - varying regime with respect to @xmath2 , which supports the validity of our general framework . here , we note that the exact solutions ( for @xmath300 ) can not be reproduced if the simple hamiltonian evolution is just applied instead of the markovian qme although one may expect that there is no need of the markovian qme in such a rapid regime . to further shed light on the underlying mechanism , in figs . [ fig : evolution](a ) and [ fig : evolution](b ) , we show the time - dependent probabilities @xmath301 and @xmath302 to find the system in the instantaneous ground and excited states , respectively . in the calculation , we set @xmath303 and @xmath296 , the situation of which corresponds to the green open squares in the rapid regime of fig . [ fig : probabilities ] . for @xmath304 in fig . [ fig : evolution](a ) , @xmath305 and @xmath306 hold true and the values are almost the same as the corresponding probabilities without dissipation ( the dotted lines ) . in this regime , therefore , neither the non - adiabatic transition nor the dissipation play significant roles . in the vicinity of @xmath307 , @xmath301 changes from one toward zero and @xmath302 from zero toward one . the zoom around @xmath307 [ fig . [ fig : evolution](b ) ] shows that this non - adiabatic transition occurs mainly in the regime @xmath308 ( the gray shaded area ) . here , one finds that the non - adiabatic transition is still close to the behavior without dissipation . this suggests that the non - adiabatic transition is essentially instantaneous with respect to the timescale of relaxation . finally , for @xmath309 in fig . [ fig : evolution](a ) , @xmath310 gradually approaches the value of @xmath311 due to the dissipation . however , in this time period , there is no non - adiabatic transition , as evidenced by the fact that the probabilities without dissipation ( the dotted lines ) do not show any change in time . as a result , we can notice that there is no time where the non - adiabatic transition and the dissipation simultaneously play significant role . this situation can be studied more carefully and clearly from the viewpoint of the individual timescales summarized in table [ tbl : conditions ] . in figs . [ fig : evolution](c ) and [ fig : evolution](d ) , we therefore show @xmath312 with @xmath313 , according to eqs . , eq . and eq . . for @xmath304 in fig . [ fig : evolution](c ) , we can find @xmath294 . this indicates that the ordinary adiabatic evolution is indeed validated , which is consistent with the above discussion . in addition , we also find that @xmath182 holds true for @xmath304 . therefore , according to table [ tbl : conditions ] , we notice that the iea is also well justified in this time regime . however , around @xmath307 , one finds that these conditions break down . as shown in fig . [ fig : evolution](d ) , the adiabatic evolution is no longer safely ensured for @xmath314 because @xmath19 falls within one order of magnitude of @xmath55 and @xmath294 is not satisfied . in particular , @xmath19 reaches its minimum value of @xmath238 at @xmath227 , which means that the instantaneous eigenstates change most rapidly at @xmath227 and its timescale is given by @xmath238 . naturally , the non - adiabatic transition becomes pronounced around @xmath308 , as we have already seen in fig . [ fig : probabilities](b ) . in a similar manner , for the validity of the iea , we can notice that @xmath0 is within one order of magnitude of @xmath2 over the entire region in fig . [ fig : evolution](d ) . hence , the iea is not ensured in this time regime ( @xmath315 ) due to the failure of @xmath182 . nevertheless , one can also find that @xmath0 is much shorter than @xmath168 , i.e. @xmath184 , in fig . [ fig : evolution](d ) . this means that the effect of relaxation is negligible compared to the non - adiabatic effect . in particular , for @xmath308 , @xmath0 is two orders of magnitude shorter than @xmath168 even though @xmath0 becomes comparable to @xmath2 . as a result , the markovian qme still works well because @xmath175 plays only a minor role around @xmath307 . this is actually what we have explained in the general discussion ( section [ subsec : beyondadiabaticregime ] ) . the important point is that such a situation is generally guaranteed as long as we are in the weak coupling regime ; the dlz model corresponds to one prominent example . thus , the dlz model has been discussed to demonstrate the ability of the markovian qme in this section . finally , it would be worth digressing from the main subject to point out that the spontaneous emission ( the vacuum fluctuation ) has a great importance in the context of the aqc and the qa . in this case , the two states in the dlz model are regarded as the ground and the first excited states . then , as illustrated in figs . [ fig : dlzmodel ] and [ fig : probabilities ] , the spontaneous emission has the ability to recover the ground state when @xmath300 . this means that the computational errors caused by the depopulation of the ground state will be naturally corrected . in other words , a self - healing mechanism is given to the system by the bath . for finite temperature , according to eq . , this mechanism can dominate when @xmath316 , whereas the thermal excitation and relaxation will activate when @xmath317 . in this sense , the mechanism is different from the thermally - assisted qa in ref . , the situation of which is focused on the latter case . in order to effectively utilize this mechanism , however , the coupling angle @xmath261 in @xmath81 [ eq . ] should be @xmath296 ( transverse coupling ) in principle for the following two reasons . first , the thermal excitation is reduced around the minimum gap because @xmath318 = 0 $ ] at @xmath240 . second , if @xmath316 is fulfilled , the spontaneous emission can dominate the relaxation process after the avoided crossing because @xmath318 \ne 0 $ ] for @xmath319 . in reality , however , it is difficult to directly control @xmath261 ( the form of @xmath81 ) in experiments . therefore , the past discussions were focused only on the diagonal coupling in most cases @xcite . nevertheless , the encoding direction of the target hamiltonian of the system would be allowed to be changed , which is physically equivalent to the change of the coupling angle @xmath261 . in this context , we conjecture that the encoding direction also plays an important role to increase the computational performance of the aqc and the qa . we have presented a detailed analysis of the markovian qme under the wca when the system hamiltonian is time - dependent . while this problem has been discussed by many authors in the past , there has been no consensus for the condition to validate the formalism . this fact indicates that the problem is non - trivial , while one may think it straightforward at first glance . in our view , one major reason was the complete lack of the way to explicitly quantify the temporal change of @xmath1 . therefore , in this paper , we first introduced the timescale @xmath0 as a measure for the adiabaticity of @xmath1 itself . here , the adiabaticity of @xmath1 is conceptually different from the ordinary adiabatic theorem , and therefore , enables us to derive the markovian qme without the adiabatic theorem . furthermore , in a broad range of situations , it was also shown that the framework is well justified even if @xmath0 is much shorter than the decay timescale of the bath correlation function @xmath2 . this result arises from the fact that there is no situation where the non - adiabatic effect and the dissipative nature play considerable roles simultaneously in time , as long as the wca is validated . we have thus clearly shown that there is no need to restrict ourselves to either the adiabatic - evolution regime or the slowly - varying regime with respect to @xmath2 . this is in stark contrast to the past understanding . as a result , the markovian qme is justified well beyond the adiabatic regime . we here remark that the presented route of the formulation and the well - defined approximations allow us to clearly understand the structure of the framework with sufficient generality . hence , the scheme in this paper is immediately applicable to a wide range of physical systems . as an example , we have applied the framework to the dlz model and illustrated the ability of the markovian qme . even in this simple model , we stressed that the expression of @xmath0 is difficult to obtain without eqs . and this has been the very origin of the inconsistent quantification of the timescale of @xmath1 . then , it was shown that the numerical results have good agreements with the exact transition probabilities . furthermore , the relationship between the relevant timescales [ figs . [ fig : evolution](c ) and [ fig : evolution](d ) ] indeed supports our general scenario . it would be worth noting that such an approach generally enables an easy estimation of whether the individual approximations are justified or not at a certain time @xmath48 . finally , a short digression has been made to discuss the importance of the spontaneous emission in the context of the aqc and the qa . we discussed the possibility that the spontaneous emission provides the self - healing mechanism that naturally corrects the computational errors . the results presented in this paper would be of interest to those who try to control quantum systems as a function of time @xcite because in reality the quantum systems can not be free from uncontrolled interactions with environmental degrees of freedom ( bath ) . although our demonstration in the dlz model was focused on the zero temperature limit , the framework is , of course , applicable to finite temperature cases . interesting directions for future research are investigations of the non - adiabatic regime of the aqc and the qa in the presence of dissipation . as conjectured from the dlz model , there are indeed possibilities to improve the computational performance . the problem is , however , still non - trivial when the system size and the number of the relevant eigenstates become large . in that case , we expect that our methodology will play a key role for the development of new schemes beyond the adiabatic regime . we thank y. yamamoto , n. nagaosa , k. kamide , y. yamada and m. bamba for fruitful discussions and comments . this work was funded by impact program of council for science , technology and innovation ( cabinet office , government of japan ) and by jsps kakenhi ( grants no . 26287087 ) . here , in order to make the paper self - contained , we describe the time - convolutionless ( tcl ) framework and show a brief derivation of eq . within the wca . we note that the derivation is the same as ref . even though the system hamiltonian @xmath1 is time - dependent . this section also plays a preliminary role for the development of the slippage technique in appendix [ slp ] . for this purpose , we first rewrite the total hamiltonian as @xmath320 , where @xmath321 is a dimensionless parameter introduced only to easily measure the order of @xmath81 . in the interaction picture , the total density operator @xmath322 evolves according to @xmath323 \equiv \lambda\vl_{\s\b}(t)\vrho(t ) , \label{app : sch.eq.}\end{aligned}\ ] ] where @xmath324 denotes the liouville superoperator in the interaction picture . for notational convenience , we will write @xmath325 and @xmath326 in the following . however , we will finally replace @xmath327 and @xmath328 after the formulation is completed , in order to eliminate the parameter @xmath321 . before discussing the framework in detail , we summarize here the liouville superoperators . we denote the liouville superoperator in the schrdinger picture by @xmath329 , \label{app : liouville_sch}\end{aligned}\ ] ] where @xmath330 is an arbitrary operator . we note that there is no time dependence in the hamiltonian @xmath331 when @xmath332 but @xmath331 is described in the time - dependent way because there is no confusion . the liouville superoperator satisfies a relation @xmath333 in the liouville space , where the adjoint superoperator @xmath334 for any superoperator @xmath335 is defined in such a way that @xmath336 } = { \text{tr}[\ox_1^{\dagger } \mathcal{a } \ox_2]}$ ] holds for arbitrary operators @xmath337 and @xmath338 . it is then natural to introduce the evolution superoperator by @xmath339 where @xmath340 denotes the chronological ( anti - chronological ) time ordering for the superoperators . the interaction picture of an arbitrary operator is , for example , easily described by using @xmath341 as @xmath342 by further introducing the interaction picture of the liouville superoperator @xmath343 , we can obtain @xmath344 , \label{app : liouville_int}\end{aligned}\ ] ] which is consistent with the definition of eq . . in the similar manner to @xmath341 , we therefore define the evolution operator for the interaction picture by @xmath345 note that @xmath322 can be written as @xmath346 when @xmath322 evolves according to eq . . we now derive the tcl form of the qme . to this end , we first decompose eq . as @xmath347 where @xmath348 and @xmath349 are the projection superoperators defined by @xmath350 } \otimes \orho_{\b}$ ] and @xmath351 . the formal solution of eq . can then be described as @xmath352 where @xmath353 . by writing @xmath354 , we can readily obtain @xmath355 with @xmath356 . then , we obtain @xmath357 here , we have assumed the existence of @xmath358 because @xmath359 and @xmath360 suggest that @xmath361 may be inverted for sufficiently small couplings and @xmath362 @xcite . substitution of eq . into eq . finally yields the tcl form of the qme , @xmath363 } + { \text{tr}_{\text{b } } [ \vi(t ) \oq \vrho(t_0 ) ] } , \label{app : tcl}\end{aligned}\ ] ] with definitions of the superoperators @xmath364 where we have used @xmath365 } = \vrho_{\s}(t)$ ] . eqs . - are exact and local in time although the superoperators @xmath366 and @xmath367 are complicated in general . however , by assuming the separable initial state @xmath368 , the second term in eq . becomes zero because of @xmath369 . in the next subsection , therefore , @xmath366 , called the tcl generator , will be estimated up to the second order in @xmath321 , which corresponds to the wca . here , the tcl generator @xmath366 will be expanded in the power series @xmath370 and truncated to the second order in @xmath321 . for this purpose , @xmath358 in eq . is expanded as @xmath371^n$ ] . then , we obtain @xmath372^n \op.\end{aligned}\ ] ] therefore , further expansion of @xmath373 in the power series of @xmath374 gives @xmath375 with @xmath376 as a result , by assuming the vanishing odd moments of the system - bath interaction hamiltonian @xmath377 } = 0 $ ] [ eq . ] , or equivalently , @xmath378 , we can find @xmath379 the tcl form of the qme [ eq . ] then becomes @xmath380 ] } .\end{aligned}\ ] ] thus , up to the second order in @xmath81 , namely within the wca , eq . has been derived . the concept of the slippage technique is first proposed by surez _ et al_. @xcite and expanded by gaspard and nagaoka @xcite in order to cure the problem of the complete positivity when the secular approximation is not applied . the slippage of the initial condition alters the original initial condition of the system @xmath381 to the slipped ( renormalized ) one @xmath382 , where @xmath383 is the slippage superoperator . in this appendix [ slp ] , the concept is generalized to the case of the time - dependent system hamiltonian and is employed to justify the markov approximation . we note that the slipped initial condition has been used to obtain all of the numerical results presented in section [ sec : dlz ] . to study the slippage superoperator @xmath383 , by using eq . , we rewrite the first term of eq . as @xmath384 } \simeq \textstyle { \int_{t_0}^{t } \dd t ' } \du_{\s}(t , t_0 ) \oc_{\s}(t , t ' ) \ou_{\s}(t , t_0 ) \vrho_{\s}(t),\end{aligned}\ ] ] where @xmath385 is the correlation superoperator defined as @xmath386}.\end{aligned}\ ] ] in the derivation , we have used @xmath387 we can therefore rewrite the tcl form of the qme [ eq . ] within the wca as @xmath388 where @xmath389 . this equation should be compared with eq . . one can then readily notice the following relation ; @xmath390 which will be used to derive the slippage superoperator in the next subsection . here , under the markovian approximation ( @xmath391 ) , eq . gives @xmath392 where @xmath393 is distinguished from @xmath98 because @xmath393 may be different from @xmath98 especially in the time region of @xmath104 . we are now ready to derive the slippage superoperator @xmath383 . to this end , we formally integrate eq . and eq . in time , respectively , @xmath394 then , by assuming @xmath395 for @xmath105 , eqs . and give @xmath396 up to the second order in @xmath81 , where @xmath397 is used since we have assumed @xmath105 . this equation means that in the markovian qme we should use a ` slipped ' initial condition @xmath398 rather than the original one @xmath381 . in other words , the markovian qme is fully justified in combination with the slipped initial condition . then , transforming the integration @xmath399 , we find @xmath400 finally , by inserting eq . into eq . , the specific form of the slippage superoperator is given by @xmath401 . \label{app : slp2}\end{aligned}\ ] ] here , it is possible to apply the iea @xmath402 [ eq . ] when @xmath403 . however , we note that , even if @xmath403 is not satisfied , there is no difficulty to numerically evaluate eq . because the integrand becomes negligible for @xmath404 due to the decay of @xmath99 . we also remark that eq . recovers the results by gaspard and nagaoka @xcite when the system hamiltonian is assumed to be time - independent . in this appendix [ sec : sa ] , we discuss the sa to obtain the dissipator @xmath405 [ eq . ] from @xmath406 [ eq . ] . as we have seen in the time - independent case [ eqs . and ] , to perform the sa , we have to extract the oscillating behaviors from @xmath407 and @xmath156 and have to average out the rapid oscillating terms in eq . . for this purpose , we formally integrate @xmath408 over @xmath409 from @xmath48 to @xmath410 , @xmath411 where @xmath412 is a certain timescale at time @xmath48 and the integration variable has been changed from @xmath409 to @xmath413 in the right - hand side . here , @xmath414 can be rewritten as @xmath415 by using eq . . in this equation , then , @xmath416 and @xmath417 are allowed if we assume @xmath418 due to eqs . and . equation therefore results in @xmath419 and , similarly , we have @xmath420 these approximations have close analogies with the procedures for the iea [ eqs . ] . as a result , one can indeed find the oscillatory behaviors in eqs . and . inserting these equations into eq . , we obtain @xmath421 there are two functions in the integrand that depend on the variable @xmath422 . one is the density operator @xmath423 , the temporal change timescale of which is characterized by the relaxation timescale @xmath168 [ eq . ] . the other one is @xmath424 that corresponds to the beat between the two different bohr frequencies @xmath144 and @xmath171 . the intrinsic timescale of the oscillation period for @xmath425 is described by @xmath169 [ eq . ] at time @xmath48 . therefore , in the integral of eq . , @xmath424 for @xmath425 oscillates rapidly when @xmath426 whereas @xmath423 remains unchanged when @xmath427 in such a situation , indeed , the non - secular terms of @xmath428 average out to zero . as a result , only the secular terms , @xmath429 , survive in the integration of eq . . then , by retracing the steps [ eqs . - ] in reverse order , we again arrive at eq . but without non - secular terms . this means that the dissipator @xmath160 [ eq . ] is well approximated by @xmath430 , where @xmath431 this equation is nothing but eq . . as is obvious from the above discussion , the sa is justified when the three conditions , @xmath432 [ eq . ] , @xmath433 [ eq . ] , and @xmath434 [ eq . ] , are all satisfied . therefore , we need @xmath435 simultaneously , to perform the sa . thus , eq . is obtained . 60ifxundefined [ 1 ] ifx#1 ifnum [ 1 ] # 1firstoftwo secondoftwo ifx [ 1 ] # 1firstoftwo secondoftwo `` `` # 1''''@noop [ 0]secondoftwosanitize@url [ 0 ] + 12$12 & 12#1212_12%12@startlink[1]@endlink[0]@bib@innerbibempty @noop _ _ ( , ) @noop _ _ ( , ) @noop _ _ ( , ) @noop _ _ ( , ) link:\doibase 10.1103/physreve.76.031115 [ * * , ( ) ] link:\doibase 10.1103/revmodphys.81.1665 [ * * , ( ) ] link:\doibase 10.1103/physreve.91.012143 [ * * , ( ) ] link:\doibase 10.7566/jpsj.83.123001 [ * * , ( ) ] link:\doibase 10.1103/physreva.90.023827 [ * * , ( ) ] link:\doibase 10.1126/science.1057726 [ * * , ( ) ] link:\doibase 10.1103/physreva.80.022303 [ * * , ( ) ] link:\doibase 10.1103/physreva.79.022107 [ * * , ( ) ] link:\doibase 10.1103/physreve.58.5355 [ * * , ( ) ] link:\doibase 10.1038/ncomms2920 [ * * , ( ) ] link:\doibase 10.1038/ncomms4243 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.100.060503 [ * * , ( ) ] link:\doibase 10.1103/physreva.65.012322 [ * * , ( ) ] http://stacks.iop.org/1367-2630/12/i=12/a=123010 [ * * , ( ) ] http://stacks.iop.org/1367-2630/14/i=12/a=123016 [ * * , ( ) ] http://stacks.iop.org/0305-4470/12/i=5/a=007 [ * * , ( ) ] link:\doibase 10.1103/physreve.65.055102 [ * * , ( ) ] link:\doibase 10.1103/physrevx.5.031044 [ * * , ( ) ] link:\doibase 10.1038/nphys1801 [ * * , ( ) ] link:\doibase 10.1038/nphys2170 [ * * , ( ) ] link:\doibase 10.1126/science.1183628 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.105.246804 [ * * , ( ) ] link:\doibase 10.1038/ncomms2412 [ * * , ( ) ] link:\doibase 10.1103/physreve.55.300 [ * * , ( ) ] http://www.sciencedirect.com/science/article/pii/s0370157398000222 [ * * , ( ) ] link:\doibase 10.1103/physrevb.84.235140 [ * * , ( ) ] link:\doibase 10.1103/physrevb.92.245422 [ * * , ( ) ] link:\doibase 10.1007/bf01343193 [ * * , ( ) ] link:\doibase 10.1143/jpsj.5.435 [ * * , ( ) ] link:\doibase 10.1007/bf01011696 [ * * , ( ) ] link:\doibase 10.1103/physreve.82.021921 [ * * , ( ) ] link:\doibase 10.1103/physreva.87.042111 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.105.030401 [ * * , ( ) ] http://adsabs.harvard.edu/abs/2015arxiv151106994d [ ] link:\doibase 10.1103/physrevlett.107.210402 [ * * , ( ) ] link:\doibase 10.1103/physreva.90.042112 [ * * , ( ) ] link:\doibase 10.1103/physreva.91.052103 [ * * , ( ) ] link:\doibase 10.1007/s11128 - 004 - 7712 - 7 [ * * , ( ) ] link:\doibase 10.1103/physreva.55.1653 [ * * , ( ) ] link:\doibase 10.1063/1.463831 [ * * , ( ) ] http://scitation.aip.org/content/aip/journal/jcp/111/13/10.1063/1.479867 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.101.150402 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.103.210401 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.105.050403 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.113.200403 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.116.120402 [ * * , ( ) ] link:\doibase 10.1103/revmodphys.88.021002 [ * * , ( ) ] link:\doibase 10.7566/jpsj.84.014001 [ * * , ( ) ] http://stacks.iop.org/1367-2630/12/i=11/a=113042 [ * * , ( ) ] link:\doibase 10.1143/jpsj.53.108 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.62.3004 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.97.200404 [ * * , ( ) ] link:\doibase 10.1103/physrevb.75.214308 [ * * , ( ) ] link:\doibase 10.1103/physrevb.76.024310 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.103.220401 [ * * , ( ) ]

the absence of diffusion of waves in disordered media was predicted by anderson 50 years ago @xcite . originally proposed in the context of electronic transport in disordered crystals , it has since been observed for different types of waves , including light and sound . recently , direct observations of the anderson localization by disorder @xcite and of a localization transition by quasiperiodic potentials @xcite of quasi one - dimensional ( 1d ) matter waves of ultracold atoms were reported . these experiments pave the way to the observation of new phenomena and shed new light on long standing problems , amongst which the question of possible anderson localization in presence of interactions . in the present paper we consider the case of an atomic vapor described as a weakly interacting bose gas in the presence of a weak disorder ( what is meant by `` weak '' here will be made quantitative in sec . [ model ] ) . in this configuration it has been shown theoretically in refs . @xcite and supported by numerical simulations in ref . @xcite that a small amount of disorder does not drastically alter the _ equilibrium _ properties of the system , but merely decreases the condensate and superfluid fractions . furthermore , even in the 1d limit considered in the present work , it has been experimentally demonstrated in refs . @xcite that one can observe global phase coherence in the presence of disorder and remain far from , say , the bose glass phase originally proposed by giamarchi and schulz and fisher _ et al . _ @xcite . here we are interested in _ transport _ properties . specifically , we study a quasi 1d , weakly interacting bose einstein condensate ( bec ) , propagating through a disordered potential . in this context , localization has been theoretically studied mainly for effective _ attractive _ interactions ( see , e.g. , @xcite and references therein ) , with less attention on the _ repulsive _ case we consider here ( see , however , refs . @xcite ) . in the absence of an external potential , ( repulsive ) interactions make the system superfluid and introduce a new characteristic speed in the system , the speed of sound @xmath2 . as mentioned above , when the speed @xmath3 of the bec relative to the external potential tends to zero , the addition of a weak random potential preserves superfluidity , although with a reduced superfluid fraction . what happens as @xmath3 increases ? this question was investigated in a previous publication @xcite , where the disordered potential , of length @xmath0 , was modeled by a series of randomly located delta peaks . for small velocities @xmath4 , perturbation theory shows that the superfluidity is preserved , e.g. , the flow is dissipationless and with a perfect transmission . in contrast , in the high speed limit @xmath5 , where the kinetic energy dominates over the interaction energy , the transport properties of the bec are deeply altered , and tend to those of the non - interacting gas , displaying an exponential damping of the transmission with length @xmath0 , a behavior characteristic of the strong anderson localization . thus , two limiting cases of stationary flow have been identified @xcite , with contrasting transport properties : superfluidity in the deep subsonic regime , and anderson localization in the deep supersonic one . in between , in the region @xmath6 where both interaction and kinetic energies are important , it was shown that stationary scattering solutions do not exist : one reaches a regime of time - dependent flows with more or less ( depending on the speed ) complex density excitations . the range of speeds around @xmath2 where this phenomenon is observed increases as the length @xmath0 increases . the different types of existing flows are summarized in fig . [ fig1 ] . through a disordered potential @xmath7 consisting in a series of uncorrelated delta peaks extending over a domain of size @xmath0 [ cf . ( [ e2a ] ) and the discussion in sec . [ sec - lstar ] ] . dark region : time dependent flow ; light gray ( light blue online ) regions : stationary flow . in the supersonic case , the yellow solid line corresponds to the threshold @xmath1 between these two domains as determined from eq . ( [ tstar - impl ] ) . the blue dashed line is the localization length @xmath8 ( [ exp1 ] ) . the supersonic region below @xmath8 denoted as `` ballistic '' corresponds to the region where the perturbation theory of sec . [ pts ] applies . note the enlarged scale for @xmath9 $ ] . [ fig1 ] ] in the present study we concentrate on the supersonic stationary region of the phase diagram [ gray ( light blue online ) @xmath10 region in fig . [ fig1 ] ] . in this domain we provide analytical and numerical evidence of anderson localization in the presence of interaction for different types of disorder . we compute analytically the interaction - dependent localization length as well as the corresponding distribution of transmission coefficients . we also explain the disappearance of the supersonic stationary flow observed at a given velocity for increasing length of the disordered sample . this onset of time dependence is an important qualitative effect revealed by our study . we show that it is directly connected to interaction effects and provide an analytical estimate of the length @xmath1 of the disordered region above which most of the realizations of the random potential lead to time dependent flows ( see fig . [ fig1 ] ) . the paper is organized as follows . in sec . [ model ] we present the model and identify its range of validity . in sec . [ seciii ] we take some time to properly define the transmission coefficient of a bose - condensed beam over an obstacle . in [ seciv ] we introduce the different types of disordered potentials studied in the present work . in sec . [ ssr ] we present analytical and numerical results showing that anderson localization is indeed possible in the supersonic regime . we consider the three possible supersonic regimes : perturbative ( sec . [ pts ] ) , anderson localized ( sec . [ non - pert ] ) and onset of time dependence ( sec . [ sec - lstar ] ) . in sec . [ experiment ] we discuss experimental strategies and possible signatures for the observation of anderson localization in an interacting bose - einstein condensate . finally we present our conclusions in sec . [ conclusion ] . some technical points are given in the appendixes . in appendix [ app1 ] we derive the probability distribution of transmission in a special case ( perturbative regime and correlated gaussian potential ) . in appendix [ app2 ] we present the derivation of the fokker - plank equation ( [ dmpk ] ) for the distribution of the transmission coefficients . we study here the transport properties of a quasi - one - dimensional ( 1d ) bose - einstein condensate formed of particles of mass @xmath11 , experiencing a repulsive effective interaction ( characterized by the 3d s - wave scattering length @xmath12 ) , in the presence of an obstacle represented by the external potential @xmath13 . the potential is not necessarily disordered at this point , the only restriction we impose throughout the present work is that it should have a finite extent , i.e. , @xmath14 when @xmath15 . the configuration we consider corresponds to the `` 1d mean field regime '' @xcite ( see also the discussion in ref . @xcite ) , where the system is described by a 1d order parameter @xmath16 depending on a single spatial variable : the coordinate @xmath17 along the direction of propagation . @xmath16 obeys the nonlinear schrdinger equation @xmath18\psi\ ; .\ ] ] in all the present work we choose to work in the `` laboratory frame '' where the condensate is initially at rest . ( [ e1 ] ) describes its 1d dynamics in the presence of an obstacle moving at constant velocity @xmath3 in this frame , which corresponds to the experimental situation where an obstacle is swept through a condensate initially at rest , see e.g. refs . @xcite . on the theoretical side , one should imagine that , from an initial static configuration where the condensate is at rest with @xmath19 , the potential intensity and speed have been slowly ramped up to a point where the condensate dynamics is described by eq . ( [ e1 ] ) . we choose @xmath20 , this corresponds to a potential moving from left to right in the laboratory frame . the reduction of the motion of the condensate to a single spatial dimension is typically achieved through a transverse harmonic confining potential of pulsation @xmath21 . we choose a normalization such that @xmath22 is the linear density of the condensate . in this case , the interaction amongst particles results in eq . ( [ e1 ] ) in the nonlinear term @xmath23 , with @xmath24 @xcite . in the stationary regime , where the flow is time independent in the frame moving with the potential , @xmath25 depends on @xmath17 and @xmath26 only through the variable @xmath27 . the appropriate boundary condition is @xmath28 ( where @xmath29 is a constant ) ( see @xcite and the discussion in section [ stationary_regime ] below ) . the condensate is then characterized by a chemical potential @xmath30 , a speed of sound @xmath31 and a healing length @xmath32 . it is customary to characterize the transverse confinement via the `` harmonic oscillator length '' @xmath33 . with @xmath34 denoting a typical order of magnitude of @xmath35 , the 1d mean field regime in which eq . ( [ e1 ] ) is valid corresponds to a density range such that @xmath36 in this domain the wave function of the condensate can be factorized in a transverse and a longitudinal part @xcite . the transverse wave function is gaussian ( this is ensured by the condition @xmath37 ) , the longitudinal one is of the form @xmath38 and @xmath16 satisfies eq . ( [ e1 ] ) @xcite . the left - hand side ( l.h.s . ) inequality in eq . ( [ s1.0 ] ) prevents the system to enter in the tonks - girardeau regime . more precisely , a general analysis of 1d bose gas shows that at zero temperature no bec is possible @xcite . this results in a algebraic decrease of the one body density matrix monitored by phase fluctuations occurring over a phase - coherence length @xmath39 @xcite . hence the results obtained using eq . ( [ e1 ] ) are valid if they describe structures with a characteristic length scale smaller than @xmath40 . the l.h.s . inequality in eq . ( [ s1.0 ] ) ensures that @xmath40 is exponentially large compared to the healing length . if one considers , for instance , @xmath41rb or @xmath42na atoms in a guide with a transverse confinement characterized by @xmath43 hz , the ratio @xmath44 is roughly of order @xmath45 and restriction ( [ s1.0 ] ) still allows the density to vary over four orders of magnitude . even if the mean field approach is legitimate in 1d , the effects of disorder have to be taken into account with some care . it may well be that the introduction of a disordered potential @xmath13 in eq . ( [ e1 ] ) modifies the properties of the ground state . this is indeed the case as shown in refs . @xcite : a disordered potential decreases the condensate and the superfluid fraction , but the effects are weak provided the intensity of the disorder remains weak ( see ref . @xcite for an extension to finite temperature ) . more precisely , in the case of a disorder formed by randomly spaced delta impurities with density @xmath46 ( see sec . [ delta ] ) one can show @xcite that , in the dilute impurity limit , at @xmath47 the non - superfluid fraction ( normal part ) is proportional to @xmath48 [ the notations are those of eq . ( [ e2a ] ) ] and thus remains small provided the dimensionless coefficient @xmath49 is small ( weak disorder limit ) . at finite @xmath3 , the normal fraction is multiplied by a factor @xmath50^{-3/2}$ ] ( see ref . @xcite ) , which diverges when @xmath51 . one thus expects the mean field approach to fail near the region @xmath52 of fig . this is supported by the numerical results presented in @xcite . hence , in the center of the time - dependent region of fig . [ fig1 ] we can not trust the results obtained from eq . ( [ e1 ] ) . however , far from this region , the 1d mean field approach is expected to be valid even in presence of ( weak ) disorder , as experimentally demonstrated in refs . @xcite . in the present work we characterize the localization properties of the condensate in the random potential by studying the transmission coefficient . ( [ e1 ] ) being non linear , the definition of transmission and reflection coefficients needs to be treated with special care . this is the purpose of the present section where we first define the stationary regime ( section [ stationary_regime ] ) and then the transmission coefficient within this regime ( section [ transmis ] ) . it is customary to perform a madelung transformation and to write @xmath53 where @xmath35 is the density and @xmath54 the local velocity . from ( [ e1 ] ) one can check that they verify the continuity equation @xmath55 the stationary regime is defined as the regime where the system is at rest in the frame moving with the obstacle . in this case , in the laboratory frame @xmath25 , @xmath56 , @xmath57 and @xmath58 are time dependent , but they depend on @xmath17 and @xmath26 only through the variable @xmath27 . it is then possible to get a first integral of ( [ e2 ] ) under the form @xmath59 in the case of subsonic ( @xmath60 ) and stationary motion , the flow is superfluid and the order parameter is only affected in the vicinity of the obstacle , with @xmath61 and @xmath62 @xcite . for @xmath63 , a regime of stationary flow also exists but in this case the obstacle induces density oscillations with a pattern stationary in its rest frame @xcite . this means that in the laboratory frame the phase velocity of these waves is identical to the velocity @xmath3 of the obstacle . on the other hand , the energy transferred from the obstacle to the fluid propagates with the group velocity , which in the case of bogoliubov excitations is greater than the phase velocity , i.e. as just argued than @xmath3 . as a consequence , radiation conditions require that the wake is always located ahead of the obstacle , i.e. , upstream , with no long - range perturbation of the fluid on the downstream side @xcite . this means that in this case the flow far in the downstream region remains unperturbed , with @xmath64 and @xmath65 . the two possible stationary configurations ( subsonic and supersonic ) are represented in fig . [ fig2 ] . ( 9,5.5 ) ( 1,0 ) ( @xmath66 ) . in both plots the potential is represented by a thick solid line ( hatched down to zero ) and the density profile is represented by a thin solid line.[fig2],title="fig:",width=302 ] ( 0.5,3.)downstream ( 6.7,3.)upstream ( 1,2.7)(1,0)7.5 ( 6.6,2.3)@xmath27 ( 1.4,1.7)@xmath67 ( 1.4,5.3)@xmath67 ( 3.2,1.)@xmath68 ( 3.8,5.)@xmath68 ( 0,1.5)@xmath29 ( 0.4,1.55)(1,0)0.5 ( 0,5.03)@xmath29 ( 0.4,5.08)(1,0)0.5 hence , in any stationary configuration ( subsonic or supersonic ) , the above reasoning fixes the integration constant in the right hand side ( r.h.s . ) of eq . ( [ e3 ] ) to its value at @xmath69 , i.e. , @xmath70 . in the stationary regime one gets from eqs . ( [ e1 ] ) and ( [ e3 ] ) @xmath71 where @xmath72 and @xmath73 \ ; .\ ] ] in this section we restrict the analysis to the stationary regime of section [ stationary_regime ] , and define the transmission of the condensate through the obstacle represented by a potential @xmath13 ( not necessarily disordered ) verifying @xmath74 . as the wave equation ( [ e1 ] ) is nonlinear , one can not , in general , properly define reflection and transmission coefficients , since it is generally not possible to disentangle incoming and reflected waves in the nonlinear flow upstream the obstacle . however , following a procedure devised in ref . @xcite ( see also @xcite ) , we will show that one can define a transmission and a reflection coefficient in the limit of small nonlinearity as well as in the limit of weak reflection and arbitrary nonlinearity . outside the scattering region , @xmath75 and one can get a first integral of eq . ( [ e4 ] ) under the form @xmath76 which defines the `` free '' asymptotic density profiles . @xmath77 in eq . ( [ t1 ] ) are integration constants . the boundary condition discussed in the previous section imposes @xmath78 and @xmath79 when @xmath80 . this fixes the value @xmath81 . the value of @xmath82 at @xmath83 has to be determined by the integration of eq . ( [ e1 ] ) ( cf . ref . ( [ t1 ] ) expresses the energy conservation for a fictitious classical particle with `` mass '' @xmath84 , `` position '' @xmath85 and `` time '' @xmath86 , evolving in a potential @xmath87 ( whose typical shape is displayed in fig . [ fig3 ] ) . this type of analysis is common in the study of nonlinear equations such as eq . ( [ e1 ] ) , see e.g. , the review @xcite ( the first time we found it used is in ref . it is employed here as a convenient tool for getting intuition about the behavior of the solution of the gross - pitaevskii equation ( see below ) . as a function of @xmath88 ( drawn for @xmath89 ) . the fictitious particle has a classical energy @xmath82 when @xmath90 . the ( red ) dashed line corresponds to an approximation of @xmath91 by @xmath92 , obtained by keeping only the first term in expansion ( [ t3a]).[fig3 ] ] from now on we restrict to the supersonic stationary regime where an imperfect transmission occurs ( in the subsonic stationary regime one has perfect transmission ) . in this case the fictitious particle is initially ( i.e. , when @xmath66 ) at rest at the bottom of the potential @xmath87 with @xmath93 . the behavior of @xmath85 for @xmath94 depends on the value of @xmath95 . a stationary solution exists only if @xmath96 remains bounded , i.e. , if @xmath97 ( where @xmath98 corresponds to the local maximum of @xmath87 , see fig . [ fig3 ] ) . in this case , the asymptotic behavior of @xmath99 corresponds to oscillations between the values @xmath100 and @xmath101 defined in fig . 3 . for future references we note that @xmath91 is zero when @xmath102 and when @xmath103 , and that the derivative @xmath104 is zero when @xmath102 and when @xmath105 , with @xmath106 at large velocity , when @xmath107 , one has @xmath108 and @xmath109 writing @xmath110 we now argue that , following ref . @xcite , one can write a perturbative version of ( [ t1 ] ) in a limit where @xmath111 we emphasize that restriction ( [ t2 ] ) corresponds to @xmath112 ( i.e. , small oscillations ) only when @xmath3 is of order of @xmath2 or smaller . the approach developed below is however able to tackle large relative density oscillations ( @xmath113 ) at large velocities ( @xmath107 ) @xcite . in this sense it will allow us to penetrate in the non perturbative regime where the upstream density oscillations are large and the transmission is low . using the variable @xmath114 , we write eq . ( [ t1 ] ) in the upstream region ( @xmath90 ) as @xmath115 where @xmath116 . a simple limited expansion around @xmath117 yields @xmath118 where @xmath119 and @xmath120 the second term in the r.h.s . of expansion ( [ t3a ] ) is small compared to the first one precisely in the limit ( [ t2 ] ) . in the following we restrict to this regime and neglect the second term of the r.h.s . of ( [ t3a ] ) . this corresponds to approximating the exact @xmath91 by the ( red online ) dashed line in fig . [ fig3 ] and to write eq . ( [ t3 ] ) under the form @xmath121 where the dimensionless parameter @xmath122 is defined by @xmath123 the solution of eq . ( [ t4 ] ) is @xmath124 where @xmath125 and @xmath126 is a phase factor . we recall that eq . ( [ t6 ] ) describes the density oscillations in the upstream region @xmath90 . these oscillations can be described as the sum of incident and reflected waves ( @xmath127 and @xmath128 ) of the form @xmath129 this analysis allows one to determine the reflection and the transmission coefficients as @xmath130 of course the sum of the incident @xmath127 and the reflected @xmath131 waves ( [ t6c ] ) is an approximate solution of the nonlinear schdinger equation ( [ e1 ] ) which is only valid in regime ( [ t2 ] ) , i.e. , in the regime of arbitrary interaction and small transmission ( @xmath132 ) , or in the regime of arbitrary transmission and small interaction ( @xmath107 ) . up to this point we presented a theory valid for any potential of finite extent . from now on we concentrate on the particular case of random potentials . we denote @xmath133 an arbitrary random potential , and use a subscript when dealing with one of the particular cases defined below . the first potential of interest , analyzed in ref . @xcite , is a series of @xmath135 randomly located identical delta peaks of the form : @xmath136 the intensity of the peaks is measured by the ( non random ) positive quantity @xmath137 . the scatterers have random uncorrelated positions @xmath138 , with mean density @xmath139 and average separation @xmath140 . hence the potential extends over a mean length @xmath141 . denoting henceforth the disorder average by @xmath142 , for @xmath17 and @xmath143 inside the disordered region one gets the mean value @xmath144 and the irreducible two - point correlation function @xmath145 where @xmath146 another commonly used model of disorder is provided by gaussian random processes with zero average . we consider here potentials which are non zero only over a region of finite extent ( with typical size @xmath0 ) and generate them in the following way ( see , e.g. , chap . 5 of ref . @xcite and references therein ) : let s consider a gaussian white noise @xmath147 extending over all the real axis , with zero mean and second moment @xmath148 . then for a given function @xmath149 one defines @xmath150 where @xmath151 is a parameter characterizing the disorder and whose meaning is explained below . if @xmath152 were a delta function , then @xmath153 would be a gaussian white noise over @xmath154 $ ] ( and zero everywhere else ) . the actual function @xmath149 has a finite extension , and this induces finite correlations in the disordered potential . from ( [ e2d ] ) it is clear that @xmath155 . if the domain of integration in the r.h.s . of ( [ e2d ] ) were extended to all @xmath156 , @xmath153 would have a gaussian distribution @xmath157}{\sqrt{2\pi\sigma^2 } } \ ; , \ ] ] where @xmath158 defining the correlation function @xmath159 as @xmath160 one would get in this case @xmath161 with a fourier transform @xmath162 where @xmath163 is the fourier transform of @xmath152 . imposing here the normalization condition @xmath164 leads to a two - point correlation function ( [ e2f0 ] ) which is as in eq . ( [ e2b ] ) of the form of @xmath165 multiplied by a function whose integral over @xmath17 equals unity [ @xmath166 in ( [ e2f0 ] ) instead of @xmath167 in ( [ e2b ] ) ] . thus , with definition ( [ e2d ] ) and normalization ( [ e2cbis ] ) , @xmath151 plays for disorder ( [ e2d ] ) the same role as @xmath168 [ eq . ( [ e2c ] ) ] for disorder ( [ e2a ] ) : it characterizes the amplitude of the fluctuations of the potential . the typical extent of @xmath149 will in turn characterize the range of the correlations . since @xmath153 as given by eq . ( [ e2d ] ) is typically non zero only over a region of finite extent , eqs . ( [ e2f0 ] ) and ( [ e2f ] ) are only correct if @xmath17 and @xmath143 are inside this region . more precisely , they should be in this region , at a distance from 0 or @xmath0 larger than the typical extent @xmath169 of the function @xmath152 . in the following we always consider the case where @xmath0 is very large compared to @xmath169 ( otherwise one could simply not speak of a disordered region ) and it is clear that the characteristics of the disorder are properly defined only inside the disordered region . we consider two special cases of correlation corresponding to different forms of @xmath152 : a lorentzian @xmath170 and a gaussian @xmath171 we denote the corresponding potentials by @xmath172 and @xmath173 . for the correlation functions one gets respectively @xmath174 and @xmath175 in both cases @xmath169 is the typical correlation radius . the choice of a lorentzian correlated disordered potential originates from experimental and theoretical results in the case of micro - fabricated guides . in this type of setting , the atoms are magnetically guided over a chip @xcite . unavoidable imperfections and irregularities in the design of the circuit induce fluctuations in the current which , in turn , result in a random contribution to the magnetic field used for guiding the atoms . thus the potential seen by the atoms has a random component which is typically lorentzian correlated , with a correlation length @xmath169 which decreases when the distance between the guide and the chip increases @xcite . the gaussian correlated potential @xmath173 is more academic but , by comparison with the results obtained with @xmath172 it allows one to check what is really specific to the lorentzian case , and what is a mere effect of finite correlation length . another experimentally relevant type of disorder is the so called speckle potential which is generated by an optical speckle field produced by a laser beam passing through a diffusing plate @xcite . the corresponding potential will be denoted by @xmath176 and may be mathematically generated as follows @xcite : @xmath177 { \rm d}y \right|^2 , \ ] ] where @xmath178 and @xmath179 are two independent gaussian white noise processes of zero mean with @xmath180 ( @xmath181 and @xmath182 or 2 ) . here also we characterize the disorder by studying its statistical properties in the limit where the domain of integration in the r.h.s . of ( [ e2l ] ) is extended to all @xmath156 . in this case one gets @xmath183 where @xmath184 is given by formula ( [ s2 ] ) ( replacing @xmath152 by @xmath185 ) . this yields @xmath186 and the correlation function defined in eq . ( [ e2f0 ] ) reads here @xmath187 \\ = & 4 \left [ \int_{\mathbb{r}}\!\!{\rm d}y\ , w_{\sss s}(x - x'+y)w_{\sss s}(y ) \right]^2 \ ; . \end{split}\ ] ] contrarily to the choice ( [ e2cbis ] ) , @xmath185 should not be normalized to unity here because from eq . ( [ e2l ] ) this is homogeneity - wise impossible . instead , the choice @xmath188 corresponds to the typical experimental situations @xcite and leads to a correlation function @xmath189 whose integral over @xmath17 equals unity and whose fourier transform is @xmath190 hence , definition ( [ e2l ] ) and choice ( [ ws ] ) correspond here also to characterizing the amplitude of the disorder s fluctuations by the parameter @xmath151 and the range of the correlations by @xmath169 . as explained in ref . @xcite , and recalled in the introduction , anderson localization in a weakly repulsive bose - einstein condensate is only possible in the supersonic regime ( cf . [ fig1 ] ) which we consider now . in the present section we first analyze the transmission across a short disordered sample , in which case perturbation theory is applicable ( sec . [ pts ] ) . we then turn to generic non - perturbative configurations ( sec . [ non - pert ] ) where anderson localization is expected . in this regime we obtain evidences of the occurrence of anderson localization in the presence of interaction . finally we discuss the upper limit of the localized regime and the onset of time dependent flows for long disordered samples ( sec . [ sec - lstar ] ) . in the supersonic stationary regime , simple perturbation theory yields @xmath191 where @xcite @xmath192 \ ; , \ ] ] and @xmath193 is given by eq . ( [ pt4 ] ) . perturbation theory always predicts a stationary density profile . this is certainly wrong when @xmath3 is close to @xmath2 ( cf . fig . [ fig1 ] ) , but in this case @xmath193 gets very small and one precisely goes out of the domain of validity of perturbation theory ( @xmath194 as given by ( [ spt1 ] ) is no longer small compared to @xmath29 ) . far ahead of the obstacle ( in a region where @xmath195 is larger than @xmath169 and @xmath196 ) ( [ spt1 ] ) gives @xmath197 the perturbative regime in which eqs . ( [ spt1],[spt2 ] ) are valid is also the one where the constant @xmath122 in ( [ t5],[t6 ] ) is small compared to unity . this can be inferred from the comparison of ( [ spt2 ] ) and ( [ t6 ] ) which indeed shows that @xmath132 in the regime where ( [ spt2 ] ) holds and that , in this case , @xmath198 ) , yielding @xmath199 from ( [ spt3 ] ) it is clear that the average reflection coefficient is @xmath200 furthermore , one can show that the corresponding probability distribution of the reflection coefficient is poissonian with @xmath201 note that for properly normalizing this probability distribution for @xmath202 $ ] , one should include in the prefactor of the r.h.s . of ( [ spt4 ] ) a correcting term of order @xmath203 which can be safely neglected in the limit ( [ spt5 ] ) . we give in appendix [ app1 ] a demonstration of result ( [ spt4 ] ) for the special case of a correlated gaussian potential @xmath153 of type ( [ e2d ] ) . below , we show that the same result holds for a potential @xmath7 of type ( [ e2a ] ) [ see eq . ( [ dmpk0 ] ) in section [ non - pert ] ] , and we checked numerically that it is also the case for the speckle potential @xmath176 ( [ e2l ] ) ( cf fig . [ fig4 ] ) . in all these cases , the average reflection coefficient reads ( up to the above discussed exponentially small correction ) @xmath204 where @xmath205 we recall that the function @xmath206 depends on the type of disorder considered . for a potential of form ( [ e2a ] ) one has @xmath207 , for the other potentials considered in this work it is given by ( [ e2j ] ) , ( [ e2k ] ) and ( [ csq ] ) . concomitantly to distribution ( [ spt4 ] ) of reflection coefficients one gets for the transmission @xmath208 from ( [ e7b ] ) [ or ( [ spt10 ] ) ] one gets @xmath209 the perturbative approach holds when @xmath210 i.e. , when @xmath211 . this corresponds to the region which is denoted as `` ballistic '' in fig . [ fig1 ] @xcite . its accuracy is shown for @xmath212 in fig . [ fig4 ] for a speckle potential @xmath176 of type ( [ e2l ] ) ( we also checked this prediction for the potentials @xmath213 and @xmath173 , with also excellent results ) . for the transmission coefficient @xmath214 in a potential @xmath176 with @xmath215 , @xmath216 and @xmath217 moving at velocity @xmath218 in a condensate of initial constant density @xmath219 . the corresponding localization length is @xmath220 . the histogram corresponds to a statistical analysis of the results of the numerical solution of eq . ( [ e1 ] ) for 10000 different random potentials . the solid line is the perturbative result ( [ e7b]).[fig4 ] ] at this stage , @xmath8 is simply a notation for expression ( [ spt11 ] ) , but it will be shown to be the actual localization length of the matter wave in a disordered potential ( in section [ non - pert ] ) . the results derived here also hold for a noninteracting gas , obtained by taking the limit @xmath221 in eq . ( [ e1 ] ) , in which case @xmath222 and @xmath223 . ( [ spt11 ] ) , with @xmath193 replaced by @xmath224 , then coincides with the antsygina - pastur - slyusarev formula for the localization length @xcite and the distribution of transmissions ( [ e7b ] ) holds , with @xmath225 . in the present work , those formulas are modified to include interactions . the generalization simply consists in replacing the wave vector @xmath226 by @xmath227 . this replacement has , as an important physical consequence , the effect of diminishing , at a given speed @xmath3 , the localization length ( there is an effective reduction of the available kinetic energy by the repulsive interactions ) . for instance , in the case of a potential @xmath7 , since @xmath228 , there is a relative difference @xmath229 between @xmath230 and @xmath231 , that is 11 @xmath232 for @xmath233 . this is illustrated in fig . [ fig5 ] , which displays the average @xmath234 as a function of @xmath0 for a disorder @xmath7 of type ( [ e2a ] ) , with and without interactions . for a potential @xmath7 ( characterized by @xmath235 and @xmath236 ) . in the interacting case @xmath237 and @xmath238 . the non - interacting case is drawn for the same velocity and corresponds to a value @xmath239 . in both cases the dashed line is the analytical result ( [ e7 ] ) and the solid line corresponds to a statistical analysis of the results of the numerical solution of eq . ( [ e1 ] ) for 15000 different random potentials . the departure of the numerical results from the dashed lines occurs when the systems leaves the perturbative regime.[fig5 ] ] when the size @xmath0 of the sample is large compared to the value @xmath8 determined in section [ pts ] , the perturbative approach fails . we now propose a non perturbative method allowing to treat both the regimes @xmath240 and @xmath241 and showing that @xmath8 , as defined in eq . ( [ spt11 ] ) , is indeed the localization length in the presence of interactions . within the framework of the non perturbative approach , we are able to provide approximate analytical results in the case of the model disorder potential ( [ e2a ] ) . this potential being zero between two impurities , one can write a series of first integrals of eq . ( [ e4 ] ) in each segment @xmath242x_n , x_{n+1}[$ ] as follows : @xmath243 = e^{(n)}_{\rm cl } \ ; .\ ] ] in the region @xmath244 the integration constant @xmath245 of eq . ( [ t1 ] ) is denoted as @xmath246 in ( [ np1 ] ) ( taking @xmath247 ) and is zero , whereas in the region @xmath248 one has @xmath249 ( and @xmath250 ) . as a function of @xmath85 ( drawn for @xmath89 ) . for @xmath244 , the fictitious particle is initially at rest at the bottom of potential @xmath87 with @xmath251 . the value of the classical energy changes from @xmath252 to @xmath253 at each impurity @xmath254 . the lower panel displays the corresponding oscillations of @xmath99 , with two impurities at @xmath255 and @xmath256 ( their position is indicated by vertical dashed lines).[fig6],title="fig : " ] as a function of @xmath85 ( drawn for @xmath89 ) . for @xmath244 , the fictitious particle is initially at rest at the bottom of potential @xmath87 with @xmath251 . the value of the classical energy changes from @xmath252 to @xmath253 at each impurity @xmath254 . the lower panel displays the corresponding oscillations of @xmath99 , with two impurities at @xmath255 and @xmath256 ( their position is indicated by vertical dashed lines).[fig6],title="fig : " ] from ( [ e4 ] ) it is a simple matter to show that the matching condition of the density at impurity position @xmath254 is @xmath257 where @xmath258 [ @xmath259 denotes the limit of the derivative @xmath260 at the left [ at the right ] of @xmath254 . relation ( [ np1a ] ) between the derivatives of the amplitude results [ from eq . ( [ np1 ] ) ] in a relation between the classical energies : @xmath261 \ ; .\ ] ] hence , eq . ( [ np1 ] ) allows to draw a classical analogous of the solution of the nonlinear schdinger equation in the presence of potential ( [ e2a ] ) formed by a series of delta peaks : the fictitious classical particle defined in sec . [ transmis ] evolves in the potential @xmath87 and experiences kicks at `` times '' @xmath254 . each kick changes the `` energy '' according to ( [ np2 ] ) , as illustrated in fig . the key point in the remaining of this section will be to derive the probability distribution of @xmath262 which then directly allows one to get the distribution of @xmath122 s and of the transmission coefficients [ through eqs . ( [ t5 ] ) and ( [ t7 ] ) ] . let us introduce the quantities @xmath263 the parameters @xmath122 and @xmath264 defined in ( [ t5 ] ) and ( [ t6b ] ) are related to the ones of eq . ( [ np2a ] ) by @xmath265 and @xmath266 ( i.e. , @xmath122 is the last of @xmath267 s ; the same holds for @xmath264 ) . denoting by @xmath268 the value of the phase @xmath126 [ appearing in([t6 ] ) ] for @xmath269x_{n-1},x_{n}[$ ] , one gets in this domain [ the derivation is exactly the same as for eq . ( [ t6 ] ) ] @xmath270 and one can rewrite eq . ( [ np2 ] ) as @xmath271 \nonumber \\ & + & \frac{4\hbar^2}{m b^2 } \lambda_{n-1 } \sqrt{\kappa^2 b^2 + 1 } \ ; \zeta_{n-1 } \ ; , \end{aligned}\ ] ] where @xmath272 \ ; .\ ] ] using definition ( [ np2a ] ) one can rewrite eq . ( [ np3 ] ) in terms of the parameter @xmath267 as @xmath273 eqs . ( [ np3],[np5 ] ) are valid provided ( [ t2 ] ) holds , _ i.e. _ , provided @xmath274 , which reads @xmath275 in the following we also impose the condition @xmath276 precisely , we neglect all the quantities of order @xmath277 . this is an important technical point . it facilitates the analysis by allowing one to get simple formulas as we now illustrate in the perturbative case : eq . ( [ np5 ] ) allows for instance to compute the average value of the reflection coefficient in the perturbative regime ( as already done in section [ pts ] , eq . ( [ spt10 ] ) ) . in this regime , additionally to condition ( [ np5b ] ) one has @xmath278 . then eq . ( [ np5 ] ) implies at leading order @xmath279 which , together with the initial condition @xmath280 , leads immediately to @xmath281 which is identical to result ( [ spt10 ] ) in the case of a potential @xmath7 for which @xmath282 [ cf . ( [ spt11 ] ) ] . let us now proceed and consider the generic non perturbative regime where @xmath283 may become large compared to unity and where ( [ t2 ] ) and ( [ np5b ] ) are still valid . taking the condition ( [ np4a ] ) into account , eq . ( [ np5 ] ) reads @xmath284 it is natural to assume that the phase of the cosine in the r.h.s . of ( [ np4b ] ) is uniformly distributed in @xmath285 $ ] and independent of the phase at step @xmath286 . this could be called a `` phase randomization '' approximation . this relies on hypothesis ( [ np4a ] ) and on the assumption that there is a large number of density oscillations over the ( random ) length between @xmath287 and @xmath288 , i.e. , @xmath289 then , the argument of the cosine in definition ( [ np4b ] ) is uniformly distributed , @xmath290 s are uncorrelated random variables , with all the same law characterized by its average @xmath291 and variance @xmath292 note that the regimes ( [ np4a ] ) and ( [ gnp2bis ] ) imply that @xmath293 which in turn implies that the kinetic energy @xmath294 is much larger than @xmath295 ; _ i.e. _ , one is exactly in the anderson regime where the incident kinetic energy is much larger than the typical value of the ( disordered ) potential representing the obstacle . hence a classical particle would flow almost unperturbed over the potential but , as we shall see , a quantum particle experiences an exponentially small transmission . let @xmath296 be the probability that @xmath267 lies in the interval @xmath122 , @xmath297 . going to the continuous limit and defining the continuous variable @xmath298 [ where @xmath299 is the parameter ( [ spt11 ] ) in the case of a potential @xmath7 ] it is shown in appendix [ app2 ] that @xmath301 verifies the following fokker - planck equation @xmath302\ ; .\ ] ] equation ( [ dmpk ] ) follows directly from eq . ( [ np6 ] ) in the regime where conditions ( [ np4a ] ) and ( [ gnp2bis ] ) hold . it is precisely the dorokhov - mello - pereyra - kumar ( dmpk ) equation @xcite for the transmission in a single disordered channel [ with @xmath303 . equation ( [ dmpk ] ) is sometimes referred to as melnikov s equation ( after ref . @xcite ) but has a much longer history ( see the discussion in refs . @xcite ) . since before entering the disordered region the particle has a classical energy @xmath251 corresponding to @xmath304 , eq . ( [ dmpk ] ) has to be solved for the initial condition @xmath305 where @xmath306 is the one - sided delta function : @xmath307 . in the limit of small @xmath26 ( i.e. , in the perturbative regime @xmath308 ) , @xmath122 remains small and one can approximate in the r.h.s . of ( [ dmpk ] ) the term @xmath309 by @xmath122 . it is then simple to verify that the solution of this approximate equation that satisfies ( [ np8 ] ) is @xmath310 this result for the small @xmath26 solution of the dmpk equation has been already obtained in ref . @xcite ( see also the discussion in ref . the distribution law ( [ dmpk0 ] ) is exactly equivalent to distribution ( [ spt4 ] ) of the reflection coefficient in the perturbative regime and this proves the validity of the poissonian distribution ( [ e7b ] ) for a potential @xmath7 of type ( [ e2a ] ) @xcite . in the general case ( i.e. , for all @xmath311 ) the solution of ( [ dmpk ] ) with the initial condition ( [ np8 ] ) is ( see , _ e.g. _ , refs . @xcite and references therein ) @xmath312 where @xmath313 . from distribution ( [ np9 ] ) , a lengthy computation or alternatively the direct use of the dmpk equation ( [ dmpk ] ) @xcite yields @xmath314 in the large @xmath26 limit , distribution ( [ np9 ] ) tends to a log - normal distribution , i.e. , the distribution of the variable @xmath315 is gaussian ( see ref . @xcite ) @xmath316 from this distribution one gets the correct average @xmath317 , eq . ( [ np10 ] ) , and a standard deviation @xmath318^{1/2 } = \sqrt{2 t}$ ] , which is in agreement with the exact result in the limit @xmath319 @xcite . at the extremity of a sample of length @xmath0 one has @xmath320 , and the distribution ( [ np11 ] ) is the log - normal distribution of transmission typical for anderson localization in the regime @xmath321 ( see , e.g. , ref . @xcite ) . as a side product of this analysis , eqs . ( [ np10 ] ) and ( [ np11 ] ) confirm that @xmath8 is indeed the localization length as was anticipated in the notation . we have tested the validity of the dmpk approach for a bose - einstein beam of interacting particles propagating in a disordered potential @xmath7 of type ( [ e2a ] ) . the numerical results for the probability distribution @xmath322 are compared on fig . [ fig7 ] with the dmpk prediction ( [ np9 ] ) . the agreement is seen to be excellent . the distribution evolves from the poissonian result ( [ e7b ] ) ( for low values of @xmath323 ) towards a distribution peaked at low @xmath214-values for large @xmath323 . in this latter case one can check that the distribution tends to a log - normal by plotting @xmath324 . of type ( [ e2a ] ) ( characterized by @xmath325 and @xmath326 ) plotted for different values of the ratio @xmath327 with @xmath328 . the black solid lines are the dmpk result ( [ np9 ] ) and the colored histograms correspond to the numerical simulations ( 50000 samples used for each value of @xmath26 ) . cases ( a ) , ( b ) , ( c ) and ( d ) correspond respectively to @xmath329 , 0.5 , 1 and 2 . the inset displays @xmath330 as a function of @xmath26 . the thick solid line is extracted from numerical simulations and the thin ( red ) solid line is the dmpk prediction ( [ np10 ] ) . they can be distinguished only around @xmath331 as shown in the blowup of the ( gray ) shaded region for @xmath332 . [ fig7 ] ] we have also checked the validity of the present approach over a sizable range of lengths of disordered region and of intensities of disordered potential by plotting in the inset of fig . [ fig7 ] the average @xmath333 as a function of @xmath323 . the agreement of the numerical results with the dmpk prediction ( [ np10 ] ) is excellent . note however the beginning of a small departure around @xmath334 ; this effect will be studied more thoroughly in sec . [ sec - lstar ] ( cf . [ fig9 ] ) . finally , we discuss numerical results obtained for the disordered potentials introduced in secs . [ gaussian ] and [ speckle ] . although we do not have an analytical derivation of the dmpk equation for these potentials , the numerical results indicate a very good quantitative agreement for a disordered potential @xmath173 and for a speckle potential @xmath176 . we display the comparison of the numerical data with the dmpk predictions for a speckle potential in figure [ fig8 ] . the same agreement is obtained for a gaussian potential @xmath173 . hence , the behavior analytically predicted for the potential @xmath7 appears to be of general validity , meaning that the above defined regime of `` phase randomization '' can probably be extended to correlated potentials , leading to a regime of single parameter scaling . however , we have noticed that , although showing an overall good agreement with the dmpk prediction , the lorentzian correlated potential @xmath172 exhibits some deviations in the tail of the distribution , the details of which will be studied elsewhere . ( characterized by @xmath335 and @xmath336 @xmath337 ) plotted for different values of the ratio @xmath327 with @xmath338 . curves ( a ) , ( b ) and ( c ) corresponds respectively to @xmath339 , 0.52 and 0.68 . for each curve , the black solid line is the dmpk result ( [ np9 ] ) and the colored histogram is the result of 10000 numerical simulations.[fig8 ] ] in the previous sections the main effect of interaction has been shown to be a renormalization of the localization length @xmath8 . interaction induces a modification of the wave vector : expression ( [ spt11 ] ) for the localization length coincides with the noninteracting one but computed for an effective interaction - dependent wave vector @xmath193 given by eq . ( [ pt4 ] ) , instead of @xmath226 . the repulsive interaction diminishes the available kinetic energy and therefore reduces the localization length with respect to the noninteracting case ( since @xmath340 ) . we now discuss another , more spectacular , effect of interactions on the localization properties of a propagating bec on a disordered potential . in the previous sections [ pts ] and [ non - pert ] , we completely neglected the presence of an upper limit for the classical energy @xmath341 , which is given by the local maximum of the effective potential @xmath91 , namely @xmath342 ( see fig . [ fig3 ] ) . trajectories that pass beyond @xmath343 would become unstable and develop singularities with infinitely large density at @xmath344 . in practice this implies , on the level of eq . ( [ e1 ] ) , that a stationary flow can not be maintained in this case and that the disorder induces time - dependent dynamics of the condensate . in the vicinity of @xmath343 , the density profile of the condensate in between two adjacent scatterers becomes quite different from the cosine shape ( [ t6 ] ) that was derived for weak nonlinearities and/or low density modulations , and resembles more to a periodic train of gray solitons @xcite . in a crude approximation , we neglect this complication and assume that the spatial evolution of the density is still given by eq . ( [ t6 ] ) for all classical energies until @xmath345 . trajectories that happen to pass beyond @xmath343 are considered to be `` lost '' , i.e. , they do no longer contribute to the probability distribution for the transmission . this formally amounts to introducing a `` sink '' in the stochastic equation ( [ np6 ] ) , namely at @xmath346 . in the corresponding fokker - planck equation ( [ dmpk ] ) , this sink is appropriately modeled by imposing the boundary condition @xmath347 as a consequence of this boundary condition , the integrated probability distribution @xmath348 is no longer conserved , but decreases with increasing @xmath26 , i.e. , increasing length @xmath0 of the disorder region . in the following , we show how this affects the dmpk predictions of section [ non - pert ] and how the `` survival probability '' , i.e. , the fraction of trajectories that remain below this boundary at given length @xmath0 , can be analytically computed in the limit @xmath107 . in this limit , from eq . ( [ t1c ] ) one gets @xmath349 and thus @xmath350 . modifications of the probability density @xmath301 due to the presence of the sink appear only when the typical value of @xmath122 is of order @xmath351 which , as just remarked , is large compared to unity in the case @xmath107 . in this case @xmath301 is already negligibly small around @xmath352 . we therefore make the approximation @xmath353 in the fokker - planck equation ( [ dmpk ] ) , which then reads @xmath354\ ; .\ ] ] using , from now on , the probability distribution @xmath355 for finding a given value of @xmath315 at fixed @xmath356 , we obtain in this limit @xmath357 where we introduce @xmath358 . clearly , the log - normal distribution ( [ np11 ] ) corresponds to a solution of eq . ( [ lognorm - eq ] ) in the absence of any additional boundaries . in the presence of the sink , which is imposed by the boundary condition @xmath359 with @xmath360 we can straightforwardly find the solution of eq . ( [ lognorm - eq ] ) by subtracting from the log - normal distribution ( [ np11 ] ) a `` mirror '' distribution centered at some @xmath361 ( namely @xmath362 ) with a suitable prefactor . this yields the distribution @xmath363\ ; , \label{lognorm - decay}\end{aligned}\ ] ] which is defined for @xmath364 . clearly , this distribution satisfies the evolution equation ( [ lognorm - eq ] ) as well as the boundary condition @xmath359 for all @xmath26 and the initial condition @xmath365 for @xmath364 . the presence of the sink at @xmath366 explains a phenomenon barely noticeable in fig . 7 , but exemplified in fig . [ fig9 ] , namely the departure of the observed average @xmath367 from the usual dmpk result @xmath368 . this departure is due to the fact that the numerically computed average only takes into account the stationary solutions which as will be seen from eq . ( [ psurv]) become less and less numerous when @xmath26 increases . hence what is computed numerically is the average of @xmath369 over the distribution ( [ lognorm - decay ] ) . this reads @xmath370 - \nonumber \\ & & e^{z_{\rm max } } \left(\frac{t}{2}+ z_{\rm max}\right)\ , \textrm{erfc } \left(\frac{t+z_{\rm max}}{2\sqrt{t}}\right ) \ ; .\label{zmoy_leak}\end{aligned}\ ] ] where the error function is defined by @xmath371 and @xmath372 . expression ( [ zmoy_leak ] ) is compared in fig . [ fig9 ] with the results of a numerical simulation performed in the case @xmath373 ( corresponding to @xmath374 ) for 10000 random potentials @xmath7 of type ( [ e2a ] ) characterized by @xmath326 and @xmath375 ( leading to @xmath376 ) . the agreement is seen to be very good . since the sink cuts the solutions which are strongly scattered by the random potential , the remaining stationary states have a higher transmission coefficient . this effect increases with the sample length @xmath0 , which explains the behavior of the curve in fig . [ fig9 ] . plotted as a function of @xmath327 in the case of random potentials @xmath7 characterized by @xmath326 and @xmath375 . the curve is drawn in the case @xmath373 . the red solid line is the numerical result and the black dashed line is the analytical result ( [ zmoy_leak ] ) . the straight ( thin dashed ) line is the usual dmpk result @xmath377 [ eq . ( [ np10])].[fig9 ] ] as an other test of the validity of our approach ( which amounts to model the upper boundary @xmath378 by a perfect sink and to neglect nonlinear deformations of the density pattern of the flow close to the threshold ) we now determine the probability for a trajectory to remain below the boundary . this survival probability reads @xmath379 \nonumber \\ & & - \frac{e^{z_{\rm max}}}{2 } \ , \mathrm{erfc } \left(\frac{z_{\rm max } + t}{\sqrt{4 t}}\right ) \ ; .\label{psurv}\end{aligned}\ ] ] as anticipated , @xmath380 clearly decreases from @xmath381 ( at @xmath382 ) to 0 ( for large @xmath26 ) . the knowledge of @xmath380 allows to determine the value @xmath1 of the length of the disordered region beyond which most of the random realizations lead to a non - stationary flow of the condensate . we can , most conveniently , define @xmath1 through the condition @xmath383 with @xmath384 . this leads to the implicit equation for the threshold value @xmath385 : @xmath386 this equation can be explicitly solved in the limiting case of large @xmath387 . as it is natural to assume that @xmath385 ought to be of the order of @xmath378 , which is the only relevant scale in this equation , we make the ansatz @xmath388 and assume ( which is to be verified _ a posteriori _ ) that @xmath389 is of the order of unity , whereas @xmath390 . this yields to lowest non vanishing order @xmath391 \ ; , \ ] ] for the left - hand side of eq . ( [ tstar - impl ] ) and @xmath392 \nonumber \\ & = & \frac{1}{\sqrt{\pi\ , z_{\rm max } } } \left [ 1 + \mathcal{o}\left(z_{\rm max}^{-1}\right ) \right ] \ ; , \label{tstar - right}\end{aligned}\ ] ] for the right - hand side of eq . ( [ tstar - impl ] ) . this finally results in @xmath393 neglecting terms of the order of @xmath394 , we therefore obtain for the threshold length @xmath395 \ .\ ] ] we emphasize that eq . ( [ lstar ] ) holds for @xmath396 , i.e. , for @xmath397 [ see eq . ( [ defzmax ] ) ] . this is much more restrictive than the condition @xmath107 which is assumed to hold true when deriving eqs . ( [ lognorm - decay ] ) and ( [ tstar - impl ] ) . plotted as a function of the length @xmath0 of the disordered region . the condensate flow through disorder potentials @xmath7 of type ( [ e2a ] ) was numerically computed for this purpose ( red solid line ) , at parameters for which @xmath398 . the black dashed line shows the analytical prediction of this survival probability @xmath399 according to eq . ( [ psurv ] ) . the vertical dashed line marks prediction ( [ lstar ] ) for the threshold length @xmath1 at which @xmath400 ( horizontal dashed line ) , namely @xmath401 .[fig10 ] ] figure [ fig10 ] shows a comparison of the analytical predictions ( [ psurv ] ) and ( [ lstar ] ) with numerical data obtained from the integration of the time - dependent gross - pitaevskii equation ( [ e1 ] ) . the condensate flows through a disorder potential @xmath7 of type ( [ e2a ] ) with @xmath398 . we see that the fraction of stationary trajectories @xmath380 is very well described by eq . ( [ psurv ] ) , and that the approximate expression ( [ lstar ] ) predicts very well the length @xmath1 at which the crossover length from stationary to time - dependent flow occurs . for velocities @xmath3 not extremely large compared to the speed of sound , the condition @xmath396 will not be fulfilled and estimate ( [ lstar ] ) will not be valid , while @xmath402 might still hold and the average evolution of the system might still be fairly well described by the simplified fokker - planck equation ( [ dmpk - large ] ) . in that case , the implicit equation ( [ tstar - impl ] ) has to be solved numerically . in figure [ fig1 ] one can see that the numerical solution of eq . ( [ tstar - impl ] ) ( yellow solid line ) provides a very reasonable estimate of the boundary between the bright supersonic region ( stationary flows ) and the dark time - dependent region , in a regime of not extremely large @xmath403 , where eq . ( [ lstar ] ) fails to properly predict the threshold length @xmath1 . the simulations are performed by solving eq . ( [ e1 ] ) numerically using a potential of type @xmath7 ( characterized by @xmath404 and @xmath405 ) . for each @xmath3 and @xmath0 we consider 100 realizations of such a potential and statistically determine the quantity @xmath406 , i.e. , the fraction of stationary solutions . @xmath406 is plotted in fig . [ fig1 ] using a gray scale ( dark , @xmath407 ; light blue / gray , @xmath408 ) as a function of the normalized variables @xmath409 and @xmath403 ( this normalization rescales interaction effects ) . the qualitative agreement of figure [ fig1 ] is made quantitative in figure [ fig11 ] . in this figure the numerical solution of eq . ( [ tstar - impl ] ) is compared with its determination extracted from numerical simulations in the supersonic regime . more precisely , the solid ( red ) line in fig . [ fig11 ] is simply the contour @xmath410 in fig . this corresponds exactly to definition ( [ tstar ] ) of @xmath1 . the agreement between the numerical result and the theory of the present section [ dashed curve , solution of eq . ( [ tstar - impl ] ) ] is seen to be excellent @xcite . as a function of @xmath3 in dimensionless units . the black dashed line corresponds to the solution of eq . ( [ tstar - impl ] ) and the red solid line corresponds to the value of @xmath1 extracted from numerical simulations for a potential @xmath7 with the same characteristics as in fig . [ fig1 ] ( see the text).[fig11 ] ] we conclude this section by emphasizing that the existence of an upper threshold @xmath1 corresponding to lengths of the disordered region beyond which most of the flows are time - dependent is a genuine nonlinear effect [ absent if one sets @xmath411 in eq . ( [ e1 ] ) ] . actually , whereas interactions only weakly modify the precise value of the localization length , the existence of the threshold @xmath1 is a remarkable qualitative effect induced by nonlinearity . moreover , as illustrated in figures [ fig9 ] and [ fig10 ] , this effect persists even in the limit @xmath107 where naively one would expect no noticeable consequence of interaction . on the basis of the results obtained in the previous section we present here what are the more favorable experimental configurations for observing anderson localization in an interacting bose - einstein beam . we also discuss a possible experimental signature of localization . in the non - interacting regime the only condition for observing anderson localization in 1d is that the size of the disordered region should be larger than the localization length . then one can observe an exponential decay of transmission with a log - normal distribution ( in the limit @xmath321 ) . the situation is more complex when interactions are turned on . what is particularly interesting is the interplay between localization and superfluidity . indeed these two phenomena are conflicting one with the other : superfluidity is the ( counterintuitive ) ability to pass over an obstacle without reflection whereas anderson localization corresponds to a large reflection in a domain where one would expect almost perfect transmission . as a result of the interplay between these two extreme phenomena , and depending on the fluid velocity and on the sample size , the flow may be stationary and superfluid , dissipative and time dependent or stationary supersonic ( and also dissipative ) @xcite . anderson localization does not occur in the superfluid region ( where the transmission is perfect ) and either does not exist or can not be clearly identified in the time dependent regime ( where interference effects are washed out @xcite ) ; but is truly observed in the supersonic stationary regime , as demonstrated in sec . [ ssr ] . in that regime , a first experimentally relevant effect is the modification of the localization length with respect to its value in the absence of interactions . this effect is very well described by renormalizing the wave vector @xmath224 to @xmath193 [ eq . ( [ spt11 ] ) ] , which means that part of the kinetic energy available to the flow is taken by interactions . however , as already discussed in ref . @xcite this effect is only sizable in a regime where @xmath3 is not too large compared to @xmath2 , and is thus relevant only in the perturbative regime ( cf . [ fig5 ] ) . a second experimentally observable effect is the modification of the localization length due to the correlations of the disordered potential . this is described by formula ( [ spt11 ] ) where @xmath206 is the fourier transform of the two - point correlation function of the disorder . for the different potentials considered here , @xmath412 for a potential @xmath7 or is alternatively given by eqs . ( [ e2j ] ) , ( [ e2k ] ) and ( [ csq ] ) for correlated potentials . explicitly this yields @xmath413 for a potential @xmath7 of type ( [ e2a ] ) ; @xmath414 for a potential @xmath173 of type ( [ e2d ] ) ; @xmath415 for a potential @xmath172 of type ( [ e2d ] ) ; and @xmath416 for a potential @xmath176 of type ( [ e2l ] ) ( when @xmath417 ) . the validity of these expressions has been tested in sec . [ non - pert ] . in the non - interacting case ( i.e. , @xmath223 ) , expressions ( [ exp1 ] ) to ( [ exp4 ] ) correspond to a high energy limit and can be obtained through a first order born expansion within the phase formalism of refs . @xcite . in all three cases , one sees that the localization length is drastically enhanced due to the non - zero correlation length with respect to the uncorrelated disorder , eq . ( [ exp1 ] ) . in the gaussian and the lorentzian cases the localization length scales exponentially with @xmath418 and @xmath419 , respectively [ see eqs . ( [ exp2],[exp3 ] ) ] . in the case of a speckle potential , the effect is even stronger : one sometimes speaks of an `` effective mobility edge '' @xcite , meaning that beyond a critical wave - vector ( or a critical velocity ) the localization length ( [ exp4 ] ) is infinite . this is an artifact of the antsygina - pastur - slyusarev formula ( [ spt11 ] ) which can be corrected by going to higher orders ( see refs . @xcite ) : the corrections to this result give a localization length which is finite , but typically larger than any other relevant scale in experimental systems . hence , in all the cases the dependence of the localization length with respect @xmath193 ( i.e. , with velocity ) is amplified by correlations . mathematically this is due to the fact that the denominator in the antsygina - pastur - slyusarev formula ( [ spt11 ] ) for the localization length in presence of correlations tends to zero when @xmath420 . in order to minimize this effect one needs to impose the following condition : @xmath421 in the r.h.s . of eq . ( [ exp5 ] ) we replaced @xmath193 by @xmath422 because in practice condition ( [ exp5 ] ) is verified in regimes where @xmath423 , i.e. , when the approximation @xmath424 is sound . note that this condition is arbitrary and is only superficially analogous to the 3d ioffe - regel criterion @xcite . the latter defines a true mobility edge that separates a metallic from a localized phase whereas eq . ( [ exp5 ] ) only requires that the localization length does not get too large . understood in this sense , the criterion ( [ exp5 ] ) is exactly equivalent to the definition of an `` effective mobility edge '' sometimes used in the literature . in the absence of interactions it is always possible ( at least theoretically ) to define a system with a length @xmath425 which verifies ( [ exp5 ] ) ; i.e. , a system where one can observe anderson localization . if we now turn on interactions , a major effect is the appearance of a length scale @xmath1 which signals the onset , for @xmath426 , of a regime of time - dependent flows ( cf . [ sec - lstar ] ) . in this regime , anderson localization disappears , and the time - averaged transmission coefficient scales as @xmath427 @xcite . this is the most spectacular effect of interactions in the transport properties of the system . in order to observe anderson localization , the system size should therefore satisfy @xmath428 . in practice , one should be in a regime of parameters such as illustrated in fig . [ fig12 ] : the crossing @xmath429 has to occur at a velocity lower than @xmath430 . and the solid line is the threshold length @xmath1 . @xmath431 is the typical velocity beyond which it is almost impossible to observe anderson localization in a realistic system ( see text ) . the ( blue ) colored zone corresponds to the region where anderson localization can be experimentally observed in presence of interaction.[fig12 ] ] based on the numerical solution of eq . ( [ tstar - impl ] ) one can show that the crossing @xmath432 occurs at a velocity @xmath433 @xcite ( see also fig . [ fig1 ] ) . this condition only allows the system to reach a ( stationary ) regime where @xmath434 . but if one wants to observe anderson localization one should be able to reach a regime where @xmath435 say , in order to get as close as possible to the domain of log - normal distribution of transmissions still remaining in the region of stationary flows . this imposes @xmath436 . this must be supplemented by condition ( [ exp5 ] ) , i.e. @xmath437 . hence the correlation length @xmath169 should be smaller or equal to @xmath438 . fig . [ fig13 ] shows the phase diagrams of a one dimensional interacting beam of condensed atoms moving through a speckle potential in this regime . for plotting this diagram one has generated 16 random potentials and studied in each case if a stationary solution exists or not . the dark blue region corresponds to a domain where no stationary solution exists while the light blue one corresponds to a domain where all the potentials admit a stationary solution ( the color code is the same as in fig . [ fig1 ] and is explained in sec . [ sec - lstar ] ) . the region between @xmath8 and @xmath1 in figure [ fig13 ] is the region where one can observe anderson localization . for a beam with velocity @xmath3 moving in a speckle disorder of extent @xmath0 . the figure has been drawn for a potential @xmath176 of type ( [ e2l ] ) characterized by @xmath335 and @xmath439 . the light blue region corresponds to a domain of stationary flow ( @xmath408 : 100 @xmath232 of the solutions are stationary , see the explanation in the text ) ; the dark region corresponds to time dependent flow . the curves indicating the values of @xmath8 and @xmath1 correspond to the analytical results ( [ exp4 ] ) and to the numerical solution of eq . ( [ tstar - impl]).[fig13 ] ] let us now evaluate the orders of magnitude of the different parameters allowing one to reach the appropriate regime . for concreteness we consider a beam of @xmath41rb atoms such as the one of the atom optics group at laboratoire charles fabry de linstitut doptique . for a correlation length of 0.26 @xmath440 m the velocity cut - off for observing anderson localization is roughly @xmath441 mm / s . note that in ref . @xcite the velocity of the expanding condensate is about 1.6 mm / s , i.e. , smaller that @xmath430 as it should . if we use the parameters of ref . @xcite ( @xmath442 , @xmath443 mm / s , @xmath444 nm , @xmath445 atoms/@xmath440 m , @xmath446 mm / s and @xmath447 @xmath440 m ) it is impossible to satisfy condition ( [ exp5 ] ) because @xmath443 mm / s @xmath448 and also because @xmath449 . however the atom optics group has recently improved the sensitivity of its detectors which is now close to being able to detect a density as small as 1 atom/@xmath440 m . this allows one to work with a smaller density and to improve the ratio @xmath450 which can be tuned down to the value of @xmath451 . then the localization length can be selected by tuning the speckle amplitude . for instance @xmath452 mm can be obtained by choosing @xmath453 hz at @xmath454 mm / s . these parameters are close to those used in fig . [ fig13 ] and are reachable experimentally . however , for observing anderson localization one needs to keep the beam stable for almost 1 s ( 0.31 s if we want @xmath455 ) whereas in the current experiment this is only possible during 0.1 s , hence it is still a matter of debate to decide if the observation of anderson localization of a bose condensed beam in the presence of interaction is within the reach of present time technology . it is also interesting to make a connection between the physics described here and the recent experiment observation of anderson localization of a condensate expanding in a disordered potential performed in the same group @xcite . contrarily to the propagation of a beam studied in the present work , ref . @xcite considers the spreading of a wave packet ( initially at rest ) in a speckle potential . after a first stage of expansion , mainly driven by interactions , the experimental cloud expands with a constant velocity @xmath456 mm / s but the particle density and the sound velocity are functions of the position . therefore it is not possible to place this experiment on a single point of the phase diagram displayed in fig . . however one can evaluate the ratio @xmath403 at different positions , that is at different @xmath0 . for instance , at the typical value @xmath434 , fig . 2 of reference @xcite allows to calculate the sound velocity as well as the healing length @xmath457 , yielding the typical experimental values @xmath458 and @xmath459 . moreover the ratio between the typical disorder amplitude @xmath460 and the chemical potential @xmath461 is the same in fig . [ fig13 ] and in experiment : @xmath462 ( note however that in the experimental case this is the value of the _ local _ chemical potential that matters ) . hence , although the experimental setup of ref . @xcite forbids a direct comparison with the results of the present work , the estimates of the typical values @xmath458 and @xmath459 indeed locate the experimental system within the regime of anderson localization of fig . [ fig13 ] . once the appropriate regime of parameters for observing anderson localization in a bose condensed beam has been determined , it is also important to identify possible experimental signatures . in our theoretical approach we use the transmission coefficient @xmath214 of the beam over the disordered region as the relevant parameter . however the measure of @xmath214 might be experimentally involved , and we propose here to use an other related quantity , namely the rate of energy dissipation @xcite : @xmath463 where @xmath464 is the drag force exerted by the beam on the obstacle . definition ( [ exp6 ] ) is quite natural : the force exerted on the obstacle is the mean value of the operator @xmath465 over the condensate wave function . it is rigorously justified by the analysis of ref . @xcite in terms of stress tensor . in the stationary case , changing integration from @xmath17 to @xmath86 in ( [ exp6 ] ) , a simple integration by parts yields @xmath466 in the r.h.s . of eq . ( [ exp10 ] ) we made use of relations ( [ e4 ] ) and ( [ t1 ] ) . it has been shown in section [ transmis ] that @xmath81 and that under assumption ( [ t2 ] ) ( small nonlinearity and arbitrary transmission or weak transmission and arbitrary nonlinearity ) one has ( see eqs . ( [ t7 ] ) and ( [ t5 ] ) ) @xmath467 which yields @xmath468 in the regime @xmath469 , using eq . ( [ spt3 ] ) one recovers from ( [ exp11 ] ) the perturbative result already obtained in ref . @xcite : @xmath470 ) is rather simple and it is worth spending some time to discuss it . consider an incident beam of particles with density @xmath471 and momentum @xmath472 moving from @xmath83 towards an obstacle at rest . part of the particles is transmitted ( a fraction @xmath214 ) and the other part is reflected ( a fraction @xmath473 ) . the collisions are elastic and each of the reflected particles experiences an exchange of momentum @xmath474 with the obstacle . during a time @xmath389 there are @xmath475 collisions and by the law of action and reaction the obstacle experiences a force @xmath476 in the r.h.s . of ( [ exp12 ] ) , one has written that @xmath477 is the flux of particles colliding with the obstacle i.e. , @xmath478 . ( [ exp11 ] ) and ( [ exp12 ] ) are identical because what we call @xmath29 is the downstream density of the beam ( cf . [ fig2 ] ) , i.e. , precisely @xmath479 . depending on which quantity is held constant ( @xmath29 as in the present paper , or @xmath471 ) eq . ( [ exp11 ] ) or ( [ exp12 ] ) is more appropriate ( cf . the discussion of the fixed input and fixed output problem in ref . this is somewhat reminiscent of the controversy on the correctness of the landauer formula ( see , e.g. , the discussion in ref . @xcite ) . on the basis of ( [ exp11 ] ) one sees that the measure of @xmath480 gives direct informations on the transmission of the interacting beam through the disordered region , allowing one to reveal in which configuration is the system . for instance in the localized regime the energy dissipation is high ( @xmath481 ) and grows exponentially with the size @xmath0 of the disordered region , whereas in the perturbative regime @xmath480 is much lower and scales as @xmath0 . in the present work we have presented an analysis of the transmission of a weakly interacting bose gas incident on a disordered potential . we have shown on the basis of numerical and analytical results that there is a regime of anderson localization in this system and proposed experimental signature of this phenomenon . in order to properly identify a `` localized regime '' we have studied the transmission coefficient and its probability distribution . the transmission coefficient @xmath214 is well defined under assumption ( [ t2 ] ) , which holds in the following regimes : ( i ) small nonlinearity and arbitrary transmission or ( ii ) weak transmission and arbitrary nonlinearity . in other cases there is no obvious way to define the transmission of the non - linear beam because one can not separate in the up - stream region an incident flow from a reflected one . however , our analysis in terms of @xmath341 and @xmath122 ( defined in sec . [ transmis ] ) is always valid , even when condition ( [ t2 ] ) is not fulfilled . this just means that , out of regime ( [ t2 ] ) , the connection ( [ t7 ] ) between @xmath122 and @xmath214 is invalid . but for instance this does not invalidate at all the analysis leading to the dmpk equation ( [ dmpk ] ) , and the experimental signature proposed in sec . [ signature ] also remains valid even when it is not possible to properly define @xmath214 . we note that the validity of the dmpk approach for non - interacting particles is a well established fact in the theory of disordered systems . what is achieved in the present work is its extension to the case of interacting particles . other studies of anderson localization in the presence of interactions have concentrated on the long time behavior of the time evolution of initial wave packets @xcite . although those results are still a matter of active debate in the community , the results of the present work produce strong evidence of the existence of anderson localization for weakly interacting bose particles ( with effective repulsive interaction ) propagating through disordered samples of finite size @xmath482 . although the present study leads to the important conclusion that anderson localization in the presence of interaction is possible , it is rather disappointing to remark that it can be clearly identified only when @xmath483 , i.e. , in a regime where interactions do not play a major role ( see the discussion of sec . [ apconf ] and also ref . @xcite ) . in this respect , the more interesting and new effect of interactions is the existence of an upper threshold @xmath1 for the length of the disordered region : for @xmath426 no stationary flow is possible . as shown in sec . [ sec - lstar ] , @xmath1 is directly connected to the probability distribution of the parameter @xmath122 and to the localization properties of the system . it would be very interesting to lead a systematic study of the transmission in the interaction - induced time dependent regime ( @xmath426 ) where the numerical results of ref . @xcite indicate a power law decay of the time - averaged transmission , a signature generally considered of loss of phase coherence and onset of ohmic behavior @xcite . work in this direction is in progress . it is a pleasure to thank b. altshuler , a. comtet , j .- l . pichard and c. texier for inspiring discussions . this work was supported by anr grants no . 05nano00802 , no . nt05242103 and no . 08blan016501 , by the ifraf institute and by the alexander von humboldt foundation . we gratefully acknowledge funding by the excellence initiative of the german research foundation ( dfg ) through the heidelberg graduate school of fundamental physics ( grant no gsc 129/1 ) and the global networks mobility measures the frontier innovation fund of the university of heidelberg . we furthermore acknowledge support through the dfg forschergruppe 760 `` scattering systems with complex dynamics '' . we give here a demonstration of the perturbative results ( [ spt4 ] ) , ( [ spt10 ] ) and ( [ spt11 ] ) in the special case of a gaussian disordered potential @xmath153 of type ( [ e2d ] ) . a simple way to obtain this results starts by noticing that any gaussian noise verifying @xmath484 and @xmath485 [ and here we are specifically interested in @xmath147 that appears in eq . ( [ e2d ] ) ] can be written as ( see , e.g. , ref . @xcite ) @xmath486 where @xmath487 s are random positions uniformly distributed on the real axis with density @xmath488 and mean spacing @xmath489 and @xmath490 is a random variable ( with @xmath491 and @xmath492 ) . in order to calculate the probability distribution of the reflection coefficient @xmath473 [ whose value is given by ( [ spt3 ] ) ] one should first consider the distribution of @xmath493 the quantity @xmath494 as given by ( [ a2 ] ) is formally equivalent to the position @xmath495 of a particle performing a random walk in the complex plane after @xmath496 iterations . the particle is initially at the origin and performs jumps of constant amplitude @xmath497 with random direction . denoting by @xmath498 the probability to find the particle in the domain @xmath499 around @xmath495 after @xmath135 steps , if @xmath500 [ which is ensured by taking the limit @xmath501 in ( [ a2 ] ) ] , the central limit theorem yields @xmath502 it is then a simple exercise to get the distribution of @xmath503 . one obtains @xmath504 \ ; , \ ] ] where @xmath505 from relation ( [ spt3 ] ) , eq . ( [ a5 ] ) immediately yields the announced probability distribution ( [ spt4 ] ) with @xmath506 for a potential @xmath153 of type ( [ e2d ] ) , @xmath507 [ see eq . ( [ e2 g ] ) ] and eq . ( [ a6 ] ) demonstrates in this case the validity of eqs . ( [ spt10 ] ) and ( [ spt11 ] ) . in this appendix , we explain how to obtain the dmpk equation ( [ dmpk ] ) starting from the discrete langevin equation ( [ np6 ] ) . let us consider a generic situation where @xmath267 obeys a stochastic recursion relation of the type @xmath508 with uncorrelated random variables @xmath290 : @xmath509 it is clear that under assumption ( [ gnp2bis ] ) , eq . ( [ np6 ] ) is of type ( [ b1 ] ) with all the odd @xmath135 averages in ( [ b2 ] ) being zero and @xmath510 [ cf . ( [ varxi ] ) ] . let @xmath511 be the probability that @xmath267 lies in the interval @xmath122 , @xmath297 . one can express @xmath512 as @xmath513 this yields @xmath514 using the fact that @xmath267 depends on the variables @xmath515 , @xmath516 @xmath517 but not on @xmath290 ( as can be seen directly from ( [ b1 ] ) ) one can write the last of eqs . ( [ b4 ] ) as @xmath518 in the case of eq . ( [ np6 ] ) one has @xmath519 and the successive moments of @xmath520 read @xmath521 @xmath522 with all the other moments being of order @xmath277 or more , i.e. , negligible in regime ( [ np4a ] ) . ( [ b5 ] ) thus reads @xmath523 = \nonumber \\ & & \nonumber \\ & & \displaystyle -\frac{\partial}{\partial\lambda}\left[(1 + 2\lambda)p\right ] + \frac{\partial^2}{\partial\lambda^2}\left[(\lambda^2+\lambda)p\right ] = \nonumber \\ & & \nonumber \\ & & \displaystyle \frac{\partial}{\partial\lambda } \left[\lambda(\lambda+1)\frac{\partial p}{\partial\lambda}\right ] \ ; .\end{aligned}\ ] ] in the continuous limit , defining @xmath524 , the l.h.s . of eq . ( [ b8 ] ) is simply the first derivative of @xmath525 with respect to @xmath26 and ( [ b8 ] ) reduces to eq . ( [ dmpk ] ) of the main text . 99 p. w. anderson , phys . rev . * 109 * , 1492 ( 1958 ) . et al . _ , nature * 453 * , 891 ( 2008 ) . et al . _ , nature * 453 * , 895 ( 2008 ) . k. huang and h. f. meng , phys . lett . * 69 * , 644 ( 1992 ) . s. giorgini , l. pitaevskii , and s. stringari , phys . b * 49 * , 12938 ( 1994 ) . g. e. astrakharchik , j. boronat , j. casulleras , and s. giorgini , phys . a * 66 * , 023603 ( 2002 ) . d. clment , ph . bouyer , a. aspect , and l. sanchez - palencia , phys . a * 77 * , 033631 ( 2008 ) . yong p. chen . _ et al . _ , a * 77 * , 033632 ( 2008 ) . t. giamarchi and h. j. schulz , phys . b * 37 * , 325 ( 1988 ) . m. p. a. fisher , p. b. weichman , g. grinstein , and d. s. fisher , phys . rev . b * 40 * , 546 ( 1989 ) . s. a. gredeskul and y. s. kivshar , phys . rep . * 216 * , 1 ( 1992 ) . n. bilas and n. pavloff , phys . lett . * 95 * , 130403 ( 2005 ) . t. paul , p. leboeuf , n. pavloff , k. richter , and p. schlagheck , phys . rev . a * 72 * , 063621 ( 2005 ) . t. paul , p. schlagheck , p. leboeuf and n. pavloff , phys . * 98 * , 210602 ( 2007 ) . c. menotti and s. stringari , phys . a * 66 * , 043610 ( 2002 ) . d. s. petrov , g. v. shlyapnikov , and j. t. m. walraven , phys . * 85 * , 3745 ( 2000 ) . c. raman _ et al . _ , phys . rev . lett . * 83 * , 2502 ( 1999 ) . _ et al._. phys . lett . * 85 * , 2228 ( 2000 ) . p. engels and c. atherton , phys . lett . * 99 * , 160405 ( 2007 ) . m. olshanii , phys . lett . * 81 * , 938 ( 1998 ) . a. d. jackson , g. m. kavoulakis , and c. j. pethick , phys . a * 58 * , 2417 ( 1998 ) . p. leboeuf and n. pavloff , phys . a * 64 * , 033602 ( 2001 ) . l. p. pitaevskii and s. stringari , j. low temp . physics * 85 * , 377 ( 1991 ) . m. schwartz , phys . b * 15 * , 1399 ( 1977 ) . d. s. petrov , ph.d . thesis , university of amsterdam , 2003 , available online at http://www-old.amolf.nl/publications/theses/petrow . a. v. lopatin and v. m. vinokur , phys . lett . * 88 * , 235503 ( 2002 ) . g. e. astrakharchik and l. p. pitaevskii , phys . a * 70 * , 013608 ( 2004 ) . t. ernst , t. paul , and p. schlagheck , arxiv:0905.4750 ( submitted to phys . a ) . v. hakim , phys . e * 55 * , 2835 ( 1997 ) . h. lamb , _ hydrodynamics _ , ( cambridge university press , cambridge , 1997 ) . p. leboeuf , n. pavloff , and s. sinha , phys . a * 68 * , 063608 ( 2003 ) . t. paul , m. hartung , k. richter , and p. schlagheck , phys . a * 76 * , 063605 ( 2007 ) . b. i. ivlev and n. b. kopnin , adv 33 * , 47 ( 1984 ) . j. s. langer and v. ambegaokar , phys . rev . * 164 * , 498 ( 1967 ) . note in particular that condition ( [ t2 ] ) is always fulfilled in absence of interaction since in this case the speed of sound is @xmath526 . b. i. shklovskii and a. l. efros , _ electronic properties of doped semiconductors _ , ( springer - verlag , 1984 ) . j. fortgh and c. zimmermann , rev . phys . * 79 * , 235 ( 2007 ) . s. kraft , a. gnther , h. ott , d. wharam , c. zimmermann , and j. fortgh , j. phys . b * 35 * , l469 ( 2002 ) . j. estve _ et al . a * 70 * , 043629 ( 2004 ) ; t. schumm _ et al . _ j. d * 32 * , 171 ( 2005 ) . wang , m. d. lukin , and e. demler , phys . rev . lett . * 92 * , 076802 ( 2004 ) . j w goodman , _ speckle phenomena in optics : theory and applications _ , roberts and compagny publishers ( greenwood village , 2007 ) . d. clment _ et al . _ , new j. phys . * 8 * , 165 ( 2006 ) . l. fallani , c. fort , and m. inguscio , adv . at . . phys . * 56 * , 119 ( 2008 ) . ( [ e2l ] ) corresponds to a typical physical situation where the potential seen by the particles is proportional to the intensity of an electric field whose components are random gaussian variables . in ref . @xcite this region has been improperly called `` ohmic '' . t. n. antsygina , l. a. pastur and v. a. slyusarev , sov . j. low temp . phys . * 7 * , 1 ( 1981 ) [ fiz . nizk . temp . * 7 * , 5 ( 1981 ) ] . i. m. lifshits , s. a. gredeskul , and l. a. pastur , _ introduction to the theory of disordered systems _ , john wiley ( new - york 1988 ) . o. n. dorokhov , jetp lett . * 36 * , 318 ( 1982 ) [ pisma zh . fiz . * 36 * , 259 ( 1982 ) ] ; p. a. mello , p. pereyra , and n. kumar , ann . ( n.y . ) * 181 * , 290 ( 1988 ) . v. i. melnikov , sov . solid state * 23 * , 444 ( 1981 ) [ fis . tela ( leningrad ) * 23 * , 782 ( 1981 ) ] . c. w. j. beenakker , rev . * 69 * , 731 ( 1997 ) . p. a. mello , j. math . phys . * 27 * , 2876 ( 1986 ) . c. w. j. beenakker and j. a. melsen , phys . b * 50 * , 2450 ( 1994 ) . in order to complete the proof of this equivalence one has to show that the mean value @xmath527 obtained from ( [ dmpk0 ] ) for a sample of length @xmath528 is identical to the result ( [ spt5 ] ) . this amounts to show that for a potential of the form ( [ e2a ] ) one has @xmath529 , which is easily obtained . a. a. abrikosov , solid state comm . * 37 * , 997 ( 1981 ) . b. a. van tiggelen in _ diffusive waves in complex media _ , j. p. fouque ed . , p. 1 ( kluwer academic publishers , dordrecht , 1999 ) . note that in ref . @xcite the theoretical determination of @xmath1 was obtained using a criterion different from ( [ tstar ] ) resulting in a less precise estimate at high @xmath403 . f. m. izrailev and a. a. krokhin , phys . lett . * 82 * , 4062 ( 1999 ) . l. sanchez - palencia _ et al . _ , phys . rev . lett . * 98 * , 210401 ( 2007 ) . l. tessieri , j. phys . a * 35 * , 9585 ( 2002 ) . e. gurevich and o. kenneth , phys . a * 79 * , 063617 ( 2009 ) . a * 80 * , 023605 ( 2009 ) . a. f. ioffe and a. r. regel , prog . semicond . * 4 * , 237 ( 1960 ) . note that the value @xmath530 does not depend on the type of correlation of the disordered potential considered since it is simply the value of @xmath403 for which eq . ( [ tstar - impl ] ) admits the solution @xmath531 . w. gurin _ et al . _ , lett . * 97 * , 200402 ( 2006 ) . n. pavloff , phys . a * 66 * , 013610 ( 2002 ) . note that there is a factor of 2 difference between the present definition ( [ pt4 ] ) of @xmath193 and the one used in ref . y. imry , _ introduction to mesoscopic physics _ , oxford university press ( oxford , 1997 ) . d. l. shepelyansky , phys . lett . * 70 * , 1787 ( 1993 ) ; g. kopidakis , s. komineas , s. flach , and s. aubry , phys . lett . * 100 * 084103 ( 2008 ) ; a. s. pikovsky and d. l. shepelyansky , phys . lett . * 100 * , 094101 ( 2008 ) ; s. fishman , a. iomin , and k. mallick phys . e * 78 * , 066605 ( 2008 ) . s. datta , _ electronic transport in mesoscopic systems _ ( cambridge university press , 1995 ) . m. v. berry and s. klein , eur . * 18 * , 222 ( 1997 ) . c. texier , ph.d . thesis , universit paris 6 , 1999 , available online at http://www.lptms.u-psud.fr/membres/texier/research.html

knowledge of the distribution of cosmic - ray ( cr ) densities within our galaxy is a key to understanding their origin and propagation . high - energy crs interact with the gas in the interstellar medium ( ism ) or the interstellar radiation field , and produce @xmath0-rays via nucleon - nucleon interactions , electron bremsstrahlung and inverse compton ( ic ) scattering . since the ism is transparent to these @xmath0-rays , we can probe crs in the local ism , beyond direct measurements performed in the solar system , as well as in remote locations of the galaxy . although much effort has been made since the cos - b era ( e.g. , * ? ? ? * ; * ? ? ? * ) , the results have been limited by the angular resolution , effective area and energy coverage of the instruments . the advent of the _ fermi _ gamma - ray space telescope enables studying the spectral and spatial distribution of diffuse @xmath0-rays and crs with unprecedented sensitivity . here we report an analysis of diffuse @xmath0-ray emission observed in the third galactic quadrant . the window with galactic longitude @xmath5 and latitude @xmath6 hosts kinematically well - defined segments of the local and the perseus spiral arms and is one of the best regions to study the cr density distribution across the outer galaxy . the region has been already studied by @xcite using egret data . the improved sensitivity and angular resolution of the _ fermi _ lat ( large area telescope ; * ? ? ? * ) and recent developments in the study of the ism allow us to examine the cr spectra and density distribution with better accuracy . we exclude from the analysis the region of the monoceros r2 giant molecular cloud and the southern filament of the orion - monoceros complex ( e.g. , * ? ? ? * ) , in @xmath7 and @xmath8 , because 1 ) star forming activity and possible high magnetic fields suggested by the filamentary structure ( e.g. , * ? ? ? * ; * ? ? ? * ) could indicate a special cr environment , and 2 ) an ob association in monoceros r2 may hamper the determination of ism densities from dust tracers ( see 2.1.2 for details ) . study of the @xmath9 conversion factor which transforms the integrated intensity of the 2.6 mm line of carbon monoxide , @xmath10 , into the molecular hydrogen column density , @xmath11 , is also possible since the region contains well - known molecular complexes . in the local arm we find the molecular clouds associated with canis major ob 1 , ngc 2348 and ngc 2632 @xcite . at a few kpc from the solar system , in the interarm , lower - density region located between the local and perseus arms , we find maddalena s cloud @xcite , a giant molecular cloud remarkable for its lack of star formation , and the cloud associated with canis major ob 2 @xcite . this study complements the _ fermi _ lat study of the cassiopeia and cepheus region in the second quadrant reported by @xcite . the paper is organized as follows . we describe the model preparation in 2 and the @xmath0-ray observations , data selection and the analysis procedure in 3 . the results are presented in 4 , where we also discuss the emissivity profile measured for the atomic gas and we compare it with predictions by a cr propagation model . a summary of the study is given in in order to derive the @xmath0-ray emissivities associated with the different components of the ism we need to determine the interstellar gas column densities separately for each region and gas phase . for atomic hydrogen we used the leiden / argentine / bonn galactic h survey by @xcite . in order to turn the h line intensities into @xmath12(h ) column densities , a uniform spin temperature @xmath13 has often been adopted in previous studies . we will consider this option to directly compare our results with the former egret analysis of the same region @xcite and other studies of the galactic diffuse emission by the lat @xcite . recent h absorption studies @xcite , however , point to larger average spin temperatures in the outer galaxy , so we have tried different choices of @xmath14 to evaluate how the optical depth correction affects the results . we will find that the emissivity per h atom and the inferred cr density is affected by up to @xmath15 50% in the perseus arm , and will take this uncertainty into account in the discussion . the integrated intensities of the 2.6 mm line of co , @xmath10 , have been derived from the composite survey by @xcite . the data have been filtered with the moment - masking technique in order to reduce the noise while keeping the resolution of the original data . figure [ fig : diagram ] shows the velocity - longitude profile of h emission in our region of interest ( roi ) . the preparation of maps accounting for the different galactic structures present along the line of sight is similar to that described in detail in @xcite and based on a sequence of three steps : 1 . preliminary separation within galactocentric rings ; 2 . adjustment of the boundaries based on the velocity structures of the interstellar complexes ; 3 . correction for the spill - over due to the velocity dispersion of the broad h lines between adjacent regions . four regions were defined in galactocentric distance , namely the local arm ( galactocentric radius @xmath16 ) , the interarm region ( @xmath17 ) , the perseus arm ( @xmath18 ) and the region beyond the perseus arm ( which hosts a faint segment of the outer arm ; @xmath19 ) . the boundaries separating these regions under the assumption of a flat rotation curve @xcite for the case of @xmath20 and @xmath21 ( where @xmath22 and @xmath23 are the galactocentric radius and the orbital velocity of the local group of stars , respectively ) are overlaid in figure [ fig : diagram ] . the preparation of the h and co gas maps started from these preliminary velocity boundaries , which were then adjusted for each line of sight to the closest minimum in the h spectrum . then , the spill - over from one velocity interval to the next ones due to the velocity dispersion for the broad h lines was corrected by fitting each h spectrum with a combination of gaussian profiles . we believe that this separation procedure provides more accurate estimates of the ism column densities of each galactic region than a simple slicing based on the rotation curve . in particular , effort was put into separating the outer arm structures from the more massive perseus arm component , especially at @xmath24 where the h lines from the two regions merge into a single broad component . for directions where a minimum in the h brightness temperature profile was not found near the @xmath25 kpc velocity boundary , we integrated the profiles on both sides of the @xmath25 kpc velocity boundary to estimate the perseus and outer arm contributions . then we inserted a line in the gaussian fitting at the outer - arm velocity extrapolated from the @xmath26 trend observed at @xmath24 to correct for the spill - over due to the velocity dispersion . given these difficulties we expect large systematic uncertainties in the outer - arm @xmath12(h ) column densities and the corresponding @xmath0-ray emissivities will not be considered for the scientific interpretation . we note that the impact on the emissivities associated with the inner regions is small , @xmath27% as described in [ sec : crgradient ] . the resulting maps are shown in figure [ fig : hi ] and [ fig : co ] . they exhibit a low level of spatial degeneracy , and thus allow us to separate the @xmath0-radiation arising from the interaction with crs in each component . it has been long debated whether the combination of h and co surveys traces total column densities of neutral interstellar matter . by comparing gas line surveys , the @xmath0-ray observations by egret and dust thermal emission , @xcite reported a considerable amount of neutral gas at the interface between the two h and co emitting phases , associated with cold dust but not properly traced by h and co observations . their finding was then confirmed by lat data for the gould belt in the second quadrant @xcite . in order to complement the h and co maps , we have prepared a map derived from the @xmath28 reddening map by @xcite . the residual point sources at low latitudes were masked by setting to zero regions of @xmath29 radius centered on the positions of potential iras point sources if the @xmath28 magnitude exceeded by @xmath30% that in surrounding pixels . the masked regions were then restored through an inpainting technique @xcite . in the course of the work , various source masking techniques have been used with negligible impact on the h and co emissivity results . the resulting map was fitted with a linear combination of the set of @xmath12(h ) and @xmath10 maps described in the operation was repeated for different choices of h spin temperature . the fit was performed over the same region as for the @xmath0-ray analysis , excluding a @xmath31 region centered around canis major ob 1 @xcite where the temperature correction applied by @xcite to construct the @xmath28 map from the dust thermal emission is highly uncertain . a preliminary fit had led to extremely negative residuals ( @xmath32 mag ) around @xmath33 , @xmath34 . therefore , the residual @xmath28 map was calculated masking this region in the fit . we are aware that the temperature corrections used by @xcite are less reliable with decreasing latitude , but the improvement we find in the @xmath0-ray fit by adding the dust residual map supports the use of their map at low latitude . the residual @xmath35 map , after subtracting the linear combination of @xmath12(h ) and @xmath10 maps , is shown in figure [ fig : ebv ] ( left panel ) . the residuals typically range from @xmath36 to @xmath37 magnitude . large regions of positive residuals are found along the galactic plane , in association with molecular / atomic clouds . they are expected to trace gas not correctly traced by h and co surveys . a remarkable region of positive residuals is detected at intermediate latitudes around @xmath38 , @xmath39 , in a region not covered by co surveys . it corresponds to positive residuals also in @xmath0-rays ( [ sec : procedure ] ) and may be due to a missing , but possibly co - bright molecular cloud ( already suggested by @xcite discussing the completeness of their survey , see figure 8 of their paper ) . the negative residuals are generally small and may result from limitations in the gas column density derivation and/or dust spectral variations . the dust residual map compares well with the @xmath0-ray residual map obtained when using only h and co to model the @xmath0-ray emission ( figure [ fig : ebv ] , right panel ) . the correlation between the spatial distributions of the dust and @xmath0-ray residuals is statistically confirmed in [ sec : procedure ] . dust and @xmath0-rays are consistent with the presence of missing gas in the positive residual clumps . the faint `` glow '' of negative residuals on both sides of the galactic plane is driven by the nearby @xmath12(h ) maps and it remains even when using the smallest possible column - densities derived in the optically thin case . it may suggest a small change in average spin temperature from the massive , compact clouds sampled in the plane to the more diffuse envelopes sampled off the plane , or it may be due to the presence of more missing gas in the plane than our templates can provide for in the fit . the dust - to - gas ratio as well as the @xmath0-ray emissivity in the h components would then be driven to higher values by the low latitude data and would slightly overpredict the data off the plane . the interpretation of the @xmath35 map in this region of the sky is complicated by the lack of distance information for the dust emission . it is not possible to unambiguously assign the residuals to any of the regions under study . since we aim at separating different regions along the lines of sight to investigate the cr density gradient in the outer galaxy , using the h and co lines is essential . we have therefore used the @xmath35 map to correct for the total gas column densities . this approach is supported by the correlation we find between the dust and @xmath0-ray data ( [ sec : procedure ] ) . we also note that , since the dust contribution linearly correlated with the h and co maps has been removed in the @xmath35 map , this procedure allows us to extract the @xmath0-ray emissivities that are actually correlated with the h and co components . to model @xmath0-ray emission not related with interstellar gas , we referred to the galprop code ( e.g. , * ? ? ? * ; * ? ? ? * ) for @xmath0-rays produced through ic scattering and to the first _ fermi _ lat catalog ( 1fgl ) for point sources @xcite . galprop @xcite is a numerical code which solves the cr transport equation within the galaxy and predicts the @xmath0-ray emission produced via interactions of crs with interstellar matter ( nucleon - nucleon interaction and electron bremsstrahlung ) and low - energy photons ( ic scattering ) . ic emission is calculated from the distribution of ( propagated ) electrons and the interstellar radiation fields developed by @xcite . here we adopt the ic model map produced in the galprop run 54_77xvarh7s in which the cr electron spectrum is adjusted to agree with that measured by the lat @xcite . this galprop model has been used in publications by the lat collaboration such as @xcite . the 1fgl catalog is based on the first 11 months of the science phase of the mission and contains 1451 sources detected at a significance @xmath40 ( the threshold is 25 in term of test statistic , ts ] where @xmath41 and @xmath42 is the maximum likelihood with and without including the source in the model , respectively . @xmath41 is conventionally calculated as @xmath43 , where @xmath44 and @xmath45 are the data and the model - predicted counts in each pixel denoted by the subscript @xmath46 , respectively ( see , e.g. , * ? ? ? ts is expected to be distributed as a @xmath47 with @xmath48 degrees of freedom if the numbers of free parameters in the model are respectively @xmath49 and @xmath50 ( 4 for sources in the 1fgl catalog ) . ] ) . for our analysis we considered 21 point sources in the roi with ts larger than 50 . following a well - established approach that dates back to the cos - b era ( e.g. , * ? ? ? * ) , we modeled the @xmath0-ray emission as a linear combination of maps tracing the column density of the interstellar medium . this approach is based on a simple , but very plausible assumption : @xmath0-rays are generated through interactions of crs and the interstellar gas , and the ism itself is transparent to @xmath0-rays . then , assuming that cr densities do not significantly vary over the scale of the interstellar complexes under study and that crs penetrate clouds uniformly to their cores we can model the @xmath0-ray intensities to first order as a linear combination of contributions from cr interactions with the different gas phases in the various regions along each line of sight . we also added the ic model map by galprop and models for point sources taken from the 1fgl catalog as described in 2.2 . to represent the extragalactic diffuse emission and the residual instrumental background from misclassified cr interactions in the lat detector , we also added an isotropic component . cr interactions with ionized gas are not explicitly included in the model . the mass column densities of ionized gas are poorly known , but their contribution is generally lower ( @xmath51 ) than that of the neutral gas and its scale height is much larger ( @xmath52 kpc compared with @xmath53 kpc ; @xcite ) . we therefore expect the diffuse @xmath0-ray emission originating from ionized gas to be largely accommodated in the fit by other components with large angular scales , such as the isotropic and ic ones , and to minimally impact the determination of the neutral gas emissivities . therefore , the @xmath0-ray intensities @xmath54 ( @xmath55 ) can be modeled as : @xmath56 where sum over @xmath46 represents the combination of the galactic regions , @xmath57 ( @xmath58 ) and @xmath59 ( @xmath60 ) are the emissivities per h atom and per @xmath10 unit , respectively . @xmath61 ( @xmath62 ) is the emissivity per unit of the @xmath35 map ( for which independent normalizations are allowed between the positive and negative residuals ; see [ sec : procedure ] ) . @xmath63 and @xmath64 are the ic model and isotropic background intensities ( @xmath65 ) , respectively , and @xmath66 represents the point source contributions . compared to the egret study by @xcite , we use two additional maps to better trace the ism : the co map in the perseus arm and the @xmath35 map . the lat on board the _ fermi _ gamma - ray space telescope , launched on 2008 june 11 , is a pair - tracking telescope , detecting photons from @xmath67 mev to more than 300 gev . details on the lat instrument and pre - launch expectations of the performance can be found in @xcite , and the on - orbit calibration is described in @xcite . routine science operations with the lat started on 2008 august 4 . we have accumulated events from 2008 august 4 to 2010 february 4 to study diffuse @xmath0-rays in our roi . during this time interval the lat was operated in sky survey mode nearly all of the time , obtaining complete sky coverage every two orbits and relatively uniform exposures over time . we used the standard lat analysis software , the _ science tools _ and selected events satisfying the standard low - background event selection ( the so - called _ diffuse _ class ; * ? ? ? * ) . we also required the reconstructed zenith angles of the arrival direction of photons to be less than @xmath68 and the center of the lat field of view to be within @xmath69 from the zenith , in order to reduce the contamination of photons from the earth limb . in addition , we excluded the period of time during which the lat detected bright grbs , i.e. , grb080916c @xcite , grb090510 @xcite , grb090902b @xcite , and grb090926a @xcite . the model described by equation ( 1 ) was fitted to the data using the _ science tools _ , which take into account the energy - dependent instrument point spread function and effective area . we have analyzed the lat data from 100 mev to 25.6 gev using 13 logarithmically - spaced energy bands from 100 mev to 9.05 gev , and a single band above 9.05 gev . we then have compared the model and data in each energy band using a binned maximum - likelihood method with poisson statistics ( in @xmath70 bins ) ; we thus did not assume an a - priori spectral shape of each model component except for the ic emission . for the other components the convolution with the instrument response functions was performed assuming an @xmath71 spectrum , and the integrated intensities were allowed to vary in each narrow energy bin . changing the fixed spectral shape index over the range from @xmath72 to @xmath73 has a negligible effect on the obtained spectrum . in the highest energy band , we have set both the normalization and the spectral index free to accommodate the wider bin width . we used a post - launch response function , p6_v3_diffuse , developed to account for the @xmath0-ray detection inefficiencies due to pile - up and accidental coincidence in the lat @xcite . we stopped at 25.6 gev since the photon statistics does not allow us to reliably separate different gas components above this energy . we started with point sources detected with high significance ( @xmath74 ) in the 1fgl catalog ; we have 14 sources in our roi for which the normalizations are set free . we also included 8 sources lying just outside ( @xmath75 ) of the region boundaries , with all the spectral parameters fixed to those in the 1fgl catalog . as a starting point we used h maps prepared for @xmath13 . we added model components step by step as described below . we first fitted the lat data using equation ( 1 ) without the @xmath35 map and the co map in the perseus arm , and then included the co map and confirmed that the fit improved significantly ; the test statistic summed over 14 bands with separate fits in each band ( i.e. , 14 more free parameters ) is 187.6 . the @xmath0-ray emission associated to the gas traced by co in the perseus arm is thus significantly detected by the lat . next we included the @xmath35 map in the analysis . we allowed the independent normalizations between the positive part and the negative part of the @xmath35 map , and found that the normalizations differ with each other . we thus will use the independent normalizations hereafter . we chose this model to better represent the lat data and constrain the cr distributions , and leave a detailed discussion about the use of dust as ism tracer to a dedicated paper . the improvement of the fit is very significant : @xmath76 for 28 more free parameters . the correlation between the @xmath35 map and the @xmath0-ray residual map obtained by the fit without the @xmath35 map , shown in figure [ fig : ebv ] , further supports the use of @xmath35 map in our analysis . we also tried a fit without the ic component to assess the systematics . the effects on the derived emissivities are typically 23% and @xmath15 5% for @xmath77 and @xmath78 , respectively . they are much smaller than the statistical errors and systematic uncertainties ( see below ) , although the inclusion of the ic map improves the fit to the lat data . therefore the uncertainties on the ic model have no significant impact on our analysis due to its rather flat distribution across the region of interest while the gas in the ism is highly structured . on the other hand , lowering the threshold for point sources down to @xmath79 yields an about twice smaller emissivity for the @xmath10 map in the perseus arm . the emissivities of other components are unchanged within the statistical errors . this is plausibly due to the very clumpy distribution of the clouds in the perseus arm as seen by a terrestrial observer , see figure [ fig : co ] , which makes it difficult to separate from that of some discrete sources . we thus use equation ( 1 ) with point sources detected at @xmath80 in the 1fgl catalog as our baseline model , but we do not consider the highly uncertain co emissivities in the perseus arm for the discussion . we summarize the results in tables [ tab:125k ] and [ tab:250k ] for @xmath13 and 250 k , respectively , and the number of counts in each energy bin in table [ tab : counts ] . the differential emissivities are multiplied by @xmath81 where @xmath82 is the center of each energy bin in logarithmic scale . they are given for each model component . we note that our isotropic term ( @xmath64 ) includes the contribution of the instrumental background and might partially account also for ionized gas ( see [ anamodel ] ) , thus it is significantly larger than the extragalactic diffuse emission reported by @xcite . to illustrate the fit quality , we give the data and model count maps and the residual map in figure [ fig : gammamap ] ( for @xmath13 ) , in which residuals ( data minus model ) are expressed in approximate standard deviation units ( square root of model - predicted counts ) . although some structures ( clustering of positive or negative residuals ) are observed , the map shows no excesses below @xmath83 and above @xmath84 . over 99% of the pixels are within @xmath85 . we thus conclude that our model reasonably reproduces the data . figure [ fig : summaryspectra ] presents the fitted spectra for each component . the emission from the h gas dominates the @xmath0-ray flux . although the emission from the gas in the co - bright phase and that traced by @xmath35 is fainter than the ic and isotropic components , their characteristic spatial structures ( see figure [ fig : hi ] and [ fig : co ] ) allow their spectra to be reliably constrained . to examine the effect of the optical depth correction applied to derive the h maps , as anticipated above we tried several choices of a uniform @xmath14 . we stress that the true @xmath14 is likely to vary within clouds , but we stick to this simple approximation exploring the following values : 100 k ( which is a reasonable lower limit in the uniform approximation ) , 250 k and 400 k ( which are the two values indicated by absorption measurements in the outer galaxy by @xcite ) , and the optically - thin approximation ( which yields the lower limit allowed on the atomic column densities ) . the results on the maximum log - likelihood values are summarized in table [ tab : lnl ] together with the integrated h emissivities obtained above 100 mev in each region . the evolution of @xmath86 with @xmath14 is plotted in figure [ fig : logl ] . the h emissivity varies by @xmath87% for the local arm , @xmath88% for the interarm region , and @xmath89% for the perseus arm with respect to the @xmath13 case . we observe an increase of @xmath86 with increasing spin temperature . considering the fact that @xmath90 is a typical value in the second quadrant of the outer galaxy according to a recent study by @xcite and because @xmath86 saturates at @xmath91 , we regard 250 k as a plausible estimate of the average @xmath14 in our roi . unfortunately , the estimates by @xcite have a rather large uncertainty ( about @xmath92 k ) in each galactocentric radius bin , and they do not cover the region in the third quadrant we are investigating ( see figure 5 of * ? ? ? * ) . in the following sections , we will concentrate on @xmath13 for comparison with previous @xmath0-ray measurements and on @xmath90 which agrees well with h absorption and the lat data . in figures [ fig : spectra_125k ] and [ fig : spectra_250k ] ( left panels ) , we report the emissivity spectra found per h atom in the local arm , interarm , perseus arm and outermost regions for @xmath93 and 250 k , respectively . for comparison with the local interstellar spectrum ( lis ) we also plot the model spectrum used in @xcite which agrees well with lat data in a mid - to - high - latitude region ( @xmath94 ) of the third quadrant ( assuming @xmath13 ) . we see that the spectral shape of the local - arm emissivity agrees well with the model for the lis and does not depend on the choice of spin temperature . the integral emissivity of the local arm is 10% lower than that reported by @xcite for the same spin temperature . this difference is not significant given the uncertainties in the kinematic separation of the gas components . the present result is also consistent with the measurement in the second quadrant @xcite . together they show that the cr density along the local arm is rather uniform within 1 kpc around the sun , both in the second and third quadrants . the comparison of the data with the model emissivity expected for the local arm region based on locally - measured crs ( figures [ fig : spectra_125k ] and [ fig : spectra_250k ] ) indicates a better fit for higher @xmath14 ; @xmath13 gives emissivities 1520% lower than the model , whereas @xmath90 shows better agreement by about 10% . although the theoretical emissivity has uncertainties due to imperfect knowledge of the cr spectrum ( see * ? ? ? * ) , the fact that a high @xmath14 value yields a better match both to the local absolute emissivity and to the spatial distribution of the diffuse emission ( figure [ fig : logl ] ) leads to larger @xmath14 than a value conventionally used in @xmath0-ray astrophysics ( 125 k ) . this is in accord with independent estimates of @xmath14 as discussed in [ sec : procedure ] . we also observe that the emissivity spectra do not vary significantly with galactocentric distances in the outer galaxy . to examine the spectral shape more quantitatively , we present the emissivity ratios of the interarm and perseus regions relative to the local arm in the right panels of figures [ fig : spectra_125k ] and [ fig : spectra_250k ] . the spectral shape in the interarm region is found to be consistent with that in the local arm ; a fit to the data for @xmath13 with a constant ratio gives @xmath95 for 13 degrees of freedom . although the fit is not fully acceptable for the perseus arm ( @xmath96 ) , the large @xmath47 is driven solely by the last bin . we note a possible interplay between the perseus - arm and the adjacent outer - arm emissivities in the highest energy bins ( see left panels of figure [ fig : spectra_125k ] and [ fig : spectra_250k ] ) . it can be due to a small but non - negligible spatial difference between the modeled templates and data and/or to the presence of unresolved point sources ( generally harder than diffuse emission ) . photon fluctuations from the structured gas components can also lead the fit to a slightly different solution in the spatial separation of the components . one would expect these possible systematic uncertainties to become important at high energy given the limited counts in the overall map . it is difficult to quantitatively test these effects without knowledge of the true model distributions , but we can note that the small deviations seen at 400560 mev and 1.62.2 gev from a constant ratio are not confirmed by the general trend of the other points . they indicate that there are systematic uncertainties not fully accounted for by the statistical errors in the fit . we thus do not claim nor deny the spectral softening of the perseus arm at high energy . a test using @xmath90 for the @xmath12(h ) maps gives the same conclusion on the spectral shape . we thus conclude that the spectral shapes are consistent with the lis in the @xmath97 gev energy band , independent of the assumed @xmath14 . considering that these @xmath0-rays trace cr nuclei of energies from a few gev to about 100 gev ( see , e.g. , figure [ fig : emissivitygradient ] of * ? ? ? * ) , lat data indicate that the energy distribution of the main component of galactic crs does not vary significantly in the outer galaxy in the third quadrant . we note that @xcite reported a possible spectral hardening in the perseus arm in the second quadrant . this might be due to the presence of the very active star - forming region of ngc 7538 and of crs having not diffused far from their sources , or to contamination by hard unresolved point sources . in fact @xcite did not rule out the possibility that their result is due to systematic effects . high - energy @xmath0-rays are a powerful probe to determine the co - to - h@xmath98 calibration ratio , @xmath9 , if the cr flux is comparable in the different gas phases inside a cloud . since the @xmath0-ray emission from the molecular gas is primarily due to cr interactions with @xmath99 , and since the molecular binding energy is negligible in processes leading to @xmath0-ray production , the emissivity per @xmath99 molecule is expected to be twice the emissivity per h atom . then , under the hypothesis that the same cr flux penetrates the h- and co - bright phases of an interstellar complex , we can calculate @xmath9 as @xmath100 . we show @xmath78 as a function of @xmath77 for the local arm and the interarm region in figure [ fig : xco ] . we do not consider the correlation in the perseus arm , because @xmath78 from this region is affected by large systematic uncertainties ( see 3.2 ) . since the emissivity associated with the co - bright gas is not well determined in the lowest energy range ( see table [ tab:125k ] and [ tab:250k ] ) because of the poor angular resolution of the lat , and the fit at very high energy is affected by larger uncertainties ( 4.1 ) , we have plotted only data in the 200 mev9.05 gev range . the linear correlation supports the assumption that galactic crs penetrate molecular clouds uniformly to their cores . it also suggests that contamination from point sources and cr spectral variations within the clouds are small . we have derived the maximum - likelihood estimates of the slope and intercept of the linear relation between @xmath78 and @xmath77 taking into account that @xmath78 and @xmath77 are both measured ( not true ) values with known uncertainties . the resulting intercepts are compatible with zero . the @xmath9 values we have obtained for @xmath90 are @xmath4 for the local arm ( @xmath101 ) and @xmath102 for the interarm region ( @xmath17 ) . decreasing the spin temperature to 125 k does not affect the @xmath9 derivation in the well resolved , not too massive , clouds of the local arm where we find @xmath103 . on the other hand , the separation in the @xmath0-ray fit between the dense h peaks and clumpy co cores becomes more difficult for more distant , less resolved clouds where h and co tend to peak in the same directions . a change in the largest @xmath12(h ) column - densities from the optical depth correction can impact the @xmath9 determination in two ways : first by changing the @xmath77 emissivity , second by modifying the @xmath12(h ) contrast within the cloud , hence the h and co separation . the global impact is mild since we find @xmath104 in the interarm region for @xmath13 . @xcite reported comparable values of @xmath9 in the second quadrant for @xmath13 : they obtained @xmath105 and @xmath106 for the local arm ( @xmath107 ) and the perseus arm . given the systematic uncertainty in @xmath9 , roughly of the order of @xmath108 , due to h optical depth correction , the results of both studies point to a rather uniform ratio over several kpc in the outer galaxy . yet , these values are twice larger than found in the very nearby gould - belt clouds of cassiopeia and cepheus , @xmath109 , @xmath13 . however , we confirm that the increase in @xmath9 beyond the solar circle is significantly lower than the trend adopted in the model of @xcite . what fraction of the gould - belt to local - arm difference in the average @xmath9 can be attributed to a difference in the spatial sampling ( resolution ) of the clouds remains to be investigated . nearly the same region has been analyzed by @xcite using egret data . the main difference from their analysis is our improved scheme for the kinematical separation of the ism components along the lines of sight and the inclusion of the reddening residual map . the @xmath9 value measured in the local arm by egret , @xmath110 ( for @xmath13 ) , is statistically compatible with ours . the fact that we excluded the region of monoceros r2 from the analysis can also explain in part this difference . because of the pile - up of different clouds along a line of sight , the derivation of individual cloud masses is beyond the scope of this study . let us just note that maddalena s cloud , with its very low rate of star formation , seems to share a quite conventional @xmath9 factor . further investigation , including higher resolution @xmath0-ray maps when more high - energy lat data become available , is required to fully understand the mass distribution in the clouds . in figure [ fig : emissivitygradient ] ( left panel ) we show the emissivity gradient found beyond the solar circle for different spin temperatures . here we do not include the results for the optically - thin approximation which is equivalent to an infinitely high @xmath14 and gives similar emissivities to @xmath111 . the typical statistical errors associated with these measurements are illustrated in the right panel for the @xmath13 case . in the right panel , a shaded area shows the characteristic systematic error due to the lat event selection efficiency , evaluated to be @xmath112 in the energy range under study . in order to evaluate the impact of the delicate separation of the gas in the outermost region , we have compared two extreme cases . the first one adopts the kinematic @xmath25 kpc boundary and applies no correction for velocity dispersion , the second assigns all the outer - arm gas to the perseus arm . the emissivity in the perseus arm differs by about 5% from the original one , and those in the local arm and interarm regions hardly change . therefore these effects are significantly smaller than the uncertainties due to the optical depth correction of the h data . we also note that the main effect of the lat selection efficiency uncertainty is to rigidly shift the profile without any significant impact on the gradient . we thus conclude that the most important source of uncertainty in the cr density gradient derivation is currently that in the @xmath12(h ) determination . this is mainly because the optical depth correction is larger for dense h clouds in the local and perseus arms than for diffuse clouds in the interarm region . the loss in contrast between the dense ( low - latitude ) and more diffuse ( mid - latitude ) h structures resulting from an increase in spin temperature affects the fit , particularly in the perseus component which is more narrowly concentrated near the plane . when probing the cr densities as the `` ratio '' between the observed numbers of @xmath0-rays to h atoms , at the precision provided by the lat the uncertainties in the ism densities are dominant . an interesting finding of the former egret analysis @xcite was an enhancement of the @xmath0-ray emissivity in the perseus arm compared with the interarm region . this possibility is relevant for models of diffuse @xmath0-ray emissions based on the assumption that cr and ism densities are coupled ( e.g. , * ? ? ? * and references therein ) . the local arm emissivity obtained by the egret study for @xmath13 is @xmath113 , which is @xmath114 larger than our lat result . however , the two studies are based on different h surveys which yield different total @xmath12(h ) column densities integrated along the lines of sight . the column - density ratios between the surveys varies from 0.6 to 1.0 within the roi , with an average value of @xmath115 . the difference is likely due to the improved correction for stray - radiation in the more recent survey , as discussed in @xcite . the egret local - arm emissivity scaled by 0.8 is in good agreement with our result for the same spin temperature . if we do not include the @xmath35 map in the fitting , we obtain an emissivity of @xmath116 which is still consistent with the down - scaled egret result within @xmath1515% . we thus conclude that our result is consistent with the previous study but is more reliable because of higher @xmath0-ray statistics , finer resolution and an improved h gas survey . we can therefore compare the present emissivity gradient ( for consistency in the case @xmath13 ) with that reported by the egret study , as summarized in figure [ fig : emissivitygradient ] ( right panel in which the egret results multiplied by 0.8 are also shown ) . although we observe good agreement between the two studies in the local and the perseus arms , this is not true for the arm / interarm contrast . the difference could be due to the simple partitioning in cloud velocity used for the egret study . the h mass obtained for clouds in the interarm region with the simple partitioning is @xmath117 larger ( for @xmath13 ) than with our separation scheme , exaggerating the amount of gas in the interarm region , and thus lowering the emissivity by the same amount . our emissivity profile is thus consistent with the previous study , but with improved precision ( smaller statistical errors ) and accuracy ( more reliable region separation method and better estimation of the point source contributions ) . we thus do not confirm a marked drop in the interarm region . low spin temperatures yield a smooth decline in h emissivity to @xmath118 in the outer galaxy , without showing a significant coupling with ism column densities . the perseus - to - interarm contrast is at most of the order of 15%20% for high spin temperatures as shown in the left panel of figure [ fig : emissivitygradient ] . these profiles are similar at all energies , in particular at high energies where the component separation is more reliable thanks to the better angular resolution . the surface density of h in the perseus arm is on average 30%40% higher than in the interarm region . therefore , even if we adopt @xmath111 which gives the largest arm - interarm contrast , the coupling scale ( or the coupling length ) between the crs and matter ( e.g. , * ? ? ? * ) required to agree with the lat data would be larger than those usually assumed for this type of model ( @xmath119 . see e.g. , * ? ? ? * figure 7 ) . whether the true emissivity profile exhibits a small contrast between the arms or smoothly declines with distance is beyond our measurement capability without further constraints on the hcolumn - density derivation . new h absorption measurements will allow us to investigate this issue with better accuracy . to compare with the second quadrant results @xcite , we have integrated the emissivities above 200 mev for @xmath13 . we find values of @xmath120 , @xmath121 and @xmath122 for the local - arm , interarm , and perseus - arm regions , respectively . the nearer value is about 20% lower than in the 2nd quadrant ( which , however , samples very nearby clouds in the gould belt ) and the outer ones compare very well with the 2nd quadrant measurement over the same galactocentric distance range . despite the uncertainties due to the optical - depth correction ( that might have a different impact in the two quadrants ) , both lat studies consistently point to a slowly decreasing emissivity profile beyond @xmath123 . let us consider the predictions by a cr propagation model to see the impact of such a flat profile on the cr source distribution and propagation parameters . we adopted a galprop model , starting from the configuration used for the run 54_77xvarh7s which we used to predict the ic contribution . the cr source distribution in this model is , @xmath124 where @xmath125 is the distance of the sun to the galactic center . as shown in figure [ fig : crsourcemodel ] , this function is intermediate between the distribution of supernova remnants ( snrs ) obtained from the @xmath126 relation @xcite and that traced by pulsars @xcite . the boundaries of the propagation region are chosen to be @xmath127 ( maximum galactocentric radius ) and @xmath128 ( maximum height from the galactic plane ) , beyond which free escape is assumed . the spatial diffusion coefficient is assumed to be uniform across the galaxy and is taken as @xmath129 , where @xmath130 is the velocity of the particle relative to the speed of light and @xmath131 is the rigidity . we adopted @xmath132 and @xmath133 ( kolmogorov spectrum ) . reacceleration due to the interstellar magnetohydrodynamic turbulence , which is thought to reproduce the observed b / c ratio at low energy , assumes an alfvn velocity @xmath134 . the cr source distribution and propagation model parameters have been used often in the literature ( see e.g. , * ? ? ? we note that the same cr source distribution and similar propagation parameters are adopted in the galprop run used by @xcite . the left panel of figure [ fig : gradientmodeldata ] compares the calculated profile ( solid line ) with lat constraints ( bow - tie plot bracketing the profiles obtained for different @xmath14 ; see the left panel of figure [ fig : emissivitygradient ] ) . the model is normalized to the lat measurement in the innermost region . despite the large uncertainties , lat data lead to a significantly flatter profile than predicted by our model ; the lat results indicate to a factor of 2 larger emissivity ( cr energy density ) in the perseus arm even if we assume @xmath135 . the higher @xmath14 makes the discrepancy larger , hence the conclusion is robust . not using the @xmath35 map in the analysis does not change the conclusion , since the emissivities in the interarm region and the perseus arm are almost unaffected by its presence . the discrepancy between the @xmath0-ray emissivity gradient in the galaxy and the distribution of putative cr sources has been known as the `` gradient problem '' since the cos - b era ( e.g. , * ? ? ? it has led to a number of possible interpretations , including , for the specific case of the outer galaxy , the possibility of a very steep gradient in @xmath9 beyond the solar circle @xcite . the emissivities in the outer galaxy were more difficult to determine in the cos - b / egret era due to lower statistics and higher backgrounds . now , thanks to the high quality of the lat data and the improved component separation technique applied to gas line data , we measure a flat h emissivity gradient in the outer galaxy together with a flat evolution of @xmath9 over several kpc , so the gradient problem requires another explanation . the most straightforward possibility is a larger halo size ( @xmath136 ) , as discussed by , e.g. , @xcite , @xcite and @xcite . we therefore tried several choices of @xmath136 and @xmath137 as summarized in the dotted lines in the same figure . the values of @xmath137 are chosen to reasonably reproduce the lis of protons and electrons , b / c ratio and @xmath138 ratio at the solar system , and are similar to those given in @xcite . all models are normalized to the lat data in the local arm . models with @xmath139 or smaller are found to give too steep emissivity gradients . a cr source distribution as in equation 2 with a very large halo ( @xmath140 ) provides a gradient compatible with the @xmath0-ray data , if we fully take into account the systematic uncertainties . we note that @xmath141 is still compatible with @xmath138 measurements ( e.g. , * ? ? ? considering the large statistical and systematic uncertainties in the snr distribution , a flatter cr source distribution in the outer galaxy also could be possible . we thus tried a modified cr source distribution , in which the distribution is the same as equation 2 below @xmath142 and constant beyond it ( see a thin solid line of figure [ fig : crsourcemodel ] as an example ) . figure [ fig : gradientmodeldata ] right shows the models with several choices of @xmath142 for @xmath128 and @xmath132 . we obtained a reasonable fit to the data using a flat cr source distribution beyond @xmath143 . such a constant cr source density in the outer galaxy is in contrast not only with the ( highly uncertain ) distribution of snrs , but also with other tracers of massive star formation and snrs , like , 1 ) co lines which trace the interstellar phase where massive stars form ( e.g. , * ? ? ? * ) , 2 ) ob star counts ( e.g. , * ? ? ? * ) , and 3 ) the @xmath144al line which is related to the injection of stellar nucleosynthesis products in the ism by snrs @xcite . however , a very large halo size and/or a flat cr source distribution just beyond the solar circle seem to be favored by the lat data . the above discussion depends on the propagation parameters and the solution is not unique . the exploration could be extended to other regions of the parameter space or to a non - uniform diffusion coefficient ( e.g. , * ? ? ? * ) , but examining propagation models in detail is beyond the scope of our study . our bottom line is that the analysis of lat data presented here and by @xcite consistently show that the cr density gradient in the outer galaxy is flatter than expectations by commonly - used propagation models . in the future , the extension to the inner part and the accurate determination of the gradient over the whole galaxy will be key to constraining the cr origin and transport . we also note that a spin temperature @xmath91 , which is favored by recent studies in the outer galaxy ( e.g. , * ? ? ? * ) , gives a small arm / interarm contrast at the 1020% level that is not fully compatible with the propagation models ( including the one we adopted here ) which predict a monotonic cr gradient . even though the present analysis includes a dust template to account for the abundant missing gas present locally at the interface between the h and co - bright phases , an alternative way to reconcile the flat emissivity profile and a marked decline in cr density in the outer galaxy is to invoke an increase in missing gas mass with galactocentric distance in the low metallicity environments of the outer galaxy ( see e.g. * ? ? ? * ; * ? ? ? * ) beyond the local correction applied here . we note that the large masses of dark gas in the outer galaxy suggested by @xcite ( outweighing that of h by a factor of 5 - 15 ) might explain our results , whereas the remarkably constant dark gas fraction of 30% with mild dependence on metallicity suggested by @xcite is not sufficient to explain the large h emissivities measured by the lat beyond the solar circle . we have studied the diffuse @xmath0-ray emission in the third galactic quadrant using the first 18 months of _ fermi _ lat science data . thanks to the excellent performance of the lat , we have obtained high - quality emissivity spectra of the atomic and molecular gas ( traced by @xmath10 ) in the 100 mev 25.6 gev energy range . at the level of accuracy allowed by the lat , the study of cr densities from @xmath0-ray observations is now mostly limited by the understanding of the ism mass tracers , notably by the uncertainties in the derivation of atomic gas column densities from h surveys and by the distribution of gas not accounted for by radio and microwave line surveys . in spite of those uncertainties , robust conclusions can be drawn concerning the ism and crs . the molecular mass calibration ratio of the local arm is found to be @xmath145 , significantly larger than that for the very local gould - belt clouds in the second galactic quadrant reported by @xcite . no significant differences of the ratio are found between the local arm and the interarm region . no significant variations in the cr spectra are found across the outer galaxy in the region studied , and no large contrast in emissivity is seen in the interarm region between the local and perseus arms ( a contrast @xmath146 1020% is allowed by data ) . the measured gradient is much flatter than predictions by a widely - used propagation model assuming that the cr source distribution largely peaks in the inner galaxy . a larger halo size and/or a flatter cr source distribution beyond the solar circle than those usually assumed are required to reproduce the lat data , while other scenarios such as a non - uniform diffusion coefficient or vast amounts of missing gas in the outer galaxy are also possible . reliable determinations of the amount of atomic hydrogen in the plane are key to better constraining the property of crs in our galaxy . the _ fermi _ lat collaboration acknowledges generous ongoing support from a number of agencies and institutes that have supported both the development and the operation of the lat as well as scientific data analysis . these include the national aeronautics and space administration and the department of energy in the united states , the commissariat lenergie atomique and the centre national de la recherche scientifique / institut national de physique nuclaire et de physique des particules in france , the agenzia spaziale italiana and the istituto nazionale di fisica nucleare in italy , the ministry of education , culture , sports , science and technology ( mext ) , high energy accelerator research organization ( kek ) and japan aerospace exploration agency ( jaxa ) in japan , and the k. a. wallenberg foundation , the swedish research council and the swedish national space board in sweden . this paper makes use of a development version of galprop provided by the galprop team to the lat collaboration solely for interpretation of the lat data . galprop development is supported by nasa grant nnx09ac15 g . ( h ) ( in unit of @xmath147 ) for the local arm ( top left ) , interarm ( top right ) , perseus arm ( bottom left ) , and outer arm ( bottom right ) regions , obtained for a spin temperature @xmath13 . the outlined area in the bottom right corner is not used in the analysis ( see 1 ) . the maps have been smoothed with a gaussian with @xmath148 for display . , title="fig:",scaledwidth=50.0% ] ( h ) ( in unit of @xmath147 ) for the local arm ( top left ) , interarm ( top right ) , perseus arm ( bottom left ) , and outer arm ( bottom right ) regions , obtained for a spin temperature @xmath13 . the outlined area in the bottom right corner is not used in the analysis ( see 1 ) . the maps have been smoothed with a gaussian with @xmath148 for display . , title="fig:",scaledwidth=50.0% ] ( h ) ( in unit of @xmath147 ) for the local arm ( top left ) , interarm ( top right ) , perseus arm ( bottom left ) , and outer arm ( bottom right ) regions , obtained for a spin temperature @xmath13 . the outlined area in the bottom right corner is not used in the analysis ( see 1 ) . the maps have been smoothed with a gaussian with @xmath148 for display . , title="fig:",scaledwidth=50.0% ] ( h ) ( in unit of @xmath147 ) for the local arm ( top left ) , interarm ( top right ) , perseus arm ( bottom left ) , and outer arm ( bottom right ) regions , obtained for a spin temperature @xmath13 . the outlined area in the bottom right corner is not used in the analysis ( see 1 ) . the maps have been smoothed with a gaussian with @xmath148 for display . , title="fig:",scaledwidth=50.0% ] ( in unit of @xmath149 ) for the local arm ( top left ) , interarm ( top right ) , and the perseus arm ( bottom left ) regions . the small box in the bottom right corner indicates the area not considered in the analysis . the maps have been smoothed with a gaussian of @xmath150 for display . , title="fig:",scaledwidth=50.0% ] ( in unit of @xmath149 ) for the local arm ( top left ) , interarm ( top right ) , and the perseus arm ( bottom left ) regions . the small box in the bottom right corner indicates the area not considered in the analysis . the maps have been smoothed with a gaussian of @xmath150 for display . , title="fig:",scaledwidth=50.0% ] ( in unit of @xmath149 ) for the local arm ( top left ) , interarm ( top right ) , and the perseus arm ( bottom left ) regions . the small box in the bottom right corner indicates the area not considered in the analysis . the maps have been smoothed with a gaussian of @xmath150 for display . , title="fig:",scaledwidth=50.0% ] map in unit of magnitudes , obtained by subtracting the parts linearly correlated with the combination of @xmath12(h ) and @xmath10 maps . the small box in the bottom right corner shows the area not considered in the analysis . the map has been smoothed with a gaussian of @xmath151 for display . ( right ) @xmath0-ray residual ( data minus model ) map obtained by the fit without the @xmath35 map ( only h and co maps ) in unit of standard deviations ( square root of model - predicted counts , saturated between @xmath152 and @xmath153 ) . the map has been smoothed with a gaussian of @xmath154 . , title="fig:",scaledwidth=50.0% ] map in unit of magnitudes , obtained by subtracting the parts linearly correlated with the combination of @xmath12(h ) and @xmath10 maps . the small box in the bottom right corner shows the area not considered in the analysis . the map has been smoothed with a gaussian of @xmath151 for display . ( right ) @xmath0-ray residual ( data minus model ) map obtained by the fit without the @xmath35 map ( only h and co maps ) in unit of standard deviations ( square root of model - predicted counts , saturated between @xmath152 and @xmath153 ) . the map has been smoothed with a gaussian of @xmath154 . , title="fig:",scaledwidth=50.0% ] and @xmath153 ) above 100 mev obtained by our analysis . point sources with @xmath80 in the 1fgl included in the fit are shown by crosses in the model map . data / model count maps are in @xmath70 pixels , and the residual map has been smoothed with a gaussian of @xmath154.,title="fig:",scaledwidth=50.0% ] and @xmath153 ) above 100 mev obtained by our analysis . point sources with @xmath80 in the 1fgl included in the fit are shown by crosses in the model map . data / model count maps are in @xmath70 pixels , and the residual map has been smoothed with a gaussian of @xmath154.,title="fig:",scaledwidth=50.0% ] and @xmath153 ) above 100 mev obtained by our analysis . point sources with @xmath80 in the 1fgl included in the fit are shown by crosses in the model map . data / model count maps are in @xmath70 pixels , and the residual map has been smoothed with a gaussian of @xmath154.,title="fig:",scaledwidth=50.0% ] is assumed and spectra in the interarm region and the perseus arm are shifted horizontally for clarity . , title="fig:",scaledwidth=50.0% ] is assumed and spectra in the interarm region and the perseus arm are shifted horizontally for clarity . , title="fig:",scaledwidth=50.0% ] and 250 k are shown in the left panel and the right panel , respectively . dotted lines show the best linear fits . each data point corresponds to an energy bin used in the @xmath0-ray analysis . ( see table [ tab:125k ] and [ tab:250k ] ) , title="fig:",scaledwidth=50.0% ] and 250 k are shown in the left panel and the right panel , respectively . dotted lines show the best linear fits . each data point corresponds to an energy bin used in the @xmath0-ray analysis . ( see table [ tab:125k ] and [ tab:250k ] ) , title="fig:",scaledwidth=50.0% ] . ( right ) emissivity gradient obtained by the lat compared with the egret results under the assumption of @xmath13 . the shaded area indicates the systematic uncertainty in the lat selection efficiency of @xmath155% . the egret points have been downscaled by 20% to account for the change in h survey data between the two studies ( see [ contpar ] ) . , title="fig:",scaledwidth=50.0% ] . ( right ) emissivity gradient obtained by the lat compared with the egret results under the assumption of @xmath13 . the shaded area indicates the systematic uncertainty in the lat selection efficiency of @xmath155% . the egret points have been downscaled by 20% to account for the change in h survey data between the two studies ( see [ contpar ] ) . , title="fig:",scaledwidth=50.0% ] relation @xcite and that traced by the pulsar distribution @xcite shown by dotted lines . the thin solid line represents an example of the modified distributions introduced to reproduce the emissivity gradient by the lat . , scaledwidth=50.0% ] ) = ( 1 kpc , @xmath156 ) , ( 2 kpc , @xmath157 ) , ( 4 kpc , @xmath158 ) , ( 10 kpc , @xmath159 ) and ( 20 kpc , @xmath160 ) . the solid line is for @xmath128 . the right panel shows different choices of the break distance beyond which a flat cr source distribution is assumed : @xmath161 with 1 kpc steps . , title="fig:",scaledwidth=50.0% ] ) = ( 1 kpc , @xmath156 ) , ( 2 kpc , @xmath157 ) , ( 4 kpc , @xmath158 ) , ( 10 kpc , @xmath159 ) and ( 20 kpc , @xmath160 ) . the solid line is for @xmath128 . the right panel shows different choices of the break distance beyond which a flat cr source distribution is assumed : @xmath161 with 1 kpc steps . , title="fig:",scaledwidth=50.0% ] ccccccccccc 0.100.14 & @xmath162 & @xmath163 & @xmath164 & @xmath165 & @xmath166 & @xmath167 & @xmath168 & @xmath169 & @xmath170 & @xmath171 + 0.140.20 & @xmath172 & @xmath173 & @xmath174 & @xmath175 & @xmath176 & @xmath177 & @xmath178 & @xmath179 & @xmath180 & @xmath181 + 0.200.28 & @xmath182 & @xmath183 & @xmath184 & @xmath185 & @xmath186 & @xmath187 & @xmath178 & @xmath188 & @xmath189 & @xmath190 + 0.280.40 & @xmath191 & @xmath192 & @xmath193 & @xmath194 & @xmath195 & @xmath196 & @xmath197 & @xmath198 & @xmath199 & @xmath200 + 0.400.56 & @xmath201 & @xmath202 & @xmath203 & @xmath204 & @xmath205 & @xmath206 & @xmath207 & @xmath208 & @xmath209 & @xmath210 + 0.560.80 & @xmath211 & @xmath212 & @xmath213 & @xmath214 & @xmath215 & @xmath216 & @xmath217 & @xmath218 & @xmath219 & @xmath220 + 0.801.13 & @xmath221 & @xmath222 & @xmath223 & @xmath224 & @xmath225 & @xmath226 & @xmath227 & @xmath228 & @xmath229 & @xmath230 + 1.131.60 & @xmath231 & @xmath232 & @xmath233 & @xmath234 & @xmath235 & @xmath236 & @xmath237 & @xmath238 & @xmath239 & @xmath240 + 1.602.26 & @xmath182 & @xmath241 & @xmath242 & @xmath243 & @xmath235 & @xmath244 & @xmath245 & @xmath246 & @xmath247 & @xmath248 + 2.263.20 & @xmath249 & @xmath250 & @xmath251 & @xmath252 & @xmath253 & @xmath254 & @xmath255 & @xmath256 & @xmath257 & @xmath258 + 3.204.53 & @xmath259 & @xmath260 & @xmath261 & @xmath262 & @xmath263 & @xmath264 & @xmath265 & @xmath266 & @xmath267 & @xmath268 + 4.536.40 & @xmath269 & @xmath270 & @xmath271 & @xmath272 & @xmath273 & @xmath274 & @xmath275 & @xmath266 & @xmath276 & @xmath277 + 6.409.05 & @xmath278 & @xmath279 & @xmath280 & @xmath281 & @xmath282 & @xmath283 & @xmath284 & @xmath285 & @xmath286 & @xmath287 + 9.0525.6 & @xmath288 & @xmath289 & @xmath290 & @xmath291 & @xmath292 & @xmath293 & @xmath294 & @xmath295 & @xmath296 & @xmath297 + ccccccccccc 0.100.14 & @xmath298 & @xmath299 & @xmath300 & @xmath301 & @xmath302 & @xmath303 & @xmath304 & @xmath305 & @xmath306 & @xmath307 + 0.140.20 & @xmath308 & @xmath309 & @xmath310 & @xmath311 & @xmath312 & @xmath313 & @xmath178 & @xmath314 & @xmath315 & @xmath316 + 0.200.28 & @xmath317 & @xmath318 & @xmath319 & @xmath320 & @xmath321 & @xmath322 & @xmath178 & @xmath323 & @xmath324 & @xmath325 + 0.280.40 & @xmath326 & @xmath327 & @xmath328 & @xmath329 & @xmath330 & @xmath331 & @xmath332 & @xmath333 & @xmath334 & @xmath335 + 0.400.56 & @xmath336 & @xmath337 & @xmath338 & @xmath339 & @xmath340 & @xmath341 & @xmath303 & @xmath342 & @xmath343 & @xmath344 + 0.560.80 & @xmath345 & @xmath346 & @xmath347 & @xmath348 & @xmath349 & @xmath350 & @xmath351 & @xmath309 & @xmath352 & @xmath353 + 0.801.13 & @xmath354 & @xmath355 & @xmath356 & @xmath357 & @xmath358 & @xmath359 & @xmath360 & @xmath361 & @xmath362 & @xmath363 + 1.131.60 & @xmath364 & @xmath365 & @xmath366 & @xmath234 & @xmath215 & @xmath186 & @xmath237 & @xmath367 & @xmath368 & @xmath369 + 1.602.26 & @xmath370 & @xmath371 & @xmath372 & @xmath214 & @xmath373 & @xmath186 & @xmath245 & @xmath374 & @xmath375 & @xmath258 + 2.263.20 & @xmath376 & @xmath377 & @xmath184 & @xmath339 & @xmath378 & @xmath379 & @xmath255 & @xmath380 & @xmath375 & @xmath381 + 3.204.53 & @xmath382 & @xmath383 & @xmath384 & @xmath262 & @xmath385 & @xmath386 & @xmath265 & @xmath387 & @xmath388 & @xmath389 + 4.536.40 & @xmath390 & @xmath391 & @xmath392 & @xmath393 & @xmath253 & @xmath394 & @xmath275 & @xmath395 & @xmath396 & @xmath397 + 6.409.05 & @xmath398 & @xmath399 & @xmath400 & @xmath281 & @xmath401 & @xmath165 & @xmath284 & @xmath402 & @xmath403 & @xmath269 + 9.0525.6 & @xmath404 & @xmath405 & @xmath406 & @xmath407 & @xmath408 & @xmath409 & @xmath294 & @xmath410 & @xmath411 & @xmath412 + cc 0.10 - 0.14 & 26673 + 0.14 - 0.20 & 71637 + 0.20 - 0.28 & 91336 + 0.28 - 0.40 & 93286 + 0.40 - 0.56 & 78330 + 0.56 - 0.80 & 61337 + 0.80 - 1.13 & 45386 + 1.13 - 1.60 & 30713 + 1.60 - 2.26 & 19351 + 2.26 - 3.20 & 11301 + 3.20 - 4.53 & 6426 + 4.53 - 6.40 & 3761 + 6.40 - 9.05 & 2095 + 9.05 - 25.6 & 2333 + ccccc 100 k & 114407.6 & @xmath413 & @xmath414 & @xmath415 + 125 k & 114480.1 & @xmath416 & @xmath417 & @xmath418 + 250 k & 114533.8 & @xmath419 & @xmath420 & @xmath421 + 400 k & 114544.5 & @xmath422 & @xmath423 & @xmath424 + optically - thin & 114552.8 & @xmath425 & @xmath426 & @xmath427 +

the recent observations of and scalability concerns with the dynamics of the bgp routing table size growth , @xcite , bring up the question of how small the routing table sizes for distributed routing on realistic massive graphs can be made _ in principle_. in other words , what are the _ fundamental _ limits of compactness of graph - theoretic routing in such networks ? answering this question involves two things . on one hand , it calls for assessment of results obtained in the area of distributed routing . since our first interest is the lower limits that can be achieved _ in principle _ , we are more concerned with idealized _ static _ routing in this paper . being the simplest and most fundamental routing model , static routing is where such limits can manifest themselves . in the more complicated dynamic case , these limits can only be higher . on the other hand , answering the above question also requires understanding of the basic properties of _ realistic _ massive growing networks , the internet being a good example of those . the structure and evolution of these networks are subjects of intensive studies these days . the above two research fields virtually do not intersect . we are not aware of any routing schemes designed specifically for scale - free graphs , and , vice versa , the literature concerned with the properties of scale - free nets has not addressed the routing problem yet . we tend to explain this lack of overlapping by the fact that understanding of the nature of large networks observed in reality started just recently . we review the results relevant to our work in the two separate subsections . since the pioneering work by kleinrock and kamoun , @xcite , the trade - off between stretch and the amount of routing information ( i.e. memory space ) required by a routing scheme has been understood , analyzed , and improved . many relatively recent `` internet routing architecture '' proposals , @xcite , are based on the ideas of @xcite . since @xcite was the first work of its type , it is not surprising that nave routing based on it would generate stretch that is far from optimal . indeed , the simple calculations presented in appendix [ sec : kk_stretch ] show that the stretch produced by @xcite on the present internet interdomain graph would be of the order of 10 . however , the high value of stretch was not the central problem with @xcite . this work along with the works that closely followed it , @xcite , concentrated on optimal hierarchical network clustering ( splitting into areas ) satisfying a set of assumptions , but no proof of existence of such clustering for generic graphs and no algorithm to find it when it exists were obtained . while several subsequent works tried to overcome these problems , all known hierarchical routing schemes eventually turned out to be inferior with respect to more recent _ direct _ routing schemes . in the work by peleg and upfal , @xcite , the trade - off between memory space and stretch for _ generic _ networks was rigourously analyzed for the first time . the work contained several issues that were addressed in many publications that followed @xcite . one of the issues was that only the total ( per network ) space was bounded , the local ( per node ) space was unbounded . another issue was that the scheme required relabelling of nodes . the fact that useful information about network topology can be embedded in node labels to reduce the space is most easily seen in the case of ring networks , where the local lower bound for shortest path routing is @xmath8 if relabelling is not allowed , and @xmath9 if nodes on the ring can be labelled sequentially . the most efficient label set for an @xmath10-node network is obviously @xmath11$].$ ] label set is relevant to a specific terminology used in the routing literature . the common term `` routing table '' denotes a direct routing scheme with labels from this set . ] the fundamental lower bound is obtained in @xcite . it is shown there that if the label space is @xmath11 $ ] , then , for _ any _ stretch ( including shortest path routing ) , there can not exist a loop - free generic routing scheme that would guarantee the local space less then @xmath12 . for _ shortest path _ routing , the lower bound turns out to be higher . the pessimistic but intuitively expected results obtained in @xcite show that for _ any _ stretch-1 routing scheme , there exists a graph with maximum node degree @xmath13 , @xmath14 , such that @xmath15 bits of memory are required at @xmath16 nodes . since the trivial upper bound for shortest path routing is also @xmath17 , @xcite effectively demonstrates _ incompressibility _ of generic shortest path routing . fortunately , the majority of graphs are slightly better . applying the kolmogorov complexity theory to routing , the authors of @xcite provide many upper and lower bounds for _ almost all _ graphs . in particular , for shortest path routing , not more than @xmath18 ( but not less than @xmath19 ) bits per node are shown to be enough for the @xmath20 portion of all @xmath11$]-labelled graphs . if labelling is relaxed to allow for @xmath21-sized labels , then the local memory space upper bound is reduced to @xmath21 . the question about the space lower bound for stretch-1 routing on almost all graphs with free relabelling remains open . while the results from the complexity theory or classical random graph theory ( @xcite ) for almost all graphs may induce some optimism , very little can be said , on the practical side , about if all the graphs from a given class of graphs are good or bad with respect to routing compactness . exploring specifics of various graph families , a number of compact routing schemes for special types of graphs have been constructed . there are several results for rings , complete networks , trees , grids , decomposable , planar , outplanar , bounded genus , chordal , etc . , graphs , but none so far for scale - free graphs ( even in the `` almost all '' context ) . it is easy to explain since the properties of scale - free networks from the graph - theoretic perspective are not fully understood yet . thus , the only currently existing tool to analyze the limits of compactness of routing on scale - free networks is generic routing schemes . generic shortest path routing is incompressible , which means that if the memory space is to be reduced , then the stretch must be increased . the memory space lower bound dependence on stretch is not `` continuous . '' as shown in @xcite , any generic routing scheme with the maximum stretch strictly less than @xmath22 must use at least @xmath23 bits of memory on some nodes of some graphs . in other words , the lower bound for generic schemes with stretch @xmath24 is the same as in the incompressible case of shortest path routing ( consider the bound of @xmath15 discussed above and take @xmath25 ) . furthermore , as shown in @xcite , the lower bound for schemes with stretch strictly less than @xmath26 is nearly the same as for shortest path routing@xmath8 bits of memory on some nodes of some graphs . the minimum stretch factor that allows for significant memory space lower bound decrease is 3 . cowen introduces a very simple direct stretch-3 routing scheme with the local memory space upper bound of @xmath27 in @xcite . the scheme uses relabelling , and labels are of size @xmath28 . in @xcite , thorup and zwick improve cowen s result and deliver a local space upper bound of @xmath29 . they also show how to implement routing decisions at _ constant _ time per node and to reduce the label size to @xmath30 . we call the above two stretch-3 schemes the _ cowen _ and _ thorup - zwick _ ( tz ) schemes respectively . the local memory space upper bound provided by the tz scheme is nearly ( up to a logarithmic factor ) optimal ( best possible ) since , as demonstrated in @xcite , any generic routing scheme with stretch strictly less than @xmath31 must use at least @xmath32 bits of memory on some nodes of some graphs ( see table [ table : stretch_1 - 3 - 5 ] ) . to the best of our knowledge , the tz scheme is the single generic stretch-3 routing scheme delivering a nearly optimal local memory upper bound today . this makes it `` exceptional '' in a sense that it delivers a nearly optimal first possibility to decrease the local space down from the shortest path routing incompressible limits . this also explains why it is a primary subject of our present work . .total space lower bounds for low stretch values . [ cols="^,^,^",options="header " , ] the results in the previous sections suggest that the average stretch depends more strongly on the characteristics of the graph distance distribution ( on its first and second moments , in particular ) than on the graph size . more specifically , we are taking the distance distribution in a graph to be gaussian , ( [ gaussian - pdf - norm ] ) , and , hence , the average tz stretch @xmath33 in ( [ stretch - average ] ) is a function of the average distance @xmath6 and the width of the distance distribution @xmath7 in a graph , @xmath34 . at this point , we wish to explore the analytical structure of @xmath35 in more detail . the natural starting point is to fix either @xmath6 or @xmath7 in @xmath35 to their observed values , @xmath36 and @xmath37 respectively , and to consider two functions , @xmath38 and @xmath39 , which are shown in fig . [ fig : minimum](a , b ) . to our great surprise , we discover that these two functions have _ unique minimums _ and that the points corresponding to the internet distance distribution ( or , simply , the _ internet points _ ) are _ very close _ to them . in other words , one may get an impression that the internet topology has been carefully crafted to result in a distance distribution that would minimize the average tz stretch . of course , this can be only an impression and not an explanation since the internet evolution has had nothing to do with stretch . the next question we have to ask is if the minimums we observe in fig . [ fig : minimum](a , b ) correspond to a true local minimum of @xmath35 . our analytical results allow us to collect enough data to draw fig . [ fig : minimum](c ) , where the stretch function @xmath35 is shown for @xmath40 $ ] . note that not all regions of @xmath41 correspond to gaussian - like distance distributions . indeed , when @xmath42 , distribution @xmath43 from ( [ gaussian - pdf - norm ] ) looks more like an exponential decay since it is cut off from the left by condition @xmath44 . also , when @xmath7 is very small and the continuous form of the gaussian distribution approaches a @xmath45-function , its discrete form ( [ gaussian - pdf - norm ] ) goes to either a constant , @xmath46}$ ] , when @xmath47 , or to a sum of two constants with equal weights , @xmath48 , when @xmath49 . this explains the peculiar peak formation in the @xmath50 area in the picture ( see also appendix [ sec : minimum_analysis ] ) . networks with such distance distributions are easily constructible ( complete networks , stars , various forms of their interconnections , etc . ) . they all have regular structure , and the exact knowledge of their structure is required for precise stretch calculations . this is why our analytical approach is slightly off in giving the precise answer for stars ( @xmath51 instead of 1 for @xmath52 , @xmath50 ) . note , however , that for the single case when @xmath7 is allowed to be strictly 0 , that is , for the complete network case , @xmath53 , @xmath54 , we obtain the correct answer for the average stretch , 2 . the @xmath41-region for distance distributions in realistic networks is , thus , @xmath55 , where we observe a concave area in a form of channel , fig . [ fig : minimum](c ) . the area is characterized by particularly low stretch values , which makes us call it the _ minimal stretch region _ ( the msr ) . the width and depth of the msr slowly _ increase _ as @xmath41 grow . in the area of smaller @xmath41 , the msr has a unique critical point , which we call the _ msr apex_. the internet point is located very close to the apex , which is characterized by the shortest distance between the sets of minimums of @xmath33along the @xmath6- and @xmath7-axes . we may express these sets as two functions , which we denote as @xmath56 and @xmath57 respectively . we find that @xmath58 and @xmath59 _ almost touch _ each other at the apex . since the intersection of these two functions would correspond to a stationary point of @xmath35 , we call the apex a _ quasi - stationary _ point emphasizing that the both derivatives of @xmath35 are _ nearly _ zero at this point . the apex can be more easily observed in fig . [ fig : minimum](d ) showing a projection of fig . [ fig : minimum](c ) on the @xmath6-@xmath7 plane . the solid lines representing the above two sets of minimums forming the msr , almost touch each other near the apex , and the internet point is very near their closest segment . an opportunity to look at the apex from yet another angle is presented in fig . [ fig : minimum](e ) showing a projection of fig . [ fig : minimum](c ) on the @xmath6-@xmath33 plane . we see that starting from the apex , as @xmath6 increases , the stretch values along @xmath58 and @xmath59 become virtually equal and slowly _ decrease _ as the average distance grows . we also note that @xmath60 graphs are far away from the apex and that they have average stretch values that are far from minimal . we can see now that the apex is indeed a critical or `` phase transition '' point since it is located at the boundary of the two regions of the average stretch function . the first region , the msr , is characterized by lowest possible stretch values corresponding to distance distributions observed in real - world graphs . the second region , with substantially higher average stretch values , corresponds to distance distributions in more regular graphs . to illustrate this point in more detail , we turn our attention back to fig . [ fig : minimum](d ) . we see that the two sets of minimums , @xmath58 and @xmath59 , are linear in the msr with sufficiently large @xmath7 , @xmath61 , where the continuous approximation of the distance distribution works particularly well . the two top dashed lines in fig . [ fig : minimum](d ) represent the linear fits of @xmath58 and @xmath59 in the area with @xmath61 . the exact location of the intersection of these fits is @xmath62 , while the two closest points on the data curves for @xmath58 and @xmath59 are @xmath63 and @xmath64 . if the linear form of @xmath58 and @xmath59 sustained in the area with smaller @xmath7 as well , then @xmath58 and @xmath59 would intersect at @xmath65 , where we would observe a true stationary point of @xmath35 , which we could then test for the presence of an extremum of the stretch function . this does not happen , however . instead , as @xmath6 and @xmath7 become small , the linear behavior breaks near the apex due to increasingly `` more discrete '' structure of the distance distribution . see more on this in appendix [ sec : minimum_analysis ] . in appendix [ sec : minimum_analysis ] , we show that linearity of @xmath58 and @xmath59 can be analytically derived from the fact that the distance distribution is taken to be gaussian . of course , this does not explain why the _ internet _ point is so close either to the msr or to its apex . the linear form of @xmath58 and @xmath59 sheds some light on a closely related issue of why the average stretch is virtually independent of @xmath1 . in fig . [ fig : minimum](d ) , the shaded area represents a set of @xmath41 , for which the average stretch is approximately the same as for the internet , @xmath66 . in other words , it is a projection of the cyan area in fig . [ fig : minimum](c ) on the @xmath6-@xmath7 plane . we see that in the msr , the @xmath67 boundaries are almost parallel straight lines . therefore , if the average stretch is to be independent of @xmath1 , which is observed in section [ sec : stretch_distribution ] , then the points representing distance distributions in power - law graphs , @xmath68 , from fig . [ fig : dd](b ) should lie along the @xmath67 boundaries , and this is what indeed happens . yet again , the linear relation between @xmath69 and @xmath70 in the power - law graphs , and the fact that this relation is just as required for the average tz stretch being virtually independent of @xmath1 , come from two seemingly disjoint domains . to finish the list of various `` coincidences , '' we construct a linear fit of ( @xmath71 ) ( the bottom - most dashed line in fig . [ fig : minimum](d ) ) . the internet point , @xmath72 , lies on this line . our numeric analysis shows that the internet value of @xmath72 is a unique value of @xmath1 minimizing the distance between the linear fit of ( @xmath71 ) and @xmath65 , which is the intersection of the linear fits of @xmath58 and @xmath59 . in other words , the internet distance distribution is the point that is _ closest _ to the msr apex , compared to distance distributions in all other scale - free graphs with power - law node degree distributions . of course , the tz scheme , reducing , in principle , the routing table size to about @xmath4 entries for @xmath5-node scale - free networks , and making routing decision running time constant , can not pretend to be a realistic internet interdomain routing scheme . first , it is static . second , addressing in interdomain routing is based on ip addresses rather than on interdomain graph node labels , that is , as numbers . third , it assumes availability of the global topology view . most importantly , in the context of our work , the scheme is not a stretch-1 scheme . indeed , interdomain routing in the internet is essentially shortest path routing . a routing scheme that would prevent a pair of ass from utilizing a peering link between them is not realistic , of course . thus , any stretch @xmath73 routing scheme applied to the internet would involve augmentation , in one form or another , of the routing information provided by the scheme with the shortest path routing information for non - shortest paths . this explains why we are concerned with the average stretch produced by a scheme . our principal finding that the average tz stretch on the internet graph is reasonably low opens a well - defined path for the future work in the area of applying relevant theoretical results obtained for routing to realistic scale - free networks ( see the next section ) . if the average stretch of even the `` exceptional '' tz scheme turned out to be relatively high , the scheme would be inapplicable _ in principle _ ( not just _ in practice _ ) , which would essentially close the above path , demonstrate impossibility to construct efficient and scalable routing for the internet , and call for searching one somewhere beyond the traditional graph - theoretical approach . as we mention in the introduction , our finding that the internet distance distribution is in a close neighborhood of the msr apex can not be explained in the present idea set since the internet , as we know it today , has nothing to do with stretch . while we lack sufficient information to show cause for this effect , we do believe it strongly suggests the analytical structure of the average stretch function may be an indirect ( or even direct ) indicator of some yet - to - be discovered processes that have influenced the internet s topological evolution . in other words , a rigorous explanation of this phenomenon would probably require much deeper understanding of the internet evolution principles ( that are far from being even known if we accept the critique of the ba model and alike ) and demonstration of a link between them and the tz scheme . we believe that an explanation of this effect will most probably have the following pattern . the internet evolution principles turn out to be such that they minimize @xmath74 , where @xmath74 is some known or yet unknown characteristic of a network . at the same time , the distance to the msr apex turns out to be a monotonically increasing function of @xmath74 . there are reasons to believe that the distance to the apex is not something random . indeed , as we have seen , the apex is a _ unique _ critical point of the average tz stretch function , and , at the same time , the tz scheme is also `` exceptional '' in a sense that it delivers a nearly optimal first possibility to deviate from incompressible shortest path routing . the outlined pattern would create a necessary link between the two seemingly disjoint domains . the list of immediate practical next questions that remain open include : * what is the average stretch produced by the _ scheme on scale - free graphs ? + obtaining analytical results for the cowen scheme is a harder problem . however , one can expect that the average stretch would be lower . indeed , it is easy to see that making high - betweenness nodes belong to the ls should decrease the average stretch . the cowen ls construction algorithm uses the greedy set cover algorithm by lovsz preferring high degree nodes , @xcite , but as observed in @xcite , betweenness is linearly correlated with node degrees in the internet . the expectation of the cowen scheme producing an average stretch that is lower than in the tz case is in a good agreement with the general stretch - space trade - off since the cowen scheme has a higher local memory space upper bound . * do any routing schemes based on the _ additive _ stretch factor deliver lower average stretch on scale - free graphs ? + the multiplicative stretch factor may be too coarse for short distances that prevail in scale - free graphs . * what is the memory space lower bound for _ shortest path _ routing on scale - free graphs ? + we know that generic shortest path routing is incompressible . however , the situation is better for almost all graphs . we do not know any bounds for scale - free graphs . answering this question has very important practical implications since ( policy - constrained ) shortest path routing seems to be a requirement for internet interdomain routing . * more specifically , can any upper bounds be obtained for the total number of stretch @xmath73 paths produced by existing @xmath73 routing schemes on scale - free graphs ? + if these upper bounds are found to be as low as the total space upper bounds , then the original routing information can be augmented with the @xmath75 information for @xmath73 paths without increasing the space upper bounds . * what bounds can be obtained for _ dynamic _ routing on scale - free graphs ? + of course , no realistic internet routing can neglect the adaptation cost considerations . while of critical practical importance , obtaining various bounds for dynamic routing on scale - free graphs seems to be the hardest problem among those we list here . furthermore , one can hardly expect such bounds to be low , taking into consideration rather pessimistic lower bounds obtained for dynamic routing on generic networks ( cf . section [ sec : routing ] ) . in addition , it is clear that the tz scheme can not be easily modified to perform well in the dynamic case since the scheme labels nodes with topology - sensitive information . in other words , the scheme is not name - independent . as soon as topology changes , nodes need to be relabelled . significant progress in construction of name - independent static low - stretch routing schemes has been recently made by arias , cowen , et al . in @xcite . on the theoretical side , which may turn out to have practical implications as well , the explanation of internet distance distribution proximity to the msr apex appears to be a very interesting problem . it would be much easier to solve , of course , if the integral in ( [ stretch - superlong ] ) could be evaluated analytically . what might be an alternative set of assumptions that would make an expression analogous to ( [ stretch - superlong ] ) analytically solvable ? can similar results be obtained for other forms of distance distributions , and , yet more importantly , for other routing schemes ( the cowen scheme , or the stretch-5 routing scheme with the @xmath76 local memory space upper bound obtained in @xcite , for example ) ? a technical issue with _ experimental _ ( vs. analytical ) studies of the msr is that we do not know an efficient algorithm to generate graphs with a given distance distribution . in any case , the explanation of internet s proximity to the apex is of great theoretical interest , as the fundamental laws governing the internet evolution remain unclear . therefore , on the practical side , a proper explanation of this effect may help us , for example , in out intent to move , @xcite , from purely _ descriptive _ internet evolution models to more _ explanatory _ ones , in the terminology of the program outlined by the authors of @xcite . the internet started as a small research network designed and fully controlled by a group of few enthusiasts ( @xcite ) . today , after a series of `` phase transitions , '' it has evolved to a huge network interconnecting tens of thousands of independent and even adversarial networks without a single point of external control . this makes the internet a `` self - governing , '' `` self - evolving '' complex system , a research subject of areas of physics ( statistical mechanics , in particular ) studying evolution of such systems in general . construction of realistic , efficient , and scalable routing for this `` new '' internet is an interesting and challenging task lying ahead of us . 10 y. afek , e. gafni , and m. ricklin . upper and lower bounds for routing schemes in dynamic networks ( abstract ) . in _ proceedings of the 30@xmath77 annual ieee symposium on foundations of computer science ( focs ) _ , pages 370375 . ieee , 1989 . w. aiello , f. chung , and l. lu . a random graph model for massive graphs . in _ proceedings of the 32@xmath78 annual acm symposium on theory of computing ( stoc ) _ , pages 171180 . acm press , 2000 . d. aingworth , c. chekuri , p. indyk , and r. motwani . fast estimation of diameter and shortest paths ( without matrix multiplication ) . , 28(4):11671181 , 1999 . r. albert and a .- statistical mechanics of complex networks . , 74:4797 , 2002 . r. albert , h. jeong , and a .- the diameter of the world wide web . , 401:130131 , 1999 . l. a. n. amaral , a. scala , m. barthlmy , and h. e. stanley . classes of small - world networks . , 97:1114911152 , 2000 . m. arias , l. cowen , k. a. laing , r. rajaraman , and o. taka . compact routing with name independence . in _ proceedings of the 15@xmath77 annual acm symposium on parallel algorithms and architectures ( spaa ) _ , pages 184192 . acm , 2003 . barabsi and r. albert . emergence of scaling in random networks . , 286:509512 , 1999 . , http://bgp.potaroo.net/. b. bollobs . . academic press , new york , 1985 . a. broido and k claffy . analysis of routeviews bgp data : policy atoms . in _ proceedings of network resource data management workshop ( nrdm ) _ , 2001 . a. broido and k claffy . complexity of global routing policies . in _ cooperative association for internet data analysis ( caida ) _ , 2001 . . a. broido and k claffy . internet topology : connectivity of ip graphs . in _ spie international symposium on convergence of it and communication _ , 2001 . a. broido , e. nemeth , and k claffy . internet expansion , refinement , and churn . , 13(1):3351 , 2002 . h. buhrman , j .- h . hoepman , and p. vitny . space - efficient routing tables for almost all networks and the incompressibility method . , 28(4):14141432 , 1999 . i. castineyra , n. chiappa , and m. steenstrup . . ietf , rfc 1992 , 1996 . h. chang , s. jamin , and w. willinger . internet connectivity at the as level : an optimization driven modeling approach . in _ proceedings of mometools _ , 2003 . q. chen , h. chang , r. govindan , s. jamin , s. j. shenker , and w. willinger . the origin of power laws in internet topologies revisited . in _ ieee infocom _ , f. chung and l. lu . the average distance in a random graph with given expected degrees . , 1(1):91114 , 2003 . r. cohen , d. dolev , s. havlin , t. kalisky , o. mokryn , and y. shavitt . on the tomography of networks and multicast trees . technical report tr49 , hebrew university , 2002 . r. cohen and s. havlin . scale - free networks are ultrasmall . , 90(05):058701 , 2003 . l. cowen . compact routing with minimum stretch . , 38(1):170183 , 2001 . s. dolev , e. kranakis , d.krizanc , and d. peleg . bubbles : adaptive routing scheme for high - speed dynamic networks . , 29(3):804833 , 1999 . d. dor , s. halperin , and u. zwick . all pairs almost shortest paths . , 29:17401759 , 2000 . a. doria , e. davies , and f. kastenholz . . irtf , internet draft , 2003 . s. n. dorogovtsev . private communications . s. n. dorogovtsev , a. v. goltsev , and j. f. f. mendes . pseudofractal scale - free web . , 65(06):066122 , 2002 . s. n. dorogovtsev and j. f. f. mendes . language as an evolving word web . , 268:26032606 , 2001 . s. n. dorogovtsev and j. f. f. mendes . evolution of networks . , 51:10791187 , 2002 . s. n. dorogovtsev and j. f. f. mendes . . oxford university press , oxford , 2003 . s. n. dorogovtsev , j. f. f. mendes , and a. n. samukhin . metric structure of random networks . , 653(3):307422 , 2003 . y. dourisboure and c. gavoille . improved compact routing scheme for chordal graphs . in _ proceedings of the 16@xmath77 international conference on distributed computing ( disc ) _ , volume 2508 of _ lecture notes in computer science _ , pages 252264 . springer , 2002 . t. eilam , c. gavoille , and d. peleg . compact routing schemes with low stretch factor . , 46:97114 , 2003 . m. elkin and d. peleg . @xmath79-spanner constructions for general graphs . in _ proceedings of the 33@xmath80 annual acm symposium on theory of computing ( stoc ) _ , pages 173182 . acm , 2001 . p. erds and a. rni . on random graphs . , 6:290297 , 1959 . a. fabrikant , e. koutsoupias , and c. h. papadimitriou . heuristically optimized trade - offs : a new paradigm for power laws in the internet . in _ proceeding of the 29@xmath77 international colloquium on automata , languages , and programming ( icalp ) _ , volume 2380 of _ lecture notes in computer science _ , pages 110122 . springer , 2002 . m. faloutsos , p. faloutsos , and c. faloutsos . on power - law relationships of the internet topology . in _ proceedings of the acm sigcomm _ , pages 251262 , 1999 . f. fich . end - to - end communication . in _ proceedings of the 2@xmath78 international conference on principles of distributed systems ( opodis ) _ , pages 3743 . hermes , 1998 . m. flammini , j. van leeuwen , and a. marchetti - spaccamela . the complexity of interval routing on random graphs . , 41(1):1625 , 1998 . p. fraigniaud and c. gavoille . lowers bounds for oblivious single - message end - to - end communication . research report rr-1289 - 03 , labri , university of bordeaux , 2003 . a. fronczak , p. fronczak , and j. a. hoyst . average path length in random networks . . c. gavoille . a survey on interval routing . , 245(2):217253 , 2000 . c. gavoille . routing in distributed networks : overview and open problems . , 32(1):3652 , 2001 . c. gavoille and m. genegler . space - efficiency for routing schemes of stretch factor three . , 61(5):679687 , 2001 . c. gavoille and m. nehz . interval routing in reliability networks . in _ proceeding of the 9@xmath77 international colloquium on structural information and communication complexity ( sirocco)_. carleton university press , 2003 . to appear . c. gavoille and d. peleg . the compactness of interval routing for almost all graphs . in _ proceedings of the 12@xmath77 international symposium on distributed computing ( disc ) _ , volume 1499 of _ lecture notes in computer science _ , pages 161174 . springer , 1998 . c. gavoille and s. prenns . memory requirement for routing in distributed networks . in _ proceedings of the 15@xmath77 annual acm symposium on principles of distributed computing ( podc ) _ , pages 125133 . acm , 1996 . i. s. gradstein and i. m. ryzhik . . academic press , san diego , sixth edition , 2000 . j. guare . . vintage books , new york , 1990 . g. huston . analyzing the internet s bgp routing table . , 4(1 ) , 2001 . g. huston . . ietf , rfc 3221 , 2001 . g. huston . scaling inter - domain routing a view forward . , 4(4 ) , 2001 . h. jeong , b. tombor , r. albert , z. n. oltvai , and a .- the large - scale organization of metabolic networks . , 407:651654 , 2000 . f. kastenholz . . irtf , internet draft , 2002 . l. kleinrock and f. kamoun . hierarchical routing for large networks : performance evaluation and optimization . , 1:155174 , 1977 . l. kleinrock and f. kamoun . optimal clustering structures for hierarchical topological design of large computer networks . , 10:221248 , 1980 . d. krioukov and d. volchenkov . an economy - based model for the internet evolution , 2003 . to appear . d. krizanc , f.l . luccio , and r. raman . compact routing schemes for dynamic ring networks . technical report 2002/26 , department of mathematics and computer science , university of leichester , 2002 . a. krzywicki . defining statistical ensembles of random networks . in _ workshop on discrete random geometries and quantum gravity _ , utrecht , 2001 . . e. kushilevitz , r. ostrovsky , and a. rosn . log - space polynomial end - to - end communication . , 27(6):15311549 , 1998 . a. lakhina , j. byers , m. crovella , and p. xie . sampling biases in ip topology measurements . in _ proceedings of ieee infocom _ , 2003 . b. leiner , v. cerf , d. clark , r. kahn , l. kleinrock , d. lynch , j. postel , l. roberts , and s. wolff . a brief history of the internet . , 40(2):102108 , 1997 . f. liljeros , c. r. edling , l. a. n. amaral , h. e. stanley , and y. aberg . the web of human sexual contacts . , 411:907908 , 2001 . l. lovsz . on the ratio of optimal integral and fractional covers . , 13:383390 , 1975 . b. mandelbrot . an informational theory of the statistical structure of languages . in _ communication theory _ , pages 486502 . betterworth , 1953 . s. milgram . the small world problem . , 61 , 1967 . m. molloy and b. reed . a critical point for random graphs with a given degree sequence . , 6:161179 , 1995 . m. molloy and b. reed . the size of the giant component of a random graph with a given degree sequence . , 7:295305 , 1998 . j. m. montoya and r. v. sole . small world patterns in food webs . technical report 00 - 10 - 059 , santa fe institute , 2000 . m. e. j. newman . the structure of scientific collaboration networks . 98:404409 , 2001 . m. e. j. newman . random graphs as models of networks . in _ handbook of graphs and networks _ , berlin , 2003 . wiley - vch . m. e. j. newman , s. h. strogatz , and d. j. watts . random graphs with arbitrary degree distributions and their applications . , 64(02):026118 , 2001 . m. e. j. newman , d. j. watts , and s. h. strogatz . random graph models of social networks . , 99:25662572 , 2002 . w. b. norton . and peering . equinix white paper . d. peleg . . siam , philadelphia , pa , 2000 . d. peleg and e. upfal . a trade - off between space and efficiency for routing tables . , 36:510530 , 1989 . r. perlman . hierarchical networks and subnetwork partition problem . , 9:297303 , 1985 . e. ravasz and a .- hierarchical organization in complex networks . , 67(02):026112 , 2003 . e. ravasz , a. l. somera , d. a. mongru , z. n. oltvai , and a .- hierarchical organization of modularity in metabolic networks . , 297:15511555 , 2002 . s. ross . . academic press , san diego , seventh edition , 2000 . , http://www.nd.edu/~networks/. g. siganos , m. faloutsos , p. faloutsos , and c. faloutsos . power - laws and the as - level internet topology . in _ acm / ieee transactions on networking _ , 2003 . to appear . l. subramanian , s. agarwal , j. rexford , and r. h. katz . characterizing the internet hierarchy from multiple vantage points . in _ proceeding of ieee infocom _ , 2002 . h. tangmunarunkit , j. doyle , r. govindan , s. jamin , w. willinger , and s. shenker . does as size determine as degree ? , october 2001 . h. tangmunarunkit , r. govindan , s. jamin , s. shenker , and w. willinger . network topology generators : degree - based vs. structural . in _ proceedings of the 2002 conference on applications , technologies , architectures , and protocols for computer communications ( sigcomm ) _ , pages 147159 . acm press , 2002 . h. tangmunarunkit , r. govindan , and s. shenker . internet path inflation due to policy routing . in _ spie itcom _ , 2001 . m. thorup and u. zwick . approximate distance oracles . in _ proceedings of the 33@xmath80 annual acm symposium on theory of computing ( stoc ) _ , pages 183192 . acm , 2001 . m. thorup and u. zwick . compact routing schemes . in _ proceedings of the 13@xmath77 annual acm symposium on parallel algorithms and architectures ( spaa ) _ , pages 110 . acm , 2001 . a. vzquez , r. pastor - satorras , and a. vespignani . internet topology at the router and autonomous system level . . a. vzquez , r. pastor - satorras , and a. vespignani . large - scale topological and dynamical properties of the internet . , 65(06):066130 , 2002 . p. verkaik , a. broido , and k claffy . atomized routing . in _ fourth international workshop on the design of reliable communication networks ( drcn ) _ , 2003 . submitted . d. j. watts . . princeton university press , princeton , nj , 1999 . d. j. watts and s. h. strogatz . collective dynamics of `` small - world '' networks . , 393:440442 , 1998 . w. willinger , r. govindan , s. jamin , v. paxson , and s. shenker . scaling phenomena in the internet : critically examining criticality . , 99(suppl . 1):25732580 , 2002 . * appendices * in this appendix , we calculate a rough estimate of the stretch factor of the kleinrock - kamoun ( kk ) hierarchical routing scheme , @xcite , applied to the observed internet interdomain topology . we find that the stretch is very high , which is consistent with the observation made in @xcite that the approach used there works reasonably well only for sparsely connected networks . the scale - free networks , on the contrary , are extremely densely connected . recall that @xcite assumes existence of a hierarchical partitioning of a network of size @xmath10 into @xmath81 levels of clusters . each @xmath82-level cluster consists of @xmath83 @xmath84-level clusters , @xmath85 , @xmath86-level clusters being nodes . the optimal clustering is achieved when @xmath87 . there are few other fairly strong assumptions about the properties of required partitioning . neither algorithm for its construction nor proof of its existence are delivered , but if it does exist then the stretch factor is shown to be @xmath88d_k,\ ] ] where @xmath6 is the network average distance and @xmath89 is the diameter of a @xmath82-level cluster . it is further assumed in @xcite that both the network diameter and average distance are power - law functions of the network size . this is certainly not true for scale - free networks with power - law _ node degree _ distributions . for the very recent results on the average distance in such networks see the references mentioned in section [ sec : scale - free_networks ] . in the numerical evaluations in this appendix , we use the value of @xmath90 observed in the internet . much fewer results are available for the network diameter . however , in @xcite , it is shown that the diameter of networks with power - law node degree distribution with exponent @xmath1 lying between @xmath91 and @xmath26 scales almost surely as @xmath92 . for the internet , @xmath2 , and since the internet size @xmath93 is relatively large , we may write the internet diameter @xmath94 as @xmath95 with some multiplicative coefficient @xmath96 . the observed value of @xmath94 , @xmath97 ( @xcite ) , defines @xmath96 then . the size of a @xmath82-level cluster is obviously @xmath98 but nothing rigorous can be said about its degree distribution since there is no procedure for its construction . thus , it is natural to assume that its degree distribution is also power - law with @xmath99 , which gives an estimate of the @xmath82-level cluster diameter as @xmath100 . substituting this in ( [ kk_stretch : input ] ) and performing summation gives @xmath101 .\ ] ] using the numerical values for @xmath10 , @xmath6 , @xmath94 , and optimal @xmath102 , we can see that the kk stretch factor on the internet interdomain topology is @xmath103 note that a 15-time path length increase in the internet would lead to as path lengths of @xmath104 and ip hop path lengths of @xmath105 . stretch factors for smaller , non - optimal values of @xmath81 are shown in fig . [ fig : kk_stretch ] . the fact that the stretch grows almost linearly with the number of hierarchical levels follows directly from ( [ kk_stretch : exact_output ] ) , which , for large @xmath10 , can be rewritten as @xmath106 using @xmath107 , @xmath108 ( @xcite ) , and optimal @xmath87 , we obtain the following estimate of the stretch factor as a function of the network size : @xmath109 in this appendix , we prove various statements made in section [ sec : analytical_results ] . a rigorous proof would involve the same type of argument as used to show that the hypergeometric distribution converges to the binomial distribution for small selections from large sets . we provide a close outline of the proof . suppose we have a set of @xmath10 objects of two types : @xmath110 objects of type 1 and @xmath111 objects of type 2 , @xmath112 and @xmath113 . recall that the hypergeometric distribution , @xmath114 gives the probability that a random sample of @xmath81 objects from the set of all objects contains @xmath115 objects of type 1 . if sampling is with replacement that is , as soon as one object is selected , it is immediately returned back to the set before the next object is selected , then the probability that @xmath115 type-1 objects have been selected after @xmath81 one - object selections is given by the binomial distribution , @xmath116 one can easily see that probabilities for small selections with and without replacement converge in the limit of large @xmath10 , @xmath117 { } p_b(x).\ ] ] suppose now that @xmath43 and @xmath118 are distance p.d.f . and c.d.f . of some graph of size @xmath10 , diameter @xmath94 , and that the graph ls @xmath119 is of size @xmath120 . the problem of finding the probability @xmath121 that the @xmath122th closest landmark is at distance @xmath13 is equivalent to the following object selection problem . the total number of objects to select from is @xmath10 , the number of types of objects is @xmath94 , and the number of objects of type @xmath13 , @xmath123 , is @xmath124 . in addition , @xmath120 random objects are marked as landmarks . since landmarks are randomly marked , @xmath121 is the probability that a randomly selected object is of type @xmath13 , which is @xmath43 by definition , times the probability that @xmath125 random objects are of ( possibly different ) types @xmath126 ( corresponding to the closer landmarks ) , times the probability that @xmath127 random objects are of ( possibly different ) types @xmath128 ( corresponding to the farther landmarks ) . that is , denoting the latter two probabilities by @xmath129 and @xmath130 respectively , we can write that @xmath131 with some normalization coefficient @xmath132 . the probability than one random object is of type @xmath133 ( @xmath134 ) is @xmath118 ( @xmath135 ) by definition , but obtaining the exact answer for @xmath129 ( @xmath130 ) involves some combinatorics resulting in a generalization of the hypergeometric distribution ( [ hypergeometric ] ) , which is left as an exercise for a rigorous reader . however , since @xmath136{}0 $ ] , sampling without replacement can be approximated by sampling with replacement , cf . ( [ hyper - to - binom ] ) , which leads to the following approximations for @xmath129 and @xmath130 : @xmath137 the analogy to the order statistics becomes straightforward now . the normalization coefficient @xmath132 in ( [ gip - p+ ] ) is defined by the normalization condition @xmath138 which can be expressed as the following lebesgue - stieltjes integral : @xmath139 in the continuous limit , evaluation of the corresponding riemann integral gives @xmath140^{-1}.\ ] ] if we denote by @xmath141 the c.c.d.f . for the distance to the closest landmark , @xmath142 then according to the cluster definition ( [ cluster_def ] ) , the average cluster size is @xmath143 using the explicit expression for @xmath144 in ( [ gi ] ) , we can calculate an upper bound for @xmath145 by representing the sum in ( [ g1-sum ] ) as the following lebesgue - stieltjes integral : @xmath146 evaluation of the corresponding riemann integral gives @xmath147 where equality is attained in the continuous limit . note that the obtained upper bound , @xmath148 , is equivalent to the c.c.d.f.for the geometric distribution , @xmath149 , with the success probability @xmath150 and number of trials @xmath151 . the reason for that becomes transparent after formulating the problem in the object selection framework considered in appendix [ sec : proofs : gi ] . substituting ( [ g1-final ] ) in ( [ average - c - sum ] ) completes the proof : @xmath152 note that clusters are originally defined as objects `` inverse '' to balls . they are `` inverse '' in the following sense : . if we denote the c.d.f . of @xmath153 by @xmath154 , then the average ball size is @xmath155 since ( [ average - c - sum ] ) can be written as @xmath156 the known fact that the average ball and cluster sizes are equal is just a consequence of integration by parts of the following lebesgue - stieltjes integral : @xmath157 note that the first case in ( [ stretchval - shortcut ] ) , @xmath158 , accounts exactly for stretch-1 paths from @xmath159 to destinations @xmath160 , ( see definition ( [ cluster_def ] ) ) . the second case , @xmath161 , accounts for the degenerate shortcut on paths from @xmath159 to @xmath162 that go through @xmath163 . on such paths , @xmath164 , the length of the shortcut portion of the total path from @xmath159 to @xmath163 is zero , and the stretch is 1 . it is easy to see that , on small - world graphs , paths from @xmath159 to @xmath162 that do not go through @xmath163 but that still have shortcuts are rare and hard to account for without knowing the graph topology . indeed , consider triangle @xmath165 in fig.[fig : shortcut - triangle ] , where @xmath166 represents source node @xmath159 , @xmath167 is destination @xmath163 , and @xmath119 is @xmath162 . shortcutting occurs when destination @xmath167 is `` in front of '' its landmark @xmath119 , which is represented by placing @xmath167 to the left of the vertical line . it is clear that on small - world graphs , characterized by low average distances and narrow distance distributions , the possibility for @xmath167 to _ not _ lie on path @xmath168 is minuscule . to see this , suppose we still want to account for cases when @xmath167 does not belong to @xmath168 by approximating such cases within the @xmath91-dimensional euclidean space . the shortcut path is then @xmath169 in fig . [ fig : shortcut - triangle ] , with @xmath170 representing the first node @xmath171 on the path from @xmath159 to @xmath162 such that @xmath172 . its position on @xmath168 is defined by @xmath173 , and the length of segment @xmath174 is denoted by @xmath175 . the stretch is then given by @xmath176 , and solving @xmath177 , we find that @xmath178 . our numerical experiments show that it does not matter if , with condition @xmath161 , we use @xmath179 or @xmath180 : the two stretch distributions are virtually identical and the average stretch difference is only in the fourth digit . this means that @xmath181 is virtually always @xmath182 as soon as @xmath161 . from @xmath183/z = 1 $ ] follows @xmath184 , which means that @xmath167 lies on @xmath168 . note , however , that condition @xmath161 is not the exact condition for the shortcut presence in the euclidean space . as mentioned above , the exact condition is that @xmath167 is to the left of the vertical line , which implies @xmath185 . in our experiments , we observe that using this condition instead of @xmath161 leads to the stretch distribution drastically distinct from the one observed in simulations . this is because any estimates based on approximation of the finite metric space in a graph by the euclidean metric space are not applicable to small - world graphs . such estimates are more applicable to grid - like graphs with wide distance distributions and average distances growing as power - law functions of the graph size . suppose the distance matrix @xmath186 is given . let us denote by @xmath187 the set of elements in @xmath186 equal to @xmath188 ( @xmath188-elements ) , @xmath189 note that the distance p.d.f . is equal to the distribution of values of elements in @xmath186 , @xmath190 we can also define the set of elements satisfying the triangle inequality , @xmath191 the distribution of @xmath188-elements in @xmath192 is @xmath193 suppose that a group of @xmath82 nodes is randomly selected in the graph , and that their indices are @xmath194 . the distribution of distances between them is also the distribution of values of elements in a @xmath195 submatrix @xmath196 obtained from @xmath186 by intersecting the @xmath197 rows and columns . since the selection is random , this distribution is also @xmath198 . if _ two _ groups of random nodes , @xmath199 and @xmath200 , are selected and we are looking for the distribution of distances between pairs of nodes belonging to the _ different _ groups , we construct a set @xmath201 by intersecting the @xmath199 rows of @xmath186 with the @xmath200 columns and _ vice versa_. this subset of @xmath186 is no longer a submatrix of @xmath186 , but the distribution of @xmath188-elements in it is still @xmath198 . our problem is to find the conditional distribution @xmath202 for distance @xmath203 between a group of @xmath204 ( on average ) random nodes at distance @xmath115 from some random node ( the landmark closest to the destination ) and another group of @xmath205 ( on average ) nodes at distance @xmath206 from the same node . the groups may overlap when @xmath207 . they define the subset @xmath208 as in the previous paragraph . the distribution we are looking for , @xmath202 , is the distribution of values of elements in this subset . since there are no correlations in the distance matrix , the only difference between this subset and truly random subsets @xmath209 from the previous paragraph is due to the triangle inequality caused by the fact that the two groups of nodes are neighbors of the same node . this means that @xmath210 is also a subset of @xmath211 defined in ( [ dabc - def ] ) , and , since @xmath210 is random in other respects , the distribution of values of its elements is the same as in @xmath211 . in other words , @xmath212 in ( [ dabc - distrib ] ) . noticing that , by the formula for the conditional probability , @xmath213 completes the proof . in this appendix we demonstrate that linearity of functions @xmath214 in the msr follows from the gaussian form of the distance distribution . we consider the continuous case , for which @xmath35 is given by ( [ stretch - superlong ] ) . the necessary condition for a local minimum of @xmath33 in the @xmath6- or @xmath7-direction is @xmath215 or @xmath216 respectively . after some algebra using @xmath217 , we obtain , with @xmath115 , @xmath206 , and @xmath203 having the same semantics as in section [ sec : analytical_results ] , @xmath218 these can be significantly simplified by approximating variables @xmath115 , @xmath206 , and @xmath203 by their means ( the `` mean field '' approximation ) : @xmath219 , @xmath220 , and @xmath221 . note that @xmath222 and @xmath223 where @xmath224 is , in fact , a function of @xmath6 and @xmath7 , @xmath225 , we can reduce ( [ ds - dd - long ] ) and ( [ ds - ds - long ] ) to , respectively , @xmath226 the direct solution of ( [ ds - ds - medium ] ) would involve resolving function @xmath231 first . however , a simpler way is to use an asymptotic form of the error function in the last term of ( [ ds - ds - medium ] ) , @xmath232 which is valid for sufficiently large @xmath7 , @xmath233 . this reduces ( [ ds - ds - medium ] ) to @xmath234 + 1 = 0,\ ] ] which has a unique solution @xmath235 . to see that @xmath214 are linear , we just need to check that @xmath231 is a linear function of its arguments in the msr . indeed , @xmath236^{q-1 } { \mathrm{d}}y.\ ] ] changing variables , @xmath237 , and using the asymptotic form of the error function , ( [ erf - asympt ] ) , we see that @xmath238 substituting this into ( [ xi - def ] ) , we find that @xmath239 numerical evaluations of @xmath240 yield @xmath241 and @xmath242 . we have a good match between @xmath243 and the data fit ( cf.fig . [ fig : minimum](d ) ) , @xmath244 . the match between @xmath245 and the data fit , @xmath246 , is worse , which suggests that the source of error is mostly in ( [ erf - asympt ] ) . we do not obtain non - zero values for the additive coefficients , @xmath247 , in @xmath248 that define the apex location . thus , we conclude that a more accurate analysis of the essentially discrete case with small @xmath7 is required to analytically obtain the apex location . note , however , that equations ( [ ds - dd - short ] ) and ( [ ds - ds - medium ] ) are consistent with the observed analytical structure of @xmath35 even for small @xmath7 . indeed , the solutions of the _ system _ of equations ( [ ds - dd - short ] ) and ( [ ds - ds - medium ] ) ( corresponding to the true stationary points of @xmath33 , @xmath249 ) exist only for @xmath250 , when the last term of ( [ ds - ds - medium ] ) goes away . then we have from ( [ xi - def ] ) with @xmath251 that @xmath252 as expected , and _ any _ @xmath6 delivers a solution . as discussed in section [ sec : minimum ] , the actual average distance can be only an integer @xmath82 or @xmath253 when @xmath250 . thus , the flat - topped peaks and narrow cracks we observe in fig . [ fig : minimum](c ) at @xmath54 are consistent with @xmath249 there . since @xmath254 only when @xmath250 , we have also effectively demonstrated , at least with the approximations we have made to obtain ( [ ds - dd - short ] ) and ( [ ds - ds - medium ] ) , that the apex _ can not _ be a _ true _ stationary point of @xmath33 . the ls size @xmath120 is a function of @xmath10 , and , hence , solutions @xmath255 are also functions of the graph size . they are shown in fig . [ fig : xistar ] . we see that @xmath256 . this sheds some light on the analytical structure of @xmath33 as a function of @xmath10 . we can see , for example , that as the network grows , the msr becomes narrower and closer to @xmath54 . since for scale - free nets , @xmath7 does approach @xmath86 as @xmath257 ( @xcite ) , we conclude that , independent of their size , the scale - free graphs are characterized by the lowest possible average stretch values .

landau fermi - liquid theory has seen considerable success in describing a wide variety of fermionic many - particle systems , such as liquid helium-3 , electrons in metals , nuclei and nuclear matter @xcite . it gives an effective description of the low - lying elementary excitations ( quasiparticles ) at low temperatures , i.e. , at a temperature @xmath1 such that the system can be considered degenerate ( @xmath2 , where @xmath3 denotes the chemical potential ) but still in the normal phase above any symmetry breaking phase transition ( @xmath4 ) , for example , to a magnetic or superconducting phase . nevertheless , the theory is not only important for the description of the above `` normal '' ( fermi - liquid ) phase , but is also vital to correctly describe the emergence of a possible ordered phase @xcite . indeed , according to the bardeen - cooper - schrieffer ( bcs ) theory the onset of superconductivity is to be viewed as an instability of the fermi liquid under an attractive interaction , which results in the formation of cooper pairs @xcite . the same is true for magnetism in the case of repulsive interactions . in particular , bcs theory implies that the high - density and low - temperature region of the phase diagram of quantum chromodynamics ( qcd ) contains a color superconducting phase of quarks @xcite . in fact , since quarks carry color , flavor and spin quantum numbers , many distinct superconducting phases are possible and are characterized by the various symmetries of the cooper - pair wavefunction . these phases of cold and dense qcd might occur in the core of neutron stars where matter is compressed to several times the nuclear density @xmath5 . the presence of such a superconducting phase is expected to have observable consequences on the cooling and magnetic fields of neutron stars @xcite . a particularly interesting phase is the so - called 2sc phase in which only two flavors of quarks are paired , while the unpaired quarks of the third flavor remain a fermi liquid . this phase further illustrates the importance of understanding the quark - gluon plasma , i.e. , the normal phase of quarks . fortunately , due to asymptotic freedom , the coupling constant @xmath6 of qcd becomes small at large chemical potential @xmath3 , such that a systematic study of the high - density and low - temperature region of the qcd phase diagram is possible using perturbation theory . at high densities the dominant interaction between quarks is that of one - gluon exchange , where the long - range behavior of the gluons is screened due to the quark - gluon plasma . while the electric gluons ( longitudinal ) are screened by a debye mass @xmath7 , the magnetic ( transverse ) gluons are only dynamically screened @xcite . this residual long - range behavior of the magnetic gluons dominates the low - temperature behavior of the system . examples of the effects of magnetic gluons can be seen in the non - standard scaling of the critical temperature of the superconducting phase @xmath8 @xcite , which depends exponentially only on @xmath9 instead of showing the expected @xmath10 behavior , or of the life - time of elementary excitations of ( grand - canonical ) energy @xmath11 near the fermi surface @xmath12 @xcite that signal a ( marginal ) non - fermi - liquid behavior at zero temperature . in the context of a possible deconfined phase of quarks inside neutron stars , the fermi system is expected to be imbalanced due to the different chemical potentials of the various quarks . this deconfined phase may contain @xmath13 , @xmath14 , and @xmath15 quarks with different densities as a consequence of their different masses and charges . in cold atomic physics , an analogous population imbalance has been realized experimentally between two spin states @xcite , which resulted in frustration of the pairing between particles . this will cause very different behavior from the normal bcs case and is currently a very hot topic in a wide variety of fields @xcite . in particular , the imbalance between different flavors of quarks is expected to significantly alter the properties of the qcd phase diagram , such as the 2sc phase mentioned above or the color - flavor locked phase where all the colors and flavors are paired @xcite . in the present paper , we therefore generalize the quasiparticle properties based on fermi - liquid theory to an imbalanced system . to do so , the effect of the dressed gluons on the thermodynamic potential , specific heat and effective mass of the quasiparticles is determined at low temperatures . the thermodynamic potential will be calculated for a two - flavor quark system in the limit of a large number of quark flavors ( large-@xmath0 ) . this approximation corresponds to the random phase approximation ( rpa ) , which has been quite successful in condensed - matter systems such as the interacting electron gas in metals @xcite . using the framework of landau fermi - liquid theory for relativistic systems , derived in section [ sec - theory ] , the effective mass of the quasiparticles can be determined from the thermodynamic potential by considering the specific heat . the result and its implications on the applicability of fermi - liquid theory is discussed . as we will show , the logarithmic dependence of the temperature , which is due to the transverse gluons , shows the breakdown of fermi - liquid theory at low temperatures . furthermore , even though within our approximations there is no interaction between the different flavors of quarks in the limit of weak coupling , the rpa correction still presents some mixing between the different flavors , which stems from the long - wavelength screening of the longitudinal gluons . above and throughout the article natural units are used , i.e. , units such that @xmath16 . other conventions and technical details of the calculation can be found in the appendices . to describe the normal state of the quark - gluon plasma at nonzero temperatures , which is a strongly interacting gas of quarks , landau fermi - liquid theory can be used . this theory takes as a starting point the non - interacting fermi gas and switches on the interaction adiabatically . as long as the temperature of the system is much higher than the critical temperature @xmath17 for superconductivity , no bound states ( cooper pairs ) will form and each state of the non - interacting fermi gas is transformed into a state of the interacting gas . therefore , the excitations of such a fermi liquid remain of a fermionic nature . for simplicity , fermi - liquid theory will be introduced here for a system at zero temperature even though the quark - gluon plasma is formally a marginal fermi liquid in that case . however , the introduced concepts turn out to be valid for nonzero temperatures much lower than the fermi temperature @xmath18 , which is expected to be around @xmath19 for quarks in the core of neutron stars . the basis of landau fermi - liquid theory is to consider the effect of a small change of the ground - state fermi distribution on the thermodynamic potential density , which can be written as @xmath20 here @xmath21 is a small change in the ground - state momentum distribution of species @xmath22 , @xmath23 is the particle density , @xmath24 is the energy of a quasiparticle that for our purposes only depends on the magnitude of the vector @xmath25 , and @xmath26 is the chemical potential . in the case of electrons the index @xmath22 specifies the spin of the electron , while for the case of quarks it specifies the spin , color and flavor of the quark . hence , the quasiparticle energy and the effective interaction between quasiparticles are defined as @xmath27 an important quantity in fermi - liquid theory is the effective mass @xmath28 of a quasiparticle of species @xmath22 near its fermi surface . it is defined in terms of the group velocity @xmath29 evaluated at the fermi momentum @xmath30 @xmath31_{k = k_\sigma}^{-1},\ ] ] such that upon linearizing the quasiparticle energy around the fermi surface we obtain @xmath32 with @xmath33 as the fermi energy of species @xmath22 . in the non - interacting limit , the effective mass reduces to @xmath34 and @xmath35 for non - relativistic and relativistic dispersions , respectively . using the fact that the pressure is an invariant under galilean or lorentz transformations ( see appendix . [ app - invariance ] ) , of which the effect can be written as a change in the distribution as shown in fig . [ fig : boosted_fermi_sphere ] , eq . ( [ thermpotential ] ) should give zero for this particular choice of @xmath36 and hence relate @xmath24 to @xmath37 . in other words , by performing an infinitesimal galilean or lorentz transformation on the fermi sphere , the effective mass can be related to the effective interaction at the fermi surface @xcite . ] to this end , consider an infinitesimal change in the fermi surface from @xmath30 to @xmath38 @xmath39 where the change in the fermi surface solely depends on the direction in momentum space denoted by @xmath40 . the corresponding change in the thermodynamic potential density can be written as @xmath41+\frac{1}{2}\sum_{\sigma,\sigma'}\iint\frac{\mathrm{d}\hat{\mathbf{k}}}{(2\pi)^3}\frac{\mathrm{d}\hat{\mathbf{k}}'}{(2\pi)^3}\intop_{k_\sigma}^{k_\sigma+\delta k_\sigma}\mathrm{d}kk^2\intop_{k_{\sigma'}}^{k_{\sigma'}+\delta k_{\sigma'}}\mathrm{d}k'k'^2f_{\sigma\sigma'}(\mathbf{k},\mathbf{k}')-\sum_\sigma n_\sigma\delta\mu_\sigma\nonumber\\ = & \sum_{\sigma}\frac{k_\sigma^3}{4\pi^2m_\sigma^*}\int\frac{\mathrm{d}\hat{\mathbf{k}}}{4\pi}\delta k_\sigma^2(\hat{\mathbf{k}})+\sum_{\sigma,\sigma'}\frac{k_\sigma^2k_{\sigma'}^2}{2(2\pi^2)^2}\iint\frac{\mathrm{d}\hat{\mathbf{k}}}{4\pi}\frac{\mathrm{d}\hat{\mathbf{k}}'}{4\pi}f_{\sigma\sigma'}(k_\sigma , k_{\sigma'},\theta)\delta k_\sigma(\hat{\mathbf{k}})\delta k_{\sigma'}(\hat{\mathbf{k}}')-\sum_\sigma n_\sigma\delta\mu_\sigma.\end{aligned}\ ] ] in the second line the integrals were expanded to second order in @xmath42 , and @xmath43 was defined as the effective interaction at the fermi surfaces between two species , with @xmath44 the angle between the directions of @xmath25 and @xmath45 . next we expand @xmath46 in terms of spherical harmonics and @xmath47 in terms of legendre polynomials as @xmath48 for an imbalanced system , the difference in density @xmath23 results in a difference in @xmath26 , which can cause @xmath49 to be different from each other even if their non - interacting bare masses @xmath50 are the same . for each species , the particle density obeys @xmath51 . inserting these expansions and using the orthogonality of the spherical harmonics , eq . ( [ thermpotexpand ] ) becomes @xmath52 in the balanced case , all fermi momenta are the same , @xmath53 , @xmath54 , then eq . ( [ reducedthermpotexpansion ] ) reduces to @xmath55,\ ] ] where the intensive quantities @xmath56 , @xmath3 , @xmath57 and @xmath58 are the same for all species , and @xmath59 is the total density , i.e. , the density summed over all species . the ( dimensionless ) landau fermi - liquid parameters @xmath58 are defined as @xmath60_{k_f}^{-1}=\sum_{\sigma'}\frac{k_fm^*f_{l,\sigma\sigma'}(k_f)}{2\pi^2}.\ ] ] note that @xmath58 is species independent , but the inter - species and intra - species interactions can be different , so the subscript @xmath22 is shown explicitly in @xmath61 but not in @xmath58 . from the above it is possible to derive the equation for the effective mass , as we first show for the simpler balanced case . consider a change in the distribution due to a lorentz transformation as shown in fig . [ fig : boosted_fermi_sphere ] . the shape of the fermi sphere is determined from the equation @xmath62 , where @xmath63 is the four - momentum , @xmath64 is the four - velocity of the fermi sphere with the lorentz factor @xmath65 . the minus sign in the previous expression and in @xmath66 are a consequence of our choice of the metric @xmath67 . in the rest frame , where @xmath68 , the above reduces to the condition @xmath69 . similarly , the shape of the fermi surface can be determined from the lorentz transformation of the momentum to a frame moving with velocity @xmath70 , @xmath71 which for an infinitesimal lorentz transformation reduces to @xmath72 and in the non - relativistic limit to @xmath73 . note that the deformation of the fermi sphere due to the lorentz contraction shows up only in higher - order terms of @xmath74 . first , consider the non - relativistic case in which an infinitesimal galilean transformation shifts all momenta from @xmath25 to @xmath75 . this results in a change of the fermi surface according to @xmath76 , such that @xmath77 . furthermore , the total energy of the system transforms as @xmath78 , where @xmath79 is the total mass of the system , which induces a change in the chemical potential @xmath80 . inserting this into eq . ( [ landaufermithermpot ] ) gives the well - known landau effective - mass relation in the balanced case @xcite @xmath81 secondly , for a relativistic system the momenta are shifted by @xmath72 such that @xmath82 . since the total energy of the system transforms as @xmath83 , the change in the chemical potential is @xmath84 . inserting these into eq . ( [ landaufermithermpot ] ) gives the relativistic landau effective - mass relation @xcite @xmath85 note that in the right - hand side @xmath86 also depends on @xmath3 , because the fermi momentum @xmath57 is a , in general complicated , function of @xmath3 . now for the more general imbalanced case . since the lorentz transformation is applied uniformly to each species , according to the symmetry of the formula , we can expect that the contribution from each species in eq . ( [ reducedthermpotexpansion ] ) should vanish , namely , @xmath87 then we get the general expression for the effective mass , @xmath88,\ ] ] where @xmath89 is the density of states at the fermi energy of species @xmath22 . the arguments in @xmath90 emphasize that , in the imbalanced case , the interaction may depend not only on the angle between the momenta but also on the absolute value of the fermi momentum of each species . in fact , the same expression of @xmath28 can be obtained more strictly by considering the addition of a single particle to the system and comparing the energy increase in two different frames . we can introduce the average effective mass , which is of practical importance , as @xmath91 with @xmath92 the total number of species . correspondingly , the average chemical potential is @xmath93 . in order to get a similar expression as in eq . ( [ rel - landaumass ] ) , we introduce the relative weight of each species as @xmath94 such that @xmath95 , and then generalize the landau effective - mass relation with average values as @xmath96,\ ] ] where the double - sum term , which can be defined again as @xmath86 , plays the role of the landau parameter in the balanced case as in eq . ( [ rel - landaumass ] ) , therefore we obtain now @xmath97 which , obviously , reduces to eq . ( [ landaufermiparameter ] ) as the system becomes balanced . in the following sections the effective mass will be derived from a microscopic rpa calculation . in particular , in sec . [ sec - thermpotential ] it will be determined from the specific heat . the specific heat ( per volume ) for low temperatures can be expressed in terms of the effective masses by @xmath98_{v , n}\nonumber\\ & \simeq\sum_\sigma\frac{k_\sigma m^*_\sigma}{6}t=\sum_\sigma\frac{\pi^2}{3}n_\sigma(0)t,\end{aligned}\ ] ] where it was used that @xmath99 and for the temperature derivative the sommerfeld expansion was used , c.f . ( [ eq : ap_sommerfeld_exp ] ) . this result is a trivial generalization of the single species or balanced case , since the effective mass already contains the effect of interactions , such that the contribution to the heat capacity of a single quasiparticle is additive . our starting point is the partition function @xmath100 containing the euclidean qcd lagrangian density fixed in a linear gauge @xmath101 , @xmath102 where @xmath103 is the quark field of flavor @xmath104 ( the color @xmath105 and spin @xmath15 degrees of freedom are not shown explicitly ) with mass @xmath106 and chemical potential @xmath107 , @xmath108 are the gluon fields , @xmath109 are the ghost fields , @xmath6 is the qcd coupling constant , @xmath110 is a gauge fixing parameter , @xmath111 are the fine - structure constants of the color @xmath112 group , the covariant derivative @xmath113 and the antisymmetric gluon field tensor @xmath114 . for large densities , the fermionic degrees of freedom become increasingly important , such that the large-@xmath0 limit could give insight into the behavior of quarks at high densities . therefore , consider the large-@xmath0 limit with a fixed rescaled coupling @xmath115 , @xmath116 integrating out the fermions and the ghosts gives @xmath117\\ & -\mathrm{tr}_c\ln\left(\partial_{\mu}f_\mu\delta^{ab}+\mathrm{g}n_f^{-\frac{1}{2}}f^{abc}a_\mu^cf_\mu\right)\\ & + \frac{1}{2}a_\mu^ad^{-1}_{0,\mu\nu}\delta^{ab}a_\nu^b\\ & + \frac{1}{2}\mathrm{g}n_f^{-\frac{1}{2}}\left(\partial_{\mu}a_\nu^a-\partial_{\nu}a_\mu^a\right)f^{abc}a_\mu^ba_\nu^c\\ & + \frac{1}{4}\mathrm{g}n_f^{-1}f^{abc}f^{ade}a_\mu^ba_\nu^ca_\mu^da_\nu^e,\end{aligned}\ ] ] where the subscripts @xmath118 explicitly indicate that the trace should also be taken over color , flavor and spin space . when we expand the first line in @xmath0 , the first - order contribution will vanish due to conservation of color , i.e. , @xmath119=0 $ ] . only the second - order expansion in @xmath0 of the first line will give a contribution . thus expansion in powers of @xmath0 gives @xmath120 where we defined the polarization tensor @xmath121 from the contribution of the quark with flavor @xmath104 as @xmath122 which is shown diagrammatically in fig . [ fig - polarizationtensor ] . the polarization tensor @xmath121 is diagonal in color space , as can be seen explicitly form eq . ( [ eq : def_polarization_tensor ] ) . upon integrating out the gluons the thermodynamic potential density is @xmath123\nonumber\\ & \left.+\frac{n_g}{2}\mathrm{tr}\ln\left(1+d_0\sum_f\pi_f\right)\right\},\label{eq : therm_pot}\end{aligned}\ ] ] where @xmath124 is the number of colors , @xmath125 is the number of gluons , and we use the qcd values @xmath126 and @xmath127 in the following . the first two terms are the ideal fermi and bose gas contributions to the thermodynamic potential density , where the term @xmath128 cancels the two unphysical degrees of freedom from @xmath129 . the last term is the rpa correction ( ring sum ) to the thermodynamic potential density , as is shown in fig [ fig - thermopotential ] . since we are mainly interested in the temperature dependence of @xmath130 and the fact that the above definition contains ( divergent ) zero - temperature contributions , the @xmath131 expression will be subtracted , i.e. , we consider @xmath132 . in sec . [ sec - weakcoupling ] an example is considered to check the expression for the landau effective mass analytically and in sec . [ sec - thermpotential ] the above thermodynamic potential density is calculated numerically for the case of a two - flavor imbalanced quark system . to the first order . ] or rpa approximation where the double gluon line signifies the dressed gluon propagator . ] in the limit of weak coupling the interaction contribution to the thermodynamic potential density is given solely by the exchange diagram @xmath133 using the definition of the free gluon propagator in the lorentz gauge , this can be written as @xmath134,\end{aligned}\ ] ] where the fermi and bose distributions are defined as @xmath135}+1},\quad n_b^s(\mathbf{p})=\frac{1}{e^{\beta[s\epsilon^g(\mathbf{p})]}-1},\ ] ] respectively , the explicit expression for @xmath136 is shown in eq . ( [ polarizationtensor ] ) , the matsubara sum over the bosonic frequencies @xmath137 was performed , @xmath138 denotes the free quark dispersion while @xmath139 is the free gluon dispersion , and the function @xmath140 was defined as @xmath141 which will soon be shown to play the role of the interaction between two quarks ( @xmath142 ) , two anti - quarks ( @xmath143 ) or a quark and an antiquark ( @xmath144 or @xmath145 ) mediated by the emission or absorption of a gluon ( depending on the sign of @xmath146 relative to @xmath147 and @xmath148 ) . the quasiparticle energy and effective interaction can be obtained by varying the thermodynamic potential with respect to the fermi distribution . in the flavor - imbalanced case , the distributions depend only on the flavor but not on the color and spin indices . to distinguish between the different contributions for each particle species it can be checked that the multiplication factor @xmath149 could be written as @xmath150 , where @xmath15 and @xmath151 are just dummy indices denoting the spin degrees of freedom , which are helpful to arrive at the following expressions . note that @xmath152 is independent of the spin index , and in fact the spin degrees of freedom has already been summed over in the expression for @xmath140 . explicitly we find for these quantities @xmath153-\sum_{s'}\frac{g^2}{4}\sum_{c'}t_{cc'}^at_{c'c}^a\sum_{s_2,s_3=\pm1}\int\frac{\mathrm{d}^3p}{(2\pi)^3}\mathcal{f}^f_{+,s_2,s_3}(\mathbf{k},\mathbf{q})\nonumber\\ & \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\times[(1+n_b^{s_3}(\mathbf{k}-\mathbf{q})(1-n_f^{s_2}(\mathbf{q}))+n_b^{s_3}(\mathbf{k}-\mathbf{q}))n_f^{s_2}(\mathbf{q})],\label{quasiparticle - energy}\\ & f_{\{c , f , s\};\{c',f',s'\}}(\mathbf{k},\mathbf{k}')\equiv\frac{\delta^2\omega}{\delta n_{c , f , s}^+(\mathbf{k})\delta n_{c',f',s'}^+(\mathbf{k}')}=\frac{g^2}{4}t_{cc'}^at_{c'c}^a\delta_{ff'}\sum_{s_3=\pm1}\mathcal{f}^f_{+,+,s_{3}}(\mathbf{k},\mathbf{k}'),\label{quasiparticle - int}\end{aligned}\ ] ] where @xmath154 and @xmath155 , according to the definitions in appendix [ app - euclidean ] . in the quasiparticle energy eq . ( [ quasiparticle - energy ] ) the first term is the noninteracting part , while the second term corresponds to the quark self - energy , as is shown in appendix [ app - quark - selfenergy ] . the angular averaged interaction parameters are defined by the coefficients of the legendre polynomial expansions of the effective interaction eq . ( [ quasiparticle - int ] ) evaluated on the fermi sphere , by denoting @xmath156 , @xmath157 which gives for the zeroth and first interaction parameters @xmath158,\nonumber\\ f_{1;f}&=\frac{g^2n_gn_s}{8n_c}\frac{3}{\mu_{0f}^2}\frac{\lambda^2 + 2m_f^2}{2k_f^2}\nonumber\\ & \qquad\qquad\times\left[1+\frac{\lambda^2 + 2k_f^2}{4k_f^2}\ln\left(\frac{\lambda^2}{\lambda^2 + 4k_f^2}\right)\right],\label{weakcouplandauparam}\end{aligned}\ ] ] where @xmath159 is the number of spin degrees of freedom , and a regulatory gluon mass @xmath160 was introduced , i.e. , @xmath161 . we see that in this approximation the interaction is color- and spin - independent . for the sake of simplicity , therefore , we from now on omit the unnecessary color and spin indices , and use @xmath162 and @xmath163 for the fermi momentum and the non - interacting fermi energy of the quark with flavor @xmath104 . the effective mass of the balanced case , eq . ( [ rel - landaumass ] ) , can be obtained from the interaction parameters eq . ( [ weakcouplandauparam ] ) , and eq . ( [ landaufermiparameter ] ) , by expanding to the first order in the coupling constant @xmath164 where it was used that @xmath165 as follows from eq . ( [ quasiparticle - energy ] ) . for the imbalanced case , a similar result can be obtained for each species according to eq . ( [ imbalancemass ] ) , @xmath166 where the kronecker delta @xmath167 in the interaction simplifies the expression considerably , such that we obtain the similar expression as in eq . ( [ effectivemassexpand ] ) . as a special case , we can use the above results to discuss a two - flavor imbalanced system , where the average effective mass and chemical potential can be defined as @xmath168 and @xmath169 with the subscripts @xmath170 and @xmath171 for the majority and minority flavors , respectively . the imbalance can be quantified as @xmath172 . for small imbalance , the factors @xmath173 can be expanded in @xmath174 with respect to the value in the balanced case , namely @xmath175 . note that , the linear coefficients of the expansion can , in general , be different for each flavor . we obtain the expression of @xmath56 for small @xmath174 @xmath176,\end{aligned}\ ] ] where the @xmath177 term is kept since the linear term may vanish for the symmetric case , namely @xmath178 , and we have replaced @xmath179 with @xmath3 because the difference is in higher order of @xmath6 . compared with eq . ( [ averageimbalancemass ] ) , the effective landau parameter for the average effective mass is @xmath180.\end{aligned}\ ] ] similarly , we can obtain the effective mass difference @xmath181 as a function of @xmath174 @xmath182.\end{aligned}\ ] ] it is quite natural to find @xmath183 in the leading order . the consistency of the above results will now be verified to the lowest order in the coupling constant using the quark self - energy . starting from the dispersion and self - energy of eq . ( [ quasiparticle - energy ] ) , the effective mass in eq . ( [ effectivemassexpand ] ) or eq . ( [ imbalancemassexpand ] ) can also be derived in a different manner . the quasiparticle pole of the quark propagator is @xmath184 where @xmath185 is the renormalized positive - energy projected self - energy . the effective mass corresponding to this pole is most easily defined from the fermi velocity of the quasiparticle @xmath186_{p = k_f}.\ ] ] inserting the definition of the quasiparticle pole into the above gives , using @xmath187 , @xmath188_{p = k_f}^{-1}+\mathcal{o}(g^4)\nonumber\\ = & \mu_f-\left[\frac{\partial\epsilon_{0f}(\mathbf{p})\sigma_{c , f , s}^+(\epsilon_{0f}(\mathbf{p})-\mu_f,\mathbf{p})}{\partial\epsilon_{0f}(\mathbf{p})}\right]_{p = k_f}+\mathcal{o}(g^4).\label{effectivemassnew}\end{aligned}\ ] ] using that the explicit form of the self - energy in the zero - temperature limit can be written as ( c.f . ( [ quarkselfenergy ] ) ) @xmath189 and the relation eq . ( [ eq : ap_eff_int_transf ] ) between the vector derivatives of the effective interaction , the derivative in eq . ( [ effectivemassnew ] ) can be rewritten as @xmath190_{p = k_f}=\frac{\mu_{0f}}{k_f}\left[\sum_{c',f',s'}\int\frac{\mathrm{d}^3q}{(2\pi)^3}\hat{\mathbf{p}}\cdot\frac{\partial\epsilon_{0f}(\mathbf{p})f_{\{c , f , s\};\{c',f',s'\}}(\mathbf{p},\mathbf{q})}{\partial\mathbf{p}}n_{f'}^+(\mathbf{q})\right]_{p = k_f}\\ = & -\frac{\mu_{0f}}{k_f}\int\frac{\mathrm{d}q}{2\pi^2}q^2\epsilon_{0f}(\mathbf{q})\delta(q - k_{f})\frac{1}{3}\left[3\int\frac{\mathrm{d}\hat{\mathbf{q}}}{4\pi}\hat{\mathbf{p}}\cdot\hat{\mathbf{q}}\sum_{c',f',s'}f_{\{c , f , s\};\{c',f',s'\}}(\mathbf{p},\mathbf{q})\right]_{p = k_f } = -\frac{k_f\mu_{0f}^2}{2\pi^2}\frac{f_{1;\{c , f , s\}}(k_f)}{3},\end{aligned}\ ] ] where partial integration was used and @xmath191 at zero temperature . therefore eq . ( [ effectivemassnew ] ) is identical to eq . ( [ imbalancemassexpand ] ) , which shows that the phenominological landau argument is indeed consistent with the microscopic diagrammatic calculation . before we start the calculation of the temperature dependence of the large-@xmath0 thermodynamic potential , it is useful to examine the dressed gluon propagator , which is a crucial ingredient of the rpa theory . the most physical gauge to study the propagator is the coulomb gauge , since its form results from considering linear response @xcite . this is also by far the most used gauge in condensed - matter theory . in this gauge the gluon propagator is @xcite @xmath192 where @xmath193 is the three - dimensional transverse projector defined in appendix [ app - green ] , @xmath194 and @xmath195 are related to the longitudinal and transverse projections of @xmath196 including the contributions of various species as well as the vacuum . more details can be found in appendix [ app - gluon - selfenergy ] . consider the spectral functions of the transverse and longitudinal propagator , which are defined by @xmath197.\ ] ] in general the spectral function depends on the gauge , however , the positions of the poles are gauge independent , and the coulomb gauge has the additional property @xmath198 for @xmath199 as required of a physical spectral function . for the balanced case , the spectral functions are shown in fig . [ fig : rhotl ] for several values of @xmath3 and in fig . [ fig : rhotltemp ] for several values of @xmath1 . the latter includes a small-@xmath1 correction to @xmath194 and @xmath195 , see eq . ( [ eq : ap_pi_temperature_correction]).-correction the spectral functions will no longer satisfy @xmath200 . ] the transverse and longitudinal plasmon modes and the large contribution due to the decay of the gluon into the particle - hole continuum ( @xmath201 ) are clearly visible . note that massless quarks are used in the limit of high density , c.f . appendix [ app - gluon - selfenergy ] . however , a small but finite quark mass is necessary to renormalize the real part of the polarization tensor of the vacuum , as shown in eq . ( [ eq : ap_renormalised_vacuum ] ) . since the vacuum contribution plays no significant role in the following calculation , we simply take this nonzero quark mass @xmath202 in the vacuum term the same for different flavors . for the two - flavor imbalanced case , it is clear from eq . ( [ eq : therm_pot ] ) that in the leading - order correction each flavor contributes separately to the polarization tensor . because of the above setting , there is no mass imbalance in the present system , therefore it is enough to consider only positive @xmath174 due to the symmetry . following the notation in sec . [ sec - weakcoupling ] , all the above results can be easily generalized to such an imbalanced system by using @xmath203 to replace @xmath204 , and @xmath205 for @xmath206 . as @xmath207 , the system reduces to the balanced case . now the low temperature condition requires @xmath208 , such that @xmath174 can not be too close to @xmath209 , namely the extremely imbalanced case . the spectral function for a two - flavor imbalanced system is shown for various @xmath174 in fig . [ fig : rhotlh ] . comparing with fig . [ fig : rhotltemp ] , we see that increasing @xmath174 has a similar effect as increasing @xmath1 . the dispersion relations of the modes can be found by solving @xmath210=0.\label{eq : re_invprop_zero}\ ] ] in the transverse case there is a single solution @xmath211 , however , in the longitudinal case there are two solutions @xmath212 , as shown in fig . [ fig : dispersion ] . note that @xmath213 is not a real propagating mode because the imaginary part in the region @xmath201 is large due to particle - hole creation processes , as explained in appendix [ app - gluon - selfenergy ] . the two plasmon modes @xmath214 and @xmath215 approach the so - called plasma frequency @xmath216 as @xmath217 . the plasma frequency of the transverse and longitudinal mode can be found by expanding the inverse propagators for small @xmath218 and small @xmath219 , which in the zero - temperature limit gives for both cases @xmath220 whose solution for small coupling constant yields @xmath221 @xcite . here @xmath222 is the gluon thermal mass whose expression is obtained in the hard dense and hard thermal loop approximation as shown in eq . ( [ thermalmass ] ) . in the limit @xmath223 all solutions reduce to @xmath224 , ( [ eq : ap_fg_large_pomega_behavior ] ) . furthermore , for the limit @xmath225 the vacuum becomes increasingly dominant such that eq . ( [ eq : re_invprop_zero ] ) has a zero at large @xmath226 called the landau pole , @xmath227 @xcite , c.f . ( [ eq : ap_fg_large_pomega_behavior ] ) , which , however , plays no role for our purpose as we are interested in the low - temperature behavior of the theory that is hardly influenced by the high - energy behavior of the gluon propagator . the dispersion relations for the case with nonzero @xmath1 or @xmath174 is similar but just with a little higher @xmath228 . in fact , as a generalization of eq . ( [ thermalmass ] ) , the thermal mass of a two - flavor imbalanced quark system reads @xmath229,\ ] ] where @xmath174 thus plays the same role as @xmath230 . the full ideal - gas contribution to the thermodynamic potential density with the zero - temperature contribution subtracted is ( c.f . appendix [ app - thermpotential ] ) @xmath231 the first part is the contribution of an ideal massless fermi gas , while the second part is the stefan - boltzmann law of an ideal bose gas . the rpa correction to @xmath130 , the last term in eq . ( [ eq : therm_pot ] ) , can be written as @xmath232 where the lorentz gauge is used , while for the coulomb gauge the second identity is not valid but the last result is still the same , which is a consequence of the gauge invariance of the thermodynamic potential density . using contour deformations to carry out the matsubara sum , as shown in appendix [ app - matsubara ] , we obtain @xmath233\\ & \times\{\im[\ln\tilde{f}(\omega_+,\mathbf{q})]+2\im[\ln\tilde{g}(\omega_+,\mathbf{q})]\},\end{aligned}\ ] ] where @xmath234 , and similar for @xmath235 by replacing @xmath194 with @xmath195 . the temperature dependence can be obtained by subtracting the zero - temperature contribution . since @xmath194 and @xmath195 contain corrections of order @xmath236 , the leading - order correction in @xmath237 comes from two parts , @xmath238\mathrm{sgn}[\im(\tilde{f})^0]\right.\\ & \left.+2\arctan\frac{\im(\tilde{g}^0)}{\re(\tilde{g}^0)}+2\pi\theta[-\re(\tilde{g}^0)]\mathrm{sgn}[\im(\tilde{g}^0)]\right)\\ & \left.\frac{1}{2}[\im(\ln\tilde{f}^t)-\im(\ln\tilde{f}^0)+2\im(\ln\tilde{g}^t)-2\im(\ln\tilde{g}^0)]\vphantom{\frac{1}{2}}\right ] , \end{aligned}\ ] ] where the superscript @xmath1 or @xmath239 means the corresponding terms are taken at nonzero @xmath1 or @xmath131 . we will refer to the first part as the @xmath240 term and the second as the non-@xmath240 term . the leading correction from the non-@xmath240 term can be shown to be proportional to @xmath236 and is not of great interest in our study , since we will concentrate on the anomalous and dominant @xmath1 dependence , which is a consequence of the @xmath240 term . in the @xmath240 term , the arctangent terms can be interpreted as contributions due to production and decay of thermal gluons , because of their dependence on the imaginary part of the gluon self - energies @xmath194 and @xmath195 , while the theta function terms are interpreted as a correction to the ideal gas law due to thermal plasmon modes . the frequency integral over the theta function can be performed explicitly . in both the transverse and longitudinal case the sign of the imaginary part is positive when the real part is negative , such that after the frequency integration the result of the integral is proportional to @xmath241 where @xmath242 and @xmath243 signifies the region where @xmath244 . in the limit of low temperature this integral will go to a constant . the theta - function contribution can thus be neglected since it is of higher order in the temperature than is of interest to us here . next we perform the integrals over the arctangents , whose structure at @xmath245 is shown in fig . [ fig : arctan ] . note that , for the study of the leading - order @xmath1 corrections , it is not necessary to include the @xmath236 term in @xmath194 and @xmath195 since the integral with @xmath246 at low @xmath1 is already in the order of @xmath236 . the dominant contribution for small temperatures ( @xmath2 ) comes from the frequency integration over the domain @xmath247 $ ] for the case @xmath248 , which is due to particle - hole creation . for the two - flavor balanced case ( @xmath249 , @xmath250 ) , it was found numerically that the integral in the limit of small temperatures is @xmath251 for the electric ( longitudinal ) gluons and @xmath252 for the magnetic ( transverse ) gluons , with @xmath253 , @xmath254 , @xmath255 , and @xmath256 . these results are quite close to the analytic results for the leading - order correction terms at small @xmath6 and @xmath1 obtained in ref . @xcite , where @xmath257/6\approx0.4460 $ ] , @xmath258 , @xmath259/18\approx0.4032 $ ] , and @xmath260 . for the imbalanced case , the above result for the transverse gluons can be generalized to @xmath261,\ ] ] where the contributions from both flavors with chemical potentials @xmath262 are additive . the longitudinal part needs some further discussion . since the static long - wavelength longitudinal modes are screened ( c.f . appendix [ app - gluon - selfenergy ] ) , the two chemical potentials contributing to @xmath263 can not be separated even in the lowest - order term . this is different from the result obtained in weak coupling , as shown in eq . ( [ quasiparticle - int ] ) , where no interaction between the two flavors is involved . therefore it is not surprising to find that a simple generalization of the balanced case , as @xmath264 , does not fit well with the numerical results . to obtain a reasonable ansatz for the longitudinal part , we integrate the corresponding arctangent term of the imbalanced case in the low temperature limit up to @xmath265 to obtain @xmath266,\ ] ] where we see that the factor @xmath267 is a mixture effect of the two flavors . this expression fits the numerical results very well with an error of only @xmath268 . furthermore , just based on the numerical data , we find another ansatz which fits the results even better @xmath269 where the mixture effect is shown implicitly in the logarithm , since the denominator becomes dimensionless by canceling with the average chemical potential @xmath270 in the numerator . the specific heat ( per volume ) at fixed volume and particle number is @xcite @xmath271 ^ 2}{(\partial n_f/\partial\mu_f)_t},\ ] ] where the entropy density @xmath272 and particle number density @xmath273 . in the low - temperature limit the second term can be neglected , therefore the specific heat can be obtained , using eqs . ( [ ct ] ) and ( [ cl ] ) , as @xmath274,\ ] ] where the specific heat @xmath275 of an ideal gas has been subtracted . the effective mass and the first landau parameter can be determined in the high - density limit by comparing eq . ( [ cv ] ) and eq . ( [ heatcapacity ] ) , @xmath276,\end{aligned}\ ] ] where we used @xmath277 in the massless limit at high density . the first term in the effective mass is due to the ideal - gas specific heat . however , as pointed out before , the factor @xmath278 shows the mixing of the two flavors , unlike the weak - coupling results obtained in sec . [ sec - weakcoupling ] , e.g. , eq . ( [ imbalancemassexpand ] ) . this is because , with the rpa correction , we incorporate the sum of an infinite chain of gluon self - energies , which incorporates interactions between different quark flavors . since the longitudinal gluon is screened in the static long - wavelength limit , the mixing is even present in the low-@xmath1 and small-@xmath6 limit . the numerical fit in eq ( [ clnew ] ) also provides another expression for the effective mass , @xmath279.\end{aligned}\ ] ] as expected , all the above results return to the balanced case as @xmath207 . as mentioned previously , the contribution from the non-@xmath240 term is not included , which acts as a constant shift on the @xmath280 and @xmath281 factors . in fig . [ fig : mass ] the change of effective mass due to the interaction is given as function of imbalance at various temperatures . finally , we emphasize again that @xmath174 should not be too close to @xmath209 even though the divergence from @xmath282 is suppressed by the prefactor @xmath283 , because in such an extremely imbalanced case the condition @xmath284 is not satisfied for the minority flavor . we have calculated the thermodynamic potential perturbatively in the large-@xmath0 limit and the effective mass of the quarks is determined by using fermi - liquid theory for an imbalanced cold dense quark system . the temperature dependence is obtained by using the gluon self - energy , from which the contributions from transverse and longitudinal gluons are explicitly shown . for the two - flavor imbalanced quark system , the effective mass is obtained both analytically within the weak - coupling limit , and numerically within the rpa approximation . we find that , in contrast to the weak - coupling result , where the effective mass of each flavor is independent of each other due to the lack of an inter - flavor interaction , the effective mass obtained from the rpa calculation depends on the chemical potentials of both flavors , which is a consequence of the static screening of the electric gluons at long wavelength . from the rpa results , the logarithmic dependence on temperature of the specific heat and effective mass signals a breakdown of fermi - liquid theory at zero - temperature . non - fermi - liquid behavior , due to unscreened long - range magnetic interactions , was already discussed several decades ago for the case of the electron gas @xcite . in the large-@xmath0 limit the qcd thermodynamic potential is essentially the same as that of qed , apart from group theory factors . the non - abelian effects of qcd only show up if gluon self - interaction corrections are included . the logarithmic behavior in the balanced case has previously been seen in analytic and numerical calculations of the qcd thermodynamic potential and specific heat using the large-@xmath0 limit , dimensional reduction , and hard - dense loop qcd perturbation theory @xcite . we expect that such a logarithmic dependence shall have important effects on the imbalanced qcd phase diagram at low temperatures and directly influence the properties of quark matter in the core of neutron stars . to better understand these effects , it is necessary to go further , such as including the gluon self - interaction and extending our discussion to the three - flavor imbalanced case . we hope progress along these directions will be achieved in the near future . this work is supported by the stichting voor fundamenteel onderzoek der materie ( fom ) and the nederlandse organisatie voor wetenschaplijk onderzoek ( nwo ) . the four - momentum vectors in euclidean space - time will be written with capital letters @xmath285 , while for the wick - rotated case ( @xmath286 ) the roman capital letters @xmath287 are used . three momentum vectors are written in bold face @xmath288 and its length as @xmath218 . and @xmath289 are sometimes written as @xmath290 for short . the euclidean gamma matrices in the standard representation are @xmath291 @xmath292 all the above entries are @xmath293 matrices and @xmath294 are the pauli spin matrices @xmath295 the gamma matrices obey the following relations @xmath296 the charge conjugation matrix is @xmath297 which satisfies @xmath298 the eight generators of the fundamental representation of @xmath299 are taken to be @xmath300 @xmath301 @xmath302 @xmath303 and have been normalized according to @xmath304=\frac{1}{2}\delta^{ab}$ ] . in general , the product of the generators is @xmath305 furthermore , one will frequently encounter the following group - theory factors @xmath306 @xmath307 in the above expressions @xmath124 is the number of colors and @xmath308 the number of gluons . fourier transforms are normalized as @xmath309 where @xmath310 are the fermionic matsubara frequencies , @xmath311 is the imaginary time phase space volume and @xmath312 and @xmath313 . the dirac equation in euclidean space is @xmath314 where @xmath315 . in momentum space the dirac equation reads @xmath316 the eigenvalues of the matrix @xmath317 are @xmath318 , since @xmath319 . the eigenspinors corresponding to these eigenvalues are @xmath320 note that @xmath317 is not a hermitian matrix and that @xmath321 satisfies the same equation as @xmath322 . however , they can also be viewed as the the eigenspinors of a hermitian matrix @xmath323 the positive and negative energy eigenspinors have the following form @xmath324 where @xmath325 and @xmath326 . they are orthonormal in the sense that @xmath327 and satisfy the completeness relation @xmath328=1.\ ] ] using the above eigenspinors the positive and negative energy projectors can be defined as @xmath329 which in the massless case reduce to @xmath330 the helicity projection operators project the spin along the momentum of the particle and read @xmath331 where @xmath332 is the spin operator . note that @xmath333=0 $ ] . the green s functions of quarks and gluons in euclidean space are defined by @xmath334 in momentum space , the free quark propagator is @xmath335 in the lorentz gauge @xmath336 , the momentum - space free gluon propagator is @xmath337=\frac{p_{\mu\nu}}{q^2}+\frac{\xi}{q^2}\frac{q_\mu q_\nu}{q^2},\ ] ] while in the coulomb gauge @xmath338 it is @xcite @xmath339 where @xmath110 is a gauge - fixing parameter and the projectors can be written as @xmath340 note that the free gluon propagator in the lorentz gauge satisfies @xmath341 , such that in the landau gauge ( @xmath342 ) it is purely four - momentum transverse , i.e. , @xmath343 . the free gluon propagator in the coulomb gauge satisfies @xmath344 , such that in the landau gauge it is three - momentum transverse . these propagators are diagonal in color space . in this section the lorentz transformation properties of some quantities are summarized , such as the thermodynamic potential , the distribution function and the effective interaction . the lorentz transformation to a frame moving with velocity @xmath74 is @xmath345 such that the four - momentum @xmath346 of a particle transforms as @xmath347 not to be confused with the dirac matrices @xmath348 , the @xmath349 used in this section is the lorentz factor @xmath65 . the invariance of the thermodynamic potential density under lorentz transformations can be shown using the stress - energy tensor . consider the change in the thermodynamic potential density under a lorentz transformation from the rest frame to a frame moving with velocity @xmath74 . since an interacting gas of quarks in the rest frame is specified only by an energy density @xmath350 and a pressure @xmath351 , the stress - energy tensor is diagonal and of the form @xmath352 such that under a lorentz transformation the thermodynamic potential density @xmath353 changes as @xmath354 the transformation of the chemical potential ( @xmath355 ) follows from the lorentz boosted total energy of the system @xmath356 , where @xmath357 is the total momentum of the system in the rest frame , which results in @xmath358 . the transformation of the density is due to a lorentz contraction in the volume @xmath359 . thus the total change in @xmath360 is @xmath361 , such that under a lorentz transformation @xmath362 . for an alternative derivation which uses that the pressure transforms the same way as a force per area , see ref . @xcite . consider the lowest - order correction to the free thermodynamic potential density @xmath364 note that the distribution function @xmath365 is a lorentz invariant , which can be easily derived from the fact that the number of particles in a volume @xmath366 of phase space is invariant under lorentz transformations @xcite , i.e. , @xmath367 where the tilde signifies the lorentz transformed quantity . subsequently , it is possible to derive a transformation law for @xmath363 by using that the distribution , the thermodynamic potential density and @xmath368 are lorentz invariant . it follows that @xmath369 should be lorentz invariant , giving @xmath370 where @xmath371 . expand @xmath372 to the lowest order in @xmath74 , using @xmath373 , @xmath374 , and @xmath375 , @xmath376\left[1-\mathbf{v}\cdot\frac{\mathbf{p}'}{\epsilon_{0\sigma'}(\mathbf{p}')}\right]+\mathcal{o}(v^2)\\ = & \tilde{f}_{\sigma\sigma'}(\mathbf{p},\mathbf{p}')-\epsilon_{0\sigma}(\mathbf{p})\mathbf{v}\cdot\frac{\partial f_{\sigma\sigma'}(\mathbf{p},\mathbf{p}')}{\partial\mathbf{p}}-\epsilon_{0\sigma'}(\mathbf{p}')\mathbf{v}\cdot\frac{\partial f_{\sigma\sigma'}(\mathbf{p},\mathbf{p}')}{\partial\mathbf{p}'}-f_{\sigma\sigma'}(\mathbf{p},\mathbf{p}')\left[\mathbf{v}\cdot\frac{\partial\epsilon_{0\sigma}(\mathbf{p})}{\partial\mathbf{p}}-\mathbf{v}\cdot\frac{\partial\epsilon_{0\sigma'}(\mathbf{p}')}{\partial\mathbf{p}'}\right]+\mathcal{o}(v^2)\\ = & \tilde{f}_{\sigma\sigma'}(\mathbf{p},\mathbf{p}')-\mathbf{v}\cdot\left[\frac{\partial\epsilon_{0\sigma}(\mathbf{p})f_{\sigma\sigma'}(\mathbf{p},\mathbf{p}')}{\partial\mathbf{p}}+\frac{\partial\epsilon_{0\sigma'}(\mathbf{p}')f_{\sigma\sigma'}(\mathbf{p},\mathbf{p}')}{\partial\mathbf{p}'}\right]+\mathcal{o}(v^2).\end{aligned}\ ] ] if it is assumed that to the lowest order the interaction does not depend on any distribution functions , i.e. , @xmath377 , the above implies @xmath378 note that the above derivation is similar to that given in ref . in this section it is summarized how to turn the minkowski quantum field theory ( qft ) of quantum chromodynamics into a euclidean statistical field theory ( sft ) suitable for studying the dynamical properties of a many - particle system . the starting point is the gauge - fixed path integral for qcd @xmath379 where @xmath380 are the gluon fields , @xmath381 and @xmath382 the quark fields , @xmath383 and @xmath384 the ghost fields and the qcd lagrangian in minkowski space fixed in a linear gauge ( @xmath385 ) is given by @xmath386\nonumber\\ & -\frac{1}{2}a_a^\mu(\partial_{\mu}\partial_\nu-\partial^2\eta_{\mu\nu})a_a^\nu\nonumber\\ & -\bar{\eta}^a\partial_\mu\partial^\mu\eta^a - gf^{abc}(\bar{\eta}^a\partial^\mu\eta^b)a_\mu^c\nonumber\\ & -\frac{1}{2}g(\partial^\mu a_a^\nu-\partial^\nu a_a^\mu)f^{abc}\eta_{\mu\sigma}\eta_{\nu\rho}a_b^\sigma a_c^\rho\nonumber\\ & -\frac{1}{4}g^2f^{abc}f^{ade}\eta_{\mu\sigma}\eta_{\nu\rho}a_a^\mu a_c^\nu a_d^\sigma a_e^\rho+\frac{1}{2\xi}(f^\mu a_\mu^a)^2,\end{aligned}\ ] ] where @xmath103 is the quark field with flavor @xmath104 , and the summation over color and spin indices are shown implicitly . the metric was chosen to be @xmath387 and the minkowski gamma matrices in the standard representation are @xmath388 the gamma matrices satisfy @xmath389 by performing a wick rotation the above quantum field theory can be transformed into a statistical field theory . a wick rotation amounts to taking an analytic continuation from real time to imaginary time ( @xmath390 ) , which turns the minkowski metric @xmath391 into the euclidean metric @xmath392 . to do this consistently the zeroth component of all four - vectors need to change accordingly . the procedure is most easily understood by considering the length of the position four - vector @xmath393 where it is seen that the minus sign of the minkowski metric is absorbed in the definition of the euclidean four - vectors @xmath394 . in euclidean space no distinction is made between upper and lower indices . a simple way to obtain the euclidean form of a vector is to multiply the contravariant vector by the matrix @xmath395 or the covariant vector by @xmath396 and set @xmath390 . note that the spatial components do not change . for example , the euclidean position four - vector , the four - divergence and the zeroth gamma matrix are in terms of their minkowski definitions eq . ( [ eq : ap_minkowski_gamma ] ) @xmath397 generalizations to tensors is straightforward and follows for instance from the example @xmath398 . using the above procedure to find the euclidean versions of all tensors , the partition function is easily found from eqs . ( [ eq : ap_minkowski_path_int ] ) and ( [ eq : ap_minkowski_qcd_lagrangian ] ) by setting the tensors to their euclidean versions , taking @xmath399 and @xmath400 . the partition function is @xmath401 where the euclidean action is defined as @xmath402 with the euclidean lagrangian @xmath403\\ & + \frac{1}{2}a_\mu^a(\partial_\mu\partial_\nu-\partial^2\delta_{\mu\nu})a_\nu^a\\ & + \bar{\eta}^a\partial^2\eta^a+gf^{abc}(\bar{\eta}^a\partial_\mu\eta^b)a_\mu^c\\ & + \frac{1}{2}g(\partial_\mu a_\nu^a-\partial_\nu a_\mu^a)f^{abc}a_\mu^ba_\nu^c\\ & + \frac{1}{4}g^2f^{abc}f^{ade}a_\mu^ba_\nu^ca_\mu^da_\nu^e-\frac{1}{2\xi}(\partial_\mu a_\mu^a)^2.\end{aligned}\ ] ] in the above the tensors are all euclidean but the index @xmath11 has been dropped for convenience , the conjugate fields are now defined as @xmath404 and @xmath405 and the chemical potential has been added as the lagrange multiplier of the density @xmath406 . additionally , to complete the connection between qft and sft , the time integration domain is changed to @xmath407 $ ] , where @xmath408 is the inverse temperature @xmath1 of the system . due to the definition of the partition function as a trace over all states , the bosonic ( fermionic ) fields are required to obey symmetric ( anti - symmetric ) boundary conditions , namely @xmath409 . at nonzero temperature one needs to calculate matsubara summations which usually can be done using contour integration . in the following an expression is derived for such summations and the interpretation of the result is discussed . consider to this end the sum over bosonic matsubara frequencies ( @xmath410 ) of the function @xmath411 @xmath412 where the contour @xmath413 is given in fig . [ fig : ap_bose_matsubara_contours ] . along the imaginary axis is used to rewrite the frequency summation as a contour integral . the contours @xmath414 are the positive and negative energy semicircular contours obtained by adding arcs at infinity to @xmath413.[fig : ap_bose_matsubara_contours ] ] if @xmath415 when @xmath416 then it is possible to close the contour by adding the arcs of @xmath414 , in which case @xmath417 . in general @xmath418 will only have poles or branch cuts on the real axis , such that the contours can be contracted along the real axis , which gives @xmath419\mathrm{d}\omega\nonumber\\ & -\frac{1}{\pi}\int_0^\infty[1+n_b(\omega)]\im[f(-\omega+i0)]\mathrm{d}\omega,\label{eq : ap_matsubara_bosons}\end{aligned}\ ] ] where @xmath420\equiv f(\omega+i0)-f(\omega - i0)$ ] and it was used that @xmath421 for the contour @xmath422 , such that it is more convenient to take @xmath423 and use @xmath424 . the first line can be interpreted as due to thermal gluons and the second line due to thermal ( anti)gluons and a vacuum contribution . the fact that the gluon is its own antiparticle will be reflected by @xmath425=-\im[f(\omega+i0)]$ ] . let us evaluate the above for the specific case @xmath426 $ ] which satisfies @xmath427=-\im[g(\omega+i0)]$ ] . expanding the logarithm in terms of its real and imaginary parts @xmath428=&\ln|g(z)|+i\arctan\left(\frac{\im[g(z)]}{\re[g(z)]}\right)\\ & + i\pi\theta(-\re[g(z)])\mathrm{sgn}(\im[g(z)]),\end{aligned}\ ] ] then eq . ( [ eq : ap_matsubara_bosons ] ) can be written as @xmath429\\ = & \frac{1}{\pi}\int_0^\infty\mathrm{d}\omega(1 + 2n_{b}(\omega))\left[\arctan\left(\frac{\im\left[g(\omega_+)\right]}{\re\left[g(\omega_+)\right]}\right)\right.\\ & \hphantom{\times}\left.+\pi\theta(-\re[g(\omega_+)])\mathrm{sgn}(\im[g(\omega_+)])\vphantom{\arctan\left(\frac{\im\left[g(\omega_+)\right]}{\re\left[g(\omega_+)\right]}\right)}\right],\end{aligned}\ ] ] which will be used in the calculation of the rpa correction to the thermodynamic potential . for fermionic matsubara frequencies a similar derivation can be done , but now the chemical potential is included by writing @xmath430 @xmath431 where the contour @xmath413 is given in fig . [ fig : ap_fermi_matsubara_contours ] and @xmath432 has poles at the fermionic matsubara frequencies with residue @xmath433 . if @xmath434 when @xmath416 , then it is possible to close the contour as given in fig . [ fig : ap_fermi_matsubara_contours ] by adding the arcs of @xmath414 and the lower- and upper - boundaries of @xmath435 . the full matsubara sum can thus be written as a contour integral over @xmath436 . again if @xmath418 only has poles and branch cuts on the real axis , contracting the contours along the real axis gives @xmath437\mathrm{d}\omega\nonumber\\ & -\frac{1}{\pi}\int_0^\mu[1-n(\mu-\omega)]\im[f(\omega+i0)]\mathrm{d}\omega\nonumber\\ & -\frac{1}{\pi}\int_0^\infty[1-n(\omega+\mu)]\im[f(-\omega+i0)]\mathrm{d}\omega.\end{aligned}\ ] ] where in the first - to - last line it was used that @xmath438 for the contour @xmath439 , such that it is more convenient to write @xmath440 . similarly for @xmath441 , where @xmath421 , we take @xmath423 in the integral and write @xmath442 . from the above it is clear that the first line corresponds to particles above the fermi sphere , the second line to those in the fermi sphere and the last line is due to anti - particles and contains a vacuum contribution . note that the above formula must be used with care , since one has to check if the above simplifications are valid on a case - to - case basis . but for the fermionic case , where @xmath435 , the contour between @xmath438 , is related to processes inside the fermi sphere.[fig : ap_fermi_matsubara_contours ] ] the gluon self - energy , or the so - called polarization tensor , is given to the lowest order by @xmath443\nonumber\\ & \equiv\delta^{ab}\pi_{f,\mu\nu}.\end{aligned}\ ] ] the polarization tensor is purely transverse , i.e. , @xmath444 , which can be easily seen using @xmath445 and the cyclicity of the trace . expanding the polarization tensor in terms of the longitudinal and transverse parts relative to the three - momentum @xmath288 gives @xmath446 where @xmath447 and @xmath448 can be obtained by @xmath449 noticing that even in the imbalanced case , the contribution from each flavor simply adds up to the total polarization tensor , therefore , throughout this section , we will not explicitly write out the subscript @xmath104 , and all the formulae presented here are applicable to any flavor . however , in the main text , we use @xmath196 , @xmath194 and @xmath195 as the summation over all flavors if the subscript @xmath104 is not explicitly shown . after performing the matsubara sum , the individual processes involved can be identified @xmath450}{i\omega_q-\epsilon_0(\mathbf{p}+\mathbf{q})+\epsilon_0(\mathbf{p})}[n(\mathbf{p}+\mathbf{q})(1-n(\mathbf{p}))-(1-n(\mathbf{p}+\mathbf{q}))n(\mathbf{p})]\right.\nonumber\\ & \qquad\qquad\quad\left.+\frac{\sum_{ss'}[\bar{v}_s(\mathbf{p})\gamma_{\mu}u_{s'}(\mathbf{p}+\mathbf{q})\bar{u}_{s'}(\mathbf{p}+\mathbf{q})\gamma_{\nu}v_s(\mathbf{p})]}{i\omega_q-\epsilon_0(\mathbf{p}+\mathbf{q})-\epsilon_0(\mathbf{p})}[n(\mathbf{p}+\mathbf{q})\bar{n}(\mathbf{p})-(1-n(\mathbf{p}+\mathbf{q}))(1-\bar{n}(\mathbf{p}))]\right.\nonumber\\ & \qquad\qquad\quad\left.+\frac{\sum_{ss'}[\bar{u}_s(\mathbf{p})\gamma_{\mu}v_{s'}(\mathbf{p}+\mathbf{q})\bar{v}_{s'}(\mathbf{p}+\mathbf{q})\gamma_{\nu}u_s(\mathbf{p})]}{i\omega_q+\epsilon_0(\mathbf{p}+\mathbf{q})+\epsilon_0(\mathbf{p})}[(1-\bar{n}(\mathbf{p}+\mathbf{q}))(1-n(\mathbf{p}))-\bar{n}(\mathbf{p}+\mathbf{q})n(\mathbf{p})]\right.\nonumber\\ & \qquad\qquad\quad\left.+\frac{\sum_{ss'}[\bar{v}_s(\mathbf{p})\gamma_{\mu}v_{s'}(\mathbf{p}+\mathbf{q})\bar{v}_{s'}(\mathbf{p}+\mathbf{q})\gamma_{\nu}v_s(\mathbf{p})]}{i\omega_q+\epsilon_0(\mathbf{p}+\mathbf{q})-\epsilon_0(\mathbf{p})}[(1-\bar{n}(\mathbf{p}+\mathbf{q}))\bar{n}(\mathbf{p})-\bar{n}(\mathbf{p}+\mathbf{q})(1-\bar{n}(\mathbf{p}))]\right\}\nonumber\\ = & \frac{g^2}{2}\sum_{s_1,s_2=\pm1}\int\frac{\mathrm{d}^3p}{(2\pi)^3}\frac{\mathrm{tr}[p_e^{s_1}(\mathbf{p})\gamma_0\gamma_{\mu}p_e^{s_2}(\mathbf{p}+\mathbf{q})\gamma_0\gamma_\nu]}{i\omega_q - s_2\epsilon_0(\mathbf{p}+\mathbf{q})+s_1\epsilon_0(\mathbf{p})}[n^{s_2}(\mathbf{p}+\mathbf{q})(1-n^{s_1}(\mathbf{p}))-(1-n^{s_2}(\mathbf{p}+\mathbf{q}))n^{s_1}(\mathbf{p})],\end{aligned}\ ] ] where @xmath451}$ ] , @xmath452 is the fermi distribution of the corresponding flavor and @xmath453 + 1}=1-n^-(\mathbf{p})$ ] is the fermi distribution for the corresponding anti - particle . the first line is related to particle - hole creation and annihilation , the second and third line to particle - antiparticle ( and hole - antihole ) production and annihilation , and the last line to antiparticle - antiparticlehole creation and annihilation . note that , as expected , the first line vanishes in the @xmath454 limit and the last line vanishes in the @xmath455 limit . the second and third line contain infinite vacuum ( @xmath456 ) contributions , which need to be renormalized . the gluon self - energy will be decomposed in terms of a matter part , a renormalized vacuum and an infinite vacuum contribution according to @xmath457 where the various functions have been defined as @xmath458,\\ \pi_{\mu\nu}^{\textrm{inf.vac.}}&\equiv\frac{p_{\mu\nu}}{3}q^2\left[\frac{\pi_{\lambda\lambda}(q,0,0)}{q^2}\right]_{q^2\rightarrow0}.\end{aligned}\ ] ] for the vacuum expressions it is possible to extract the projection matrix @xmath459 , since the polarization tensor is purely transverse and for @xmath456 there is no preferential frame such that the tensor is built up out of only two possible quantities @xmath460 and @xmath461 . to include the dynamical properties of the vacuum expression a renormalization procedure is necessary . in the above the renormalized vacuum expression was obtained by extracting the infinite contributions from the vacuum polarization tensor at the renomalization point @xmath462 . the reason for extracting the factor @xmath463 from the vacuum @xmath464 before setting @xmath462 is because in the vacuum the dressed gluons are massless , i.e. , @xmath465 , c.f . ( [ eq : ap_renormalised_vacuum ] ) . therefore @xmath462 is still a pole for the vacuum expression , at which the residue is @xmath209 after the renormalization . in conclusion , the following full renormalized polarization tensor will be used @xmath466 in the high - density limit ( @xmath467 ) , a good approximation is to evaluate the matter part for the case of massless quarks . thus for massless quarks in the zero - temperature limit the matter parts of @xmath468 and @xmath469 can be calculated explicitly @xmath470\ln\left[1-\frac{2s_1s_2\mu}{q - is_1\omega_q}\right]\right.\nonumber\\ & \qquad\qquad\qquad\qquad\qquad\qquad\qquad\left.+\sum_{s_2=\pm1}\left[s_2\frac{\mu(3q^2 - 4\mu^2)}{24q}-\frac{\mu^2}{4}\frac{i\omega_q}{q}\right]\ln\left[\frac{1-\frac{2s_2\mu}{q - i\omega_q}}{1+\frac{2s_2\mu}{q+i\omega_q}}\right]\right\},\label{eq : ap_pi_mat_00}\\ \pi_{\mu\mu}^{\mathrm{mat}}(t=0)=&\frac{g^2}{\pi^2}\left[\frac{\mu^2}{2}-\frac{q^2}{8q}\left\{\mu\sum_{s_1=\pm1}s_1\ln\left[\frac{1+\frac{2s_1\mu}{q - i\omega_q}}{1-\frac{2s_1\mu}{q+i\omega_q}}\right]+\frac{1}{2}\sum_{s_1,s_2=\pm1}(q - is_1\omega_q)\ln\left[1+\frac{2s_1s_2\mu}{q - is_1\omega_q}\right]\right\}\right].\label{eq : ap_pi_mat_mumu}\end{aligned}\ ] ] these expressions give the same result as can be found in refs . @xcite . performing an analytic continuation to real time ( @xmath471 ) , the above expressions have branch cuts from @xmath472 and @xmath473 , the origins of which can be found from eq . ( [ polarizationtensor ] ) and are due to particle - hole processes and finite-@xmath474 contributions to particle - antiparticle production and annihilation , respectively . because lorentz invariance is broken due to the presence of the gas , @xmath475 is a function of @xmath476 and @xmath218 separately since it can be a function of @xmath477 and @xmath478 , where @xmath479 defines the rest frame of the system @xcite . the renormalized vacuum part @xmath480 has been found in ref . @xcite @xmath481\nonumber\\ \simeq&p_{\mu\nu}\frac{g^2q^2\left[\frac{5}{3}+\ln\left(\frac{m^2}{q^2}\right)\right]}{24\pi^2}.\label{eq : ap_renormalised_vacuum}\end{aligned}\ ] ] where it was assumed that the quarks have equal masses and in the last line it was expanded for small masses . in real time this expression has a branch cut for @xmath482 , which is due to particle - antiparticle production and annihilation . using a sommerfeld expansion temperature corrections to the gluon self - energy can be obtained . the sommerfeld expansion can be summarized as @xmath483 the derivation is analogous to that of the non - relativistic sommerfeld expansion @xcite . one has to keep in mind that the above is not a complete expansion for small temperatures since also the chemical potential depends on temperature . in this manner the @xmath484 correction to the gluon self - energy is found to be @xmath485,\nonumber\\ \pi_{\mu\mu}^{\mathrm{mat}}-\pi_{\mu\mu}^{\mathrm{mat}}(t=0)=&-\frac{g^2}{2}\frac{1}{6}t^2\frac{8\mu^2[3(i\omega_q)^2-(2\mu - q)(2\mu+q)]}{[(i\omega_q)^2-(2\mu - q)^2][(i\omega_q)^2-(2\mu+q)^2]}.\label{eq : ap_pi_temperature_correction}\end{aligned}\ ] ] using eqs . ( [ fg]-[eq : ap_pi_temperature_correction ] ) the behavior of @xmath194 and @xmath195 can be found in several limits . keeping the ratio @xmath486 fixed and expanding up to zeroth order in @xmath219 , the hard dense and hard thermal loop ( hdl / htl ) expressions are re - obtained @xcite @xmath487\right|_{x = q/\omega},\\ \lim_{\omega\rightarrow0}&g(\omega , q = x\omega)\\=&m_g^2\left.\left[\frac{1}{x^2}-\frac{1}{2x}\left(1-\frac{1}{x^2}\right)\ln\left(\frac{1-x}{1+x}\right)\right]\right|_{x = q/\omega},\end{aligned}\ ] ] where @xmath488 is the gluon thermal mass . the reason that the above limit returns the hdl and htl expressions , is due to the fact that the hdl / htl approximation ( @xmath489 and @xmath490 ) takes into account only the low - energy processes around the fermi surface , namely particle - hole processes . some other useful limits are @xmath491 where it can be seen that in the static long - wavelength limit ( @xmath492 ) the electric gluons ( @xmath194 ) are screened , while magnetic gluons ( @xmath195 ) are dynamically screened . furthermore , the large momenta and frequency behavior of the matter and vacuum parts are separately seen to be @xmath493}{24\pi^2}.\label{eq : ap_fg_large_pomega_behavior}\end{aligned}\ ] ] the vacuum expressions are proportional to @xmath494 since in the vacuum the gluons remain massless and thus @xmath495 is still a pole of the propagator . the logarithmic @xmath494 dependence is usually absorbed into the vertex and subsequently interpreted as the varying of the coupling constant with the energy scale @xmath226 . consider first the well - known ideal fermi gas term of a single species @xmath496\\ & = -\frac{1}{\beta}\sum_{\omega_n}\int\frac{\mathrm{d}^3p}{(2\pi)^3}2\ln[-(i\omega_n+\mu_0)^2+\epsilon_0 ^ 2(\mathbf{p})].\end{aligned}\ ] ] using eq . ( [ eq : ap_matsubara_fermions ] ) with @xmath497 $ ] and using the principle value logarithm with a branch cut on the negative real axis @xmath498&=\lim_{\eta\downarrow0}\im[\ln(-\omega^2+\epsilon_0 ^ 2(\mathbf{p})-2i\omega\eta)]\\ & = \pi\theta[\omega^2-\epsilon_0 ^ 2(\mathbf{p})]\mathrm{sgn}(-\omega),\end{aligned}\ ] ] the ideal fermi gas contribution becomes @xmath499\\ & \qquad\qquad\qquad\times\theta[\omega^2-\epsilon_0 ^ 2(\mathbf{p})]\mathrm{d}\omega\\ = & -2\beta^{-1}\int\frac{\mathrm{d}^3p}{(2\pi)^3}\left\{\ln\left[1+e^{-\beta(\epsilon_0(\mathbf{p})-\mu_0)}\right]\right.\\ & \qquad\left.+\ln\left[1+e^{-\beta(\epsilon_0(\mathbf{p})+\mu_0)}\right]\right\}+2\int\frac{\mathrm{d}^3p}{(2\pi)^3}\int_{\epsilon_0 ^ 2(\mathbf{p})}^{\infty}d\omega.\end{aligned}\ ] ] in the above it was assumed that the arc at infinity vanishes , which is the case if the time - ordering in the path - integral is taken into account properly by multiplying the integrand by @xmath500 and taking the limit @xmath501 . the infinite vacuum contribution is clearly visible and should be subtracted , giving for massless fermions @xmath502.\ ] ] the ideal bose gas follows from @xmath503,\ ] ] which will lead to the same expression in both the lorentz and coulomb gauges . in the following , the coulomb gauge is taken ( @xmath504 ) , @xmath505 - 2\mathrm{tr}\ln q^2-\mathrm{tr}\ln\frac{1}{\xi}\right)\\ & = \frac{1}{2v\beta}\sum_{\omega_n , q}\left(\ln\left[\frac{q^2}{\xi}q^2(q^2)^2\right]-2\ln q^2-\ln\frac{1}{\xi}\right)\\ & = \frac{1}{v\beta}\sum_{\omega_n , q}\ln[-(i\omega_n)^2+q^2].\end{aligned}\ ] ] the unphysical degrees of freedom clearly drop out due to the ghost contribution @xmath506 . using eq . ( [ eq : ap_matsubara_bosons ] ) with @xmath507 , the ideal bose gas contribution is @xmath508-\int\frac{\mathrm{d}^3q}{(2\pi)^3}\int_q^\infty\mathrm{d}\omega.\ ] ] again the vacuum term is clearly present and will be subtracted , giving the stefan - boltzmann law @xmath509 for completeness also the quark self - energy will be derived in the lorentz gauge to lowest order . @xmath510.\ ] ] performing the matsubara sum gives @xmath511}{i\omega_p+\mu-\epsilon_0(\mathbf{q})-\epsilon^g(\mathbf{p}-\mathbf{q})}[(1+n_b(\mathbf{p}-\mathbf{q}))(1-n(\mathbf{q}))+n_b(\mathbf{p}-\mathbf{q})n(\mathbf{q})]\right.\nonumber\\ & \qquad\qquad\quad\left.+\frac{\sum_s[\gamma_\mu u_s(\mathbf{q})\bar{u}_s(\mathbf{q})\gamma_\mu]}{i\omega_p+\mu-\epsilon_0(\mathbf{q})+\epsilon^g(\mathbf{p}-\mathbf{q})}[n_b(\mathbf{p}-\mathbf{q})(1-n(\mathbf{q}))+(1+n_b(\mathbf{p}-\mathbf{q}))n(\mathbf{q})]\right.\nonumber\\ & \qquad\qquad\quad\left.+\frac{\sum_s[\gamma_\mu v_s(\mathbf{q})\bar{v}_s(\mathbf{q})\gamma_\mu]}{i\omega_p+\mu+\epsilon_0(\mathbf{q})-\epsilon^g(\mathbf{p}-\mathbf{q})}[(1+n_b(\mathbf{p}-\mathbf{q}))\bar{n}(\mathbf{q})+n_b(\mathbf{p}-\mathbf{q})(1-\bar{n}(\mathbf{q}))]\right.\nonumber\\ & \qquad\qquad\quad\left.+\frac{\sum_s[\gamma_\mu v_s(\mathbf{q})\bar{v}_s(\mathbf{q})\gamma_\mu]}{i\omega_p+\mu+\epsilon_0(\mathbf{q})+\epsilon^g(\mathbf{p}-\mathbf{q})}[n_b(\mathbf{p}-\mathbf{q})\bar{n}(\mathbf{q})+(1+n_b(\mathbf{p}-\mathbf{q}))(1-\bar{n}(\mathbf{q}))]\right\},\nonumber\\ \delta\sigma(i\omega_p,\mathbf{p})=&-g^2\frac{n_g}{4n_c}\sum_{s_2,s_3=\pm1}\int\frac{\mathrm{d}^3q}{(2\pi)^3}\left[\frac{s_3}{\epsilon^g(\mathbf{p}-\mathbf{q})}\frac{\gamma_{\mu}p_e^{s_2}(\mathbf{q})\gamma_0\gamma_\mu}{i\omega_p+\mu - s_2\epsilon_0(\mathbf{q})-s_3\epsilon^g(\mathbf{p}-\mathbf{q})}\right]\nonumber\\ & \qquad\qquad\qquad\qquad\qquad\times[(1+n_b^{s_3}(\mathbf{p}-\mathbf{q}))(1-n^{s_2}(\mathbf{q}))+n_b^{s_3}(\mathbf{p}-\mathbf{q})n^{s_2}(\mathbf{q})-i_{s_2,s_3}].\label{eq : ap_quark_self_en}\end{aligned}\ ] ] in the last line the indicator function @xmath512 is exactly the vacuum contribution such that @xmath513 is the vacuum subtracted self - energy , i.e. , @xmath514 . to first order the self - energy for a single quark and antiquark can be found by using the free quark energy projectors and helicity projectors @xmath515.\ ] ] after projecting and evaluating the self - energy at @xmath516 the function in the first square brackets in eq . ( [ eq : ap_quark_self_en ] ) reduces to @xmath517 defined in eq . ( [ functionf ] ) . in the zero - temperature limit the self - energy for a single quark then reduces to @xmath518 note that the final result is independent of helicity @xmath15 .

when customers develop a persistent loyalty to one group of competing suppliers , or when suppliers target a specific segment of their potential customer base , we regard this as an instance of segregation of traders . in each case the interaction is between fixed ( or largely fixed ) groups of buyers and sellers and each group is segregated into subgroups : buyers have specific needs and these are met by specific products or suppliers . such segregation can confer benefits , e.g. shorter exploration time when buying , but it may also add vulnerability to a system , for example by reducing market liquidity . segregation - like phenomena have been observed in a number of studies ; examples we are aware of include emergent loyalty in a fish market @xcite , segmentation in credit markets in the philippines @xcite , persistence in customer visitation patterns @xcite and specialization and herding in financial markets @xcite . some of the previous work has focused on the buyer side , and has studied how segmentation can be identified in the space of customers , and how suppliers may respond to it adequately once it is detected @xcite . this branch of work can be thought of as a continuation of adam smith s idea that specialization of production leads to better exploitation of the market @xcite . apparent paradoxes have also emerged , such as _ hotelling s law _ @xcite : contrary to popular belief rational trading strategies may lead to minimal differentiation . in hotelling s original example sellers that minimise their total distance from all potential customers , who are uniformly positioned along the beach , end up positioning themselves all in the same place in the middle of the beach . much less work has been done to explore the conditions under which segregation occurs in markets , and what factors promote or suppress it . when the buyer and seller populations are heterogeneous due to external factors such as geographical location or individual wealth , this can be a natural trigger for segregation . here , our aim is to investigate whether and when segregation can emerge _ spontaneously _ in initially homogeneous populations . in an _ in silico _ experiment based on cat market design tournaments @xcite such segregated states were observed in the co - adaptation of markets and traders . some markets ended up focusing on attracting specific traders , and some traders preferred to trade at a specific market although no such fixed propensities were imposed from the beginning . outside the context of double auction markets , an effect similar to segregation was noticed in multi - asset minority games @xcite . specifically , in a system of agents who have the option of copying a winning ( minority ) strategy from their neighbourhood with some probability @xmath0 , a self - organized grouping of strategies arises when the probability @xmath0 is large enough . grouping was however only found when the underlying interaction network is well connected , and when agents have perfect information about the success of their neighbours . in another study based on the minority game @xcite it was noted that `` ... a population of competing agents with similar capabilities and knowledge will tend to self - segregate into opposing groups characterized by extreme behavior '' . however , the variant of the minority game used in this study supposes that agents can choose to follow a global strategy based on collective memory ; this is the primary driver of segregation . our approach is simpler in that the agents learn entirely from their own market returns and the only interaction between agents is via the market . while the cat tournament results referred to above provide a clear impetus to study spontaneous segregation , they are difficult to analyse theoretically due to the complexity of both the agents trading strategies and the market mechanisms of the competing markets . we therefore devise and analyse a stylized individual - based model of a double auction . we focus on agents using experience - weighted attraction learning ( ewa ) @xcite to learn from the payoffs received for past actions and thus optimise future trading . this general framework is similar to the one employed by hanaki et al @xcite , though in a rather different context . hanaki et al show that when agents are repeatedly given a choice of parking spots , some _ learn to be lucky _ by focussing their search on to the spots very close to the city center , while the others end up with the opposite strategy . by developing a stylized model we aim to obtain an intuitive and analytically tractable tool for understanding whether and when segregation can emerge spontaneously in a system of competing double auction markets . in the results section we first describe the model and then provide numerical results , summarizing the main findings from our numerical simulations . in the following subsection we develop an analytical description in the large market limit , using a fokker - planck approach to investigate the steady states of the system mathematically and in particular to develop insights into the properties of the segregated state . in the discussion section we summarize our main findings , discuss their robustness to e.g. variations in the model and set out possible directions for future research . to the best of our knowledge , there is no other research that models spontaneous segregation in the context of double auction markets . to address the question of spontaneous segregation , we study a stylized model of a population of adaptive traders and two double auction markets . we hypothesize that segregation can arise as a product of co - adaptation of traders , and construct on this basis a minimal model for both traders and markets . we investigate this in detail by numerical simulation , and provide a full theoretical understanding and characterisation of the observed segregation effects . a brief summary of the basic features of the model and initial simulation results can be found in @xcite . following the works of gode and sunder @xcite and ladlay @xcite , we populate our system with agents without sophisticated trading strategies , essentially zero - intelligence traders . this is done because our aim is to investigate whether market loyalty can arise as an intrinsic property of a system of interacting agents without reliance on complex trading strategies . trading operates in discrete rounds , where at each turn the prices of buy and sell orders ( bids and asks ) that the traders submit to a market are independent identically distributed random variables ( thus zero - intelligence agents ) . agents learn whether to buy or to sell from previous experience ( see below ) . the numerical values of bids and asks ( i.e their magnitudes ) , however , are assumed to be unrelated to previous trading success or any other information . we assume that bids , @xmath1 , and asks , @xmath2 , are normally distributed ( @xmath3 and @xmath4 ) , and that their means satisfy @xmath5 . in the spirit of the work of @xcite , @xmath6 can be thought of as cost and redemption values for each trader . these values are private to each trader but correspond to order prices if submitted bids and asks are truthful and reflect the actual valuation of goods by the agents . traders will not accept any price from the market : sellers will trade only if the trading price @xmath7 is no less than their asking price ( @xmath8 ) , and buyers require that the trading price is no greater than what they bid ( @xmath9 ) . ( this is in line with gode and sunder @xcite where the traders that turned out to be more similar to human traders were the zero - intelligence ones with budget constraint , i.e. traders who were not allowed to trade at loss higher than the redemption value for buyers and lower than the cost value for sellers . ) after each round of trading each agent receives a score , reflecting their payoff in the trade . scores of agents who do trade are as follows : buyers value paying less than they offered ( @xmath1 ) , and so their score is @xmath10 . sellers value trading for more than their ask ( @xmath2 ) , and so @xmath11 is a reasonable model for their payoff . we note that these scores are based on a linear model and do not reflect effects such as diminishing returns . traders who do not get to trade in a given round receive return @xmath12 . the role of a market is to facilitate trades so we define markets in terms of their price - setting and order matching mechanisms ; for an in - depth review of possible double auction market mechanisms see @xcite . we consider a single - unit discrete time double auction market where all orders arrive simultaneously and market clearing happens once every period after the orders are collected . ( note that each period consist of one round only , which is why we will talk only in terms of periods . ) we also assume that a uniform price is set by the market once all orders have arrived , these are used to determine demand and supply ( see fig . [ fig1:marketmech ] ) . the price at which demand equals supply is the equilibrium price . to make the model more flexible , we consider more generally uniform prices of the type @xmath13 where @xmath14 is the equilibrium trading price , from which the market price can deviate towards the average bid ( @xmath15 ) or the average ask ( @xmath16 ) ; the parameter @xmath17 thus represents the bias of the market towards buyers or sellers . in a setup like ours where the bids and asks are gaussian random variables , the equilibrium trading price is @xmath18 when the variances are equal as we assume here and below . once the trading price has been set , all bids below this price , and all asks above it , are marked as invalid orders as they can not be executed at the current trading price . the remaining orders are executed by randomly pairing buyers and sellers ; the execution price is @xmath7 . note that we assume here that each order is for a single unit of the good traded . an example of a discrete time double auction market is the opening auction of new york stock exchange which is used to determine the opening prices on the market . at the opening auction , once all the orders are submitted , the trade occurs at the single price set by the market that maximises the volume of trades . in the terminology of our model , this price corresponds to @xmath19 as that setting is the one that maximises the number of possible trades . ; the dashed line is the trading price for @xmath20 . right : example distributions of returns from validated orders at @xmath20 . in the situation shown , where buyers are in the minority after validation of orders , all buyers find a trading partner ( red ) . if all sellers could trade they would achieve the `` desired '' return distribution ( light blue ) . but only a randomly selected fraction finds a trading partner , leading to the `` realized '' return distribution ( dark blue ) ; the other sellers receive a return of zero . the realized sellers return distribution has the same total area as the buyers return distribution as both represent the total number of trades . ] [ fig1:marketmech ] shows graphically how the trading price is set and how the return distributions look after invalid bids and asks are eliminated . if , as in the example in the figure , the trading price is closer to the average bid ( @xmath21 ) there will be an excess of supply : only a fraction of the sellers who submitted valid asks gets to trade , but with the higher trading price so those that do trade receive higher returns . the buyers who submitted valid bids are in a minority compared to the sellers and so determine the number of trades . they can all trade , but receive relatively low returns on account of the high trading price . so far we have described how agents interact at a given market . we next define how they decide _ how to trade _ ( to buy or to sell ) , and _ where to trade _ ( at which market ) . agents trade repeatedly in our model , and they adapt their preferences for the various choices from one trading period to the next . we focus on the case of two markets , though the model can easily be extended to an arbitrary number of markets . we assume that each agent decides on an action at the beginning of each trading period , only based on his or her past experience . to formalize this we introduce a set of attractions @xmath22 for each player , one for each action @xmath23 : buy at market 1 , sell at market 1 , buy at market 2 , sell at market 2 . the attractions will generally differ from player to player , but we suppress this in the notation for now . the attractions are updated after every trading period , @xmath24 , using the following reinforcement rule : @xmath25 the quantity @xmath26 is the score gained by taking action @xmath23 in the @xmath24-th trading period . the length of the agents memory is set by @xmath27 : agents weight returns obtained @xmath28 time steps ago with exponentially decaying factors @xmath29 , effectively corresponding to a sliding window of length of order @xmath30 for the weighted averaging of past returns . the update rule above is a special instance of a more general experience - weighted attraction rule @xcite , which has been shown to be in reasonable agreement with experimental data on human learning in repeated games . many special cases of this rule are in common use in evolutionary biology and in the game - theoretic literature . one important variant of ewa is a case in which _ all _ actions are updated with their returns , no matter whether that action was actually taken or not . this assumes that an agent can calculate or at least estimate the return ( s)he would have obtained from actions @xmath23 that ( s)he did not choose to play @xcite . we argue that an agent would not normally have sufficient information to do this in the context of a double auction market : ( s)he would need access to the current price , and to the numbers of valid bids and asks submitted . this is unrealistic , which is why we posit that scores of unplayed actions are updated with an effective payoff of zero . one plausible alternative that does _ not _ rely on estimation of returns from unplayed actions would be an update rule where the attractions to unplayed actions are forgotten with a different rate ( or not forgotten at all ) ; we return to this in the discussion at the end . it only remains to specify how agents choose their actions based on the attractions . this is done in line with the experience - weighted attraction literature @xcite , simply by converting the attractions to probabilities using the so - called _ softmax _ or _ logit _ function . explicitly , each agent takes action @xmath23 in trading period @xmath24 with probability @xmath31 , where @xmath32 is the _ intensity of choice _ @xcite and regulates how strongly the agents bias their preferences towards actions with high attractions . for @xmath33 the agents choose the option with the highest attraction , while for @xmath34 they choose randomly with equal probabilities among all options . the parameter @xmath32 is sometimes denoted @xmath35 in the literature and has been referred to for example as the _ response sensitivity _ @xcite and the _ learning sensitivity _ @xcite . we begin our exploration of the two - market setup defined above with results from numerical simulations . unless specified otherwise , parameters were set as detailed in table [ parametertable ] in the methods section . all agents start with equal preferences for all four choices , i.e. @xmath36 where @xmath37 buy and @xmath38 sell . to aid visualization we project the four - dimensional space of attractions @xmath22 down to two coordinates , the overall attraction to buying as against selling , defined as @xmath39 , and the attraction to market 1 as against market 2 , @xmath40 . due to the nonlinearity of the _ softmax _ function , a single - peaked distribution of attractions with a non - zero spread will , for large enough intensity of choice @xmath32 , become a multimodal distribution in the space of preferences . this effect can not be regarded as genuine segregation . for this reason we will avoid representing agents in the space of their preferences @xmath41 , and use the underlying attractions instead . in fig . [ fig2:simulation ] ( left ) we present the steady state attraction distribution for a population of traders with intensity of choice @xmath42 . the initially narrow , _ delta - peaked _ distribution of attractions ( all initialized at @xmath43 ) has been broadened due to diffusion arising from the random nature of returns and from the stochasticity of the agents actions . the steady state shown in the left - hand panel of fig . [ fig2:simulation ] represents an unsegregated population of traders . while this population does include some traders with moderately strong preferences for one of the actions , preferences remain weak on average . the population as a whole remains homogeneous in the sense that there is no split into discernible groups . . the colour scale in the central scatter plots shows the probability density of agents.@xcite for low enough intensity of choice ( left , @xmath42 ) , an unsegregated steady state arises , while higher @xmath32 ( right , @xmath44 ) causes traders to segregate into four discernible groups . the histograms along the axes show the marginal distributions of @xmath45 and @xmath46 , respectively . from the marginals we note that in the segregated state roughly half of the population specialize in buying ( similarly half prefers market 1 ) , although the fractions of buyers ( sellers ) at different markets are not the same . other system parameters are given in table [ parametertable ] . the data shown are taken from 100 trading periods of 100 independent runs . ] [ fig2:simulation ] ( right ) contrasts this scenario with the steady state of a system with exactly the same set of parameters but at the higher intensity of choice @xmath44 . the population of traders now splits into four groups , with the agents _ persistently trading _ preferentially at one of the markets and with a preferred buy or sell action . we refer to this state as _ segregated_. the markets shown in this example ( fig . [ fig2:simulation ] ) are biased symmetrically ( @xmath47 ) : if an agent buys at market 1 , or sells at market 2 ( actions @xmath48 or @xmath49 ) , and if they manage to trade , ( s)he is awarded with a higher score on average than if they were to use the other two actions ( c.f . [ fig1:marketmech ] , where sellers are the group with the higher average return ) . in this symmetric market setup there are therefore two different kinds of behaviour among the four segregated groups . the agents in two of the groups ( @xmath48 and @xmath49 ) specialize in actions that award them higher average return , we call those _ `` return - oriented''_. we will refer to the agents in the other two groups ( @xmath50 or @xmath51 ) as__``volume - oriented '' _ _ for reasons which we explain shortly . if all traders were return - oriented , they would have no partners to trade as there would be agents playing @xmath50 or @xmath51 . consequently everyone would receive zero returns . the fact that some traders have developed persistent preferences for placing orders that give them a lower average return ( here @xmath50 , @xmath51 ) can thus be interpreted as cooperative , trade - enabling behaviour . due to the market mechanism , the agents who choose one of @xmath50 or @xmath51 will have their orders rejected as invalid more often ( see the group of buyers in fig . [ fig1:marketmech ] ) . however , those whose orders are accepted will form a minority group in the market and will always find a trading partner . hence , when submitting valid orders , these traders trade more often so we will call them _ `` volume - oriented''_. there is a tendency for agents to cluster around the diagonal and anti - diagonal in the two - dimensional projection shown in fig . [ fig2:simulation ] . inspection of the underlying attractions @xmath22 shows that this arises because many agents have one large attraction , for their preferred action , while their other attractions are close to zero . for example if a return - oriented agent , say a buyer at market 1 , has attractions @xmath52 , then this projects to the diagonal @xmath53 . similarly a volume - oriented trader who prefers to sell at market 1 is projected to the anti - diagonal @xmath54 if her other three attractions are close to zero . the fact that the attractions of non - preferred actions are often small comes from the fact that , within our reinforcement learning dynamics , agents gradually forget the scores of actions they only use rarely . + [ [ persistence - and - time - correlation . ] ] persistence and time correlation . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + the above results indicate that segregation is seen above a critical intensity of choice , @xmath55 ; we defer a discussion of how @xmath56 depends on the parameters of the model to the section analytical description below . one has to bear in mind though that the data shown in fig . [ fig2:simulation ] represents the state of the system at a given time . it therefore does not tell us whether agents really develop loyalty in the sense that they stay in one of the segregated groups for a long time , or whether they switch frequently between groups . to understand which of these two alternatives applies , and to characterise the dynamics of group switching events , one can analyse the persistence times of agents within each of the four peaks in fig . [ fig2:simulation ] . consistent with the intuitive meaning of a segregated state one finds @xcite that at high enough intensity of choice @xmath32 the agents develop `` loyalty '' to a certain market and a choice of buying and selling : their persistence times are much longer than the timescale of the small short - term fluctuations in preferences that every agent experiences . these short - term fluctuations occur on the time scale of the memory - loss , @xmath30 . on the other hand the agents are _ not _ frozen , i.e. the persistence times are finite and the agents change loyalties on longer timescales . therefore the steady state we observe is well defined rather than a consequence of agent preferences frozen - in from early fluctuations in their trading history . and forgetting rate @xmath27 are different from the standard values in table [ parametertable ] : here @xmath57 , @xmath27 as shown in the legend . this is to highlight the separation of timescales in the segregated state . a larger than usual initial number of trading periods was simulated ( @xmath58 trading periods ) to ensure the systems reached steady state at such high @xmath32 . ] what emerges from the above discussion is that a key feature of a segregated state is a separation of timescales in the dynamics of the agents . to quantify this further we consider the autocovariance function @xmath59 where @xmath60 indicates an average over the attractions of players in the population , and over the time @xmath61 in the stationary state . this is a matrix capturing cross - correlations in time between the attractions to the various actions an agent can take ; here as a summary statistic we look at the trace @xmath62 , i.e. the sum of the autocovariances . in fig . [ fig3:covariance ] we show how this autocorrelation function depends on time for various forgetting rates @xmath27 at fixed @xmath57 . we observe two separate decays , which is consistent with the intuition described above : first a fast decay of correlations that occurs as an agent moves around one of the peaks of the distribution @xmath63 , caused by randomness in returns at a given market and the fact that the decisions of each agent remain stochastic . this fast decay does not lead to full decorrelation but rather to a plateau in the autocorrelation function : the attractions of an agent stay in the same loyalty group " ( i.e. within the same peak of @xmath63 ) , and the agent s actions remain correlated . the decay from this plateau defines the slow timescale , and it is a measure of how long agents typically stay in one loyalty group before moving to another peak of @xmath63 . because the slow timescale increases very rapidly as @xmath27 is decreased the theoretical description below suggests exponential growth with @xmath30 we show a rather smaller range of @xmath27 in the figure than elsewhere in the paper . comparing the curves for different @xmath27 we note the increase of the initial value of the autocorrelation function with @xmath27 , which makes sense as the variance of the fluctuations within each peak of @xmath63 grows roughly proportional to @xmath27 . we also see that the fast time scale is of the order of unity in the rescaled time units , @xmath64 , used in the figure . this reflects the fact that the effective memory of each agent is @xmath30 trading periods . [ [ summary - and - characterisation - of - segregated - state . ] ] summary and characterisation of segregated state . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + summarizing so far , the existence of a plateau in the autocorrelation function is clear evidence of the segregation of agents into groups that remain loyal over extended periods of time to a certain set of preferences defining a group of agents , while at the same time exhibiting small and much faster fluctuations around these typical preference . this represents the most intuitive definition of segregation that we can think of . however , the existence of a plateau quantitatively , a turning point in a plot of autocorrelation versus log time would not be easy to utilize in practice to detect segretation . this is because already for @xmath65 the amplitude of the initial fast decay becomes so small that it merges into the plateau , while at the same time the slow timescale for switching between loyalty groups outgrows the range of times that can easily be explored computationally . for practical purposes we therefore stick to multimodality of @xmath63 as our criterion for segregation . our simulation results suggest that even our simplified trading system shows rich and interesting behaviour . a more detailed analysis ( discussed below ) reveals a threshold intensity of choice @xmath66 such that for values @xmath67 the system segregates , i.e. the initially homogeneous population of traders splits into groups that persistently choose to trade at a specific market . the pattern of these market loyalties is co - operative , in that some traders the volume - oriented ones forego potentially higher returns and instead choose actions that enable trade for everyone . we will see below that this behaviour has clear payback , in giving average returns across all traders that exceed those for the obvious nash equilibrium . an obvious question that remains open at this point is how robust the observed segregation behaviour is to variation in our model parameters and setup ; we defer this issue to the discussion section . to understand in more detail how segregation arises , and to characterise the nature of the transition to the segregated state , a theoretical analysis of the model dynamics would evidently be useful . mathematically our model is markovian . to capture the full dynamics one needs to keep track of the attractions @xmath68 to all actions @xmath69 of all agents @xmath70 , or equivalently the attraction vectors @xmath71 for all agents . one can write down a master equation for the evolution of the joint distribution of these @xmath72 variables . this equation is however difficult to work with and not easily analysed further . in order to make progress it is useful to realise that the payoff an agent receives when they take a particular action only depends on aggregated quantities , but not on the detailed actions of individual other agents . more specifically the return for a given action is determined by ( i ) the bid or ask the agent places ; ( ii ) the validity of the bid or ask placed and ( iii ) whether or not a suitable trading partner is found . all quantities ( i)-(iii ) are random objects and so the return received for a given action at any one time will be a random variable itself . properties ( ii ) and ( iii ) only depend on the macroscopic statistics of bids and asks placed by the population of traders in its entirety . we now focus on the limit of large ( formally infinite ) populations , that is we take the limit @xmath73 . the distribution of bids and asks submitted in any trading round will then follow gaussian distributions with mean @xmath74 and standard deviation @xmath75 for asks , and analogously for bids ( @xmath76 ) . by definition in our model the returns that flow from these bids or asks are non - negative , implying a truncation at zero of the return distributions . these distributions can then be written as @xmath77 the return an agent receives depends on the action @xmath23 the agent chooses , i.e. on the market , @xmath78 , they trade in , and whether they chose to buy , @xmath79 , or sell , @xmath80 . this is reflected in our notation above . the first term in each of the above expressions describes the case of a non - zero return . this occurs with probability @xmath81 , where @xmath82 denotes the probability that an agent s order is valid ; once validated , it is executed with probability @xmath83 , which is the probability of finding a suitable trading partner . if the order is executed an agent receives a return @xmath84 drawn from a gaussian distribution , truncated to allow only non - negative payoffs ; if the order was invalid , or was valid but not executed , the agent receives @xmath12 . this occurs with probability @xmath85 , to complete the above description of the single agent dynamics for large @xmath86 one needs expressions for the market price @xmath87 as well as the probabilities @xmath82 and @xmath83 . in the deterministic limit , @xmath73 , the expression in eq . ( [ price ] ) reduces to @xmath88 as the means of the bids and asks submitted in any one round become the population means @xmath89 and @xmath74 by virtue of the law of large numbers . in the case when @xmath90 , the bid and ask distributions are mirror images of each other around @xmath91 and so the quantity @xmath92 introduce above is the equilibrium price where demand equals supply . ( for @xmath93 the real equilibrium trading price is not given by this simplified formula , but in our model the market price is still calculated from average bids and asks and we retain the symbol @xmath94 for the average of the two means . ) the probabilities that an order is valid , @xmath82 , are calculated from @xmath95 these expressions reflect the requirement for a bid or ask to be on the correct side of the market price , and are based on our assumption of gaussian bid and ask distributions . the integrals can be carried out in closed form and expressed in terms of error functions . the trading probabilities @xmath96 , finally , can be written as @xmath97 where @xmath98 is the total number of agents taking an action @xmath23 and submitting a valid order , @xmath99 in this expression @xmath100 is given by the appropriate _ softmax _ function applied to an agent s vector of attractions @xmath101 . the average @xmath102 is over the distribution @xmath103 of attraction vectors across all agents . the expressions ( [ tradingp ] ) can be understood as follows : if a trader submits an order that is valid , ( s)he will always be able to trade if ( s)he is in the minority group , otherwise his probability of being able to trade is the size ratio of the minority and majority groups . we can now write an evolution equation for the distribution @xmath104 of attraction vectors across the population of traders . writing @xmath105 for the distribution at the end of trading period @xmath24 , we have @xmath106 with @xmath107 a transition kernel that encodes the dynamics of the system . it is of the form @xmath108 where @xmath109 is a four - dimensional vector with an entry of @xmath110 for action @xmath23 and entries @xmath43 otherwise . the learning rule of eq . ( [ reinfrule ] ) is enforced through the delta function . it is worth noting that ( [ mastereq ] ) is not a standard linear chapman - kolmogorov equation as the right - hand side is nonlinear in the distribution @xmath111 . this arises because the kernel @xmath112 depends on the trading probabilities @xmath83 as given in ( [ tradingp ] ) , which in turn depend on @xmath105 . the nonlinearity arises because we have effectively projected from a description in terms of all @xmath72 attractions to one involving only four single - agent attractions . from our reasoning so far , eq . ( [ mastereq ] ) should constitute an exact description of the model in the limit @xmath73 . it can , at least in principle , be solved numerically starting from our chosen initial condition @xmath113 . the presence of the @xmath114-peaks at zero returns @xmath12 makes the kernel awkward to deal with numerically , however . we therefore make one further simplification and transform to a fokker - planck description . this is appropriate for small @xmath27 , i.e. agents with long memory . note to this end that the change in attraction @xmath115 in any one trading period is directly proportional to @xmath27 , see equation ( [ kernel ] ) . the kramers - moyal expansion ( see e.g. @xcite ) of equation ( [ mastereq ] ) is of the form @xmath116 + \frac{r^2}{2}\partial_\mathbf{a}^2 \left[\mathbf{m}_2(\mathbf{a})p(\mathbf{a})\right ] + \ldots , \label{km}\ ] ] where the scaled jump moments @xmath117 are of order @xmath118 . the generic @xmath119-th order term on the rhs of equation ( [ km ] ) comes with a factor @xmath120 so for small @xmath27 one can proceed by neglecting higher - order terms beyond the first two . next , it is useful to introduce a re - scaled time @xmath121 . a unit time interval in @xmath122 then corresponds to @xmath30 trading periods and hence the memory length of the agents . with this replacement , equation ( [ km ] ) reduces to a fokker - planck equation in the limit @xmath123 . specifically we have in this limit @xmath124 + \frac{r}{2}\partial_\mathbf{a}^2\left[\mathbf{m}_2(\mathbf{a})p(\mathbf{a})\right ] . \label{adaptivefokkerplanck}\ ] ] as before and to keep the notation compact , we have not written out the various components of the derivatives and jump moments ; e.g. @xmath125 $ ] is to be read as @xmath126 $ ] . segregation behaviour can in principle be characterised by studying the steady - state solution of the above fokker - planck equation ( [ adaptivefokkerplanck ] ) . specifically one would investigate the conditions under which it is multimodal ( i.e. , it has several peaks ) . similar to the kernel @xmath112 , the drift and diffusion coefficients @xmath127 and @xmath128 depend on the trading probabilities @xmath83 . so finding the steady state requires an iterative approach : @xmath129 initialize @xmath104 , e.g. with delta peaked distributions , corresponding to agents without preferences ; @xmath130 calculate the number of traders taking the various actions [ eq . ( [ numberoftraders ] ) ] and thus the trading probabilities @xmath83 [ eq . ( [ tradingp ] ) ] ; @xmath131 find the steady state solution of the fokker - planck equation ( [ adaptivefokkerplanck ] ) for these @xmath83 . steps ( ii ) and ( iii ) are then repeated until a self - consistent solution is obtained , i.e. until the @xmath83 no longer change . we note that finding the stationary distribution of the fokker - planck equation ( step ( iii ) ) is non - trivial , details will be discussed below . once we have the steady - state values of the trading probabilities @xmath83 , we can switch point of view and think of the fokker - planck equation ( [ adaptivefokkerplanck ] ) as describing the dynamics of individual agents within a large population with fixed average properties . as we argue that our system is ergodic , the steady state distribution of attractions can then be re - interpreted as the distribution of a single agent s attraction sampled over a long enough time interval . zeros of the drift velocity @xmath132 are fixed points of the single agent dynamics , and for small amounts of diffusive noise @xmath27 the single agent will spend most of its time near ( stable ) fixed points , causing local maxima of @xmath104 . to detect segregation we therefore look for multiple stable fixed points of the single agent - dynamics in the steady state population . finding the steady state solution of the fokker - planck equation even with given trading probabilities is a non - trivial task in the four - dimensional space of attraction vectors @xmath101 . this is true particularly because the drift and diffusion coefficients @xmath127 and @xmath128 do not in general define a time - reversible single agent - dynamics , for which determining the steady state would be much simpler ( see for example @xcite ) . we thus proceed with one more simplification to produce a theoretical description directly comparable to simulations . we fix the agents preferences for buying or selling , i.e. each agent now carries a fixed probability @xmath133 with which they buy . they sell with probability @xmath134 . this probability may vary from agent to agent , but crucially it remains fixed in time for each trader . agents thus have a single decision left to make , namely , where to trade . in the case of two markets the single variable that we then need to track for every agent is the difference or relative attraction @xmath135 . this makes a full numerical analysis possible . the resulting fokker - planck equation is one - dimensional , and one can find a closed - form solution . locating all single agent fixed points can be achieved e.g. by a bisection method . as it turns out we do find segregation in this simplified scenario , so the reduced model is useful in its own right , and provides evidence of the robustness of segregation behaviour . hence in the sections below we focus on this simplified model , comparing numerical predictions for nonzero @xmath27 directly to simulations ; whenever possible we state the comparison with the original fully adaptive model . in the limit of large population size , the analogue of @xmath104 in the simplified model is the probability distribution @xmath136 . the quantity @xmath137 is the distribution of attraction differences , @xmath138 , among agents with buying preference @xmath133 . the distribution over preferences for buying @xmath139 is fixed as part of the specification of the model . as a simple case , we investigate a population with @xmath140 , consisting of equal numbers of two types of agents with buying preference @xmath141 and @xmath142 , respectively . agent @xmath143 chooses market @xmath110 with probability @xmath144 $ ] and independently chooses to buy with probability @xmath145 or @xmath146 depending on her type . as in the case of fully adaptive agents we can formulate the master equation for the process , and derive a fokker - planck equation in the limit of small memory - loss rates @xmath27 . we obtain : @xmath147 + \frac{r}{2}\partial_\delta^2\left[m_2(\delta|p_\mathcal{b}^{(g)},t_\gamma)p(\delta|p_\mathcal{b}^{(g)})\right ] \label{fokkerplanck}\ ] ] where @xmath148 ( for type or `` group '' ) labels the agent type . the jump moments @xmath149 and @xmath150 couple the two types of agents via the set of trading probabilities @xmath151 . the generic steady state solution of the fokker - planck equation ( [ fokkerplanck ] ) reads@xcite @xmath152 as the notation emphasizes , the stationary probability distribution is dependent on the trading probabilities , and these are themselves function(al)s of the probability distributions @xmath153 . accordingly we use the iterative procedure described after equation ( [ adaptivefokkerplanck ] ) to find a solution , repeating steps ( ii ) and ( iii ) until the trading probabilities @xmath83 remain stable to an accuracy of @xmath154 . from simulation and fokker - planck theory ( the corresponding distributions for the second type of agents @xmath155 are mirror images with respect to the y - axis ) . left : @xmath156 , right : @xmath157 . forgetting rate is @xmath158 , other system parameters as specified in table [ parametertable ] . ] [ fig4:ssbuyers ] compares the steady state distributions @xmath159 obtained from direct numerical simulations and from the fokker - planck theory described above . the buying preferences @xmath160 are symmetric about 1/2 and so we do not show @xmath161 , which would be the mirror image about @xmath162 . the distribution obtained for low @xmath32 ( fig . [ fig4:ssbuyers ] left ) shows the behaviour expected for an unsegregated population , with a single - peaked distribution of relative attraction @xmath138 for each agent type ; the mean of the distribution essentially coincides with the fixed point of the single agent dynamics , i.e. the solution of @xmath163 . in the high-@xmath32 regime , the agents of each type segregate into two groups corresponding to the two peaks of the distribution of the attraction differences @xmath138 . as in the case of fully adaptive agents , one group is return - driven , i.e. prefers the market that awards them with higher returns , although they are not always able to trade there . the agents in the other group are volume - driven ; they settle for the market where returns are lower on average but where they can trade more regularly . the fokker - planck theory successfully reproduces the qualitative transition from unsegregated behaviour at low values of the intensity of choice to segregated steady states at higher values of @xmath32 . the quantitative agreement with numerical simulations is good , remarkably so given that the latter were obtained for relatively small systems ( @xmath164 agents ) and for moderate @xmath158 while we developed the theory for the combined limits of large @xmath86 and small @xmath27 . the agreement between theory and numerical experiment also suggests that segregation is not a finite - size effect . to track the change in the shape of the relative attraction distributions from unimodal to bimodal as @xmath32 is changed we consider the so - called binder cumulant ( see for example @xcite ) . this quantity is defined as @xmath165 and it is a good indicator of segregation . it has different limiting values for unimodal and bimodal distributions . the quantity @xmath166 is in principle dependent on the type of agent considered , @xmath167 ; but in the situations we consider where the two types have distributions of score differences @xmath138 that are mirror images of each other , this dependence disappears . for numerical simulation data we show @xmath168 . the binder cumulant takes the value @xmath169 for a gaussian distribution , while @xmath170 for a distribution consisting of two sharp peaks ( the precise value depends on the relative weight of the peaks ) . for small @xmath27 , where the @xmath138-distributions become sharp around their peak(s ) according to equation ( [ stationarydistribution ] ) these are therefore the values of @xmath166 we expect for @xmath32 below and above the segregation threshold @xmath66 , respectively . for finite values of the learning rate @xmath27 , we expect a smooth transition between the two limiting values . ) ) of the relative attraction distribution for different values of @xmath27 as shown in the legend ( large to small @xmath27 from bottom to top on the right of the plot ) . inset : numerical simulations ( blue ) versus fokker - planck theory ( orange ) for @xmath158 . simulated system size @xmath164 traders , with @xmath138 distributions obtained from the last @xmath30 trading periods of @xmath171 independent simulation runs . ] in fig . [ fig5:binders ] we show the predictions of our fokker - planck theory for the binder cumulant of the reduced model with fixed buy / sell - preferences as a function of @xmath32 , and for different values of @xmath27 . these confirm the expectations set out above , with the segregation transition become increasingly sharp as @xmath27 decreases . the limit @xmath172 of the theory can be worked out by a separate procedure and gives a sharp transition at a well - defined segregation threshold @xmath66 ( for the system presented in fig . [ fig5:binders ] , @xmath173 in the limit @xmath172 ) . the inset of the fig . [ fig5:binders ] compares numerical simulations and theory for @xmath158 and again shows very good agreement . we attribute the remaining deviations to finite - size effects and to the fact that simulations necessarily operate at nonzero @xmath27 while the fokker - planck theory is derived in the limit of small @xmath27 . the qualitative behaviour of the binder cumulant , i.e. a transition between the expected theoretical values for small and large values of @xmath32 that becomes sharper with decreasing @xmath27 , is exactly the same in the fully adaptive model ; see @xcite . there are of course quantitative differences ; e.g. the threshold value @xmath174 ( for @xmath172 ) is higher @xcite . [ [ further - characterisation - of - segregation - dynamics . ] ] further characterisation of segregation dynamics . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + so far we have successfully constructed a mathematical description that reproduces the segregation effects seen in simulations . we now use the theory to look more closely at the emergence of segregation and the properties of the segregated state . to this end we consider average returns across the population of agents . these allow us detect whether segregation brings population - level benefits even though all agents make decisions on a purely individual basis . given that persistence times are finite , the population - averaged returns also give the long - time average returns for any agent and so they tell us about the benefits of segregation for single agents . in fig . [ fig6:returns ] we plot the average return obtained by agents in the steady state against intensity of choice @xmath32 for populations with different forgetting rates @xmath27 . the data shown are from the fokker - planck analysis of the reduced two - strategy model described above . remarkably , the average return is a non - monotonic function of @xmath32 : it has a minimum close to the segregation threshold @xmath66 , with a height that decreases as @xmath27 is reduced . the qualitative non - monotonic trend is also found in numerical simulations ( see inset ) . it is harder to detect there as the absolute changes in returns are fairly small , but appears in both the reduced and the fully adaptive model . to put the returns for our segregated steady states into context , we compare them to two benchmark values . the dashed line in fig . [ fig6:returns ] indicates the first of these , which is the average return of an unsegregated population ( in the limit @xmath172 limit , to be discussed in more detail shortly ) . while this homogeneous population return decreases monotonically with @xmath32 , segregation avoids this decrease for intensities of choice @xmath67 and in fact converts it to an increase . for different values of @xmath27 ( see legend ; @xmath27 decreases from top to bottom ) , as predicted by fokker - planck theory ( circles denote @xmath56 for the given value of @xmath27 ) . black dashed line : average return of an unsegregated population for @xmath172 . horizontal dotted line : envy - free nash equilibrium ( see text ) . the segregated solution achieves higher returns than either the unsegregated population at the same @xmath32 , or the nash equilibrium . insert : comparison between numerical simulations and fokker - planck theory for @xmath158 and population size @xmath164 . ] as a second baseline we show in fig . [ fig6:returns ] the average return in an envy - free nash equilibrium . a nash equilibrium in general is a state in which no agent can increase their payoff by unilaterally changing strategy . in our system , a state would be specified by the probability @xmath175 for each agent @xmath143 to choose market 1 as the set of actions is @xmath176market 1 , market 2@xmath177 . the strategy of an agent is defined by @xmath178 and a nash equilibrium is reached when no agent c@xmath143 an achieve higher average return by changing @xmath175 , at fixed strategies of all other players . this still leaves potentially many different nash equilibria @xcite and so we focus on what we will refer to as the _ envy - free _ nash equilibrium . these are equilibria in which no agent is envious of any other agent s return because they all achieve the same payoff on average over time . this is realized in our system when the average returns for all four distinct actions ( buy or sell , at market 1 or 2 ) are the same . this condition allows us to identify a unique set of trading probabilities and from these the envy - free nash equilibrium return . as fig . [ fig6:returns ] shows , this is always _ lower _ than the average return of a segregated population ( both in theory and simulations , see the inset ) . a segregated steady state is thus better in terms of returns than both the homogeneous steady state at the same @xmath32 and the envy - free nash equilibrium . what is notable here is that the segregated state is not envy - free over short time horizons : an agent in the volume driven group obtains a lower return than one from the return driven group . however , as emphasized above , on very long time scales agents can change their loyalties , i.e. change group , so that in a long - time average all achieve the same return . this again emphasizes the co - operative nature of the segregated state . [ [ estimating - the - segregation - threshold - beta_rm - s - the - rrightarrow-0-limit . ] ] estimating the segregation threshold @xmath66 : the @xmath172 limit . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + we now proceed to study the segregation threshold @xmath66 of the intensity of choices and how it depends on the parameters of the model . this is easiest in the limit @xmath123 where the segregation transition is sharp , see fig . [ fig5:binders ] above . we first focus on the regime @xmath179 , i.e. the unsegregated phase , in which the distributions @xmath180 will be single - peaked . in the limit @xmath123 , equation ( [ stationarydistribution ] ) indicates that ( i ) this peak becomes infinitely sharp , so that the distributions are of the form @xmath181 ; and ( ii ) the location @xmath182 of the peak for each group of players is determined by the zero - drift condition @xmath183 the next step is to solve these two equations for @xmath184 and @xmath185 , which is the @xmath123 analogue of finding the steady state of the fp equation for general @xmath186 . as before one needs to make the solution self - consistent so that the trading probabilities @xmath83 appearing in ( [ zero_drift ] ) are those calculated from the distributions @xmath181 themselves . the iterative approach explained after the equation ( [ adaptivefokkerplanck ] ) can again be used to find such a self - consistent solution , starting from @xmath187 . we briefly explore the properties of this homogeneous steady state before considering how to detect the onset of segregation . we focus on agent types with symmetric biases toward buying and selling , @xmath188 . for this choice we find non - zero @xmath189 even in the limit @xmath190 . this indicates that agents recognize the more rewarding option : agents that are more likely to buy have a preference for the market that is good for buyers , and similarly for sellers . of course @xmath190 means that agents nevertheless choose randomly between the markets . as @xmath32 increases , the relative attractions @xmath189 become more pronounced , i.e. they move away from zero , and agents start to choose the `` better '' market more frequently . this is the reason for the decay of the average return in the homogeneous steady state with @xmath32 , as shown in fig . [ fig6:returns ] : as agents of each type increasingly focus on `` their '' market , buyers congregate in one market and sellers in the other ; trading opportunities are reduced and thus the average return ( which includes zero returns for trading periods where an agent can not trade ) decreases . the more definitive choices agents make are the consequence of the two effects : ( 1 ) the fixed points @xmath189 increase in absolute value ( and they correspond to the difference in the average score an agent receive in the two markets ) ; ( 2 ) the increase of @xmath32 makes the choices more definitive . quantitatively , we find that ( 2 ) is the stronger effect . as the intensity of choice @xmath32 is increased further , we expect a transition to a segregated steady state . to detect this transition we can follow the general logic explained in the beginning of analytical description : a segregated state must have more than one fixed point of the single - agent dynamics . the peak position @xmath184 is always a fixed point for agent type 1 from eq . ( [ zero_drift ] ) , and similarly for type 2 . to detect the onset of segregation we therefore need to check when _ additional _ fixed points appear , i.e. additional solutions of the zero - drift condition @xmath191 . in looking for these alternative fixed points we need to keep the trading probabilities @xmath83 _ fixed _ at their values calculated for the homogeneous steady state , because we are considering the _ single - agent _ fixed points . note that in general one needs to search for alternative fixed points globally , i.e. across all possible @xmath138 . this is because for most parameter settings the new fixed points appear far from @xmath184 or @xmath185 , respectively , as pairs of stable and unstable fixed points . , assuming the other agent type has the opposite preference @xmath192 . different sets of symmetric market parameters are compared ; non - trivially , the region of segregation is largest when both markets are fair . ] in figures [ fig7:pbdiffthtcsurf ] and [ fig8:pb1pbtcsurf ] we present the segregation thresholds obtained by the method above , for various parameter settings . in fig . [ fig7:pbdiffthtcsurf ] specifically we show how @xmath66 changes with @xmath193 for various symmetric markets . one observes that @xmath66 does not depend very strongly on the exact preferences for buying of the agents , except for the region of parameters where both agent types have almost even preferences for buying and selling . a full analysis of the change in monotonicity in that region exceeds the scope of this publication and will be presented elsewhere . in contrast , the effect of the market biases follows a simple trend : the region of segregation shrinks as the difference between the two markets increases , suggesting that segregation is not a trivial effect of market biases . in the fig . [ fig8:pb1pbtcsurf ] we show two contour plots to compare the trends in @xmath66 when the two agent types buying preferences are varied independently , for two different choices of market parameters . on the left is the symmetric markets setup @xmath194 that we have already used several times . from the plot we observe that the segregation threshold @xmath66 is lowest when the subgroups are symmetric with respect to buy - sell preferences , i.e. @xmath195 ; the variation of @xmath66 along that line is presented in fig . [ fig7:pbdiffthtcsurf ] . we also note that when two agent types have similar preferences ( e.g. both prefer buying over selling ) then @xmath66 is on average higher than in the case where the types have opposite preferences . this qualitative behavior we see also in the case of two fair markets that set the trading price at the equilibrium price , as shown in fig . [ fig8:pb1pbtcsurf](right ) . we do not show similar plots for asymmetrically biased markets ; the qualitative behaviour is similar there , but the line of minimal @xmath56 is no longer @xmath195 but @xmath196 , where @xmath197 is a constant that depends on the market parameters . when preferences for buying @xmath145 and @xmath146 of both agent types are varied independently . left : symmetric markets @xmath198 ; right : two fair markets ( @xmath199 ) . contours are presented in terms of @xmath200 only for the better visual accessibility . ] finally , we can adapt the above method of calculating the segregation threshold to the original model of fully adaptive agents choosing among all four possible actions ( e.g. @xmath201 ) . because of the difficulties of finding the steady state solution numerical for finite @xmath27 , this analysis is carried out only in the limit @xmath123 . this is feasible because we only need to find zeros of the drift @xmath149 , rather than solve for a full distribution @xmath104 that is a stationary solution of the fokker - planck equation . as argued before , the first moment ( @xmath149 ) is a function of the whole distribution , but in the @xmath172 limit and for @xmath202 this distribution is a delta distribution , which simplifies the calculation . however , as the @xmath172 limit is not the most realistic one ( it assumes agents with infinite memory ) , we focused most of our previous analysis on the simplified model where we can say more about segregation for nonzero @xmath27 . in fig . [ fig9:adaptivetcsurf ] we show a contour plot of the segregation threshold against the two market biases . as in fig . [ fig7:pbdiffthtcsurf ] , for agents with fixed buy - sell preferences , we notice that the segregation threshold @xmath56 increases when the difference in the symmetric market biases increases ; this can be seen specifically by looking at the symmetric markets diagonal of fig . [ fig9:adaptivetcsurf ] and fig . [ fig7:pbdiffthtcsurf ] . additionally in the system of fully adaptive agents we notice that this conclusion extends to the case when at least one market is fair , i.e. segregation arises for smaller instances of @xmath32 if at least one market is fair . in fig . [ fig9:adaptivetcsurf ] we note also a four fold symmetry , which we conjecture is a consequence of the fact that agents have more choices : a fully adaptive agent does not have a preferred market initially , so all four action choices are equal . this is not the case for an agent with a fixed preference for buying and selling because returns are buyer / seller - specific . in a system of fully adaptive agents , against the market bias parameters @xmath203 . along the lines where at least one market is fair , @xmath66 is lowest so that segregation occurs for the widest possible range of intensities of choice @xmath32.,scaledwidth=50.0% ] [ [ multiple - steady - states . ] ] multiple steady states . + + + + + + + + + + + + + + + + + + + + + + + when iteratively finding a self - consistent steady state of the fokker - planck equation as discussed around equation ( [ adaptivefokkerplanck ] ) , we always used as initial condition a distribution of all agents having zero attractions . while the dynamics of the iterative solution is not identical to the real dynamics , this choice was sensible to obtain steady states that match the ones from simulations of the real dynamics as closely as possible . after further analysis we found that in the numerical simulations for small values of @xmath27 one obtains two qualitatively different classes of segregated states . to investigate this notion further , we now explore whether there can be more than one self - consistent solution of the fokker - planck equation . these additional solutions could in principle for example be accessible by using our iterative procedure , but starting from other initial conditions . the simplest way to perform this exploration systematically is to realize that the trading probabilities @xmath83 from equation ( [ tradingp ] ) only depend on the demand - to - supply ratios @xmath204 and @xmath205 at the two markets ; we define @xmath206 for each market @xmath78 as the ratio of the number of buy and sell orders that arrive at the market . specifying @xmath207 thus tells us all trading probabilities , and hence determines a unique steady state solution of the fokker - planck equation . from this steady state solution we can recalculate @xmath204 and @xmath205 , and plot in the @xmath207-plane the two lines where the new and old @xmath204 ( @xmath205 , respectively ) coincide . the intersections of these two lines are then the self - consistent steady states we are after . we find numerically , in the range of parameter values that we have explored , that there is either one such state or there are three . and @xmath32 . * number and nature of the steady state solutions of the fokker - planck equations in the @xmath208-plane ( u : unsegregated state , s : strongly segregated state , w : weakly segregated state ) . other system parameters are set to their default values ( see table [ parametertable ] ) . to the left ( small @xmath27 ) of the orange line there are three steady state solutions , while to the right there is only one . the blue line separates the region where at least one of the solutions is segregated , i.e. bimodal . empty circles represent @xmath123 endpoints of the two lines , calculated independently . ] in fig . [ fig10:rtphasespace ] we show the resulting phase diagram in the @xmath209-plane and indicate the number and nature of self - consistent steady state solutions . where solutions are segregated we differentiate between two possibilities , strong and weak segregation . by strong segregation we mean a solution branch where the two peaks in @xmath210 remain of comparable height as @xmath123 . we call a solution branch weakly segregated when one peak weight becomes exponentially small for @xmath123 , so that the solution effectively reduces to an unsegregated one . in the limit , all agents then typically prefer the same market . the empty circles in the figure are results from a separate analysis of the deterministic ( @xmath172 ) theory , and are consistent with the extrapolation of the results for nonzero @xmath27 . numerical simulations confirm the existence of all three solution types ( u , s , w ) in the relevant regions of the phase space , where necessary starting from appropriately tuned initial conditions . where the strongly segregated solution competes with others it appears to become unstable but long lived in small system simulations . ( details are beyond the scope of this paper and will be given elsewhere . ) the blue line in figure [ rtphasespace ] is the segregation threshold @xmath66 discussed previously . it is nearly constant , increasing only very slowly with @xmath27 ; this shows that the @xmath66 values obtained via the @xmath123 analysis will be good estimates also for larger @xmath27 . a second interesting feature of fig . [ fig10:rtphasespace ] is that there is a threshold value of @xmath211 above which the strongly segregated solution is the only possible option , for any @xmath32 large enough for a segregated state to exist at all : this state is then the genuine steady state . where agents have only moderate memory , with @xmath30 of the order of 10 trading periods , steady states where one market comes to attract most agents thus disappear , and the emergence of persistent market loyalties becomes the norm . for the fully adaptive model we have preliminary evidence for a similar phase diagram as in fig . [ fig10:rtphasespace ] and in particular for the existence of weakly segregated states , but defer further analysis of their behaviour to a separate publication . the increasing fraction of global trades that now takes place online in the form of high - frequency algorithmic trading ( around @xmath212 of equity trading in the uk and around @xmath213 in the usa @xcite ) raises many challenges for market regulations . to make such regulation as informed as possible , research is needed to understand the possible long - run states of such systems . we focused on one particular long - run state the segregation of traders . as argued in the introduction , it is easy to imagine that such a state can bring benefits to the system , but also disadvantages ; both make it important to investigate and understand the phenomenon and its consequences . we have developed a stylized model of double - auction markets serving a large number of adaptive traders ; this has helped us to shed light on the benefits of segregation for the population as a whole . we have considered two variants of our model , one populated with agents who can adaptively tune their preferences for buying and selling along with their preferences for the two markets ; in the other model agents have fixed preferences for buying and hence also selling . the two models share our main qualitative result : above a ( model - dependent ) threshold value of intensity of choice @xmath66 the agents segregate , i.e. develop a long lasting loyalty to one action , or one market . the onset of segregation is signalled both in simulations and in the analytical description by the emergence of multiple peaks in the distribution of agent preferences , as summarized e.g. in the binder cumulant . these peaks are accompanied by long persistence times for the agents to remain within each peak @xcite . in the model with fixed buying preferences we have shown that one can carry out the analytical description to a significant level of detail . we find that even though individual agents do not explicitly try to maximise the well - being of the entire population , the strongly segregated state is in effect cooperative : it is more beneficial for every individual player , and the population as a whole , compared to possible unsegregated and envy - free nash equilibrium states . the segregated state is neither envy - free the traders that specialize to what we called volume - driven behaviour have lower returns than the return - driven ones nor a nash equilibrium . in this sense the segregated state is stabilized by incomplete information , of each agent about the precise returns to be expected from each action and about the average returns achieved by others . the main parameter that separates between the two states the intensity of choice @xmath32 plays effectively the same role as the scale of returns : multiplying either by some factor has the same effect on agents preferences . when the possible returns are small an agent plays randomly , while if the stakes are high an agent will try to take into account information from previous trades as much as possible . in this interpretation our main result states that there is a critical scale for single transaction returns above which the preferred state of the agent population is the segregated one . finally , it is important to discuss simplifications we have made in our analysis and the way they might have affected our main results . for example , we have made various assumptions about the bid and the ask distributions . it turns out that if these are relaxed it changes our main results only quantitatively but not qualitatively . for example , we have shown all data for the case where the mean bid is greater than the mean ask as this enables more trades at the markets ; qualitatively identical results are obtained in the opposite case , however , with the only change being the in the specific vale of the threshold @xmath66 . more generally , the assumptions we made on the shape of bid / ask distribution can be relaxed . in fact , within the fokker - planck description only the first and the second moment of the truncated bid / ask distribution appears and the precise distribution shape is otherwise immaterial . another important assumption about the agents is their learning rule . as we discussed when first setting out the model , there are several variants in the literature of the rule we used . an important class of these is given rules where an agent updates the score for unplayed actions with the fictitious score . we have argued that the assumption that agents can estimate fictitious scores is not realistic in our set up ; however , changing the way an agent forgets the score of the unplayed actions might be reasonable . this leads to a generalized reinforcement rule : @xmath214 where the parameter @xmath215 can be tuned to vary between two extremes : @xmath216 is the case we have discussed so far , where an agent forgets the score of the unplayed actions with the same rate as the scores of the played action , effectively imputing a score of 0 for actions that were not taken . on the other hand , @xmath217 corresponds to the case where an agent does not forget the attractions of the unplayed actions , which effectively means imputing a score equal to the average score up to that time . we have explored in our analytical description the entire range of @xmath215 for both variants of our model ( fully adaptive and fixed buy - sell preference ) , and find generally that the segregation threshold @xmath66 increases as @xmath215 decreases towards zero . it remains finite for any @xmath218 , so that segregation continues to take place , in a manner qualitatively similar to our vanilla model ( @xmath216 ) . the case @xmath217 , which presupposes that agents have effectively infinite memory to scores of unplayed action , is special and will be explored further elsewhere . one question that remains , and which we can not address directly here , is whether the emergence of segregation is an intrinsic property of systems with adaptive agents , or a consequence of our specific stylized model . the simplicity of the model itself argues for the former , as we did not need to make exotic assumptions to find segregation . but clearly there is still much to do from here to reach a detailed understanding of segregation in real markets , at a level that can directly influence policies . we have carried out preliminary research into the more realistic _ continuous _ double auction markets , which certainly supports the robustness of segregation as an emergent phenomenon . further directions for investigation include making not just agents but markets adaptive within the model , to understand whether the competition between markets that this will engender suppresses or promotes segregation . there is clearly also scope for capturing more complex trading strategies in the model , something we have not pursued here as we were motivated by the simplicity of high - frequency trading algorithms and more broadly by the aim to develop a baseline model on which further extensions can be built . the parameters of the system are summarized in table [ parametertable ] with their definitions and the standard values that are used in the main text unless otherwise specified . for all numerical simulations ( adaptive ) agents were initialized with @xmath36 and the trading dynamics was run until a stationary state was reached . similarly in the simplified model with fixed preferences for buying , @xmath219 for all agents initially . the time required to reach a steady state was highly dependent on the chosen parameters ( longer for low @xmath27 and high @xmath32 ) , but for the results presented in this paper was in the range between @xmath220 and @xmath221 trading periods . to be sure that the system reached the steady state we measured average returns , attractions and higher cumulants of the attraction distributions ; when the average of these observables became independent of time we assumed stationarity and started collecting the data for the analysis . statistics presented in the paper are gathered for each parameter setting from @xmath171 independent runs of the stochastic dynamics . time averages are usually taken over @xmath222 trading periods at the end of each simulation run . llr + parameter & description & typical value + @xmath223 & number of markets & @xmath224 + @xmath225 & market biases - usually assumed to be symmetric , @xmath226 & @xmath227 + @xmath86 & number of traders & @xmath228 + @xmath229 & mean and standard deviation of the ask distribution ( sellers ) & @xmath230 + @xmath231 & mean and standard deviation of the bid distribution ( buyers ) & @xmath232 + @xmath27 & forgetting rate ; range @xmath233 $ ] & @xmath158 + @xmath32 & intensity of choice ; simulation range @xmath234 $ ] & + @xmath235 & preferences for buying of the two types of agents of the toy model & ( 0.8,0.2 ) + we acknowledge funding by the engineering and physical sciences council epsrc ( uk ) , grant reference ep / k000632/1 ( network plus , towards consensus on a unifying treatment of emergence and systems far from equilibrium ) . ps acknowledges the stimulating research environment provided by the epsrc centre for doctoral training in cross - disciplinary approaches to non - equilibrium systems ( canes , ep / l015854/1 ) . we thank yuzuru sato for useful discussions . aa , ps and pm conceived the study . aa and ps carried out the research , aa developed the simulations , carried out the analytical calculations and prepared the figures . aa , ps and tg interpreted the data and shaped the study . all authors wrote and reviewed the manuscript . 10 kirman ap , vriend nj . learning to be loyal . a study of the marseille fish market . in : gatti d , gallegati m , kirman a , editors . interaction and market structure . vol . 484 of lecture notes in economics and mathematical systems . springer berlin heidelberg ; 2000 . p. 3356 . smith a. an inquiry into the nature and causes of the wealth of nations . cannan e , editor . library of economics and liberty [ online ] available from http://www.econlib.org/library/smith/smwn.html ; accessed 19 october 2015 ; 1904 . alori a , sollich p , mcburney p. spontaneous segregation of agents across double auction markets . in : amblard f , miguel fj , blanchet a , gaudou b , editors . advances in artificial economics . vol . 676 of lecture notes in economics and mathematical systems . springer international publishing ; 2015 .

an optical lattice is a crystal structure of off - resonant laser light , wherein atoms are trapped due to the optical dipole force . this creates a periodic potential in which the motional states of a single atom exhibits many of the usual textbook solid state characteristics , such as energy band structure , bloch oscillations and wannier - stark ladders @xcite . in an interacting many - body system the periodic potential constitutes an ideal setting for realizing a wide range of condensed matter phenomena such as the superfluid to mott insulator transition @xcite . likewise controlled atomic collisions in an optical lattice potential is a promising approach to implement quantum gate operations @xcite . in the context of ultracold atomic collisions a feshbach resonance arises , when a bound state of a closed scattering channel is coupled to the scattering continuum in the entrance channel @xcite . in an applied magnetic field the separated atoms are described by eigenstates of the hyperfine and zeeman - hamiltonian , and the coupling between the different channels is a result of the atomic interaction , which is not diagonal in the hyperfine basis . the position of the resonance is tunable by varying the strength of the applied magnetic field , since the open and closed channels have different magnetic moments , and therefore experience different zeeman shifts . when the closed channel bound state is tuned to be degenerate with the colliding atoms in the entrance channel the scattering cross section is resonantly enhanced . a feshbach resonance may thus be used to regulate the atomic interactions , which in the ultracold regime is characterized entirely by the @xmath0-wave scattering length . in addition , since the feshbach resonance provides an adiabatic connection between bound dimer states and the continuum , a pair of unbound atoms may be converted into a molecule by slowly ramping the magnetic field across the resonance position . to study two - particle physics on a lattice experimentally , a state with either zero or two atoms at each lattice site may be prepared by adiabatic magnetic field sweeps across a feshbach resonance starting from a pure atomic cloud @xcite . this prepares a state with at most two atoms in each lattice well , since sites with more than two atoms are eliminated by inelastic collisions . the unbound atoms may be subsequently removed by an optical purification pulse . the sample thus prepared contains only isolated pairs , whose binding can then be controlled by further adjustments of the magnetic field . a special situation arises in an optical lattice due to the band structure . the periodic potential leads to a structured continuum with scattering states grouped in energy bands separated by band gaps . the depth of the lattice controls the width of the bands and their mutual separation . in this paper we concentrate on the lowest band , where the band structure then implies that the continuum has both a lower and an upper edge . this has fundamental implications for both the bound states and the scattering states of the system . a counter - intuitive effect of the lattice band structure is the existence of bound pairs composed of atoms , which repel each other @xcite . such repulsively bound pairs exist , when the repulsive interaction is strong enough to lift the state of two atoms on the same lattice site out of the continuum and into the band gap , thereby preventing the pair from dissociating and making it a bound state of the system @xcite . in one dimension repulsively bound pairs exist for an arbitrarily small repulsion . near a feshbach resonance further intriguing two - body effects become possible . in particular , the feshbach resonance gives rise to motionally bound molecules , which are stable only at finite center of mass momentum with respect to the lattice , while decaying if at rest . it is thus possible to tune the location of the scattering resonance at a fixed magnetic field by controlling the molecular motion @xcite . here we derive analytical results fully characterizing both the bound and scattering states of an atom pair in an optical lattice near a feshbach resonance . we represent the lattice physics in a discrete model with nearest neighbor hopping . to describe the feshbach resonance we adopt a two - channel model with one open channel and one closed channel , which is adequate to describe the physics of both the scattering continuum and the bound states of the system for an isolated resonance @xcite . contrary to @xcite we include the background interaction in the entrance channel , which plays an important role both in scattering and in the bound state dynamics . the direct scattering process in the open channel interferes with the resonant scattering proceeding via the closed channel resonance state leading to a fano profile in the transmission probability . in the bound state spectrum the presence of an additional bound state leads to avoided crossings that are important when considering adiabatic transitions induced by ramping the magnetic field . the same fundamental physics of a discrete state embedded in a structured continuum occurs in the fano - anderson model , which describes autoionizing states in atoms @xcite and localized electron states in solids @xcite . from a many - body perspective the combination of an optical lattice with resonant control of the atomic interactions makes a variety of spin models attainable @xcite . in a previous paper @xcite we described the basic results of the discrete lattice model of the feshbach resonance physics , and suggested how the salient features of the system may be probed in experiments . here we present the full analytic theory of feshbach scattering in a one - dimensional lattice ( sec . [ sec : scattering ] ) . we introduce a generalized , one - dimensional scattering length , which is defined at both the upper and the lower threshold of the lowest continuum band in the lattice . this generalized scattering length shows a characteristic feshbach divergence when a bound state of the system crosses either band edge , and we identify the magnetic field positions for these divergences as a function of the lattice depth for a specific feshbach resonance . in sec . [ sec : bound - states ] we consider the bound state solutions of the combined lattice - feshbach problem and derive analytic expressions for the binding energies and the wavefunctions of the dimer states . we also show that the molecules exhibit cooperative tunneling with a tunneling rate consistent with that measured in experiments @xcite . a spectral analysis of the unbound and dimerized atom pair states is presented in sec . [ sec : spectral - analysis ] , where we demonstrate that threshold effects lead to a non - trivial peak structure of the molecular spectral function inside the band . our discrete lattice model includes only the lowest continuum band and neglects couplings to higher bands in the lattice . in sec . [ sec : validity ] we discuss when the lowest band approximation is valid and how this restriction may be relaxed . we find that a moderate interband coupling only modifies our results on a quantitative level . throughout we give quantitative results for a specific feshbach resonance for which our model is valid @xcite . we consider atoms with mass @xmath1 in a cubic optical lattice potential @xmath2 created by off - resonant laser light with wavelength @xmath3 . the lattice period is , @xmath4 , and the transverse , @xmath5 , and longitudinal depths , @xmath6 , are measured in units of the photon recoil energy , . for a cubic lattice the motion separates along the three spatial directions , and the hamiltonian , @xmath7 , is diagonalized by products of three bloch functions , @xmath8 , one for each spatial direction @xmath9 , where @xmath10 is the quasi - momentum in the range @xmath11 $ ] , and @xmath12 is the band index . the energy eigenstates form the usual band structure , splitting the continuum into bands separated by energy gaps that decrease with increasing @xmath12 . the band gaps grow and the widths of the bands decrease , as the lattice is made deeper by increasing the laser intensity . we will assume that the lattice depth along the transverse directions , @xmath5 , is much larger than the depth of the longitudinal lattice , @xmath6 . this allows us to only include the ground state along the two perpendicular directions , leaving an effectively one - dimensional lattice . as the scattering energy is increased consecutive transverse channels open , however we only consider scattering at energies much smaller than the transverse level splitting , thus justifying the neglect of the transverse degrees of freedom . a confinement induced resonance occurs at the lower edge of the continuum when the ( free space ) scattering length @xmath13 and the width of the transverse confining potential @xmath14 are related by @xmath15 @xcite . however , this resonance may be avoided by not choosing the transverse confinement too tight , and for the quantitative results we present here , it plays no role . the bloch waves are delocalized , but can be combined into orthonormal localized basis functions , wannier functions , which in the longitudinal direction take the form @xmath16 in addition to the band index they are labeled by a site index , @xmath17 , specifying at which lattice sites , @xmath18 , the wannier functions are localized . a similar wannier basis can be constructed in the two transverse directions . when it is not clear from the context , which wannier basis we refer to , we will add a superscript @xmath19 or @xmath20 to the wannier functions indicating to which lattice direction they belong . the dynamics of a single atom in the @xmath12th longitudinal band , _ i.e. _ the kinetic energy and lattice potential , is characterized by the tunneling amplitude , , between lattice sites separated by a distance @xmath21 . here @xmath22 only depends on the coordinate along the weak lattice direction , and @xmath1 is an integer . in the transverse directions tunneling can be disregarded , due to the much larger lattice depth . in this work we consider only the motion in the lowest band of the lattice , and we disregard tunneling between non - adjacent sites , assuming @xmath23 for all . in the basis of the wannier states the single particle hamiltonian for motion in the longitudinal direction then becomes @xmath24 with @xmath25 . we have set the zero of energy to be the expectation value of the energy for an atom in a lowest band wannier state , _ i.e. _ @xmath26 . the hilbert space is now represented by a discrete basis of the localized wannier functions , and accordingly we have translated the continuum description of the system into a discrete lattice model . by our definition @xmath27 corresponds to the center of the lowest longitudinal bloch band , and the single - particle energies of the discrete lattice model are @xmath28 . the approximations made are the omission of higher bands , and the neglect of tunneling beyond nearest neighbor sites . both of these restrictions may be relaxed in a systematic way @xcite . we use the longitudinal wannier functions as a basis for the two - particle states on the lattice . hence the two - particle wavefunction @xmath29 will denote the amplitude of finding the atoms in the wannier orbitals @xmath30 and @xmath31 localized around the discrete positions @xmath32 and @xmath33 . considering only the lowest bloch band we have @xmath34 . to analyze the motion of the two particles in the lattice we introduce center of mass , @xmath35 , and relative , @xmath36 , coordinates . for the center of mass motion a plane - wave ansatz @xmath37 is possible , with the first brillouin zone for the center of mass momentum , @xmath38 , running from @xmath39 to @xmath40 . this separation is pertinent , since states with distinct center of mass momenta are not coupled by the interactions or the lattice , and we can thus obtain separate results for the relative motion for each @xmath38 . taking the interaction potential @xmath41 to be purely on - site with strength @xmath42 ( defined in appendix [ sec : calc - param ] ) , the relative motion of two atoms in the open channel is then described by the hamiltonian , @xmath43 , where the relative motion lattice hamiltonian , @xmath44 , contains the discrete laplacian @xcite @xmath45 , \label{lattice_laplacian}\ ] ] and the center of mass energy , @xmath46 . the eigenstates of @xmath47 are discrete plane waves , @xmath48 . the corresponding energies , @xmath49 , are the sum of the single - particle energies of atoms with momenta @xmath50 , @xmath51 $ ] . the ( retarded ) green s function for non - interacting atoms on the lattice is @xmath52^{-1}$ ] , where @xmath53 is a positive infinitesimal . in momentum space it is diagonal @xmath54 while its coordinate space form may be found from @xmath55 due to the delta function interactions we only need to consider @xmath56 . for energies inside the band ( @xmath57 ) the solution is propagating @xmath58 with @xmath59 , while the the solution outside the band ( @xmath60 ) falls off exponentially : @xmath61^{z / a}. \label{g0_z_out}\ ] ] here @xmath62 . for energies above the band the sign of @xmath63 alternates between lattice sites . the green s function related to the open channel hamiltonian , including the interaction , @xmath41 , is found by explicitly solving the dyson equation @xmath64 in coordinate space the result has the simple form @xmath65 defining the diagonal elements of the non - interacting momentum space green s function to be @xmath66^{-1}$ ] the interacting open channel green s function has the momentum space form @xmath67 which follows from the matrix elements of ( [ dyson ] ) in the relative momentum eigenbasis , using that for an on - site interaction the matrix element , @xmath68 , is constant for all @xmath69 . with the expressions for the green s functions the formal solution of the schrdinger equation for the relative motion becomes explicit . below we derive the transmission probability for open channel collisions and analyze the near threshold scattering in terms of a generalized scattering length . this section introduces the concepts and quantities central to the ensuing discussion of two - channel scattering . we first present results relevant to scattering of distinguishable particles , and then discuss how to treat the case of identical bosons or fermions . choosing the situation with the incident wave entering from the left with quasi - momentum @xmath59 , the scattering state of two particles colliding under the influence of the entrance channel potential , @xmath41 , is given by the solution of the lippmann - schwinger equation @xmath70 for our on - site interaction of strength , @xmath42 , the scattering wavefunction becomes @xmath71 where we have used that @xmath72 . we note that for the contact potential @xmath73 equals its asymptotic form @xmath74 everywhere , hence we may readily identify the scattering amplitude for the background interaction : @xmath75 which is unit - less in a one - dimensional system . for the one - dimensional problem and distinguishable particles it is natural to recast this in terms of transmission and reflection amplitudes . these are defined from the left and right asymptotic form of the wavefunction @xmath76 by comparing with ( [ psi_e_bg_asymp ] ) it follows that the transmission amplitude is @xmath77 , while the reflection amplitude coincides with the scattering amplitude , @xmath78 . the resulting transmission probability is @xmath79 as one might expect , the transmission vanishes in the limit where @xmath80 , and approaches unity as the strength of the on - site interaction is diminished . as usual , the reflection coefficient is @xmath81 , which is unity at the edges of the band , where propagation ceases . the minimum value of the reflection probability , @xmath82 , is attained in the middle of the band , @xmath27 . we emphasize that the transmission is independent of the sign of the interaction . this is a special property of the delta - function potential . as usual the scattering amplitude may also be related to the phase shift , @xmath83 , of the scattered wave . in one dimension the relation is @xcite @xmath84 in general there will be two partial waves , corresponding to even and odd solutions of the relative motion problem . for the delta - function interaction the odd partial wave has vanishing phase shift due to symmetry , and the reflection and transmission probabilities are then given by @xmath85 and @xmath86 , respectively . the preceding derivation of the one - dimensional scattering in the lattice assumed that the two colliding atoms are distinguishable . if the collision involves identical bosons or fermions the scattering state @xmath87 must be an even or an odd function of the separation @xmath88 , respectively , to make the wavefunction either symmetric or anti - symmetric under interchange of the two atoms . for fermionic atoms in the same internal state this implies that the wavefunction vanishes for @xmath89 and the contact interaction does not produce any scattering . for identical bosons the properly symmetrized scattering state may be written as @xmath90 when identical particles collide in one dimension the transmission and reflection coefficients have no meaning , since the incoming and outgoing fluxes are identical , however the scattered wave still experiences a phase shift given by ( [ eq : f_and_delta ] ) . even though the transmission profile calculated above does not play a role in the scattering of identical bosons , it may nonetheless be mapped out spectroscopically . consider transitions from a deeper bound state @xmath91 in the molecule to an energy eigenstate inside the band induced by applied rf- or photodissociation fields , which are described by some transition operator @xmath92 . if the initial state is well localized with respect to the lattice spacing , such dissociative transitions towards the state ( [ psi_bg_sym ] ) occur with a probability @xmath93 , obtained above . when the relative quasi - momentum approaches the center or the edges of the brillouin zone , the two - particle energies @xmath94 tend to @xmath95 and the scattering amplitude attains a limiting form @xmath96 with @xmath97 at the bottom of the band and @xmath98 at the top of the band . these limits allow a natural definition of a generalized one - dimensional scattering length in the lattice @xmath99 from the limit @xmath100^{-1}$ ] of the scattering amplitude as @xmath101 @xcite . an equivalent definition is through the derivative of the phase shift with respect to the relative momentum @xmath102 in analogy with the usual free space scattering in three dimensions . our lattice scattering length is a generalization of the usual concept , in that it is defined for collision energies at both the lower _ and _ the upper edge of the continuum . the sign of @xmath103 at the upper edge of the continuum is a convention , which will be justified for the two - channel case below . our generalized scattering length depends on the center of mass motion of the pair , which is a crucial feature in the lattice , especially when we discuss the two - channel case below . for @xmath104 a bound state is situated below the continuum , making @xmath103 positive as expected . in this case the pole of ( [ f_bg_lim ] ) lies along the positive imaginary axis , @xmath105 , and the analytic continuation of the plane wave @xmath106 is the dying exponential of the bound state . conversely , for a repulsively bound pair state lying above the continuum , @xmath107 , and @xmath103 is negative . the pole of the scattering amplitude is then at @xmath108 , corresponding to a wave function decaying exponentially with the separation between the two atoms and a phase factor @xmath109 , which alternates between @xmath110 and @xmath111 from one lattice site to the next , in accordance with the behavior of the green s function for the relative motion at @xmath112 ( [ g0_z_out ] ) . as @xmath113 the scattering length diverges , and the bound state approaches the edge of the continuum . our intuition from scattering theory in three dimensions about the relation between weakly bound states and the scattering length thus holds both below and above the continuum band however , since @xmath42 is proportional to the free space background scattering length , @xmath114 , the generalized one - dimensional scattering length and its three - dimensional equivalent are inversely related @xcite . we now consider scattering in the presence of a feshbach resonance mixing the entrance channel with an energetically closed channel through a coupling , @xmath115 . this gives rise to a set of two coupled equations for the relative motion of an atom pair @xmath116 @xmath117 due to the large splitting between the bound states in the closed channel the resonance physics is faithfully represented by considering only a single bound state , the resonance state @xmath118 , of @xmath119 with an energy @xmath120 with respect to the center of the open channel band . this amounts to a single - pole approximation for the closed channel green s function @xmath121 the resonance state energy @xmath120 is composed , in part , of the free space resonance energy @xmath122 , which is linearly tunable with an applied magnetic field , due to the difference @xmath123 between the separated atoms and the closed channel magnetic moments , and , in part , of the displacement of the closed channel relative to the open channel bloch band as illustrated in fig . [ fig : energydiagram ] . at all magnetic field strengths we take the zero of energy to be the center of the open channel continuum band . in free space the bare resonance energy can be related to the magnetic field position of the zero energy resonance , @xmath124 , as outlined in appendix [ resonance_shift ] . is the energy of @xmath125 relative to the open channel threshold . applying an optical lattice changes the position of the open and closed channel by @xmath126 and @xmath127 , respectively , where @xmath128 is the average atomic energy and @xmath129 is the molecular spectrum of the first bloch band both along direction @xmath91 . note that the position of the open channel thresholds in the lattice depends on the quasi - momentum , @xmath38 , and therefore we choose @xmath27 to be the center of the open channel band . hence , the thresholds are located at @xmath95 . the difference in energy between the closed channel bound state and the center of the open channel band in the presence of the lattice is @xmath130 . ] as a function of continuous position variables the longitudinal lattice potential for each atom is @xmath131 with @xmath132 , and therefore the combined lattice potential for a pair of atoms in the resonance state is @xmath133 where we have introduced the center of mass , @xmath134 , and relative , @xmath135 , coordinates . since the size of the closed channel bound state is much smaller than the lattice spacing , we may consider the limit of @xmath136 for this state . consequently , the resonant state has a center of mass motion corresponding to a lattice potential twice as deep as that governing the motion of the individual atoms , while its mass is the sum of the two atomic masses . this leads to a tunneling rate for the resonant state , @xmath137 , which is much reduced in comparison with the rate at which the free atoms tunnel through the lattice , @xmath138 . this bare molecular tunneling leads to a modulation of the resonance energy with the center of mass momentum of the pair , which we may parametrize as @xmath139 the band averaged resonance energy , @xmath140 , incorporates the relative energy shifts of the atomic and molecular bands in the lattice . with the explicit expression for the closed channel green s function ( [ g_cl_approx ] ) the closed channel part of the scattering problem ( [ coupled_eq2 ] ) may be solved formally : @xmath141 where we have introduced a coupling matrix element , @xmath142 , incorporating the structure of the resonance state . for a typical resonance in alkali collisions the coupling arises due to the difference between the triplet and singlet molecular potentials , and it is therefore inherently of short range compared with the lattice spacing . it is therefore reasonable that the coupling only acts between atoms localized at the same lattice site . the open channel component of the scattering state in the presence of the resonance may then be expressed as @xmath143 \label{psi_op_res}\ ] ] for all @xmath144 . again the relative quasi - momentum of the scattering atoms is @xmath59 , and @xmath145 is the molecular self - energy , which for the onsite coupling may be written as : @xmath146 . from the form of the scattered wave ( [ psi_op_res ] ) the total transmission coefficient including both background and resonant contributions is found to be of the fano form @xcite @xmath147 the fano parameters @xmath148 and @xmath149 depend on the resonance shift and width functions , which are related to the real and imaginary parts of the molecular self - energy , @xmath150 where the real part @xmath151 describes the self consistently determined shift of the resonance position from @xmath152 to @xmath153 , and @xmath154 is the decay width of the resonance state due to the coupling to the continuum . outside the continuum @xmath155 , and the molecules are stable against dissociation . the real part of the molecular self - energy outside the band @xmath156 determines the energies of the bound states for a given resonance energy as described in section [ sec : bound - states ] . the width and shift of the resonance are illustrated in fig . [ fig : shift_and_width ] along with the fano parameters for @xmath157 . throughout the paper we illustrate our analytical results for a set of parameters where @xmath158 and @xmath159 . as outlined in appendix [ sec : calc - param ] this corresponds to @xmath160rb atoms in a lattice with longitudinal and transverse depths of @xmath161 and 30@xmath161 , respectively , in an applied magnetic field tuned close to the feshbach resonance near 414 g @xcite . and width , @xmath162 of the resonance for @xmath157 where the band edges are at @xmath163 ( a ) . the real part of the molecular self - energy is continuous across the band edges and has a pole at the uncoupled bound state energy , @xmath164 ( see section [ sec : bound - states ] ) . the fano parameters , @xmath165 and @xmath10 , parametrize the shape of the transmission resonance ( b ) . here @xmath166 . ] we note that @xmath167 is independent of the coupling , @xmath168 . it is related to the background transmission probability through the relation @xmath169 , which is a consequence of inversion symmetry in the problem @xcite . the reflection maximum of the fano profile occurs at @xmath170 and is nearly independent of the center of mass motion , due to the weak tunneling of the closed channel molecules . this point of total reflection only occurs , when the bare resonance state lies inside the continuum . just as in the non - resonant case the reflection probability is unity at the band edges . hence the reflection profile has two minima , when @xmath171 . when the bare resonance state lies outside the band , @xmath172 has a single minimum . the two - channel reflection profiles are plotted in figs . [ fig : r_vs_k](a)-(d ) and [ fig : r_vs_eres](a ) as colormaps and in figs . [ fig : r_vs_k](e)-(h ) and [ fig : r_vs_eres](c ) as solid lines . for the parameters used in the quantitative calculations throughout this paper @xmath173 , and hence the resonance energy appears nearly independent of @xmath38 in figs . [ fig : r_vs_k](a)-(d ) . for a fixed @xmath157 . the horizontal dashed line is the repulsively bound pair state existing above the band for @xmath174 , while the diagonal dashed curve is the band averaged resonance energy , @xmath175 . ( b ) closed channel population of the bound states . ( c ) cuts showing the reflection probability at @xmath176 0 , -3 , 10 , and @xmath177 ( solid lines from top to bottom at @xmath27 ) . the dashed curve indicate the reflection profile in the absence of the coupling . ] the condition for total transmission is @xmath178 , corresponding to @xmath179 . as @xmath152 is ramped upwards ( downwards ) starting below ( above ) the band , this point of destructive interference between the direct and resonant scattering amplitudes first appears at the magnetic field , where a bound state of the coupled channels system enters the continuum to become a scattering resonance . as we shall see below , this occurs at two critical values where @xmath180 - exactly the points delineating the region , where the condition @xmath178 can be satisfied . the reflection profile will thus only have both a fano maximum and a fano minimum if the conditions @xmath171 and @xmath181 are simultaneously met . neglecting the background interaction corresponds to the limit @xmath182 , or equivalently @xmath183 . in this case the resonance shift vanishes and the reflection probability assumes a simple breit - wigner - like profile @xcite @xmath184 ^ 2+\frac{\hbar^2\gamma_0 ^ 2(e , k)}{4}},\ ] ] where @xmath185 . the opposite limit of @xmath186 on the other hand corresponds to @xmath187 and @xmath188 , giving rise to a breit - wigner - like form of the transmission coefficient @xmath189 ^ 2+\frac{\hbar^2\gamma_\infty^2(e , k)}{4}},\ ] ] corresponding to a dip in the reflection probability at the resonance , which now occurs at @xmath190 . the resonance shift , @xmath191 , is independent of energy and momentum , while the width is given by @xmath192 . both the shift and the width approach zero as the strength of the onsite interaction is increased . the observations regarding indistinguishable particles in section [ sec : identical ] can be directly carried over to the situation with more than one scattering channel . again , the transmission coefficient @xmath193 has no straightforward meaning in the scattering of identical bosons . but following fano @xcite it is clear that the transition probability from an initial state @xmath91 to the stationary scattering state @xmath194 under the action of some transition operator @xmath92 is @xmath195 . therefore , the resonance profile is accessible spectroscopically by scanning the energy of the final state across the continuum . starting from a deeper bound state of the system the dissociated atom signal will have the same energy variation as @xmath193 . the scattering amplitude in the coupled channels case consists of a background and a resonant contribution , @xmath196 , where the resonant part @xmath197 ^ 2}{e - e_{\rm{res}}(b , k)-\sigma_m(e , k ) } \label{f_res}\ ] ] is obtained by comparing ( [ psi_op_res ] ) with @xmath198 for @xmath199 . the phase shift is related to the scattering amplitude in the same way as in the single - channel case . as in the case of single - channel scattering we define the one - dimensional scattering length , @xmath200 , from the limit of the scattering amplitude , @xmath201^{-1}$ ] , at the extrema of the relative energy , @xmath202 . after some algebra we arrive at the simple expression @xmath203 \label{a_1d}\ ] ] for the scattering length , with the upper and lower sign corresponding to the top and the bottom edge of the continuum , respectively . this lattice scattering length is of the standard feshbach form @xcite , and reduces to the non - resonant expression ( [ a_bg ] ) , when the coupling @xmath168 is set to zero . however , the generalized , magnetic field dependent scattering length at the top and the bottom of the continuum differ , in that they do not diverge at the same value of the resonance energy . the motivation for the chosen sign convention of the scattering length at the upper band edge is guided by the rationale that the two scattering lengths then become equal in the limit of a deep lattice , where @xmath204 goes to zero . ) . solid lines show @xmath205 at @xmath206 0 , 0.4 , and 0.7 @xmath40 ( from the left to the right ) , while the dashed lines correspond to @xmath207 at the same @xmath38-values ( increasing from the right to the left ) . ( b ) the corresponding binding energies below the band ( solid lines ) and above the band ( dashed lines ) and the limiting form ( [ e_bind_universal ] ) , valid near the feshbach resonances in @xmath208 ( thin dot - dashed lines ) . ] it is convenient to define a pair of detunings @xmath209 , where @xmath210 . we thus get two resonance positions , one for scattering near the upper band edge , and one for scattering near the lower band edge , as shown in fig . [ fig : a_1d ] . we caution that the relative position of @xmath211 and @xmath212 depends on the sign of the magnetic moment difference , @xmath123 . if @xmath213 ( @xmath214 ) , then @xmath211 lies above ( below ) @xmath212 . the lattice resonance positions @xmath211 and @xmath212 are shown in fig . [ fig : b0 ] as function of the lattice depth , @xmath215 . they are separated by @xmath216 and approach each other as @xmath215 is increased , since @xmath217 decreases when the lattice becomes deeper . furthermore , in the limit where both the transverse and the longitudinal lattice depths are large , _ i.e. _ , @xmath218 , the wannier orbitals approach harmonic oscillator states , and we have that @xmath219 , while @xmath220 and @xmath221 . this gives the following asymptotic behavior of the magnetic field resonance position @xmath222 + \frac{e_{\rm r}}{\delta\mu}\sqrt{v_0^{\parallel } } + \frac{2e_{\rm{r}}}{\delta\mu}\sqrt{v_0^\perp}\ ] ] where @xmath223 is the feshbach resonance position in the absence of the lattice , and the dimensionless function @xmath224 is defined in appendix [ resonance_shift ] . the asymptotic behavior is indicated by the black dash - dotted curves in fig . [ fig : b0 ] . . @xmath211 and @xmath212 are plotted for @xmath225 ( dashed curves ) and @xmath226 ( solid curves ) . in the example @xmath227 is positive so both @xmath211-curves lie above their respective @xmath212-curves . the left axis shows @xmath228 relative to the feshbach resonance position in the absence of the lattice , and the right axis shows the actual magnetic field position for the specific feshbach resonance considered @xcite . the black dash - dotted curves are given by eq . ( [ eq : asymptoticb ] ) . ] as shown in sec . [ sec : binding_energy ] the divergence of the lattice scattering length is associated with a bound state entering or leaving the continuum . hence the lattice gives rise to a @xmath38-dependent shift of the magnetic field threshold for the formation of two - body bound states away from the resonance position in free space . this fact , illustrated in fig . [ fig : a_1d ] , was also observed by orso _ et al . _ @xcite who solved the two - body physics in a one - dimensional optical lattice within a single - channel model . the zero - crossing of the scattering length occurs for @xmath229 , which corresponds to magnetic field values , @xmath230 , shifted from the resonance positions . the bound state solutions of the coupled - channels problem have energies , @xmath231 , coinciding with the poles of the scattering amplitude @xmath232 . the pole locations satisfy the equation @xmath233=\frac{{\mathcal{w}}^2{\rm{sgn}}(e_b)}{e_b - e_{\rm{res}}(b , k ) } , \label{e_bind}\ ] ] and can be determined as a subset of the roots of a quartic polynomial . the bound state energies are plotted as the solid lines in figs . [ fig : r_vs_k](a)-(d ) and [ fig : r_vs_eres](a ) . if @xmath234 the entrance channel bound state energy is @xmath235 with a wavefunction proportional to @xmath236 . this single - channel bound state is indicated by the sinusoidal dashed line in fig [ fig : r_vs_k](a)-(d ) and the horizontal dashed line in fig . [ fig : r_vs_eres](a ) . an important consequence of the structured continuum in the lattice potential is that the stability of a molecular bound state may depend on its center of mass momentum . for a given magnetic field , the scattering resonance only exists for a range of center of mass motion bloch states , as shown in fig . [ fig : r_vs_k](c)-(d ) @xcite . outside this range there is instead a true bound state of the system . such motionally bound states are stabilized by their kinetic energy , which displaces the pair state into the band gap for the relative motion , where the pair can not disintegrate . hence , for a fixed magnetic field , where a molecule at rest is unstable , accelerating the lattice provides a new avenue for tuning the feshbach resonance . the bound state pairs can tunnel together through the lattice with an effective tunneling rate , which depends on their binding energy . we now determine the tunneling matrix element of the pairs directly from their dispersion relation . in the absence of atomic tunneling processes , @xmath237 , and the energies of the bound states are given by @xmath238 ^ 2}{4}+{\mathcal{w}}^2}. \label{ebound_no_tunneling}\ ] ] with the inter - channel coupling turned off these zero tunneling bound state energies reduce to the uncoupled pair state energies for @xmath239 ; @xmath175 and @xmath42 . with a small non - vanishing single - particle tunneling amplitude we may expand ( [ e_bind ] ) to lowest order in @xmath240 and find @xmath241 comparing the momentum dependent part of the bound state energy with the dispersion relation for non - interacting pairs , @xmath242 , we see that in the limit of a deep lattice the bound atom pairs tunnel through the lattice with an effective tunneling rate @xmath243 as one would expect from arguments based on second order perturbation theory @xcite . we emphasize that the molecule dispersion depends on the applied magnetic field . cooperative tunneling of atom pairs consistent with an effective pair hopping matrix element given by @xmath244 has been demonstrated experimentally by measurement of the frequency of population transfer in a double well @xcite and by demonstrating that the transport properties of an attractive fermi gas in an optical lattice depends strongly on the formation of local pairs @xcite . instead of the bound state energy it is often instructive to consider the _ binding energy _ , defined as the distance of the bound state energy from the ( upper or lower ) band edge : @xmath245 the binding energy is positive for a bound state lying above the band and negative for a bound pair below the continuum . from fig . [ fig : r_vs_k](a)-(d ) we observe that the binding energy increases with increasing @xmath246 in agreement with the single - channel calculation in @xcite , which only discussed bound states with negative binding energy as no transverse confinement was included . as @xmath231 approaches either of the band edges , the binding energy vanishes in a particularly simple way . to see this we observe that @xmath247 for @xmath248 . replacing @xmath231 with @xmath249 in the denominator on the right hand side in ( [ e_bind ] ) it then follows that @xmath250 ^ 2},\ ] ] as the bound state energy approaches @xmath249 . the denominator is exactly the resonance shape of the generalized scattering length ( [ a_1d ] ) , and the binding energy therefore approaches the universal form @xmath251 when the scattering length is tuned to be large with respect to the lattice spacing @xmath252 in the vicinity of the feshbach resonance . we have introduced the effective _ reduced _ mass @xmath253 of the atom pair with center of mass quasi - momentum , @xmath38 , in the lattice . its limits at the top and the bottom of the band are @xmath254 , with the upper ( lower ) sign applying for @xmath255 ( @xmath256 ) . these limits are plotted in fig . [ fig : m_eff ] as a function of @xmath38 . from the expression for the effective reduced mass it is clear that the sign of the universal binding energy expression ( [ e_bind_universal ] ) adheres to the definition above . the criteria of applicability for the quadratic approximation is @xmath257 , or @xmath258 . the universal expression ( [ e_bind_universal ] ) is compared with the exact binding energies in fig . [ fig : a_1d](b ) both above and below the continuum . due to the proximity of the zero crossing of @xmath259 , where ( [ e_bind_universal ] ) diverges , the agreement below the band is worse than above the band . from eqs . ( [ a_1d ] ) and ( [ e_bind_universal ] ) it is clear that the binding energy vanishes when @xmath260 . this condition for the disappearance of the feshbach bound state coincides with the critical resonance energy at which the fano profile develops a point of total transmission , as anticipated above . experimentally the binding energy of the dimer states can be measured by modulating the depth of the lattice . the bound pairs dissociate , when the frequency of modulation satisfies a resonance condition , such that the dimers can exchange energy with the undulating lattice . the resonance is marked by the appearance of unbound atoms @xcite . the formal solution of the bound state problem consistent with the single - pole approximation for the closed channel green s function is @xmath261 where the closed channel population , @xmath262 , is determined by the normalization condition : @xmath263 . in coordinate space the open channel component of the bound state wavefunction therefore becomes @xmath264 while the closed channel component of the bound state is described by @xmath265 . here we have used that the coupling is of zero range , and that the bare resonance state is localized on a single lattice site . the closed channel weight of a bound atom pair with energy @xmath231 is @xmath266 ^ 2}\frac{|e_b|}{(e_b^2-e_k^2)^{3/2}}\right]^{-1 } , \label{z}\ ] ] which is plotted in figs [ fig : r_vs_k](i)-(l ) and [ fig : r_vs_eres](b ) . as the bound state approaches the upper or the lower band edge , it follows from the limiting form of the binding energy that @xmath267 , and consequently @xmath262 vanishes as @xmath268 in the limit where @xmath269 . in momentum space the open channel part of the bound state wavefunction is given by @xmath270 which is peaked near the center of the brillouin zone ( @xmath271 ) when the bound state energy is negative , and peaked near the edges of the brillouin zone ( @xmath272 ) for @xmath273 , making it possible to discern the bound states with @xmath274 from those with @xmath273 in time of flight imaging @xcite . this reflects that @xmath275 is predominantly comprised of the @xmath276-states with energies , @xmath94 , in proximity of @xmath231 . for @xmath273 ( @xmath274 ) this corresponds to the states near the top ( bottom ) of the band . the height of the peaks is reduced as @xmath38 approaches @xmath277 , where all relative motion quasi - momentum states are degenerate in our model . the closed channel component of the bound state is independent of the relative momentum : @xmath278 . , where a weakly bound molecular state ( solid line , narrow bars ) at rest is situated just below the band . the repulsively bound pair state lying above the band ( dashed line , thick bars ) , is peaked at the edge of the brillouin zone , and its sign alternates from one site to the next . the first column shows the momentum space wavefunction ( [ phi_mom ] ) in units of @xmath279 , the second column @xmath280 . ] the size of the bound state @xmath281 ^ 2}\frac{e_k^2}{(e^2_b - e^2_k)^2},\ ] ] reflects the magnitude of the binding energy of atom pairs . in particular , it diverges when the bound state enters the continuum from above or from below : @xmath282 in this limit the open channel part of the bound state wavefunction assumes a universal form @xmath283^{z / a}.\ ] ] the bound state wavefunctions are shown in fig . [ fig : bound_state_wf ] in both coordinate and momentum space for @xmath284 . at this resonance energy a bound molecular state exists just below the band for @xmath157 ( lower bound state branch in fig . [ fig : r_vs_k](b ) ) . with a small binding energy , this state extends over several lattice sites as shown in fig [ fig : bound_state_wf ] . for the repulsively bound pair state lying above the band ( upper bound state branch in fig . [ fig : r_vs_k](b ) ) the sign of the coordinate space wavefunction alternates between lattice sites . we have previously shown that a molecule has a finite probability of surviving a sweep through the continuum band when the applied magnetic field is varied across the feshbach resonance @xcite . the survival probability increases the faster the sweep and the weaker the coupling to the continuum . given the existence of motionally bound states , an alternative experiment to investigate the passage of a molecule through the continuum may then be conducted by starting with the applied magnetic field held at a value such that @xmath285 , and molecules at rest prepared in the bound state with energy below @xmath175 ( represented by the the lower branch of bound states in figs . [ fig : r_vs_k ] and [ fig : r_vs_eres ] ) . accelerating the lattice imparts momentum @xmath286 to the molecules . if the magnetic field is then ramped adiabatically to a new value , where @xmath287 the moving molecules remain stable , while molecules at rest are now subject to dissociative decay . with a further acceleration of the lattice , either in the opposite direction , or in the same direction , but letting @xmath38 cross one or several brillouin zones , the molecules may be swept through the continuum to a final state with center of mass momentum @xmath288 , where the molecules are stable provided @xmath289 . the probability of a molecule surviving the center of mass momentum ramp through the continuum depends on the rate of lattice acceleration and the strength of the coupling to the continuum states . in this section we will perform a spectral analysis based on the full green s functions of both the open channel atoms and closed channel molecules . in general , given a retarded green s operator , @xmath290 , the corresponding spectral function is the imaginary part of the diagonal green s function , @xmath291 , in a basis @xmath292 of orthonormal states . for the unperturbed system ( @xmath42=@xmath168=0 ) the spectral function is @xmath293 where the density of states is @xmath294 for @xmath57 . the unperturbed system has no bound states , and hence the spectral function is zero at all energies outside the continuum band . inside the band the scattering states are equally distributed over the whole lattice , making the spectral function translationally invariant and directly proportional to the density of states , which diverges at the band edges , @xmath295 . both for the single - channel and the two - channel case we will need the entrance channel green s function , @xmath296 , _ i.e. _ , the propagator from one relative coordinate in the lattice , @xmath297 , to another , @xmath88 . when we consider just a single scattering channel this can be obtained from the dyson equation ( [ dyson ] ) : @xmath298 where we have used that the unperturbed green s function is translationally invariant . this expression for @xmath299 reduces to eq . ( [ gu_z ] ) when @xmath88 or @xmath297 is equal to zero . for the spectral function we have to treat @xmath57 and @xmath60 separately . outside the continuum @xmath63 is real and we obtain the spectral function @xmath300 where @xmath301 is the decay constant for the bound state . this gives the probability distribution for finding a bound state atom pair @xmath88 lattice sites apart , when the bound state energy is @xmath302 . inside the band we find @xmath303 where @xmath73 is the scattering wave defined in eq . ( [ psi_e_bg ] ) . changing the sign of @xmath88 is equivalent to changing the direction of the incoming wave in the scattering wave from @xmath276 to @xmath304 . these two waves are degenerate and therefore it is natural to expect the average of the two for a given energy in the spectral function . we have used that the density of states is unchanged for elastic scattering , _ i.e. _ @xmath305 is also the density of states for atoms scattering under the influence of @xmath41 . it is instructive to calculate the total spectral weight residing inside the continuum band : @xmath306 with the addition of the bound state contribution ( [ eq : a_u_bound ] ) the sum rule , @xmath307 , is then seen to be satisfied . when we have two coupled channels there is a spectral density for both the open channel atoms and the closed channel molecules . we start by addressing the latter . the dressed green s function for the closed channel is found by summing all the terms in the dyson series based on the bare propagator ( [ g_cl_approx ] ) . the result is @xmath308 where @xmath309 is the molecular self - energy ( [ sigma_cl ] ) . if @xmath310 is outside the continuum , @xmath311 is real , and the spectral function is only non - zero at the poles of @xmath312 , which coincide with those of the scattering amplitudes , _ i.e. _ the bound states of the coupled channels problem ( [ e_bind ] ) . the spectral weight at the poles is given by the corresponding residues , which are just the probability of finding the bound state pair in the closed channel , @xmath262 : @xmath313 for energies inside the continuum , we obtain the usual structure , @xmath314 ^ 2 + \frac{\hbar^2\gamma^2(e , k)}{4 } } , \label{eq : a_m_continuum}\ ] ] where the energy dependent shift , @xmath315 , and decay rate , @xmath316 , are defined in eqs . ( [ eq : delta ] ) and ( [ eq : gamma ] ) , respectively . in the middle of the band the shift and width are slowly varying functions of energy , and the molecular spectral function has an approximately lorentzian shape with a peak position given approximately by the resonance energy @xmath152 . however , near the band edges significant deviations from this simple form arise due to the strong interference with the continuum threshold . this leads to a double peak structure of @xmath317 , when the bound state is about to enter the continuum from below , as is evident from fig . [ fig : a_m ] , which shows @xmath317 as a function of the energy and @xmath152 for a molecule at rest . . inside the band @xmath317 is a continuous function given by ( [ eq : a_m_continuum ] ) . outside the band the delta function peaks at the bound state poles of the propagator , @xmath312 , are indicated by vertical lines with a height of @xmath318 . offset below is a contour plot of @xmath317 inside the band , with bright colors indicating high spectral density . the solid lines in this plane are the bound state energies , while @xmath175 is indicated by the dashed line . ] the molecule lifetime also follows from fermi s golden rule , @xmath319 . though the density of states diverges at the edges of the continuum ( van hove singularities ) the decay rate goes to zero at both thresholds , due to vanishing overlap of the background scattering state with the bare resonance state when @xmath320 . for the open channel the full green s function in the resonant case reads @xmath321 where @xmath322 $ ] is the self - energy for the relative motion of an atom pair in the open channel . this gives the spectral function @xmath323 ^ 2\delta(e - e_b)\ ] ] for energies outside the band . the open channel green s functions has exactly the same poles as the molecular propagator ( [ eq : g_m ] ) , but the residues at the atomic poles depend on the separation between the atoms in the pair . however , using the identity @xmath324 one may show that @xmath325 such that the combined spectral weight of the open channel and closed channel at a pole is unity . we must sum over all possible separations for the open channel weight , since the component of the bound state wavefunction in the open channel is distributed over the lattice according to ( [ eq : psi_b_op ] ) . thus far we have only investigated the dynamics in the lowest bloch band of the lattice potential . in this section we analyze the neglected coupling to higher bands . first we estimate the size of the second order energy shift from the excluded higher bands and derive a criterion for the validity of our theoretical model . next , in sec . [ sec : incl - high - bloch ] we discuss how a modest coupling to higher atomic bands is expected to modify the central conclusions of this work . in this section we set up a criterion for using the theory , by comparing the second order energy shift from the omitted atomic and molecular exited bloch bands to the relevant energy scales . the dominant contribution to the energy shift comes from the bare resonance state coupling to atom pairs in higher bands , while the shift from open channel atoms coupling to higher molecular bloch bands are several orders of magnitude smaller for two reasons . firstly , the coupling matrix elements between an atom pair in the lowest band and an excited band bare molecule are much smaller than the coupling to a molecule in the lowest band ( see appendix [ sec : mol - couplings ] ) . secondly , the energy splitting between the ground and excited molecular bands is larger than the atomic band gap , since the molecules experience a deeper lattice potential . hence we approximate the unperturbed wavefunction by @xmath326 , the resonance state in the lowest molecular band . for simplicity we will furthermore assume the width of the bloch bands to be small compared to the separation from the lowest band , so that the energy can be approximated by the average energy of the band . under these assumptions the second order energy shift is @xmath327 where the sum is over pairs of atoms in the @xmath12th and @xmath1th bloch band , @xmath328 is the inter - band coupling matrix element defined in appendix [ sec : calc - param ] , and @xmath329 is the average energy of the @xmath12th band . if the atoms are identical bosons , we can take the sum to be over @xmath330 to avoid double counting . we will require the shift to be small compared to the relevant energy scales . if @xmath331 is comparable to the gap between the ground and first excited band @xmath332 , the higher bands will become populated , and the lowest band approximation of our model will break down @xcite . on the other hand a stronger condition can be derived by requiring that the energy shift be small in comparison with the width of the ground band , @xmath333 , which sets the relevant energy scale for the dynamics in our model . by fitting the respective energies from numerical band structure calculations at different lattice depths we arrive at the condition @xmath334 where @xmath335 when comparing @xmath331 with the band gap , and @xmath336 when requiring the energy shift to be much smaller than the width of the lowest band @xcite . the scaling factors are the bohr radius , @xmath337 , the bohr magneton , @xmath338 , gauss ( g ) , the atomic mass unit ( @xmath339 ) , and nanometers . if both criteria are met , coupling to higher bands can be safely neglected . otherwise , if the energy shift incurred from the coupling to atoms in higher bloch bands is significant with respect to the band width yet negligible with regards to the band gap , we might expect quantitative changes to our results , but as we discuss below , the qualitative features of our model will remain . in the case where the lowest band approximation fails coupling to higher bands can be systematically accounted for in the tight binding limit @xcite , leading to an effective single - band hamiltonian with a non - local coupling between dressed molecules and unbound atom pairs @xcite . both of the above stated conditions can be fulfilled for the feshbach resonance studied by syassen _ _ in @xmath160rb near 414 g @xcite . in the entrance channel both atoms are in the @xmath340 hyperfine state . the resonance is narrow , with @xmath341 mg , and has a very small difference in magnetic moments between the two channels , @xmath342 khz / g , which is preferable if eq . ( [ eq : validity_condition ] ) is to hold . a final condition for our results to be experimentally relevant is that the variation of the resonance energy due to magnetic field noise @xmath343 is much smaller than the bandwidth @xmath333 . with a lattice wavelength of 830.44 nm @xcite the recoil energy is @xmath344 khz ( equivalent to @xmath345 ) and for the particular resonance considered here @xmath34632 mg . for a longitudinal lattice depth of @xmath161 we therefore require @xmath347 according to table [ tab : parameters ] . in ref . @xcite the magnetic field noise was found to be less than 4 mg . as discussed in the previous subsection the major correction to our model comes from the coupling to higher atomic bands . we now include these couplings into our model . for each atomic band we apply the analysis of sec . [ sec : entr - chann - greens ] to obtain band specific green s functions @xmath348 and @xmath349 for the relative motion of one atom in bloch band @xmath12 and another in band @xmath1 . as for the first band @xmath350 is the green s function for noninteracting particles and @xmath351 is the green s function for particles with an on - site interaction , @xmath41 . we determine @xmath348 and @xmath349 by replacing @xmath94 by @xmath352 $ ] in eq . , and replacing @xmath42 by the band specific interaction strength @xmath353 in eq . ( for the definition of @xmath353 see appendix [ sec : calc - param ] ) . we now introduce the coupling @xmath115 between the first molecular band and all the atomic bands . this gives the molecular self - energy @xmath354 which reduces to ( [ sigma_cl ] ) , when we exclude contributions from higher bands . because we only consider energies well outside the excited bands , @xmath355 is real and slowly varying with energy for @xmath356 . therefore we conclude that the lifetime @xmath316 of the molecules is unchanged and that the change in the resonance shift @xmath357 must vary slowly . a change of the resonance shift by an amount @xmath358 will affect some of the results in secs . [ sec : two - chann - scatt ] , [ sec : bound - states ] and [ sec : resonant - case ] . the changes amount to simply replacing @xmath359 by @xmath360 , a renormalized resonance energy . for @xmath27 and @xmath157 we obtain for the parameters in @xcite an approximate relationship for the change in resonance shift @xmath361 due to the higher atomic bands @xcite . for the lattice depth used throughout this paper @xmath358 is only 2% of the bandwidth @xmath333 , and it is reasonable to neglect higher bands . a direct coupling between ground and excited atomic bands may be included by renormalizing the local interactions to include the effect of virtual transitions to excited levels through band changing collisions . this leads to an energy dependent correction to the coupling parameters @xcite . but since the correction is inversely proportional to the gap separating the ground and excited bands , it is also expected to be small . we have assumed onsite interactions and only nearest - neighbor tunneling in all atomic bands which is a questionable assumption for the higher bands , but eq . ( [ eq : deltachange ] ) still provides an estimate of the magnitude of the correction to the real part of the molecular self - energy . we have presented a complete description of the feshbach physics of an atom pair in an optical lattice represented by a discrete two - channel hamiltonian . we have demonstrated how the structured continuum due to the periodic potential modifies the resonance line shape and position as well as the properties of the bound states outside the continuum . the lattice potential facilitates repulsively bound pair states in the gap between the two - particle bloch bands , and near a feshbach resonance a further class of exotic bound states may be created , which are only stable when their center of mass momentum exceeds a critical value , thus enabling us to tune the feshbach physics by controlling the molecular motion . ignoring the small modulation of the resonance energy due to tunneling of the bare molecules , the feshbach molecules are stable bound states of the system if their center of mass quasi - momentum in the first brillouin zone satisfies @xmath362 , where @xmath363 if @xmath364 the pair state lies inside the continuum and decays with a rate @xmath162 , which we have determined analytically . for magnetic fields such that @xmath365 the feshbach molecules are stable for all @xmath38 . note that while the decay rate near threshold is proportional to the density of states , which diverges at the band edges , @xmath162 goes continuously to zero as @xmath310 approaches the edge of the continuum , since the overlap between the bare resonance state and the open channel continuum states is suppressed due to the background interactions in the open channel . an additional effect of the background interactions is to shift and modify the resonance profile . with a non - zero onsite interaction @xmath42 the position of the resonance is shifted away from the bare resonance energy @xmath120 , and the transmission of atoms in the open channel is described by a fano profile , with a shape determined by the energy dependent shift and width of the resonance . the energy dependence is most pronounced near the band edges , where the coupling to the continuum is strongest . this gives rise to threshold effects , when the resonance energy is tuned close to the band edge , as evidenced e.g. in the double peak structure of the molecule spectral function . we have outlined how the energy dependence of the scattering can be mapped out spectroscopically by inducing dissociative transitions from molecular bound states to two - particle scattering states inside the band . from the expansion of the scattering amplitude as the collision energy approaches the band edges we have defined a generalized scattering length , which exhibits a feshbach resonance at both the lower and the upper threshold of the continuum . the diverging scattering length coincides with the appearance / disappearance of the feshbach bound state . in the limit of a deep lattice potential the lowest bloch band narrows , and these two resonances coalesce . since the bare molecules and the entrance channel atoms experience different lattice induced energy offsets , the position of both the upper and the lower threshold feshbach resonance is shifted relative to the position of the feshbach resonance in free space . for the specific resonance we have considered here , this shift exceeds the width of the resonance . for the interaction between the atoms in the open channel we assume that @xmath41 is a contact potential of strength @xmath366 , where @xmath367 is the ( free space ) background scattering length . therefore the on - site interaction between a pair of atoms in bands @xmath12 and @xmath1 becomes @xmath368 ^ 4 \nonumber \\ & & \times\int dx_3 \ , \big [ w_{n0}^\parallel(x_3)w_{m0}^\parallel(x_3 ) \big]^2,\end{aligned}\ ] ] under the assumption that only the lowest band in the transverse lattice is occupied . since all lattice wells are identical , the integrals can be performed over wannier functions centered on any site . the coupling to the closed channel is of short range , since it is proportional to the difference between the atomic triplet and singlet potentials , which have the same long range form @xcite . in addition , the closed channel bound state has a size ( typically tens of ) , which is small compared with the lattice spacing ( hundreds of nm ) . we thus assume a purely on - site coupling , which in free space would couple molecular and atomic plane wave states with a strength @xmath369 . here @xmath123 is the difference between the closed and open channel magnetic moments and @xmath370 is the magnetic field width of the feshbach resonance in the absence of the lattice . in the discrete lattice model this results in the matrix element @xmath371 for the coupling of a pair of atoms in bands @xmath12 and @xmath1 to a bare molecule in band @xmath372 . the matrix elements are zero whenever @xmath373 is even due to parity of the wannier functions . a molecule experiences twice as deep a lattice as the atoms , see eq . ( [ v_lat_mol ] ) , and has twice the atomic mass . consequently , separate wannier functions , @xmath374 , have to be calculated for the center of mass motion of the resonance state . with the exception of sec . [ sec : validity ] we will consider exclusively the lowest bloch band for the atoms and molecules . therefore , we will only need three parameters to characterize the dynamics in the lattice : @xmath375 , @xmath376 and @xmath377 . as discussed in section [ sec : validity ] our discrete lattice model is valid for a narrow feshbach resonance . as a particular example we use the resonance in @xmath160rb near 414 g , which has been studied by syassen _ et al . _ @xcite . for this resonance @xmath341 mg and @xmath342 khz / g , while the background scattering length is 100.8 bohr radii . for all the plots in this paper we consider a lattice with @xmath378 and @xmath379 . this results in interaction parameters @xmath380 and @xmath381 . table [ tab : parameters ] gives the relevant model parameters for this resonance and a range of lattice depths . .discrete lattice parameters for the @xmath160rb feshbach resonance near 414 g. the recoil energy is @xmath344 khz . [ cols="^,^,^,^,^,^",options="header " , ] in free space the coupling between the collision channels induces a shift from the magnetic field @xmath382 , where the resonance state is degenerate with the entrance channel threshold [ @xmath383 , to the position @xmath384 of the scattering length divergence . this offset may be approximately evaluated by @xmath385 here @xmath386 $ ] is a dimensionless function of the ratio @xmath387 between the background ( free space ) scattering length and the so - called mean scattering length @xmath388 , which is a characteristic length scale of the long range @xmath389 part of the molecular potential @xcite . in the lattice the energy of the closed channel resonance state relative to the zero of energy ( center of the lowest band ) is then given by @xmath390+e_m(k)-e_a,\ ] ] where the lattice induced energy shifts @xmath391 and @xmath392 of the open and closed channel , respectively , are defined in fig . [ fig : energydiagram ] . the molecular wannier states , @xmath393 , where @xmath394 and @xmath395 refers to the molecular bloch band index and lattice site index along direction @xmath91 , respectively , constitute a complete set on which we can expand the atomic function @xmath396\times w_{n0}^\parallel(x_3)w_{m0}^\parallel(x_3)$ ] . if we denote the expansion coefficients @xmath397 , we have @xmath398 and @xmath399 , where @xmath400 is the @xmath401-norm . now a simple norm argument gives the following identity for the neglected molecular couplings to higher bands @xmath402 where @xmath403 for the 414 g resonance in @xmath160rb . for a lattice with @xmath404 and @xmath226 this results in the estimate @xmath405 for an atom pair in the lowest band ( @xmath406 ) . we therefore neglect the coupling of ground band atoms to molecules in higher bands in comparison with the matrix elements , which couple molecules in the lowest bloch band with excited atom pairs . m. ben dahan , e. peik , j. reichel , y. castin , and c. salomon , phys . lett . * 76 * 4508 ( 1996 ) . s. r. wilkinson , c. f. bharucha , k. w. madison , q. niu , and m. g. raizen , phys . rev . lett . * 76 * , 4512 ( 1996 ) . m. greiner , m. o. mandel , t. esslinger , t. hnsch , and i. bloch , nature ( london ) * 415 * , 39 ( 2002 ) . i. bloch , j. dalibard , and w. zwerger , rev . * 80 * , 885 ( 2008 ) . o. mandel , m. greiner , a. widera , t. rom , t. w. hnsch , and i. bloch , nature ( london ) * 425 * , 937 ( 2003 ) . t. khler , k. gral , and p. s. julienne , rev . phys . * 78 * , 1311 ( 2006 ) . g. thalhammer , k. winkler , f. lang , s. schmid , r. grimm , and j. hecker denschlag , phys . . lett . * 96 * , 050402 ( 2006 ) . t. volz , n. syassen , d. m. bauer , e. hansis , s. drr , and g. rempe , nature phys . * 2 * , 692 ( 2006 ) . k. winkler , g. thalhammer , f. lang , r. grimm , j. hecker denschlag , a. j. daley , a. kantian , h. p. bchler , and p. zoller , nature ( london ) * 441 * , 853 ( 2006 ) . c. ospelkaus , s. ospelkaus , l. humbert , p. ernst , k. sengstock , and k. bongs , phys . lett . * 97 * , 120402 ( 2006 ) . p. a. f. da veiga , l. ioriatti , and m. ocarroll , phys . e * 66 * , 016130 ( 2002 ) . s. m. mahajan and a. thyagaraja , j. phys . a * 39 * , l667 ( 2006 ) . j. hecker denschlag and a. j. daley , proceedings of the international school of physics `` enrico fermi '' , course clxiv , ultra - cold fermi gases , e - print arxiv : cond - mat/ 0610393 . m. grupp , r. walser , w. p. schleich , a. muramatsu , and m. weitz , j. phys . b. * 40 * , 2703 ( 2007 ) . n. nygaard , b. i. schneider , and p. s. julienne , phys . a * 73 * , 042705 ( 2006 ) . u. fano , phys . rev . * 124 * , 1866 ( 1961 ) . p. w. anderson , phys . rev . * 124 * , 41 ( 1961 ) . duan , phys . lett . * 95 * , 243202 ( 2005 ) . l. d. carr and m. j. holland , phys . rev . a * 72 * , 031604(r ) ( 2005 ) . f. zhou , phys . b * 72 * , 220501(r ) ( 2005 ) . n. nygaard , r. piil , and k. mlmer , phys . a * 77 * , 021601(r ) ( 2008 ) . s. flling , s. trotzky , p. cheinet , m. feld , r. saers , a. widera , t. mller , and i. bloch , nature ( london ) * 448 * , 1029 ( 2007 ) . n. strohmaier , y. takasu , k. gnter , r. jrdens , m. khl , h. moritz , and t. esslinger , phys . lett * 99 * , 220601 ( 2007 ) . n. syassen , d. m. bauer , m. lettner , d. dietze , t. volz , s. drr , and g. rempe , phys . lett . * 99 * , 033201 ( 2007 ) . j. u. nckel and a. d. stone , phys . b * 50 * , 17415 ( 1994 ) . m. olshanii , phys . lett . * 81 * , 938 ( 1998 ) . t. bergeman , m. g. moore , and m. olshanii , phys . * 91 * , 163201 ( 2003 ) . v. peano , m. thorwart , c. mora , and r. egger , new . j. phys . * 7 * , 192 ( 2005 ) . j. i. kim , j. schmiedmayer , and p. schmelcher , phys . a * 72 * , 042711 ( 2005 ) . v. a. yurovsky , phys . a * 71 * , 012709 ( 2005 ) . v. a. yurovsky , phys . rev . a * 73 * , 052709 ( 2006 ) . r. piil and k. mlmer , phys . a * 76 * , 023607 ( 2007 ) . s. a. gurvitz and y. b. levinson , phys . b * 47 * , 10578 ( 1993 ) . j. eberly , am . j. phys . * 33 * , 771 ( 1965 ) . a. j. moerdijk , b. j. verhaar , and a. axelsson , phys . a * 51 * , 4852 ( 1995 ) . g. orso , l. p. pitaevskii , s. stringari , and m. wouters , phys . * 95 * , 060402 ( 2005 ) . r. micnas , j. ranninger , and s. robaszkiewicz , rev . * 62 * , 113 ( 1990 ) . a. b. kuklov and b. v. svistunov , phys . * 90 * , 100401 ( 2003 ) . duan , e. demler , and m. d. lukin , phys . lett * 91 * , 090402 ( 2003 ) . d. petrosyan , b. schmidt , j. r. anglin , and m. fleischhauer , phys . a * 76 * , 033606 ( 2007 ) . t. miyakawa and p. meystre , j. low . phys * 148 * , 429 ( 2007 ) . m. khl , h. moritz , t. stferle , k. gnter , and t. esslinger , phys . lett . * 94 * , 080403 ( 2005 ) . the fit was performed for @xmath407 and @xmath408 , and the sum in ( [ eq:2nd_order_shift ] ) was taken over the lowest 35 bands . d. b. m. dickerscheid , u. al khawaja , d. van oosten , and h. t. t. c. stoof , phys . a * 71 * , 043604 ( 2005 ) . k. b. gubbels , d. b. m. dickerscheid , and h. t. c. stoof , new . j. phys * 8 * , 151 ( 2006 ) . r. b. diener and t .- l . ho , phys . rev . lett . * 96 * , 010402 ( 2006 ) . s. drr , t. volz , and g. rempe , phys . a * 70 * , 031601(r ) ( 2004 ) . this fit was performed for @xmath409 and @xmath410 , and the contribution from the first 35 bands were included in ( [ eq : sigma_m_higher ] ) .

the phenomena of diffusion and transport over a potential energy landscape play a key role in a number of processes in physics , chemistry and biology @xcite . josephson tunneling junctions , superionic conductors , phase - locked - loop frequency control systems , charge density waves are a few examples of systems in which these processes in periodic potential are important @xcite . in recent years the interest has been growing to experimental studies of manipulating the particle diffusion by external fields . one can effectively control the diffusion processes by varying the field parameters . for instance , huge enhancement in diffusivity ( exceeding its value in the same system without extra forcing by a factor of order hundreds ) was observed in studying particle diffusion in colloids with optical traps ( optical vortices ) @xcite . large increase in diffusivity was also observed in studies of paramagnetic particle diffusion on surfaces of garnet ferrites influenced by external time - periodic magnetic field @xcite . in the same manner diffusivity growth with shaking strength was noticed in experiments with granular gas @xcite . it turned possible to increase the diffusivity of ions in membrane channels by varying the external electromagnetic field @xcite . one can increase the rates of diffusion - limited physical processes , and effectively separate micro- and nano - particles of different nature by varying diffusion coefficients in different directions @xcite . despite these remarkable achievements , quantitative description of such enhanced diffusion under the influence of external forces remains fragmentary to date . special cases of underdamped and overdamped physical systems are often considered . in overdamped systems inertial effects may be neglected , significantly simplifying the mathematical treatment of the problem . in underdamped systems on the other hand the viscous decay of oscillations occurs on times long compared with the oscillation period in the potential ; such systems are harder to investigate , and the progress in studying these is more limited compared with overdamped ones . first detailed studies of a brownian particle motion in a washboard potential ( _ i.e. _ a periodic potential tilted with an additional constant force field ) were carried out by h. risken and colleagues @xcite , for all values of the friction . it was shown that at low friction the appearance of locked ( in which the particle oscillates around one minimum of the potential ) and running ( where the particle travels through many potential periods , not getting trapped at separate minima ) solutions was important for the particle ensemble behavior . at small temperatures running solutions appear abruptly at the external force value exceeding certain critical @xmath3 ; at smaller forces a single class of solutions ( locked ) is realized . systematization of the results of these studies may be found in authoritative monograph @xcite . approximate solution of the fokker - planck equation ( fpe ) for the coordinate and velocity distribution function was the central moment of these studies . two different approaches were used , that were checked to produce compatible results . in the first of these @xcite by expanding the distribution function in hermite functions in velocity and fourier - series in space coordinate , the fokker - planck equation was transformed into an infinite set of coupled equations for the expansion coefficients . these equations were solved by a matrix continued fraction method . particle flux was then expressed through the expansion coefficients , after appropriate series truncation . at smaller friction progressively more terms must be kept , making this method less efficient . another approach @xcite is best suited for low friction , when the energy becomes a slow variable ( changing on friction dissipation timescales ) . the distribution function was rewritten in the energy and configuration space coordinate variables , fpe was solved in these new variables with the help of physically - motivated self - consistent ansatz . many illuminating results were obtained for the distribution functions at various parameter values , stability domains of locked and running solutions , plots of the mean particle velocity , etc . diffusion properties however remained not studied . this matter was partly addressed in other works , based on simulations of the stochastic langevin equation . early systematic study of diffusion by means of numerical simulation of the langevin equation in periodic potential with extra constant forcing was undertaken in @xcite ( similar simulations were reported in @xcite for forcing with a time - periodic component ) . costantini & marchesoni analyzed both overdamped and underdamped settings . at low dissipation they observed significant enhancement of the spatial diffusion near the critical force @xmath3 ( @xcite , fig . 1 : diffusivity ten times larger than its post - maximum value at @xmath4 , at the temperature equal to 0.6 of the potential well depth @xmath5 ) , and related that to features of the particle jump statistics and locked - to - running transition . further progress in studies of the diffusion under the action of constant force was made in @xcite . time - dependence of the particle ensemble dispersion was studied in @xcite . it was shown for the first time that in underdamped systems a special regime of dispersionless transport was realized , in which dispersion virtually does not change with time , on certain limited interval of time . the authors explained this phenomenon , they showed that strongly non - equilibrium distribution of particles in space ( with steep front and exponential tail ; formed as the particles first exit their original potential well ) persists for long times , before eventually broadening and assuming normal gaussian shape . the temperature dependence of the diffusivity was studied as well @xcite . the authors observed that the maximal diffusivity @xmath6 was achieved near the @xmath3 . at low friction @xmath6 was shown to depend abnormally on the temperature increase with decreasing temperature @xmath7 . fitting ( for the three temperatures studied ) yielded @xmath8 relation . the presence of such abnormal temperature dependence is one important aspect in which underdamped systems differ from overdamped ones . in fact , certain peculiarities in diffusivity at low temperatures are observed ( for driven brownian particles in space - periodic potentials ) in overdamped situation as well . p. reimann and colleagues @xcite showed analytically that for a sinusoidal in space potential at external constant force value near @xmath9 , the value required for direct particle pull over the potential barrier ( eq . [ eq : flat ] below , p. ) , the ratio of @xmath10 to the diffusivity value @xmath11 in the viscous medium without lattice and bias forces grows at temperature decreasing @xmath12 . yet the diffusivity itself still vanishes at @xmath13 , as @xmath14 @xcite , and grows monotonically for all @xmath7 . regions with @xmath15 were obtained in overdamped problem in @xcite ( cf . fig . 6 ) for specially crafted exotic potential , flat apart from narrow peaks . the authors linked this abnormal diffusivity behavior to the large ratio of relaxation to escape time in such a system . similar abnormal @xmath16 behavior was found in @xcite ( cf . 24 ) for piecewise linear potential at extreme asymmetry , essentially sawtooth . so , while possible at large friction , special atypical conditions must be met for existence of parameters @xmath17 at which @xmath15 . on the contrary , such parameters exist universally in the problem with small friction ( @xcite , more below ) . another distinction of the diffusivity in overdamped problem is that the maximum in @xmath18 is achieved close to the @xmath9 . whereas at low friction @xmath19 is achieved at ( typically much smaller ) @xmath20 . experimental data for diffusion of colloidal particles in a tilted potential created with laser traps @xcite , fig . 1 , agreed reasonably well with this analytic @xcite @xmath18 dependence . the potential barrier height and the free diffusion coefficient were treated as fit parameters . similarly , diffusivity in @xcite ( fig . 7 ) showed peaked at @xmath9 behavior , and agreed fairly well with analytic predictions from @xcite with the potential form inferred from independent study of the colloidal particle spatial distribution function . in this work colloidal particles diffused over the potential created by the bottom grid of colloidal spheres , and extra constant forcing was due to gravity ( controlled by the sample inclination angle ) . different limiting cases of analytic predictions @xcite for the mean particle velocity and diffusivity were compared with experimental results at different settings , and showed good agreement . the physical reason behind the abnormal temperature dependence of diffusivity in underdamped systems is traced to increasing jump length of running particles ( before getting retrapped at another potential minimum ) with the temperature decrease @xcite . this phenomenon being absent in overdamped systems ( as these have no proper running states ) . the maximal diffusivity is achieved at forcing near @xmath20 , as at this value populations of locked and running particles are about equal , and it is the mutual motion between these two populations ( with temperature - independent running particle speed ) that leads to fast spreading of the particle packet , manifested as giant diffusion @xcite . more thorough analysis of ( the temperature- and force - dependence of ) the diffusion in underdamped systems was carried out in @xcite . it was shown by numeric simulations of the langevin equation that the phenomenon of diffusivity growth with the temperature decreasing ( in the current work called tad , `` temperature - abnormal diffusivity '' ) was only manifested in a narrow interval of applied external forcing . study in a wide range of temperatures demonstrated that the diffusivity grew with decreasing temperature as @xmath21}$ ] for certain @xmath22 , checked valid for @xmath23 $ ] ; whereas the power - law fit @xmath8 proposed in @xcite could approximate the results only in the narrow temperature range studied by the authors . we showed that the main exponential growth resulted from such of the correlation time at the temperature decreasing ( found from the simulation results ) . the diffusion properties were further elucidated and systematized for a broad range of the forces @xmath24 , friction coefficients @xmath25 and temperatures @xmath7 in @xcite , in sinusoidal 1d potential in underdamped setting . we introduced a velocity potential @xmath26 , which in part governed time evolution of certain slow ( not changing much on times of order one period @xmath27 of small oscillations in the potential , rather evolving on viscous dissipation timescales ) velocity variable , defined through a discrete quantity , maximal velocity in such consecutive oscillations . with certain reservations that @xmath26 also defined the velocity distribution function @xmath28 agreeing well with the one found numerically . the latter showed clear bimodal structure with maxima at @xmath29 ( locked states ) and @xmath30 ( running ones ) at intermediate forcing values . in this region of forcing values tad was observed . analytic expressions for the mean particle velocity @xmath31 and diffusivity were obtained with the model @xmath28 , good agreement was shown with direct langevin equation simulation results . simple fits were proposed determining the model parameters [ entering @xmath26 and @xmath28 ] through @xmath24 and @xmath25 , which apart from usefulness for experimentalists / applications allowed one to comprehend the features of transport and diffusivity in all the physical parameter space . physical reason for the @xmath33}$ ] scaling was explained . more detail in the introductory part of sec . [ s : results ] , p. below . in practical applications periodic in time external forces ( electromagnetic , acoustic , etc . ) are more straightforwardly realized than constant forcing . it is natural to study how the findings for the constant forcing problem are modified for the time - periodic forcing setup , both from theoretical and practical point of view . by changing the diffusivity with the aid of periodic external fields it is possible to create surface structures with prescribed properties ; influence the dynamics of point defects , under irradiation in part ; alter the dislocation motion dynamics ; increase chemical reaction rates in selected directions ; manipulate the efficiency of processes in biological systems , to mention just a few applications @xcite . most of the works on periodic forcing deal with ratchet - type systems @xcite or stochastic resonance . studies focusing on diffusion properties are scarce . we are not aware of any works that systematically study the combined dependence of the diffusivity on the temperature , the forcing amplitude and frequency in a representatively broad range of these parameters . dependence of the diffusivity on the interplay between the external periodic ( piecewise constant , eq . [ f : pcp ] on p. below ) forcing and the lattice force was studied in @xcite in 2d overdamped setup . the forcing frequency was fixed at one value . it was shown that in a system with a sum of yukawa potentials at square grid nodes the particle diffusivity was significantly enhanced for certain intricate set of bands ( `` a - windows '' ) of the forcing amplitudes and orientations with respect to the grid . at special forcing arrangement tad was observed for some diffusivity tensor components , and one of its principal values increased exponentially with inverse temperature . the other diffusivity tensor principal value decreased exponentially at the same time . it is likely that multi - dimensionality effects were crucial for tad . instead of having two , locked and running , populations in 1d accounting for ( already sophisticated in this case ) diffusion features , flows in higher dimensional problems are far more complex , known to result in further peculiarities in transport and diffusion @xcite . underdamped 1d system was studied in @xcite , with a sinusoidal potential ( same as eq . [ ux ] below ) and sinusoidal in time forcing . diffusivity enhancement by orders of magnitude was seen . tad was observed for some ( low ) forcing frequencies , at the forcing amplitude @xmath34 studied . dependence of the diffusivity on the forcing frequency was studied in more detail in @xcite , and was shown to be nonmonotonic at large forcing amplitudes ( @xmath34 and @xmath35 , at @xmath36 k , @xmath37 . ) at @xmath34 the maximum in the frequency dependence of the diffusivity @xmath38 persisted for all the temperatures studied , corresponding frequency @xmath39 increasing from @xmath40 to @xmath41 ( in units of the potential eigenfrequency ) at @xmath7 varying from 120 to 360 k , @xmath42 $ ] . the goal of the present work is to investigate in detail the dependence of the diffusivity on the amplitude and the frequency of uniform time - periodic external forcing in a 1d underdamped space - periodic system , and conditions under which tad is realized . we mainly study piecewise constant periodic ( pcp : eq . [ f : pcp ] ) forcing as this case is more straightforward to relate to the results for constant forcing . the equations solved and the methods used are described in sec . [ s : setup ] . [ s : results ] we show how the diffusivity changes with the pcp forcing frequency @xmath0 , starting from the understood results at constant forcing , @xmath43 . we observe that tad is realized for some ( intermediate ) forcing amplitudes . [ s : dvst ] is devoted to studying the @xmath16 behavior at fixed @xmath0 . the results for the diffusivity in the case of sinusoidal in time forcing are shown in sec . [ s : fsin ] . we conclude in sec . [ s : concl ] . in this paper the 1d dynamics of a particle subject to external forcing @xmath44 is described by the langevin equation : @xmath45 where @xmath46 is the time , @xmath47 is the particle coordinate , @xmath48 is its mass , @xmath25 is the friction coefficient . overdot stands for time differentiation . @xmath49 represents thermal fluctuations , described by gaussian white noise with correlation @xmath50 here @xmath51 is the boltzmann constant , @xmath7 is the temperature . the potential energy @xmath52 is taken @xmath53 where @xmath54 is the lattice constant and @xmath5 is the potential barrier height . the lattice force @xmath55 acting upon the particle is @xmath56 quantity @xmath57 corresponds to the minimal external force needed to drag the particle over the potential barrier separating potential minima on the 1d lattice at large friction . we use the same physical parameters as in @xcite , typical for adatom diffusion on close - packed metal surfaces . namely , values @xmath58 for the activation barrier and the lattice constant are adopted . two possibilities for the external periodic forcing are studied : piecewise constant periodic ( * pcp * ) driving @xmath59\ ] ] and sinusoidal one , @xmath60 @xmath0 is the angular frequency of the external force and @xmath61 is its amplitude . it is convenient to use dimensionless time @xmath62 and space coordinate @xmath63 , normalizing the physical values to the period of small oscillations @xmath64 at the potential minimum in the absence of friction , and to the lattice spatial period @xmath54 respectively : @xmath65 dimensionless temperature @xmath66 and friction coefficient @xmath67 are further introduced : @xmath68 definitions ( [ dimtx],[dimtgamma ] ) differ from such in @xcite by extra factors of 2 appearing in the definition of @xmath62 , and consequently in @xmath66 and @xmath67 as well . current definition makes interpreting the results easier ( @xmath69 corresponds to one oscillation period at small @xmath67 ) . for more straightforward comparison with our previous works we also use another dimensionless temperature in the figures , @xmath70 the parameters of the dimensionless problem are @xmath71 , @xmath72 , @xmath67 , @xmath66 . in the new variables the equation of motion has form eq . 1 with unit new mass , potential @xmath73 with no prefactors and the external forcing amplitude made dimensionless with a @xmath74 multiplier . below lowercase @xmath75 s ( frequencies ) , and corresponding periods @xmath76 denote dimensionless quantities . physical frequencies and periods are obtained by respectively dividing and multiplying these by the the period @xmath77 of small oscillations in the lattice potential at @xmath78 . overdamped problem is characterized by @xmath79 , whereas @xmath80 ( equivalently , @xmath81 ) corresponds to underdamped dynamics . we study the underdamped case here . friction coefficient @xmath82 we use corresponds to the same physical value as in earlier studies @xcite . we solve stochastic equation ( [ langevin],[xi_cor ] ) numerically using a verlet - type algorithm @xcite with a time step @xmath83 somewhat shorter than @xmath84 of the period of oscillation @xmath77 . the statistic averaging is performed over the ensemble consisting of at least @xmath85 particles . to verify the modeling consistency some computations were performed with @xmath86 . the initial conditions are set as follows . each particle is placed at @xmath87 , its velocity chosen at random , with maxwellian distribution for a given @xmath7 . thermalization over 100 oscillation periods , in the potential @xmath52 but without the @xmath88 , is then used to get equilibrium particle distribution over both coordinate and velocity . tests show that the distribution function does not change after that time . in the process of such thermalization the particles can jump to nearby elementary cells of the lattice . such relocated particles are returned back to the initial cell by translation over an integer multiple of the lattice constant , to get diffusion of all particles starting from the initial cell of the lattice . the diffusivity is computed based on the dispersion @xmath89 of the particle distribution in the limit of large times : @xmath90 angle brackets @xmath91 with no subscripts mean averaging over the ensemble . @xmath92 and @xmath93 are dimensionless in our notation . the physical diffusivity is @xmath94 . for each diffusivity calculation we find time @xmath95 , after which the dispersion grows linearly with time ( if averaged over the forcing period ) . the @xmath92 is calculated as @xmath96 at @xmath97 . it is known that in systems with low dissipation transient regimes of anomalous diffusion can be realized @xcite , characterized by @xmath98 the process with @xmath99 is referred to as superdiffusion , whereas subdiffusion is characterized by @xmath100 . stages of anomalous diffusion were observed in our simulations . the choice of initial conditions above aided in having these stages in clear form . we determine the exponents @xmath101 by the least - square linear fitting between @xmath102 and @xmath103 on the time - interval of interest . in @xcite it was shown that depending on the value of the external force ( constant in that study ) different functional dependences of the diffusivity on the temperature were realized . plots of the diffusivity as a function of the external force at several temperatures ( based on @xcite , recomputed for the current units ) are shown in fig . [ f : difconstf ] . there are three regions of @xmath61 values ( i iii ) , with qualitatively different dependence of the diffusivity @xmath104 on @xmath7 . in region i diffusion is enhanced as external force @xmath61 value increases , whereas in region iii diffusion is inhibited as @xmath61 increases . in regions i and iii @xmath104 grows with @xmath7 increasing . contrary to this ( normal temperature behavior ) in region ii @xmath104 grows as @xmath7 decreases . this is the region of `` temperature - abnormal diffusivity '' ( tad ) . the _ diffusion process _ at sufficiently late times is normal , @xmath105 , however the _ diffusion coefficient _ @xmath106 ( abnormally ) increases as the temperature decreases . . 1 : @xmath107 [ @xmath108k for the potential characteristic of the h - adatom diffusion on metal surface ( [ u0a ] ) ] , 2 : @xmath109 , 3 : @xmath110 , 4 : @xmath111 , 5 : @xmath112 , 6 : @xmath113 , 7 : @xmath114 . + bottom : the effective velocity potential in the three force regions . left , middle and right diagrams correspond to @xmath115 , 0.095 and 0.25 respectively.,title="fig:",scaledwidth=47.0% ] + . 1 : @xmath107 [ @xmath108k for the potential characteristic of the h - adatom diffusion on metal surface ( [ u0a ] ) ] , 2 : @xmath109 , 3 : @xmath110 , 4 : @xmath111 , 5 : @xmath112 , 6 : @xmath113 , 7 : @xmath114 . + bottom : the effective velocity potential in the three force regions . left , middle and right diagrams correspond to @xmath115 , 0.095 and 0.25 respectively.,title="fig:",scaledwidth=12.3% ] @xmath116 . 1 : @xmath107 [ @xmath108k for the potential characteristic of the h - adatom diffusion on metal surface ( [ u0a ] ) ] , 2 : @xmath109 , 3 : @xmath110 , 4 : @xmath111 , 5 : @xmath112 , 6 : @xmath113 , 7 : @xmath114 . + bottom : the effective velocity potential in the three force regions . left , middle and right diagrams correspond to @xmath115 , 0.095 and 0.25 respectively.,title="fig:",scaledwidth=10.0% ] @xmath116 . 1 : @xmath107 [ @xmath108k for the potential characteristic of the h - adatom diffusion on metal surface ( [ u0a ] ) ] , 2 : @xmath109 , 3 : @xmath110 , 4 : @xmath111 , 5 : @xmath112 , 6 : @xmath113 , 7 : @xmath114 . + bottom : the effective velocity potential in the three force regions . left , middle and right diagrams correspond to @xmath115 , 0.095 and 0.25 respectively.,title="fig:",scaledwidth=10.0% ] @xmath117 in @xcite it was seen that tad is also possible at time - periodic external forcing . at such forcing tad was only observed in a limited interval of the forcing frequencies . studies were carried out at @xmath118 , that would correspond to region iii for the same amplitude of constant forcing ( see fig . [ f : difconstf ] ) , so tad would not be observed at such @xmath61 value were the forcing @xmath88 constant . to understand physical reasons of the abnormal temperature dependence of the diffusivity in these two different ( time - periodic , and constant ) forcing regimes we consider the periodic forcing regime in more detail below . for piecewise constant periodic ( pcp ) forcing one can use intuition gained in the constant forcing problem for the time intervals in which @xmath44 stays constant , the findings are thus easier to interpret . most of the results in this work are for the pcp forcing setup . in @xcite we showed that the transport properties of an ensemble of underdamped particles in a washboard potential can be found by considering a simpler problem of the overdamped motion in effective potential @xmath26 in velocity space . the problem showed similarities with the problem of active brownian particle motion @xcite . whether the particle belongs to locked [ tending with time to @xmath29 , one minimum of @xmath26 ] or running [ tending to @xmath30 , another minimum of the asymmetric double - well potential @xmath26 ] population in the absence of noise is determined by the particle s initial state ( its velocity , if placed at the potential minimum at @xmath120 ) . it was shown that region ii of the force values , the region in which tad is observed , is precisely the region where both locked and running solutions are simultaneously possible . the @xmath26 was also instrumental for proposing a model distribution function @xmath28 , and for defining diffusivity through kubo relation . the analytic predictions for @xmath104 and @xmath31 based on the model @xmath28 agreed well with direct numerical simulation results for all values of @xmath121 , temperatures in the range @xmath122 $ ] , friction @xmath123 $ ] ( in the definitions of dimensionless quantities adopted here , eqs . ( [ dimtx],[dimtgamma ] ) ; agreement getting progressively better at smaller @xmath67 ) . in the expression for the diffusivity , ( the analytic prediction for ) the correlation time @xmath124 factor was shown to be responsible for the main features in @xmath125 dependence , numerically found in earlier study @xcite . namely , @xmath124 showed proper nonmonotonic dependence on @xmath24 , with a maximum near @xmath20 , progressively more pronounced at smaller @xmath7 . @xmath124 increased at temperature decreasing @xmath126 $ ] ( for certain @xmath22 ) ; this increase at low @xmath7 reflected the kramers rate for transitions in the double - well velocity potential @xmath26 , between the wells corresponding to locked and running states . schematics of the effective velocity potential in the three regions of the applied force are shown in the bottom insets in fig . [ f : difconstf ] . the first inset corresponds to @xmath127 , at which only locked solutions exist . in the second inset @xmath128 , corresponding to the maximal diffusivity ; both types of solutions coexist at this force value . finally @xmath129 corresponding to region iii is shown in the third inset ; at this @xmath61 only running solutions are realized . the boundaries between regions i iii depend on @xmath25 and @xmath48 ( on @xmath67 in the dimensionless problem ) . at smaller @xmath67 these boundaries move to the left , to the lower force values , and the width of region ii decreases . in the current study we use the same fixed values of @xmath25 and @xmath48 as in @xcite , and investigate the features of @xmath130 at the three values of the external forcing amplitude listed above . it may be expected that the same picture for different temperature behaviors of the diffusivity , with analogous three regions i iii of the external force amplitudes , holds for pcp driving , at least at small frequencies @xmath0 . the description with the effective potential is applicable , the potential @xmath26 keeping constant form for each of the two half - periods of the external forcing [ on which @xmath44 stays constant , this @xmath24 value defining corresponding @xmath131 . ] below we show that in fact the situation is more subtle . similar three regions of the external force amplitudes @xmath61 exist , differing in qualitative @xmath16 dependence . tad is realized in certain range of intermediate @xmath61 , smoothly dependent on @xmath0 . at decreasing temperatures the dependence @xmath16 starts progressively deviating from such at constant forcing . for any fixed @xmath0 at temperatures below certain critical @xmath132 , normal temperature dependence of the diffusivity is restored . we demonstrate these features by analyzing @xmath130 in sequence for the above values of @xmath61 ( @xmath115 , 0.095 , 0.25 ) , that correspond to regions i , ii , iii in the constant external force problem . we observe transient anomalous diffusion phase , and see how its characteristics relate to the diffusivity in the asymptotic late time regime , when normal diffusion regime sets in . the properties of this anomalous diffusion stage are different in the three regions of the applied force amplitudes ; this translates into different functional dependence of the diffusivity on the temperature . here we investigate how the diffusivity changes with ( dimensionless ) frequency @xmath72 of the external forcing at forcing amplitude @xmath127 . we show that the diffusivity stays nearly constant at small frequencies , up to reciprocal of the superdiffusion stage duration . at frequencies above this the diffusivity drops according to a power - law with exponent related to the exponent of the superdiffusion . this power - law dependence holds up to frequencies a few times less than the lattice potential eigenfrequency . the diffusivity grows with the temperature ( normal temperature behavior ) , related to the superdiffusion stage duration increasing with @xmath7 . figure [ f : sigma006 ] shows the growth of the particle dispersion with time . at constant external forcing this growth demonstrates a clear superdiffusion stage , @xmath133 , @xmath134 , that switches to normal diffusion ( same power - law growth but with @xmath135 ) after certain time @xmath136 . the @xmath136 as well as the superdiffusion exponent @xmath101 depend on @xmath71 and other problem parameters , @xmath67 and @xmath66 . for the values used @xmath137 , @xmath138 . lowercase @xmath139 s everywhere are dimensionless , normalized to the period of small oscillations at the lattice potential minimum at @xmath140 , @xmath141 @xmath89 and @xmath92 are dimensionless dispersion and diffusivity . ( region i ) for different force frequencies . @xmath136 denotes the end of the superdiffusion stage at constant forcing . @xmath76 is the period of the external periodic forcing . 1 : @xmath143 , 2 : @xmath144 , 3 : @xmath145 , 4 : @xmath146 ( constant external force . ) straight dashed and dotted lines show the power - law fitting behavior of the dispersion.,scaledwidth=47.0% ] curves 13 correspond to three different frequencies . oscillations in @xmath147 are clearly seen with double forcing frequency , @xmath148 . these are particularly large at @xmath75 around @xmath149 ( curve 2 ) . at these values , interestingly , intervals of @xmath150 decreasing with time are observed . this effect can be seen in earlier works @xcite . at frequencies below @xmath149 the curves @xmath150 closely follow that at @xmath146 ( constant external force , curve 4 ) . at higher frequencies the superdiffusion phase effectively ends after one period @xmath139 of the external force , normal diffusion sets in ( in a sense that averaged over the forcing period @xmath151 ) . higher frequencies of the forcing correspondingly result in lower diffusivity , as the stage of @xmath89 growing with time faster than linearly ends earlier than at lower @xmath75 s . these results suggest the following approximations for the smoothed over the forcing period dispersion @xmath152 , valid not too close to the transition moments between different diffusion regimes : @xmath153 here @xmath154 ( for the approximately boltzmannian initial distribution of the particles in @xmath155 well of the potential ) , @xmath156 , @xmath157 is close to the first potential well exit time . constants are chosen in such a way that the spline approximation ( [ sigma2fiti ] ) is continuous . quantities @xmath101 , @xmath158 depend on the problem parameters , @xmath66 , @xmath71 and @xmath67 . at fixed values of these parameters , the last approximation yields the dependence of the diffusivity @xmath159 on the forcing period @xmath139 @xmath160 or on its frequency @xmath161 @xmath162 the diffusivity thus should stay about constant at @xmath163 , and at larger forcing frequencies drop with @xmath75 according to the power law , with exponent equal to @xmath164 , determined by the superdiffusion exponent . figure [ f : domega006 ] shows simulation results for @xmath165 . dotted lines show approximations eq . [ domegafiti ] . dashed lines show the diffusivity at constant forcing with the same @xmath61 value . these results confirm the features suggested by eq . [ domegafiti ] , away from transition frequency @xmath149 and up to @xmath166 . @xmath167 at any @xmath75 , and decreases with @xmath75 , up to the frequency values a few times smaller than the potential eigenfrequency . on the forcing frequency at @xmath168 ( triangles ) and @xmath169 ( circles ) . the dashed line asymptotes show the diffusivity values at constant forcing with the same @xmath115 . the slanting dotted line shows the power law drop of @xmath92 at intermediate @xmath75 s . , scaledwidth=47.0% ] further ramification might have been added in eqs . ( [ sigma2fiti][domegafiti ] ) for the case @xmath170 , changing @xmath157 to @xmath171 ; that would indeed lead to @xmath172 increasing at @xmath173 . with our parameters , however , resulting approximations are an overstretch ; the diffusion gets markedly nonequilibrium at such high @xmath75 s , resulting in the complex @xmath172 behavior at @xmath75 approaching @xmath174 . more on this in sec . [ s : highomega ] . finally , fig . [ f : sigmat006 ] shows @xmath150 for a range of temperatures at constant forcing . despite the superdiffusion exponent @xmath101 is seen to decrease with the temperature , the duration of the superdiffusion stage @xmath175 grows fast with the temperature . as a result the dispersion is larger for larger temperatures at any given time . at late times this translates into diffusivity being larger at larger temperatures ; so normal ( increasing ) @xmath16 dependence is realized at force amplitudes in region i. ) . 1 : @xmath176 , 2 : @xmath168 , 3 : @xmath142 , 4 : @xmath177 . the short - dashed straight lines show power - law fits for the transient superdiffusion stage , and the late - time normal diffusion regime.,scaledwidth=47.0% ] here we repeat analysis of the previous section for periodic forcing with amplitude @xmath128 ( the value at which the maximal diffusivity is achieved at constant forcing ) . in the same manner as for the forcing amplitude in region i , transition from initial superdiffusion stage to normal diffusion at late times is observed , at time @xmath136 . at forcing frequency @xmath75 growing , the diffusivity again decreases only slightly from @xmath179 till @xmath75 reaches @xmath180 . it decreases fast , according to the same power law @xmath181 , at larger @xmath75 s , up to @xmath182 . contrary to the case of weaker forcing considered in the previous section , superdiffusion duration @xmath175 now _ decreases _ with the temperature . this leads to diffusivity growing at temperature decrease ( tad ) , at moderately low frequencies . this is the main effect we intended to find in the periodic driving setup , extending our previous findings at constant forcing @xcite . figure [ f : sigma0095 ] shows growth of the particle dispersion with time . again the superdiffusion stage is observed , and it ends after one period of pcp forcing or after @xmath136 ( the end time of superdiffusion at constant forcing ) , whichever is smaller . thus the same approximations eq . ( [ sigma2fiti][domegafiti ] ) are applicable for the ( averaged over @xmath139 ) dispersion growth @xmath152 and dependence of the diffusivity on the forcing frequency @xmath165 . simulation results for @xmath165 ( fig . [ f : domega0095 ] ) again agree with prediction ( eq . [ domegafiti ] ) . ( region ii ) for different force frequencies . curves 15 , bottom to top : ( dimensionless ) @xmath183 @xmath184 , @xmath185 , @xmath186 , @xmath187 , 0 ( constant forcing ) . , scaledwidth=47.0% ] on the external force frequency @xmath75 , at @xmath168 ( triangles ) and @xmath169 ( circles ) . the dashed line asymptotes show the diffusivity values for constant forcing with the same @xmath178 . dotted line shows the power law drop of @xmath92 at higher @xmath75 s . the dot - dashed horizontal line asymptote ( for high @xmath75 ) is at the diffusivity value in the lattice periodic potential in the absence of extra driving , @xmath188 , for @xmath142.,scaledwidth=47.0% ] as @xmath172 does not vary much at @xmath163 , tad must be observed at small frequencies , at it is observed at constant forcing ( @xcite : @xmath189 decreases with the temperature when @xmath61 is in region ii ) . such an abnormal dependence on the temperature is due to longer jumps the particles perform on average at lower temperatures . this in turn follows from lower probability at lower temperatures to turn the running particle into a locked state : such a transition requires certain threshold energy transfer to such a particle , and at lower temperatures the probability of such a transfer decreases exponentially with inverse temperature . hence the running particle travels at about the stationary speed @xmath190 ( the speed fluctuating around this average due to the lattice potential ) for increasingly longer distances before getting trapped again in some next well of the potential , resulting in larger diffusivity at smaller temperatures . this mechanism is not applicable in region i , where only locked solutions are realized . in terms of the dispersion evolution features , this phenomenon of the diffusivity growing at the temperature decreasing results from the superdiffusion duration @xmath191 growing at temperature decreasing , as fig . [ f : sigmat0095 ] demonstrates . this qualitatively differs from the situation at forcing amplitude in region i ( end of the previous section ) . we note that @xmath101 increases with the temperature decreasing , same as in region i. ) . 1 : @xmath176 , 2 : @xmath168 , 3 : @xmath142 , 4 : @xmath177 . @xmath178.,scaledwidth=47.0% ] at higher frequencies , on the other hand , the diffusivity starts increasing with the temperature ( normal behavior ) . this results from @xmath136 decreasing with the temperature , as can be seen in fig . [ f : sigmat0095 ] . despite @xmath179 is larger at smaller temperature @xmath192 , as @xmath75 grows the @xmath193 starts decreasing according to the power law earlier [ at @xmath75 s near the smaller @xmath194 than @xmath195 for @xmath196 does . and the exponent of this power - law decreases with @xmath7 decreasing , @xmath197 therefore at larger frequencies , commensurate with @xmath198 , the diffusivity starts increasing with the temperature , _ tad disappears_. as it is the case for the forcing amplitude in region i , @xmath172 decreases monotonically with @xmath75 from @xmath199 , up to the frequency values @xmath200 ( at frequencies so high the particle transport becomes substantially non - equilibrium , that affecting the functional dependence @xmath172 , as discussed in sec . [ s : highomega ] below ) . for strong forcing ( region iii , in which the diffusivity decreases with @xmath61 in constant forcing setup , and increases with the temperature ) we demonstrate in this subsection that the dependence of dispersion on time shows new features . after initial superdiffusion stage at times @xmath202 , prolonged dispersionless stage sets in , till @xmath203 . in this stage the dispersion stays nearly constant . as at weaker external forcing ( regions i , ii ) , both these stages may be interrupted earlier in the case of high forcing frequency ( after about a period of the forcing ) . at late times normal diffusion takes place . dependence @xmath172 is nonmonotonic here , agreeing with findings in @xcite . @xmath104 reaches a maximum near @xmath204 . tad is observed at intermediate frequencies . figure [ f : sigma025 ] shows the particle dispersion as a function of time . at low frequencies of the forcing , superdiffusion is observed till @xmath205 . dispersionless stage is realized later , at @xmath206 . physically @xmath136 is the time at which the particle distribution function in velocities assumes its stationary form . the distribution in space is still strongly non - equilibrium at @xmath136 , it has exponential tail and sharp front @xcite ( in the direction of @xmath24 ) . it takes till @xmath207 for the distribution function to assume approximately gaussian shape in space ( if averaged over the potential spatial period ) . on interval @xmath208 dispersion does not increase siubstantially . after @xmath207 normal diffusion sets in , @xmath209 at @xmath210 . ( region iii ) for different force frequencies . curves 15 : ( dimensionless ) @xmath183 @xmath211 , @xmath184 , @xmath185 , @xmath186 , @xmath187 . dotted lines show power - law fitting behavior of the dispersion.,scaledwidth=47.0% ] the whole curve @xmath147 differs insignificantly from @xmath147 in constant forcing setup if the forcing frequency @xmath212 . this @xmath147 dependence is modified at higher frequencies of the external forcing ( in a similar way to the behavior discussed in sec . [ s : regioni][s : regionii ] at weaker forcing ) . for @xmath213 the dispersionless stage is interrupted after about one period of external force , after which time approximately linear dependence ( if averaged over times @xmath214 ) sets in , @xmath215 , @xmath92 equal to its asymptotic ( at @xmath216 ) value . at yet higher forcing frequencies , @xmath217 , dispersionless stage is not realized at all , normal diffusion sets in right after the prematurely interrupted superdiffusion stage . similarly to the previous sections , described above dependence of @xmath218 on @xmath219 can be schematized by the following continuous piecewise linear approximation ( meaningful not too close to switching time - points @xmath220 ) : @xmath221 here @xmath222 , @xmath156 ; @xmath223 assumed . as before , @xmath172 can be found from this , showing the same power law decay as in eq . [ domegafiti ] at @xmath224 , and additional , linearly growing with @xmath75 feature at intermediate frequencies : @xmath225 simulation results for @xmath172 are shown in fig . [ f : domega025 ] for two temperatures . they indeed show the features suggested in eq . [ domegafitiii ] . at high frequencies , @xmath224 , the superdiffusion stage switches to normal diffusion earlier than at @xmath136 , at time @xmath226 . this behavior is the same as for the forcing amplitude in regions i ii . thus in the same manner the diffusivity grows as @xmath75 decreases , as the longer interval ( till @xmath227 ) of superdiffusive dispersion growth is used at the forcing period @xmath139 getting longer . however at smaller frequencies , @xmath228 , it is the dispersionless phase that gets interrupted prematurely ( _ i.e. _ it is not realized in full , till @xmath207 , as it would have at @xmath146 ) , so the earlier such a switch occurs ( _ i.e. _ the shorter the @xmath139 ) , the higher diffusivity is obtained in the late - time normal diffusion stage . this explains the growth of @xmath172 with @xmath75 at @xmath229 . the maximum in @xmath172 is thus observed , at @xmath230 . on the forcing frequency , at @xmath168 ( circles ) and 0.194 ( triangles ) . the dashed lines show approximately linear growth at @xmath231.,scaledwidth=47.0% ] this differs from monotonically decreasing @xmath172 in regions i , ii . such a nonmonotonic dependence @xmath172 was observed in @xcite ; the forcing studied was sinusoidal in time , its amplitude ( indeed ) corresponded to region iii in terminology of this work . @xmath232 together with @xmath233 increase with the temperature . the maximal value of the diffusivity ( achieved at @xmath232 ) increases as the temperature decreases , due to the dispersionless stage duration @xmath234 growing fast at temperature decreasing . and so does @xmath172 at a given @xmath75 near @xmath198 in a limited range of temperatures . so abnormal temperature dependence of the diffusivity ( tad ) is realized in intermediate region of the forcing frequencies around @xmath198 . the temperature dependence of the diffusivity is studied in more detail in sec . [ s : dvst ] below . as the simulation results suggest , the power law decay of @xmath172 at large @xmath75 s ( figs . [ f : domega006 ] , [ f : domega0095 ] , [ f : domega025 ] ) does not hold at frequencies that are too high , above @xmath235 of the eigenfrequency at the lattice potential minima . this is due to the diffusion process becoming markedly non - equilibrium . the distribution function in velocities differs significantly from such at constant forcing at all times ; description in terms of the velocity potential @xmath26 is no longer useful . to analyze this we introduce kinetic temperature @xmath236 ( _ i.e. _ , made dimensionless with the same multiplier @xmath237 as @xmath238 ) , and study its evolution at different @xmath75 s : fig . [ f : tkin ] , bottom . at low frequency , @xmath239 does not change ( only shows random noise ) after one period @xmath76 of the forcing . at larger @xmath240 regular oscillations in @xmath239 are observed at all times , with amplitude of order 5% of the average @xmath239 value . that average is about 10% higher than the average @xmath239 at @xmath145 . @xmath239 averaged over the forcing oscillation period grows for about @xmath241 until it reaches its stationary value . @xmath242 and @xmath185 . @xmath115 . note that time normalized to one _ forcing _ period @xmath139 ( rather than to @xmath77 as in the previous @xmath46-dependence graphs ) is shown along the abscissa axis.,title="fig:",scaledwidth=24.1354% ] @xmath242 and @xmath185 . note that time normalized to one _ forcing _ period @xmath139 ( rather than to @xmath77 as in the previous @xmath46-dependence graphs ) is shown along the abscissa axis.,title="fig:",scaledwidth=24.0% ] together with @xmath239 , @xmath147 also shows transient behavior at @xmath243 . we illustrate this in the top graphs in fig . [ f : tkin ] , by plotting @xmath244 in settled `` normal diffusion '' regime this quantity must coincide ( in average over the forcing period of duration @xmath139 ) with the diffusivity , as in such a regime theoretically @xmath245 for @xmath246 . these results in part mean that the rule of superdiffusion or dispersionless stage ( however these are modified in the high - frequency regime ) terminating and turning into normal diffusion after one period of forcing no longer applies at high forcing frequencies . this leads to the @xmath172 derivation based on approximations eqs . ( [ sigma2fiti],[sigma2fitiii ] ) progressively less accurate at higher @xmath75 s , the power - law decay first slowing down , and then going into a completely different regime at @xmath75 approaching @xmath174 . at @xmath247 the diffusivity tends to the diffusivity of particles in the lattice periodic potential in the absence of extra driving @xmath88 . such an asymptote is shown by the horizontal dot - dashed line in fig . [ f : domega0095 ] . having studied how the temperature behavior of diffusion changes at gradual forcing frequency increase , here we investigate in more detail how the diffusivity changes with the temperature at fixed frequency . besides theoretical importance , the results are of interest to experimentalists . oftentimes changing the frequency in the system is problematic , while changing the temperatures is simpler ; so the findings in this section are more straightforward to test . @xmath248 at + ( left ) : @xmath145 , force amplitudes @xmath249 0 ( squares , and linear fit ) , 0.06 ( circles ) , 0.095 ( diamonds ) , 0.25 ( triangles ) ; + ( right ) : @xmath178 , top to bottom : @xmath183 0 ( squares ) , @xmath186 ( triangles ) , @xmath185 ( diamonds ) , @xmath250 ( circles ) , @xmath184 ( pluses).,title="fig:",scaledwidth=23.8% ] at + ( left ) : @xmath145 , force amplitudes @xmath249 0 ( squares , and linear fit ) , 0.06 ( circles ) , 0.095 ( diamonds ) , 0.25 ( triangles ) ; + ( right ) : @xmath178 , top to bottom : @xmath183 0 ( squares ) , @xmath186 ( triangles ) , @xmath185 ( diamonds ) , @xmath250 ( circles ) , @xmath184 ( pluses).,title="fig:",scaledwidth=24.0% ] @xmath251 @xmath252 @xmath117 @xmath117 @xmath253 left part of fig . [ f : dtomega0001f ] shows how the diffusivity changes with @xmath238 at fixed @xmath254 . for the forcing amplitude in region i the diffusivity monotonically grows with the temperature , this ( normal ) behavior agreeing with such at constant forcing @xcite . for @xmath61 in regions ii and iii tad is observed , as already noted in sec . [ s : regionii][s : regioniii ] . tad is confined to limited ( from below ) temperature intervals , contrary to the tad behavior observed in @xcite for constant forcing . in @xcite the diffusivity was numerically checked growing ( only in region ii at @xmath146 ) at temperature decreasing down to @xmath176 . such a behavior was understood theoretically , dependence @xmath255 $ ] was predicted at low temperatures going all the way down to 0 , and agreed well with simulation results . in the current setup with periodic forcing , similar temperature dependence is observed too , @xmath256 but only above certain temperature . the growth slows down as the temperature keeps decreasing , @xmath104 reaches maximum @xmath257 at certain @xmath258 , and starts decreasing with further decrease in @xmath7 . the behavior at lower temperatures is well approximated by @xmath259 in the absence of forcing we observe @xmath260 ( corresponding dimensionful quantity @xmath261 ) classical arrhenius temperature behavior of the diffusivity . @xmath262 decreases with @xmath61 , whereas @xmath263 increases . @xmath258 and @xmath10 increase with @xmath61 as well . temperature behavior of the diffusivity is related to such of the superdiffusion stage duration @xmath264 and @xmath157 . these times may be inferred from simulations of the particle dispersion evolution with time , fig . [ f : sigma3f ] . for the forcing amplitude in region i , @xmath157 ( the time at which superdiffusion starts , about the potential well escape time ) increases with the temperature decreasing . @xmath175 at the same time decreases . this is seen to result in the diffusivity growing with the temperature , agreeing with the results shown in the @xmath145 part of fig . [ f : dtomega0001f ] . in region ii @xmath157 increases with the temperature decreasing as well . @xmath264 however also increases with the temperature decreasing ( middle graph in fig . [ f : sigma3f ] , curves 35 ) contrary to the behavior for @xmath61 in region i. tad is realized at these temperatures . at temperature @xmath265 ( @xmath266 in this graph ) @xmath136 becomes equal to @xmath76 , the period of forcing . at this temperature the maximal diffusivity is achieved , corresponding to the longest actual superdiffusion stage at the given forcing frequency . tad is only realized at temperatures above @xmath267 : even though @xmath264 ( superdiffusion duration at constant forcing ) still grows at further temperature decrease together with @xmath136 , in periodic forcing setup superdiffusion effectively ends earlier , at @xmath227 . it is this superdiffusion interruption at @xmath227 that makes @xmath16 behavior at lower temperatures to diverge in the periodic forcing setup from that at constant forcing . in the constant forcing problem , based on the same logic , tad must be realized all the way down to @xmath268 ; in agreement with conclusions in @xcite . it is clear from this argument that there is no uniform convergence with @xmath75 of @xmath269 curves to @xmath270 . at @xmath75 getting smaller @xmath269 gets close to such a @xmath16 curve for constant forcing on longer @xmath7-interval , down to progressively lower temperatures . however at any @xmath75 there exists @xmath271 such that normal temperature dependence of the diffusivity is restored at @xmath272 . this non - uniform convergence is shown in the right plot in fig . [ f : dtomega0001f ] . the same behavior is observed in region iii . the maximal diffusivity is achieved at @xmath267 corresponding to the longest superdiffusion stage at the given @xmath75 . the lowest temperature at which tad is realized is the one at which @xmath136 reaches the period of external force @xmath139 ( more precisely , the time of superdiffusion interruption at a given @xmath75 ) . at this temperature @xmath273 crosses zero value . the interval of temperatures , in which tad is observed , is wider than at forcing in region ii , due to the presence of dispersionless phase ( curve 4 in the right plot in fig . [ f : sigma3f ] ) . experimentally it is simpler to observe tad at larger @xmath61 , at which within the period of @xmath88 the superdiffusion stage is used to larger extent ; and the temperature interval of tad is thus wider . @xmath117 @xmath274 @xmath274 @xmath242 + @xmath242 @xmath275 @xmath275 @xmath242 presented study of the diffusion under external forcing can be summarized by the following schematics , fig . [ f : dtomega ] . force amplitudes fall into three regions , i iii . in region i the diffusivity grows with the temperature increasing . in regions ii iii the dependence @xmath16 is nonmonotonic , it shows a maximum at certain @xmath258 . below @xmath258 the temperature dependence of the diffusivity is normal , @xmath276 . temperature interval of tad exists above @xmath258 , @xmath15 . in the forcing frequency dependence of the diffusivity , on the other hand , it is regions i and ii that look alike : at the frequency decreasing from a fraction of the potential eigenfrequency the diffusivity grows monotonically , getting to a flat region at @xmath163 , in which it virtually reaches its limiting @xmath179 value . in region iii the @xmath172 curve passes through a maximum , near @xmath204 , corresponding to the end time of superdiffusion stage @xmath136 at constant forcing . at frequencies getting smaller , the forcing period contains progressively longer interval of `` would - be '' ( in the constant forcing problem ) dispersionless and ( at yet smaller frequencies ) normal - diffusion time - interval , resulting in @xmath104 decreasing at such further @xmath75 decrease . @xmath277 @xmath242 @xmath242 @xmath277 here we show the frequency dependence of the diffusivity for sinusoidal in time external forcing . we observe the same features in @xmath172 as we saw for pcp forcing . main conclusions of the sections above thus appear not to depend critically on specific form of time - dependence of the ( smooth enough , symmetric ) forcing . figure [ f : dfsin ] presents a comparison between diffusivities at sinusoidal in time and pcp forcing . at small frequencies the diffusivity stays close to its value at @xmath146 . in regions i and ii of the external forcing amplitude @xmath172 starts decaying according to a power law at @xmath278 , @xmath136 being the superdiffusion stage ending time at constant external forcing with the same amplitude @xmath61 . in region iii @xmath172 grows linearly with @xmath75 at intermediate @xmath75 s , reaches a maximum , then shows a power - law decay at larger frequencies . for all force amplitudes the power law decay switches to more complex behavior at @xmath75 reaching about @xmath279 of the potential eigenfrequency . presence of periodic forcing of any form of time - dependence leads to significant enhancement in diffusivity . this enhancement is somewhat smaller for sinusoidal forcing in regions i , ii ( can be understood , as at the same @xmath61 the average over period forcing @xmath280 is smaller for sinusoidal than for pcp forcing ) . in region iii the maximum in diffusivity is achieved at smaller @xmath232 , and the value of @xmath10 is larger than such a maximal diffusivity at corresponding @xmath281 under pcp driving with the same amplitude @xmath61 . this can be understood from the same argument , as in region iii the diffusivity _ decreases _ with the forcing increasing , all other parameters kept constant . we investigated enhancement of diffusion in 1d space - periodic underdamped systems by external time - periodic fields . we showed that the diffusion can be enhanced by orders of magnitude at certain choice of temperatures @xmath7 , forcing amplitudes @xmath61 and frequencies @xmath75 . three regions of @xmath61 exist , i iii , in which dependences of diffusivity @xmath104 on @xmath7 and @xmath75 differ qualitatively . at all @xmath61 there is an interval of small frequencies , in which @xmath104 depends only weakly on the frequency , approaching the limiting value @xmath179 . this value coincides with @xmath104 at constant bias force with the same value @xmath61 in the case of piecewise constant in time periodic ( pcp ) forcing , the form @xmath88 of forcing we focused on in this study . at larger frequencies , from @xmath149 corresponding to the superdiffusion stage ending time @xmath136 , to about @xmath282 ( dimensionful frequency @xmath283 of the potential eigenfrequency @xmath284 ) @xmath285 decays with @xmath0 according to a power - law . the exponent of this decay is related to the superdiffusion stage exponent @xmath101 . at yet larger frequencies non - equilibrium effects slow down the power - law decay of @xmath285 , and , after a local maximum in @xmath285 near @xmath286 , @xmath104 tends to its asymptotic value coinciding with the diffusivity in the absence of external forcing . at small forcing amplitudes , in regions i and ii , @xmath285 decreases monotonically with frequency @xmath0 from its value at constant forcing , all the way till the non - equilibrium effects become critical , at @xmath287 . at stronger forcing , region iii , @xmath172 is nonmonotonic , a maximum is reached at @xmath288 . @xmath172 grows monotonically on @xmath289 , this growth is approximately linear on certain interval to the left of @xmath198 . we studied the temperature dependence of the diffusivity at fixed @xmath75 . contrary to the constant forcing problem @xcite the diffusivity increases with the temperature for all @xmath61 when the temperature is below certain frequency - dependent threshold value . limited temperature intervals exist for the force amplitudes in regions ii and iii , in which the diffusivity decreases with the temperature ( `` temperature - abnormal diffusivity '' , tad ) . the physical reason behind the strong diffusivity enhancement is emergence of two populations of particles under the action of external force , locked and running ones . at optimal @xmath61 and @xmath7 the number of particles in the two populations is comparable , long flights of the running particles relative to the locked population take place with significant probability , resulting in giant enhancement of diffusion . for comparison we studied diffusivity under sinusoidal in time driving . we saw that the same features in @xmath172 are observed in the three regions of force amplitudes as in pcp driving case . we thus conjecture that the qualitative features of @xmath290 behavior are insensitive to specific functional dependence of the ( symmetric , smooth enough in @xmath46 ) external force on time . the effects investigated allow for simple experimental verification . diffusion of particles on solid body surface is one natural arena for this . such systems are characterized by low dissipation @xcite . another field to study tad in is in propagation of magnetic particles on non - magnetic substrate , acted upon with electromagnetic fields . abnormal diffusion enhancement could be manifested in , _ e.g. _ , enhanced growth of islands of the new phase at decreasing temperatures . once verified experimentally , the effect of abnormal diffusion enhancement can find applications in a number of new technologies : in sorting of particles , manufacturing surface structures with required properties , controlling penetration of particles through biological and artificial membranes , in memristors , devices with charge density waves , etc . 50ifxundefined [ 1 ] ifx#1 ifnum [ 1 ] # 1firstoftwo secondoftwo ifx [ 1 ] # 1firstoftwo secondoftwo `` `` # 1''''@noop [ 0]secondoftwosanitize@url [ 0 ] + 12$12 & 12#1212_12%12@startlink[1]@endlink[0]@bib@innerbibempty @noop _ _ ( , , ) @noop _ _ ( , , ) , , and , eds . , @noop _ _ , , vol . ( , , ) , ed . , link:\doibase 10.1007/978 - 1 - 4419 - 1543 - 6 [ _ _ ] ( , , ) link:\doibase 10.1016/s0370 - 1573(98)00029 - 5 [ * * , ( ) ] link:\doibase 10.1039/c0sm01051k [ * * , ( ) ] link:\doibase 10.3762/bjnano.7.31 [ * * , ( ) ] , link:\doibase 10.1088/0034 - 4885/76/4/046602 [ * * , ( ) ] , link:\doibase 10.1103/revmodphys.81.387 [ * * , ( ) ] , link:\doibase 10.1140/epjst / e2013 - 02071 - 2 [ * * , ( ) ] , @noop _ _ ( , , ) link:\doibase 10.1103/physrevlett.96.190601 [ * * , ( ) ] , link:\doibase 10.1103/physrevlett.105.230602 [ * * , ( ) ] , link:\doibase 10.1103/physrevlett.104.248001 [ * * , ( ) ] link:\doibase 10.1002/adma.201203500 [ * * , ( ) ] link:\doibase 10.1007/s10404 - 013 - 1291 - 9 [ * * , ( ) ] link:\doibase 10.1007/bf01325627 [ * * , ( ) ] link:\doibase 10.1007/bf01323506 [ * * , ( ) ] link:\doibase 10.1016/0375 - 9601(79)90384 - 0 [ * * , ( ) ] link:\doibase 10.1007/bf01321244 [ * * , ( ) ] link:\doibase 10.1007/bf01352745 [ * * , ( ) ] link:\doibase 10.1007/bf01485833 [ * * , ( ) ] link:\doibase 10.1209/epl / i1999 - 00510 - 7 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.76.4874 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.82.2820 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.98.020602 [ * * , ( ) ] , link:\doibase 10.1103/physrevlett.92.250601 [ * * , ( ) ] , link:\doibase 10.1103/physreve.80.021123 [ * * , ( ) ] http://stacks.iop.org/1367-2630/7/i=1/a=029 [ * * , ( ) ] link:\doibase 10.1103/physrevlett.87.010602 [ * * , ( ) ] link:\doibase 10.1103/physreve.65.031104 [ * * , ( ) ] @noop * * , ( ) link:\doibase 10.1016/j.physa.2004.04.019 [ * * , ( ) ] link:\doibase 10.1140/epjb / e2013 - 40866 - 7 [ * * , ( ) ] link:\doibase 10.1103/physreve.77.041107 [ * * , ( ) ] link:\doibase 10.1039/c4sm02387k [ * * , ( ) ] link:\doibase 10.1209/0295 - 5075/100/50005 [ * * , ( ) ] , , and , eds . , @noop _ _ , vol . ( , , ) , ed . , @noop _ _ ( , , ) @noop _ _ ( , , ) link:\doibase 10.1209/0295 - 5075/97/60004 [ * * , ( ) ] , link:\doibase 10.1140/epjst / e2010 - 01270 - 7 [ * * , ( ) ] link:\doibase 10.1134/s0021364012030083 [ * * , ( ) ] link:\doibase 10.1088/1742 - 6596/514/1/012045 [ * * , ( ) ] link:\doibase 10.1103/physreve.70.051104 [ * * , ( ) ] , @noop _ _ ( , , ) link:\doibase 10.1103/physrevlett.101.190603 [ * * , ( ) ] link:\doibase 10.1103/physreve.80.062102 [ * * , ( ) ] , link:\doibase 10.1103/revmodphys.79.1381 [ * * , ( ) ] link:\doibase 10.1080/00018732.2012.706401 [ * * , ( ) ]

we consider in this review article a series of works @xcite , completed in collaboration with some colleagues of ours , that all share the following common denominator . the task to perform is , possibly repeatedly , to approximate numerically the solution to a partial differential equation that has some random character . in most of our works , the equation has the simplest possible form : it is scalar - valued , elliptic , linear , non degenerate , in divergence form . typically , and with self - explanatory notations , it reads as @xmath0 on a domain @xmath1 , and for a certain right hand side @xmath2 . the random character of the problem can be encoded in the coefficient @xmath3 , and/or in the right - hand side @xmath2 , and/or in the domain @xmath1 itself . needless to say , there exist a number of successful approaches to explicitly treat randomness in such partial differential equations . recent years have witnessed an explosion of the number of methods invented in this extremely lively topic , in particular motivated by the field of _ uncertainty quantification_. stochastic finite elements , spectral methods , sparse tensor products methods , reduced basis techniques , quantization , all methods based one way or another on karhunen - loeve , polynomial chaos , or other types of similar or not economical decompositions of the random functions present in the equation , both as parameters and unknown functions , have been increasingly studied and considerably improved lately . some accessible general references in the field are the textbooks @xcite . review articles on each of the many categories of approaches , such as @xcite , are also available . the _ rationale _ behind all these methods is the reduction of the dimension of the problem , considered as a problem in a high - dimensional space consisting of the ambient physical space where the original problem is posed _ augmented _ by the space of approximation for the random dependency . a simplification of the random dependency follows , and the problem becomes amenable to efficient computational techniques . we ourselves have used , in @xcite , and reviewed , in @xcite , some of these methods ( specifically , reduced basis type methods ) in the context of equations with random parameters . in the present review article , we would like to concentrate on a somewhat alternate strategy : _ not attempt to simplify the dependency upon randomness , but embrace the difficulty arising from it_. of course , this may only be achieved in some specific situations , sufficiently general to be of broad interest , but certainly not covering the immense spectrum of applications in the engineering and life sciences , and perhaps not with the same generic character as the above mentioned general purpose methods . to some extent , the cases we consider must be slightly simpler than , and not as general as , the cases targeted by the above methods . our ambition and our achievements are more modest . on the other hand , it turns out that the cases we consider arise from a context that brings an additional level of complexity : they originate from multiscale modeling . in that respect , the mix between the presence of randomness , the multiscale feature , and the wish to compute accurately and efficiently lead to an essentially unsolved difficulty ( despite the many efforts of outstanding contributors ) . in multiscale computational science , a number of techniques exist , and are improved constantly . but , most of the times , and despite some overly optimistic claims , they do not marry so well with randomness when it comes to practical computations . conversely , the extremely efficient toolbox available for random problems has essentially no intersection with multiscale science . well beyond the somewhat limited purpose of this review article , our intention is thus to attract the attention of the community to the state of the art : _ efficiently computing when randomness and many scales are simultaneously present is still an essentially open , considerably challenging , issue_. [ [ the - two - situations - considered - similarities - and - differences . ] ] the two situations considered : similarities and differences . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + we will consider below two prototypical situations . as we mentioned , we encountered both situations in our research efforts devoted to multiscale science , and more precisely in our endeavor to study and improve computational approaches in materials science : ( i ) the approximation of the homogenized tensor in the context of stochastic numerical homogenization , and ( ii ) the multiscale finite element computation of the solution to an ( harmless ) equation posed on a randomly perforated domain . in both situations , randomness originates from geometry . but , in either situation , geometry is encoded differently in the equation . in ( i ) , it is encoded in the heterogeneities of the coefficient . in contrast , in ( ii ) it is encoded in the computational domain itself . the similarity of the two problematics is evident . by penalization , the second problem may even be viewed as a particular case of the first problem . however , the techniques we employ are different in nature , and have a different purpose . the former problem ( approximation of the homogenized tensor ) consists in the repeated resolution of an elliptic equation ( the celebrated `` corrector equation '' ) on an as large as possible bounded domain , truncation of the whole space @xmath4 , typically for @xmath5 , or @xmath6 , in practice . the equation is of the above type and will be made explicit below , see . the purpose of that repeated solution procedure is to compute an expectation ( thus , in practice an empirical mean ) from the solutions obtained for various realizations of the local environment ( mathematically encoded in the coefficient @xmath3 of ) . the reader may think of different microstructures of the material , different inclusions in the medium , etc . the task of computing an average , that is a single , deterministic output , sounds simple . in particular , the very process of averaging , performed here in the context of a stationary ergodic problem ( see the details below ) , is essentially of the same nature as the law of large numbers . this suggests that the random character of the problem progressively vanishes when the number of realizations considered increases . however , the practical monte carlo approach ( generate random environments and average out an outcome based on the computed solutions ) is plagued by variance issues . the rate of convergence of the approximation , in terms of the size of the truncated computational domain asymptotically covering the whole ambient space , is universal . it is dictated by the central limit theorem . the prefactor appearing in the error estimate is related to the variance of the problem . efficient computational approaches consist in designing tools to reduce that variance , thus the statistical error in the approximation . that error largely dominates the bias ( the deterministic part of the error ) , and thus is the critical quantity that governs the overall quality of the numerical approach . our series of works addresses various techniques to reduce the variance : antithetic variables , control variate , selection approach , the latter being somewhat in the spirit of stratified sampling . they are imported from several different contexts and are adjusted to the specific context of numerical homogenization . the latter problem ( multiscale finite element computations on randomly perforated domains ) differs from the former problem in several respects . as mentioned above , the randomness of the geometry is now encoded not in the coefficients of the equations ( which are constant , for simplicity ) , but in the domain where the equation is posed . we deal with perforations of that domain that are randomly located . note however that , as briefly mentioned above , the perforations of the domain can be , in the numerical discretization , treated by penalization , in which case the two situations we consider become closer to one another . another , more substantial , difference is that this second example is a representative case of modern multiscale techniques . in contrast with classical techniques which ( i ) aim at computing an equivalent `` homogenized '' medium , and ( ii ) achieve this in the asymptotic limit of vanishing length scales of the oscillations originally present ( this is the case of our former situation ) , modern techniques attack the multiscale problem ( i ) directly and ( ii ) at the actual length scales . homogenization theory is then seen as a guideline to construct suitable basis functions for the discretization which , in a second stage , are used to expand the numerical solution . now , in the first stage of construction of these basis functions ( the _ ad hoc _ finite elements ) , the geometry of the computational domain matters . well before being able to average solutions and/or outputs based on solutions , the issue arises to be able to compute efficiently and accurately each and every single solution , in a robust way that does not require regenerating a set of basis functions for each realization of the random environment . this will be the issue we investigate , and solve using adequate ( crouzeix - raviart type ) boundary conditions for the multiscale finite element basis set and an adequate enrichment ( using bubble functions ) of that basis set . we wish to emphasize the following fact . not only the two problems we consider are relevant beyond the context of multiscale computing , and the methods we put in action can be useful elsewhere , but the second problem considered ( robustness of a numerical approach with respect to some wild modifications of the geometry ) is not specific to random modeling and covers the generality of all _ disordered _ geometries , random or deterministic . not every non periodic problem is indeed random . techniques literally based upon perfectness of the disorder ( such as the approaches we briefly mentioned at the beginning of this introduction , and also the variance reduction techniques we use ) can be inefficient in the presence of only imperfect , partial disorder . it is indeed true that the random setting often comes with assumptions such as stationarity , etc , that make randomness tractable at the price of somehow _ `` rigidifying the disorder '' _ , if we allow ourselves this abuse of language . thus the interest of considering other techniques than techniques probabilistic in nature . other works of ours ( see @xcite for a review ) proceed further in that direction . before we get to the heart of the matter , let us expose the problems in slightly more details , and announce the plan of our review . [ [ an - elliptic - problem - with - random - coefficients . ] ] an elliptic problem with random coefficients . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + the first problem we consider ( in section [ sec : variance ] below ) arises in the numerical approximation of oscillatory / heterogeneous solutions of pdes in the context of classical homogenization . the specific setting is that of random ergodic homogenization . the theoretical basis is briefly summarized in subsection [ ssec : recall ] , both in the deterministic ( periodic ) context and in the random context . the reader willing to familiarize him or herself with that celebrated theory is referred to the classical textbooks @xcite , and also to our introductory , computationally oriented lecture notes @xcite . the central problem to solve is the corrector problem , which is of the type and more precisely reads as an elliptic equation with a random parameter @xmath7 @xmath8 = 0\ ] ] ( where @xmath9 is a fixed vector ) set on the _ whole _ space @xmath4 . the precise mathematical formulation is the problem below . from the solution @xmath10 , an expectation is to be evaluated . in practice , the equation is set on a large , truncated domain , subset of @xmath4 . several realizations of @xmath7 are considered , and an empirical mean is employed to approximate the expectation built from the set of @xmath10 . this is the essence of the monte carlo approach in this particular context . details are given below . making this monte carlo approach efficient requires to reduce the variance . this is the purpose of our series of works @xcite , which is overviewed in subsection [ ssec : variance - reduction ] . well known variance reduction techniques , adapted to the specific context of homogenization , are described and tested . notice that our research on such topics has already been partially reviewed in @xcite where the reader may find more detailed elements of context and some additional comments . in particular , one will also find in @xcite a review of some of our works @xcite dealing with problems that are , in a sense made precise there , only `` weakly random '' . such problems are small random perturbations of periodic problems , and can be approximated by dedicated methods . these `` weakly random '' problems and methods have a great interest in their own right . they are not addressed in the present review . our target is `` fully random '' problems . some of these methods will however be briefly mentioned below , for different purposes , namely either constructing auxiliary , surrogate models ( in section [ sssec : controlvariate ] ) , or deriving selection rules for the draws ( in section [ sssec : sqs ] ) . both tools are useful for the reduction of variance in `` fully random '' problems . in essence , all variance reduction methods aim at eliminating part of the noise intrinsically present in the draws of the realizations of the random variables at play . noise can be eliminated in a variety of ways . one may try and bias the draws , thereby , in again an informal language , `` twisting the arbitrary arm '' of randomness . this should be made in a very subtle way . too small a bias in these draws has no effect on the variance . too large a bias may indeed succeed in reducing the variance but might affect the average and lead to an incorrect numerical result . alternately , one may first keep the noise as is , leaving the draws unaffected , but try to cancel the noise out in a second stage , subtracting ( in some sense to be made precise ) another simulation to the original one . the variance reduction techniques we make use of , which we did not invent but borrowed from other contexts and adapted to ours , proceed in either of the above directions . notice however that the book of possible variance reduction methods is thick . we have certainly not yet followed all possible tracks . the first variance reduction technique we employ is relatively elementary . it is the classical , general purpose technique of _ antithetic variables _ , presented in subsection [ sssec : antithetic ] . the efficiency of that technique is substantial , but is also , not unexpectedly , limited in particular because the technique does not exploit much the specifics of the problem considered . we present in subsection [ sssec : controlvariate ] the technique of _ control variate _ , which requires a better knowledge of the problem . a problem simpler to simulate and close to the original problem , in a sense that is made precise below , has to be considered and concurrently solved . the technique uses that knowledge to get a much better reduction of the variance . specifically , the auxiliary problem used as control is built upon an approximation method that we developed for the `` weakly random '' problems we were alluding to above . in some sense , this problem is a reduction of the original problem . in subsection [ sssec : sqs ] , we expose yet another approach , specifically imported from solid state physics , namely that of _ special quasi - random structures_. it consists in selecting , in the spirit of another well - known technique , _ stratified sampling _ , some configurations of the random environment that are more suitable than generic configurations to compute the empirical averages , so as to again reduce the variance . similarly to the previous approach of control variate , a reduced model is also useful in the approach . [ [ an - elliptic - problem - on - a - randomly - perforated - domain . ] ] an elliptic problem on a randomly perforated domain . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + our second framework concerns a pde set on a perforated domain . randomness manifests itself in the problem via the randomness of the location of the perforations ( and , possibly , their size and shape ) . again , we consider the simplest possible equation : scalar - valued , elliptic , non degenerate , linear and in divergence form . more precisely , we consider a bounded domain @xmath11 and a set @xmath12 of perforations within this domain . the perforations are supposedly small and in extremely large a number . the parameter @xmath13 encodes the typical small distance between the perforations . we denote by @xmath14 the perforated domain ( see figure [ fig : perforation ] in section [ sec : msfem ] ) . we then consider the following problem : find @xmath15 , solution to @xmath16 where @xmath17 is a given function , assumed sufficiently regular on @xmath1 . it is important to note that the homogeneous dirichlet boundary condition on @xmath18 ( and hence on the boundary @xmath19 of the perforations ) is a crucial feature of the problem we consider . the consideration of this particular academic problem stems from our interest in various physically relevant problems , for instance in fluid mechanics , atmospheric modeling , electrostatic devices , etc . equation can also be seen as a step toward the resolution of advection - diffusion equations ( see our ongoing work @xcite for that particular category of equations ) , that of the stokes problem ( see e.g. @xcite ) , or that of the navier - stokes problem , all on perforated domains . in fluid dynamics applications , homogeneous dirichlet boundary conditions on the perforations are typical . we introduce and study a dedicated multiscale finite element method ( msfem ) . the variant of msfem we consider uses _ crouzeix - raviart type _ finite elements @xcite . we have introduced this approach in @xcite on a multiscale elliptic problem ( exactly of the form ) with a highly oscillatory coefficient ( the coefficient @xmath3 in ) , and then extended it in @xcite to the case of problem . in the presence of perforations , we indeed have to improve the previously introduced approach by the addition of _ bubble functions _ to the finite element basis set . both ingredients , crouzeix - raviart type finite elements on the one hand , and addition of bubble functions on the other hand , allow for robustness of the approach , a critical issue when the geometry of the computational domain is random . it is well known that , for the construction of multiscale finite elements , boundary conditions set on the edges ( or facets in 3d ) of mesh elements for the definition of the basis functions play a critical role for the eventual accuracy and efficiency of the approach . using crouzeix - raviart type elements ( see @xcite for their original introduction in the context of classical meaning non - multiscale finite elements ) gives a definite flexibility . in short , the continuity of the finite element basis set functions across the edges of the mesh is enforced only in a weak sense by requiring that the average of the jump vanishes on each edge . the nonconforming approximation obtained in this manner proves to be very effective . the above issue regarding boundary conditions on the mesh elements is all the more crucial when dealing with perforated computational domains , and may be of paramount importance for _ randomly _ perforated computational domains . the approach is meant to be as insensitive as possible to the possible intersections between element edges and perforations . the perforations can then be very heterogeneously distributed . ad hoc _ construction of a mesh that as much as possible avoids intersecting the perforations is indeed prohibitively difficult . our approach does _ not _ require a mesh of this type . the second ingredient of our approach is the addition of bubble functions to the finite element space . as illustrated using a simple one - dimensional analysis ( see @xcite ) , and demonstrated with the extensive set of numerical tests performed in @xcite , some of which will be reproduced here , bubble functions definitely bring an added value for the overall accuracy of the approach . in the vast literature devoted to similar types of problems and techniques , we refer the reader unfamiliar with the general context to @xcite for some background on finite element methods in general , to the review article @xcite and the book @xcite ( and references therein ) for the general background on msfem , and the works @xcite specifically addressing problems on perforated domains , either from a theoretical or a numerical standpoint . our contribution is reviewed in section [ sec : msfem ] . that section begins with some generalities about our msfem approach , contained in subsection [ ssec : presentation2d ] . in particular , we briefly mention an error estimate that we have established in the case of periodic perforations . we are unfortunately unable to extend it to the case of random perforations . it is however useful to illustrate the nice properties of the approach . subsection [ ssec : numerical - tests ] is then devoted to some numerical tests that show the announced robustness of the approach in the presence of random perforations . all the numerical tests we report below have been performed using the finite elements software freefem++ @xcite . we first recall some basic ingredients of elliptic homogenization theory in the periodic setting . we consider , in a regular domain @xmath20 in @xmath4 , the problem @xmath21 = f \quad \text{in $ { \cal d}$ } , \\ \noalign{\vskip 5pt } u^{\varepsilon}=0 \quad \text{on $ \partial { \cal d}$ } , \end{array } \right.\ ] ] where the matrix @xmath22 is symmetric and @xmath23-periodic . we manipulate for simplicity _ symmetric _ matrices , but our discussion carries over to non symmetric matrices up to slight modifications . the corrector problem associated to ( [ eq : pb0-per ] ) reads , for @xmath9 fixed in @xmath4 , @xmath24 it has a unique solution up to the addition of a constant . then , the homogenized coefficients read @xmath25_{ij } = \int_q e_i^t a_{\rm per}(y ) \left(e_j + \nabla w_{e_j}(y ) \right)dy$ ] , where @xmath26 is the unit cube and @xmath27 is the canonical basis of @xmath4 . the main result of periodic homogenization theory is that , as @xmath13 goes to zero , the solution @xmath28 to ( [ eq : pb0-per ] ) converges to @xmath29 solution to @xmath30 = f \quad \text{in $ { \cal d}$ } , \\ \noalign{\vskip 5pt } u^\star=0 \quad \text{on $ \partial { \cal d}$}. \end{array}\right.\ ] ] the convergence holds in @xmath31 , and weakly in @xmath32 . the correctors @xmath33 , @xmath34 , may then also be used to `` correct '' @xmath29 in order to identify the behavior of @xmath28 in the strong topology @xmath32 . practically , at the price of only computing @xmath35 periodic problems ( [ eq : cor - per - intro ] ) , the solution to problem ( [ eq : pb0-per ] ) can be efficiently approached for @xmath13 small . the random setting is , for the simple equation , a straightforward extension of the periodic setting . skipping all technicalities related to the definition of the probabilistic setting , which we assume discrete stationary and ergodic ( we refer e.g. to @xcite for all details ) , we now fix @xmath36 a square matrix of size @xmath35 , which is assumed stationary in the sense that @xmath37 ( where @xmath38 is an ergodic group action ) and which is assumed to enjoy the classical assumptions of uniform ellipticity and boundedness . then we consider the boundary value problem @xmath39 standard results of stochastic homogenization @xcite apply and allow to find the homogenized limit of problem ( [ eq : pb0-stoc ] ) . the solution @xmath40 to ( [ eq : pb0-stoc ] ) converges to the solution to ( [ eq : pb0-star ] ) where the homogenized matrix is now defined as @xmath41_{ij } = { \mathbb{e}}\left(\int_q e_i^t a(y,\cdot ) \ , \left ( e_j+\nabla w_{e_j}(y,\cdot)\right ) \,dy\right),\ ] ] where @xmath26 is the unit cube and where , for any @xmath42 , @xmath43 is the solution ( unique up to the addition of a random constant ) to @xmath44 = 0 \quad \hbox{\rm a.s . on $ { \mathbb{r}}^d$ } , \\ \\ \nabla w_p \quad \mbox{is stationary in the sense of ( \ref{eq : stationnarite - disc } ) } , \\ \\ { \displaystyle}{\mathbb{e}}\left(\int_q \nabla w_p(y,\cdot)\,dy\right ) = 0 . \end{array } \right.\ ] ] a striking difference between the stochastic setting and the periodic setting can be observed comparing ( [ eq : cor - per - intro ] ) and ( [ eq : correcteur - random ] ) . in the periodic case , the corrector problem is posed on a bounded domain ( namely , the periodic cell @xmath26 ) , since the corrector @xmath43 is periodic . in sharp contrast , the corrector problem ( [ eq : correcteur - random ] ) of the random case is posed on the whole space @xmath4 . in the numerical practice , truncations of that problem have to be considered , typically on large domains @xmath45 ( say , @xmath46 ) and using periodic boundary conditions : @xmath47 = 0 , \\ \\ w_p^n(\cdot,\omega ) \mbox { is $ q_n$-periodic}. \end{array } \right.\ ] ] the random matrix @xmath48 defined by @xmath49_{ij } = \frac{1}{|q_n| } \int_{q_n } e_i^t a(y,\omega ) \ , \left ( e_j+\nabla w_{e_j}^n(y,\omega)\right ) \,dy\ ] ] converges ( almost surely ) to @xmath50 when @xmath51 ( see e.g. @xcite ) . important theoretical questions about the quality and the rate of the convergence in terms of the truncation size arise ( see , in particular , the pioneering works by a. bourgeat and a. piatnitski @xcite and a recent series of works by a. gloria , f. otto and their collaborators @xcite ) . the error @xmath52 \big ) + \big ( { \mathbb{e}}\left [ a_n^\star \right ] - a_n^\star(\omega ) \big)\ ] ] is the sum of a systematic error and of a statistical error ( the first and second terms in the above right - hand side , respectively ) . it is observed , and theoretically demonstrated in some cases , that the statistical error decays with a slower rate with respect to @xmath53 than the systematic error . for large values of @xmath53 , the statistical error is therefore dominating the systematic error . in what follows , we focus on reducing the statistical error , and describe approaches to more efficiently compute @xmath54 $ ] , for a given truncated domain @xmath45 . the direct approach for numerical random homogenization consists in using empirical means approximating the expectation @xmath55 of the homogenized coefficient computed on a truncated domain of size @xmath53 . variance issues are central . [ [ description - of - the - approach . ] ] description of the approach . + + + + + + + + + + + + + + + + + + + + + + + + + + + + the variance reduction technique known as the technique of _ antithetic variables _ has been adapted to the homogenization context in @xcite . it consists in concurrently considering two sets of configurations for the random material instead of only one set . the two sets of configurations are deduced one from the other ( see figure [ fig_b12 ] for a simple example ) . fix @xmath56 and consider @xmath57 i.i.d . copies @xmath58 of @xmath59 . construct next @xmath57 i.i.d . _ antithetic _ random fields @xmath60 , @xmath61 , from the @xmath58 . the map @xmath62 transforms the random field @xmath63 into another , so - called _ antithetic _ , field @xmath64 . the transformation is performed in such a way that , for each @xmath65 , @xmath64 has the same law as @xmath63 , namely the law of the matrix @xmath66 . somewhat vaguely stated , if @xmath66 was obtained in a coin tossing game ( using a fair coin ) , then @xmath64 would be _ head _ each time @xmath63 is _ tail _ and vice versa . then , for each @xmath61 , we solve two corrector problems of the type . one is associated to the original @xmath63 , the other one is associated to the antithetic field @xmath64 . using its solution @xmath67 , we define the _ antithetic homogenized matrix _ @xmath68 , the elements of which read , for @xmath69 , @xmath70_{ij } = \frac{1}{|q_n| } \int_{q_n } e_i^t b^m(y,\omega ) \ \left(e_j + \nabla v_{e_j}^{n , m}(y,\omega ) \right ) \ , dy.\ ] ] next we set , for any @xmath61 , @xmath71 since @xmath63 and @xmath64 are identically distributed , so are @xmath72 and @xmath73 . thus , @xmath74 is unbiased ( that is , @xmath75 ) . in addition , it satisfies @xmath76 almost surely , because @xmath64 is ergodic . the hope ( which is indeed confirmed by theory in simple cases and by numerical observation in all the tests we performed ) is that the new approximation @xmath74 has a smaller variance than @xmath72 . it is indeed the case under appropriate assumptions . together with its antithetic field @xmath77 for the famous random checkerboard ( reproduced from @xcite ) . [ fig_b12],width=336 ] [ [ foundation - of - the - approach . ] ] foundation of the approach . + + + + + + + + + + + + + + + + + + + + + + + + + + + some theoretical elements that give a foundation to our approach of variance reduction are collected in @xcite . of course , as always in this context , the one - dimensional setting , being explicit , may be entirely analyzed . two particular higher dimensional cases were also analyzed . the first one is a `` genuinely '' random setting , the second one is a `` weakly random '' case . in all such cases , under appropriate assumptions , the approach is shown to qualitatively reduce the variance . a quantitative assessment of the reduction is however out of reach . only numerical tests can provide some information . [ [ numerical - tests . ] ] numerical tests . + + + + + + + + + + + + + + + + the tests performed concern three different `` input '' random fields @xmath36 in , some i.i.d . , some correlated , with various correlation lengths . we have investigated variance reduction on a typical diagonal @xmath78_{11}$ ] , or off - diagonal @xmath78_{12}$ ] entry of the approximate homogenized matrix @xmath79 , as well as on the eigenvalues of the matrix , and the eigenvalues of the associated differential operator @xmath80 $ ] ( supplied with homogeneous dirichlet boundary conditions on @xmath81 ) . we show a typical result on figure [ fig_b12_result ] , for a two - dimensional test case where @xmath82 is given by @xmath83 where @xmath84 , @xmath85 is the identity matrix , and @xmath86 is an i.i.d sequence of random variables such that @xmath87 , with @xmath88 and @xmath89 . we observe on figure [ fig_b12_result ] that the confidence interval obtained with our approach is smaller than that obtained with the standard monte carlo ( mc ) approach . the gain in variance ( at equal computational cost ) is essentially insensitive to the size of the computational domain . it is here between 6 and 10 . at fixed computational cost , the approach improves the accuracy by a factor @xmath90 . equivalently , for a given accuracy , the computational cost is reduced by a factor @xmath91 . ( with confidence interval ) with respect to @xmath92 ( in red , the classical mc strategy , in green the antithetic variable strategy , for an equal computational cost ; reproduced from @xcite ) . [ fig_b12_result],width=336 ] the approach has been extended in @xcite to nonlinear convex stochastic homogenization problems . the highly oscillatory problem , the euler - lagrange equation of which replaces , is then given by @xmath93 for some appropriate @xmath94 . we assume that @xmath95 is strictly convex with respect to its third argument , and that , for any @xmath96 , the random field @xmath97 is stationary in the sense of . we also assume that @xmath95 satisfies a standard growth condition ( see @xcite for details ) . the associated homogenized problem , the euler - lagrange equation of which replaces , is @xmath98 where , for any @xmath99 , the homogenized energy density @xmath100 is given by @xmath101 where @xmath102 as in the linear case for the corrector problems and , we approximate in practice @xmath103 by @xmath104 $ ] for some large @xmath53 , the latter quantity being itself approximated by an empirical mean . we show in @xcite that the exact same antithetic variable technique we used in the linear case can be used to compute @xmath104 $ ] ( for any @xmath53 and any @xmath99 ) more efficiently . to summarize , our numerical results show that the technique may be applied to a large variety of situations and has proved efficient whatever the output considered . variance is systematically reduced . we have observed , however , that the rate of reduction is not spectacular . this has motivated the consideration of other approaches . the _ control variate approach _ is a variance reduction technique known to be potentially much more efficient than the antithetic variable technique . it however asks to have beforehand a better information on the random variables simulated . in the context of homogenization , the work @xcite presents a first possible investigation of the efficiency of this technique . as briefly mentioned in the introduction , the approach aims at `` subtracting '' part of the noise , using an auxiliary problem close to the original problem . we describe how we derive such an auxiliary problem in the next paragraph . next , we make the setting specific , and present the variance reduction approach . [ [ a - useful - weakly - random - setting . ] ] a useful `` weakly random '' setting . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + one perturbative approach among many other possible ones , described in full details in @xcite , and addressing the random material as a small perturbation of a periodic material , consists in considering a coefficient of the form @xmath105 where , with evident notation , @xmath22 is a periodic matrix modeling the unperturbed material , and @xmath106 is a periodic matrix modeling the perturbation . consider then the case @xmath107 where the @xmath108 are , say , independent identically distributed scalar - valued random variables . one particularly interesting case ( see @xcite for other cases ) is when the common law of the @xmath108 is assumed to be a bernoulli law of parameter @xmath109 ( see figure [ fig : arnaud - bernoumat ] ) . we now _ formally _ explain how we derive an expansion of the problem in `` powers '' of @xmath109 . the approach has been recently proven to be rigorous in @xcite . 5truemm the basic observation consists in noticing that , in the corrector problem @xmath110 = 0,\ ] ] the only source of randomness comes from the coefficient @xmath111 . therefore , in principle , if one knows the law of this coefficient , one knows the law of the corrector function @xmath112 and therefore may compute the homogenized coefficient @xmath113 , the latter being a ( quite intricate but well defined ) function of this law . when the law of @xmath114 is an expansion in terms of a small coefficient , so is the law of @xmath43 . consequently @xmath113 can be obtained as an expansion . heuristically , on the cube @xmath115^d$ ] and at order 1 in @xmath109 , the probability to get the perfect periodic material ( entirely modeled by the matrix @xmath22 ) is @xmath116 , while the probability to obtain the unperturbed material on all cells except one ( where the material has matrix @xmath117 ) is @xmath118 . all other configurations , with more than two cells perturbed , yield contributions of orders higher than or equal to @xmath119 . this gives the intuition ( confirmed by the mathematical proof ) that the first order correction indeed comes from the difference between the material perfectly periodic except on one cell and the perfect material itself . the homogenized matrix reads as @xmath120 where @xmath121 is the homogenized matrix for the unperturbed periodic material and @xmath122,\ ] ] where @xmath123 is the @xmath26-periodic corrector for @xmath22 ( solution to ) , and @xmath124 solves @xmath125 = 0 \quad \text{in $ q_n$ } , \quad \text{$w_i^n$ is $ q_n$-periodic}.\ ] ] the approach has been extensively tested . it is observed that , using the perturbative approach , the large @xmath53 limit for cubes of size @xmath53 is already very well approached for small values of @xmath53 . the computational efficiency of the approach is clear : solving the two periodic problems with coefficients @xmath22 and @xmath126 for a limited size @xmath53 is much less expensive than solving the original , random corrector problem for a much larger size @xmath53 . when the second order term is needed ( and we will actually use it below ) , configurations with two defects have to be computed . they all can be seen as a family of pdes , parameterized by the geometrical location of the defects . reduced basis techniques have been shown to be useful and allow for a definite speed - up in the computation , see @xcite . [ [ variance - reduction - in - the - fully - random - setting . ] ] variance reduction in the `` fully random '' setting . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + when the parameter @xmath109 of the bernoulli law is _ not _ taken small in the definition - of the coefficient inserted in the equation , and thus the corrector problem , the expansion technique we have just described is not correct , and actually we observe experimentally that it is inaccurate . it can however serve for the construction of a control variate , useful to reduce the variance . we recall that the field @xmath127 writes as in . determining the field @xmath82 , given by , on the truncated domain @xmath45 amounts to drawing @xmath128 in each cell @xmath129 in @xmath45 . this allows to compute the associated ( approximate ) homogenized coefficient @xmath48 from the solution to the corrector problem ( with @xmath130 ) , see . in parallel to this task , we _ reconstruct _ from the specific realization of the set of @xmath131 a field that is used as a control variate . more precisely , we set @xmath132 \right).\ ] ] in this formula , @xmath133 where @xmath134 is the deterministic coefficient corresponding to the case of _ one _ defect located at position @xmath135 in @xmath45 ( it is actually independent of @xmath135 thanks to the periodic boundary conditions in ) . the parameter @xmath136 is a deterministic parameter , a traditional ingredient of control variate techniques , which is optimized in terms of the estimated variances of the objects at play . it is crucial to note that the expectation of @xmath137 is analytically computable . since by construction @xmath138 , the technique then consists in approximating the former ( thus the latter ) by an empirical mean . the theoretical study and the numerical tests in @xcite show that the variance of @xmath139 is smaller than that of @xmath48 , and hence that the quality of the approximation is improved . as an illustration , we use a similar case as in subsection [ sssec : antithetic ] , which reads as with @xmath140 . applying the above strategy provides the results on the left part of figure [ cv_result ] , where the variance is reduced by a factor close to 6 , that is , comparable to the technique of antithetic variables . one may next improve on the previous results using a _ second order _ expansion with respect to @xmath109 and including in the control variate the deterministic coefficients corresponding to the case of both one and two defects . additional parameters playing the role of @xmath136 above are needed , in order to ensure a substantial variance reduction ( see the details in @xcite ) . the variance reduction of such a case , of the order of 40 , is represented on the right part of figure [ cv_result ] . together with its confidence interval ( computed using @xmath141 i.i.d . realizations ) . left : we use the classical mc approach ( in blue ) and the control variate approach based on ( in black ) . right : we use the classical mc approach ( in blue ) and the second - order control variate approach ( in red ) ( reproduced from @xcite ) . [ cv_result],title="fig:",width=278 ] together with its confidence interval ( computed using @xmath141 i.i.d . realizations ) . left : we use the classical mc approach ( in blue ) and the control variate approach based on ( in black ) . right : we use the classical mc approach ( in blue ) and the second - order control variate approach ( in red ) ( reproduced from @xcite ) . [ cv_result],title="fig:",width=278 ] the variance reduction approach we now overview has been originally introduced by other authors for a slightly different purpose in atomistic solid - state science @xcite . it carries the name sqs , abbreviation of _ special quasirandom structures_. the idea there is to sample random atomistic systems by selecting systems of finite size that exhibit all the suitable statistical properties usually only obtained in the asymptotic , bulk limit . for instance , one imposes on each and every finite size sample considered to have the correct volume fraction ( proportion ) of species ( atoms , molecules , ions , etc ) , and/or the correct correlations present in the infinite limit . to some extent , the approach is actually a particular instance of a general approach , known in many different contexts , and popularized in political sciences because it is at the foundation of many techniques used for opinion polls : _ stratified sampling_. the approach has been adapted to the homogenization context in @xcite to which we refer the reader for a more detailed presentation . [ [ mathematical - setting . ] ] mathematical setting . + + + + + + + + + + + + + + + + + + + + + before explicitly formulating the selection mechanism employed in the method , let us motivate it . consider a formal expansion of the coefficient @xmath66 present in and in terms of a small parameter ( even though , as we did in the previous section , we will soon forget about the smallness assumption ) . assume @xmath66 reads as @xmath142 for some presumably small scalar coefficient @xmath109 , where we assume that @xmath143 and @xmath144 have nice properties . since @xmath109 is small , @xmath114 is intuitively a perturbation of the matrix - valued field @xmath143 . at least formally ( and all this has been mathematically justified in our works ) , we may expand all quantities of homogenization theory in powers of the small parameter @xmath109 . in particular , the approximations @xmath145 and @xmath146 of , respectively , the corrector @xmath147 and the homogenized matrix @xmath148 on the truncated domain @xmath45 , can be expanded in powers of @xmath109 : @xmath149 inserting the expansion in the truncated corrector problem , one easily obtains a hierarchy of equations , the first three of which reading as @xmath150 & \quad \text{in } q_n , & \quad \text{$u^n_1 $ is $ q_n$-periodic } , \\ \noalign{\vskip 3pt } - \hbox{\rm div}\,c_0 \nabla u^n_2 = \hbox{\rm div}\,\left [ \chi c_1 \nabla u^n_1 \right ] & \quad \text{in } q_n , & \quad \text{$u^n_2 $ is $ q_n$-periodic}. \end{array } \right . \ ] ] we next deduce from and that holds , with @xmath151 we may now introduce the conditions that we use to select particular configurations of the environment within @xmath45 . for finite fixed @xmath53 , we say that a configuration @xmath152 satisfies the sqs conditions of order up to @xmath135 if , for any @xmath153 , the coefficient @xmath154 in the expansion exactly matches the corresponding coefficient @xmath155 of the analogous expansion of the exact homogenized matrix coefficient @xmath148 . more explicitly , we speak about the sqs condition of * order 0 if @xmath156 , that is to say , for any @xmath99 , @xmath157.\ ] ] * order 1 if @xmath158 , that is to say , for any @xmath99 , @xmath159.\end{gathered}\ ] ] * order 2 if @xmath160 , that is to say , for any @xmath99 , @xmath161.\end{gathered}\ ] ] it is easily observed that , using such particular configurations that satisfy the sqs conditions of order up to @xmath135 , we have , in the perturbative setting considered here , @xmath162 . taking the expectation over such configurations therefore formally provides a more accurate approximation of @xmath148 . of course , the purpose is to apply the approach _ beyond _ the perturbative setting . it can be expected ( and it is indeed observed ) that selecting the configurations using these conditions may reduce the variance . to make the computation of the right - hand sides of the above conditions practical ( since in theory they can only be determined using an asymptotic limit , and are therefore as challenging to compute in practice as @xmath50 itself ) , we restrict the generality of our setting . we assume that , in , @xmath163 is a deterministic , _ constant _ matrix , @xmath164 is a deterministic , @xmath23-periodic matrix , and that @xmath165 , where @xmath166 are identically distributed , not necessarily independent , bounded random variables . for simplicity , we also assume here that @xmath167 = 0 $ ] , and refer to @xcite for more general cases . after a tedious but not complicated calculation , we obtain that the two conditions rewrite as @xmath168 i_k^\infty,\end{aligned}\ ] ] respectively , where @xmath169 and @xmath170 . in these expressions , @xmath171 and @xmath172 solve @xmath173 = \hbox{\rm div}\,\left [ \mathbf{1}_q c_1p\right ] $ ] in @xmath4 , and @xmath174 = \hbox{\rm div}\,\left [ \mathbf{1}_{q } c_1p \right ] $ ] in @xmath45 with periodic boundary conditions , respectively . the conditions are called the sqs 1 and sqs 2 conditions . on the other hand , in the particular setting chosen , condition ( sqs 0 , in some sense ) is easily seen to be systematically satisfied when @xmath53 is an integer and the truncated approximation of that is chosen is the periodic approximation . [ [ selection - monte - carlo - sampling . ] ] selection monte carlo sampling . + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + the classical monte carlo sampling consists in successively generating a random configuration @xmath175 , solving the truncated corrector problem for that configuration , and computing the empirical mean @xmath176 as an approximation for @xmath50 . in our selection monte carlo sampling , we systematically test whether the generated configuration satisfies the required sqs conditions , up to a certain tolerance , and reject it if it does not , _ before _ solving the corrector problem for that configuration and letting it contribute to the empirical mean . in full generality , the cost of monte carlo approaches is usually dominated by the cost of draws , and therefore selection algorithms are targeted to reject as few draws as possible . in sharp contrast ( and this is a very important point which we can not emphasize enough ) , in the present context where boundary value problems are to be solved repeatedly , the cost of draws for the configuration is negligible compared to the cost of the solution procedure for such boundary value problems . likewise , evaluating the quantities present in is inexpensive . therefore , the purpose of the selection mechanism is to limit the number of boundary value problems to be solved , even though this comes at the ( tiny ) price of rejecting many configurations . we also note that , as for any selection procedure , our selection may introduce a bias ( i.e. a modification of the systematic error in ) . in practice , we have observed that the gain in variance dominates the bias possibly introduced by the selection approach , and that the approach provides a better approximation of @xmath50 than the standard monte carlo approach . we have studied the approach theoretically in @xcite . we have shown that the estimator provided ( at least in the simplest variant of our approach ) converges to the homogenized coefficient @xmath50 when the truncated domain progressively extends to the whole space . the efficiency of the approach is also theoretically demonstrated for some particular and simple situations ( such as , again , the one - dimensional setting ) . a comprehensive experimental study of the approach has been completed . we include here a typical illustration of the efficiency of the approach . we again use a similar case as in section [ sssec : antithetic ] . considering only configurations that exactly satisfy , we obtain the results shown on the left part of figure [ fig : sqs_un ] . it is also possible , among the configurations that exactly satisfy , to select configurations that satisfy as best as possible the condition . in practice , we generate 2000 configurations that exactly satisfy and select among them the 100 configurations for which the difference between the left and the right - hand sides of is the smallest . we then obtain the results shown on the right part of figure [ fig : sqs_un ] . together with its confidence interval ( computed using @xmath141 i.i.d . realizations ) as a function of @xmath53 . left : we use the classical mc approach ( in black ) and the sqs approach based on ( in red ) . right : we use the classical mc approach ( in black ) and the sqs approach based on and ( in blue ) ( reproduced from @xcite ) . [ fig : sqs_un],title="fig:",width=278 ] together with its confidence interval ( computed using @xmath141 i.i.d . realizations ) as a function of @xmath53 . left : we use the classical mc approach ( in black ) and the sqs approach based on ( in red ) . right : we use the classical mc approach ( in black ) and the sqs approach based on and ( in blue ) ( reproduced from @xcite ) . [ fig : sqs_un],title="fig:",width=278 ] in the case considered here ( for which the contrast in the field @xmath66 is equal to 3 ) , the variance is reduced by a factor 20 when using configurations that exactly satisfy , and by a factor 300 if is enforced as well . to compare this variance reduction approach with the two previous ones , it is however needed to consider a case for which the contrast in @xmath66 is similar . in that case , the variance is reduced by a factor of 9 when using configurations that exactly satisfy , and by a factor of 60 if is enforced as well . in all the test cases we have considered ( see @xcite for details ) , we have observed that the systematic error is kept approximately constant by the approach ( it might even be reduced ) , while the variance is reduced by several orders of magnitude . such an efficiency is achieved at almost no additional cost with respect to the classical monte carlo algorithm . we consider the problem in two dimensions , both for the analysis and for the numerical tests . we however emphasize that , of course , the approach can be applied to the three dimensional context and that , most likely , the theoretical analysis we review here can also be extended to the three dimensional case . interestingly , our analysis in @xcite on a similar problem was performed in both the two and three dimensional settings . we also assume , for simplicity , that the domain @xmath1 is a polygonal domain . we define a mesh @xmath177 on @xmath1 , i.e. a decomposition of @xmath1 into polygons each of diameter at most @xmath178 , and denote by @xmath179 the set of all the internal edges of @xmath177 . note that we mesh @xmath1 and not the perforated domain @xmath180 . this allows us to use coarse elements ( independently of the fine scale present in the geometry of @xmath180 ) , and leaves us with a lot of flexibility . the mesh does not have to be consistent with the perforations @xmath12 . some nodes may be in @xmath12 , and likewise some edges may intersect @xmath12 . see figure [ fig : perforation ] for a representation of the perforated domain @xmath180 . contains ( possibly random ) perforations @xmath12 . the perforated domain is @xmath14 . the boundary of @xmath180 is the union of @xmath181 ( the part of the boundary of the perforations that is included in @xmath182 ) and of @xmath183 . [ fig : perforation ] ] to construct our msfem approach , we introduce the space @xmath184 = 0 \text { for all } e\in \mathcal{e}_h , \quad u=0 \text { in $ b_{\varepsilon}$ and on $ \partial { \mathcal d}$ } \end{array } \right\}\ ] ] where @xmath185 $ ] denotes the jump of @xmath186 across an edge , and the subspace @xmath187 of @xmath188 . we define the crouzeix - raviart msfem space @xmath189 as the orthogonal complement of @xmath190 in @xmath188 , where by orthogonality we mean orthogonality for the scalar product @xmath191 . it is easy to see that we may write @xmath192 where the functions @xmath193 and @xmath194 are respectively defined as follows . first , for any mesh element @xmath62 that is not a subset of the perforations @xmath12 ( i.e. @xmath195 ) , we define @xmath196 as the solution to @xmath197 with , for each edge @xmath198 of @xmath62 which is an internal edge of @xmath177 , @xmath199 and @xmath200 on @xmath198 for some constant @xmath201 . for any edge @xmath198 of @xmath62 which belongs to @xmath202 , we set @xmath203 on @xmath198 . second , for any internal edge @xmath204 that is not a subset of the perforations @xmath12 , we denote by @xmath205 and @xmath206 the two triangles sharing this edge , set @xmath207 , and consider @xmath208 solution to @xmath209 with , for each edge @xmath210 which is an internal edge of @xmath177 , @xmath211 and @xmath212 on @xmath213 for some constant @xmath214 and @xmath215 and @xmath216 on @xmath204 for some constant @xmath217 ( with an a priori different constant on the two sides of @xmath204 ) . for any edge @xmath218 , we set @xmath219 on @xmath213 . although this is not the focus of the present review , which specifically addresses approaches in the presence of randomness , we briefly mention for completeness that the approach we have just introduced enjoys suitable theoretical properties , at least in the periodic setting . one major reason why we mention this theoretical analysis is that , although there exist many works in the literature dealing with error estimates for msfem type approaches , none of them seems to apply ( to the best of our knowledge ) to problems set on perforated domains . in the periodic setting ( and unfortunately we do not know how to extend our analysis to the random setting ) , we have established in @xcite ( in dimension 2 , under technical conditions that are made precise in @xcite , and using in particular arguments already present in @xcite ) the following error estimate : @xmath226 for some universal constant @xmath227 independent from @xmath178 , @xmath13 and @xmath2 , but depending on the geometry of the mesh and other parameters of the problem . the norm used in the left - hand side of is the energy norm associated with the form @xmath228 , namely @xmath229 for any @xmath230 . some important comments are in order . first of all , the size @xmath13 of the right - hand side of owes to the fact that , as it is well known theoretically ( see @xcite ) , the size of the exact solution @xmath186 ( and thus that of the corresponding approximation @xmath231 ) is @xmath13 in @xmath232 norm . taking this scaling factor into account , the actual rate of convergence for the numerical approach we design is therefore given by @xmath233 . second of all , if the bubble functions @xmath193 are not included in the discretization space , the estimate is replaced by @xmath234 and the relative error does not converge to 0 in the limit when @xmath235 and next @xmath236 , which is the relevant limit for our numerical experiments . this variant without bubbles is considered in our numerical tests for the sake of comparison . the proof of is based upon a series of somewhat classical ingredients . we write @xmath237 where the function @xmath238 is defined as @xmath239 \ \phi_e(x),\ ] ] where the functions @xmath193 and @xmath194 are defined by and respectively and @xmath240 is defined as the @xmath241-orthogonal projection of @xmath2 on the space of piecewise constant functions . the most delicate contribution in the error to estimate is the first one in . given the standard finite element interpolation result @xmath242 and the fact that @xmath238 satisfies @xmath243 we expect @xmath244 to be close to the exact solution @xmath186 . we establish that @xmath245 \cdot \nabla \phi + \sum_{t\in \mathcal{t}_h } \int_{{\mathcal d}_{\varepsilon}\cap t } \left[-\delta ( { \varepsilon}^2 w(\cdot/{\varepsilon } ) f - v_h ) \right ] \phi \\ + { \varepsilon}^2 \sum_{e \in \mathcal{e}_h } \int_{e \cap { \mathcal d}_{\varepsilon } } [ [ u - v_h ] ] \ n \cdot \nabla ( w(\cdot/{\varepsilon } ) f ) , \label{eq : decompo_error}\end{gathered}\ ] ] where @xmath246 and @xmath247 denotes the periodic corrector associated to the periodic homogenization problem , that is the periodic solution to the problem @xmath248 in @xmath249 , @xmath250 in the perforation @xmath251 ( where @xmath252 is the unit cell ) . we refer to @xcite for more background on homogenization theory . * the first term is small because of a classical homogenization result ( see again @xcite ) that states that @xmath253 is indeed an accurate approximation of @xmath186 , * the second term is small because , at the leading order term in @xmath13 , the first factor in the integrand is equal to @xmath254 which is small due to , * the third term is estimated as an integral involving highly oscillatory periodic functions . we now solve for some particular settings , comparing our approach with other existing msfem type methods . our primary focus is the possible intersections of the perforations with the edges of mesh elements . we address this question in section [ ssec : num - rob ] , and check there the robustness of our approach with respect to the ( temporarily deterministic , but still variable ) location of the perforations : the fact that the mesh intersects , or does not intersect , the perforations has a very little influence on the ( good ) accuracy of our approach , in contrast to other approaches . our numerical tests culminate with a couple of test - cases that show the excellent performance of our approach in the presence of randomness . for the sake of completeness , we begin our numerical tests with a comparison ( for a given set of fixed perforations ) with existing numerical msfem type approaches , in both cases when we enrich or not the variational space with bubble functions ( in the case of our crouzeix - raviart msfem approach , the functions @xmath193 defined by ) . this is the purpose of subsection [ ssec : num - comp ] . we mesh @xmath1 and not the perforated domain @xmath180 . some nodes may thus be in the perforations @xmath12 , some edges may intersect @xmath12 , etc . we then face difficulties when using the msfem variant with linear boundary conditions ( a variant that we compare here with our approach ) to directly solve . indeed , properly defining the msfem basis functions e.g. when edges intersect @xmath12 would not be straightforward . similar difficulties arise for the oversampling variant . for this reason , we do not consider , but the penalized version of that problem as mentioned above . in contrast , note that the crouzeix - raviart variant we introduce here can be used either on or on its penalized version . we solve on the domain @xmath255 , with the right - hand side @xmath256 , and we take @xmath12 the set of discs of radius @xmath257 periodically located on the regular grid of period @xmath258 . the reference solution is computed on a mesh of size @xmath259 . * the standard q1 finite element method on the coarse mesh of size @xmath178 . of course , we do not expect that method to perform well for this multiscale problem and we only consider it as a `` normalization '' . * the msfem with linear boundary conditions . although this method is now a bit outdated , it is still considered as the primary msfem approach , upon which all the other variants are built . * the msfem with oscillatory boundary conditions . this variant is restricted to the two - dimensional setting . it uses boundary conditions provided by the solution to the oscillatory ordinary differential equation obtained by taking the trace of the original equation on the edge considered . the approach performs fairly well on a number of cases , although it may also fail . * the variant of msfem using oversampling . this variant is often considered as the `` gold standard '' , although it includes a parameter ( the oversampling ratio ) , the value of which should be carefully chosen . when this parameter is taken large , the method becomes ( possibly prohibitively ) expensive . in addition , we consider for each of those approaches , and for our specific crouzeix - raviart type approach , two variants : one with , and the other without a specific enrichment of the basis set elements using bubble functions . for all approaches but the crouzeix - raviart type approach that we propose , the bubble @xmath260 on the quadrangle @xmath26 is defined as the solution to @xmath261 for our crouzeix - raviart approach , we recall that the bubble @xmath260 is defined by . for a given mesh size @xmath178 , the cost for computing the basis functions ( offline stage ) varies from one msfem variant to the other . however , for a fixed @xmath178 , all methods without ( respectively , with ) bubble functions essentially share the same cost to solve the macroscopic problem on @xmath1 ( online stage ) . more precisely , for a given cartesian mesh , and when using variants including the bubble functions , there are 1.5 times more degrees of freedom in our crouzeix - raviart approach than in the three alternative msfem approaches mentioned above . since a logarithmic scaling is used for the x - axis in the figures below , this extra cost does not change the qualitative conclusions we draw below . the numerical results we have obtained in the regime where the meshsize @xmath178 is of the order of , or larger than , the parameter @xmath13 are presented on figure [ fig : errors ] . for all values of the meshsize @xmath178 , and for both the @xmath241 and the broken @xmath232 norms , a definite superiority of our approach over all other approaches is observed , and the interest of adding bubble functions to the basis set is , for each approach , also evident . a side remark is the following . on figure [ fig : errors ] , we observe that , when using bubble functions , the error decreases as @xmath178 increases . this might seem counterintuitive at first sight . note however that , when @xmath178 increases , the cost of computing each basis function increases , as we need to solve a local problem ( discretized on a mesh of size @xmath262 controlled by the value of @xmath13 ) on a larger coarse element . in contrast to traditional fem , increasing @xmath178 does not correspond to reducing the overall computational cost . for msfem approaches , increasing @xmath178 actually corresponds to decreasing the online cost but increasing the offline cost . the regime of interest is that of moderate values of @xmath178 , for which the offline stage cost is acceptable . we only show the right part of figure [ fig : errors ] ( corresponding to large values of @xmath178 , leading to a prohibitively expensive offline stage ) for the sake of completeness . , left , and @xmath232-broken , right ) errors with various approaches in the regimes @xmath263 and @xmath264 : fem the standard q1 finite elements , no os msfem with linear boundary conditions , osc msfem with oscillatory boundary conditions , os msfem with oversampling ( where the size of the quadrangles used to compute the basis functions is @xmath265 ) , cr the crouzeix - raviart type msfem approach we propose . results for all these methods are represented by solid lines . the dashed lines correspond to the variants of these methods where we enrich the finite element spaces using bubble functions ( reproduced from @xcite ) . [ fig : errors],title="fig : " ] , left , and @xmath232-broken , right ) errors with various approaches in the regimes @xmath263 and @xmath264 : fem the standard q1 finite elements , no os msfem with linear boundary conditions , osc msfem with oscillatory boundary conditions , os msfem with oversampling ( where the size of the quadrangles used to compute the basis functions is @xmath265 ) , cr the crouzeix - raviart type msfem approach we propose . results for all these methods are represented by solid lines . the dashed lines correspond to the variants of these methods where we enrich the finite element spaces using bubble functions ( reproduced from @xcite ) . [ fig : errors],title="fig : " ] in this section , we perform a series of tests with the specific purpose of testing the robustness of the approach with respect to intersections of the mesh and the perforations . given the flexibility of crouzeix - raviart type finite elements in terms of boundary conditions , we expect our approach to be particularly effective ( and therefore considerably superior to other approaches ) when some edges of the mesh happen to intersect perforations of the domain . the more frequent such intersections , the more important the difference . we solve on the domain @xmath255 , with a constant right - hand side @xmath266 , and we take @xmath12 the set of discs of radius @xmath267 periodically located on the regular grid of period @xmath268 . we compute the reference solution , and consider 3 variants of msfem : the linear version , the oversampling version and the crouzeix - raviart version . the last three approaches are implemented in the variant that includes bubble functions in the basis set and applied with a mesh of size @xmath269 . we perform two sets of numerical experiments . they are identical except for what concerns the relative position of the mesh with respect to the perforations . the difference between the two sets of tests is that , from one set of tests to the other one , the perforations are shifted by @xmath270 in the directions @xmath271 and @xmath272 . in our test 1 , no edge intersects any perforation , while , on our test 2 , many edges actually intersect perforations ( see figure [ fig : decalage - test1 ] ) . to some extent , the situation of test 1 is the best case scenario ( where as few edges as possible intersect the perforations ) and the other situation is the worst case scenario . the numerical solutions are shown on figure [ fig : decalage - test1 ] . the numerical errors , computed both in @xmath241 and @xmath232-broken norms , are correspondingly displayed on tables [ table : test1 ] and [ table : test2 ] , respectively . more than the actual values obtained for each case , this is the trend of difference between table [ table : test1 ] and table [ table : test2 ] that is the practically important feature . a comparison between the two tables indeed show that , qualitatively and in either of the norms used for measuring the error , the linear version and the oversampling version of msfem are both much more sensitive to edges intersecting perforations than the crouzeix - raviart version of msfem . in particular , the gain of our approach with respect to the linear version of msfem is much higher in our test 2 ( which is , from the geometrical viewpoint , the worst case scenario ) than in test 1 . this confirms the intuition of a better flexibility of our approach . this also allows for expecting a much better behavior of that approach for non - periodic multiscale perforated problems for which it is extremely difficult , practically , to avoid repeated intersections of perforations with mesh edges . this is confirmed by the numerical experiments we perform in the next paragraph . we have tested several examples of randomly distributed perforations , two of them being shown on figure [ fig : ex - non - per ] . for each of them , the domain @xmath273 is meshed using quadrangles of size @xmath178 , with @xmath274 . the reference solution is again computed on a mesh of size @xmath259 . according to the uniform distribution . left : the perforations are 100 rectangles , the width and height of which are uniformly distributed between 0.02 and 0.05 . right : the perforations are 60 rectangles , the width ( resp . the height ) of which is uniformly distributed between 0.02 and 0.04 ( resp . 0.02 and 0.4 ) ( reproduced from @xcite ) . [ fig : ex - non - per],title="fig : " ] according to the uniform distribution . left : the perforations are 100 rectangles , the width and height of which are uniformly distributed between 0.02 and 0.05 . right : the perforations are 60 rectangles , the width ( resp . the height ) of which is uniformly distributed between 0.02 and 0.04 ( resp . 0.02 and 0.4 ) ( reproduced from @xcite ) . [ fig : ex - non - per],title="fig : " ] errors are shown on figure [ fig : errors - non - per3 ] ( resp . figure [ fig : errors - non - per5 ] ) for the test - case shown on the left ( resp . right ) part of figure [ fig : ex - non - per ] . our approach provides results at least as accurate as , and often more accurate than the msfem approach with oversampling on quadrangles of size @xmath265 . , left , and @xmath232-broken , right ) errors with the same approaches as on figure [ fig : errors ] for the test - case shown on the left part of figure [ fig : ex - non - per ] ( reproduced from @xcite ) . [ fig : errors - non - per3],title="fig : " ] , left , and @xmath232-broken , right ) errors with the same approaches as on figure [ fig : errors ] for the test - case shown on the left part of figure [ fig : ex - non - per ] ( reproduced from @xcite ) . [ fig : errors - non - per3],title="fig : " ] , left , and @xmath232-broken , right ) errors with the same approaches as on figure [ fig : errors ] for the test - case shown on the right part of figure [ fig : ex - non - per ] ( reproduced from @xcite ) . [ fig : errors - non - per5],title="fig : " ] , left , and @xmath232-broken , right ) errors with the same approaches as on figure [ fig : errors ] for the test - case shown on the right part of figure [ fig : ex - non - per ] ( reproduced from @xcite ) . [ fig : errors - non - per5],title="fig : " ] * acknowledgements : * the work of the two authors is partially supported by eoard under grant fa8655 - 13 - 1 - 3061 and by onr under grant n00014 - 15 - 1 - 2777 . the authors would like to thank their many friends and collaborators on the issues presented here , in particular x. blanc ( paris 7 ) , p .- l . lions ( collge de france ) , a. lozinski ( besanon ) and w. minvielle . , _ introduction to numerical stochastic homogenization and the related computational challenges : some recent developments _ , in _ multiscale modeling and analysis for materials simulation _ , w. bao and q. du eds . , lecture notes series , institute for mathematical sciences ( national university of singapore ) , vol . 22 , 2011 , pp . 197 - 272 . , _ mixed multiscale methods for heterogeneous elliptic problems _ , in _ numerical analysis of multiscale problems _ , i.g . graham , t.y . hou , o. lakkis and r. scheichl eds . , lecture notes in computational science and engineering , vol . 83 , springer , 2011 , pp . 243 - 283 . , _ variance reduction in stochastic homogenization : the technique of antithetic variables _ , in _ numerical analysis of multiscale computations _ , b. engquist , o. runborg and r. tsai eds . , lecture notes in computational science and engineering , vol . 82 , springer , 2012 , pp . 47 - 70 . , _ variance reduction in stochastic homogenization using antithetic variables _ , markov processes and related fields 18 ( 2012 ) 31 - 66 ( preliminary version available at http://cermics.enpc.fr/@xmath275legoll/hdr/fl24.pdf ) . , _ a reduced basis approach for variational problems with stochastic parameters : application to heat conduction with variable robin coefficient _ , comput . methods appl . mech . eng . 198(41 - 44 ) ( 2009 ) 3187 - 3206 . , _ approximation numrique dune classe de problmes en homognisation stochastique [ numerical approximation of a class of problems in stochastic homogenization ] _ , c. r. acad . srie i 348 ( 2010 ) 99 - 103 . , _ some numerical approaches for `` weakly '' random homogenization _ , springer lecture notes in computational science and engineering , g. kreiss et al . ( eds . ) , numerical mathematics and advanced applications 2009 , pp . 29 - 45 . , _ msfem la crouzeix - raviart for highly oscillatory elliptic problems _ , in * partial differential equations : theory , control and approximation , in honor of the scientific heritage of jacques - louis lions * , ph.g . ciarlet , t. li and y. maday eds . , springer , 2014 , pp . 265 - 294 . , _ nonconforming multiscale finite element method for stokes flows in heterogeneous media . part i : methodologies and numerical experiments _ , siam multiscale modeling & simulation 13(4 ) ( 2015 ) 1146 - 1172 .

the success of the perturbative aspect of the standard model is truly impressive , the theory being now tested at higher than one - loop accuracy . it is natural that over the past years various methods have been proposed to calculate higher loop graphs . beyond one loop , there are situations in which one can neglect internal masses , or exploit the kinematics to evaluate amplitudes analytically at some special point where the feynman diagrams become simpler . these special situations , while ingenious , cover only certain classes of physical processes . there have been attempts to formulate solutions for two - loop feynman diagrams , which ideally should be as general as the one - loop solutions are . for the massless case a lot of progress has been reached recently @xcite . on the other hand , for the general massive case , it has become clear that some numerical treatment appears unavoidable because of the complexity of the scalar integrals involved . in refs . @xcite a reduction scheme was found for treating massive self - energy diagrams at two - loop , with the resulting master scalar integrals being evaluated numerically . in ref . @xcite we investigated the problem of a general massive two - loop algorithm , which would deal with multi - leg diagrams as well . once we subscribe to a semi - numerical approach , some demands must be made , which are based on the generality , the universality , the effectiveness and the accuracy that a particular formalism engenders . we have shown that for any external kinematics and internal masses , we can reduce every two - loop amplitude for a process due to any renormalizable interactions into a set of ten scalar integrals and their derivatives . it is not hard to see how a result like this is obtained if we consider first a scalar theory with trivial interactions @xcite . then , there would be two internal momenta , @xmath4 and @xmath5 , and by the use of feynman parameters to combine sets of denominators in which either @xmath4 or @xmath5 or @xmath6 runs through , and to make shifts in @xmath4 and @xmath5 , we have an integrand proportional to @xmath7^{\alpha_1}(q^2+m_2 ^ 2)^{\alpha_2 } [ ( p+q)^2+m_3 ^ 2]^{\alpha_3}\ ] ] in which a four - vector @xmath8 is left over , which is a linear function of the external momenta and feynman parameters . the `` masses '' @xmath9 are functions of external momenta , internal masses and feynman parameters . these are to be integrated over @xmath4 , @xmath5 and a set of feynman parameters . by partial differentiating with respect to @xmath9 , there is in principle only one basic function arising from @xmath10 we need to know , although for convenience one may add @xmath11 it is clear that different interactions and different graphs will give different polynomials of feynman parameters to the numerators , and also different @xmath8 and @xmath9 . an extension of this construction to tensor integrals will give us the main result mentioned earlier . in the following section , we discuss the steps and arguments needed to perform a tensorial decomposition . we show that there is a small set of basic functions which are needed , which have simple , one - dimensional integral representations . their analytic structure is easy to see , and therefore the integration contour can be extended into the complex plane and optimized for rapid numerical convergence . in section 3 , we shall give a detailed exposition of the numerical techniques involved . given that two - point functions are usually simpler to calculate , many calculations existing in the literature used unitarity cuts of self - energies to calculate lower loop - order inclusive decay rates . we would like to stress that , given a process , our formalism allows us to calculate the amplitudes individually , as opposed to using unitarity cuts . as an immediate consequence , it becomes possible to obtain rates for exclusive processes , without the need to integrate over the whole final - state phase space . in section 4 , for illustration , we discuss in detail the application of our method to the important physical process @xmath12 . the starting point of calculating a massive two - loop diagram is to express it in terms of integrals of a standard type . topologically , any generic two - loop diagram can be represented in the form shown in figure 1 . by introducing a set of feynman parameters @xmath13 to combine all propagators which have the same loop integration momentum @xmath4 , @xmath5 , or @xmath14 , the graph can always be represented as an integral over a sunset - type two - loop integral . this is illustrated in figure [ fig : intosunset ] . all dependence on the external momenta @xmath15 and internal masses @xmath16 is now contained in the variables @xmath17 , @xmath18 , @xmath19 , and @xmath8 , which also depend on the set of feynman parameters @xmath13 . = 12 cm the original feynman diagram is written at this point as an integral over tensor functions of the following type ( all momenta are rotated into euclidean ) : @xmath20^{\alpha_{1 } } \ , ( q^{2}+m_{2}^{2})^{\alpha_{2 } } \ , ( r^{2}+m_{3}^{2})^{\alpha_{3 } } } ~~ . \label{eq:1}\ ] ] by casting the graph in this form , our strategy is to develop a uniform treatment of all possible sunset - type functions which can be generated from generic two - loop feynman graphs . the further tensor reduction and decomposition into standard scalar functions is done by using standard formulae common for all diagrams . therefore it can be automatized in the form of an algebraic manipulation program . at the end , the remaining integral over the feynman parameters @xmath13 , represented in figure [ fig : intosunset ] , is performed numerically . tensor integrals of the type in eqn . [ eq:1 ] need to be decomposed into scalar integrals . by lorentz covariance , this two - loop integral is a tensor constructed from the external momentum @xmath21 and the metric tensor @xmath22 . given that , one way of obtaining the tensor decomposition would be to write down all tensor structures allowed by the symmetry of the integral , use appropriate projectors , and solve the resulting equations . another way to do this is by decomposing the loop momenta @xmath4 and @xmath5 into components parallel and orthogonal to the external momentum @xmath21 : @xmath23 after this decomposition , the tensor decomposition is obtained by noticing that the functions with an odd number of transverse loop momenta @xmath24 and @xmath25 vanish , while the even functions are transverse to @xmath21 @xcite . in ref . @xcite we have shown that the resulting scalar coefficients of the tensor decomposition are integrals of the following form : @xmath26^{\alpha_{1 } } \ , ( q^{2}+m_{2}^{2})^{\alpha_{2 } } \ , ( r^{2}+m_{3}^{2})^{\alpha_{3 } } } ~~ . \label{eq:2}\ ] ] we give in the following the tensor decompositions for all tensor integrals of the type of eq . [ eq:1 ] of rank 1 , 2 , and 3 : tensor decompositions for more than three lorentz indices are derived in a similar way . in the formulae above , a loop integration @xmath28 is understood , and we used the following notations : and : in the expressions above , we introduced a set of ten scalar integrals @xmath31 , which are related to the @xmath32 functions , and are defined as follows : @xmath33 ^ 2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_2 & = & \int d^{n}p\,d^{n}q\ , \frac { p \cdot k } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_3 & = & \int d^{n}p\,d^{n}q\ , \frac { q \cdot k } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_4 & = & \int d^{n}p\,d^{n}q\ , \frac { ( p \cdot k)^2 - \frac{1}{n } k^2 p^2 } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_5 & = & \int d^{n}p\,d^{n}q\ , \frac { ( p \cdot k ) ( q \cdot k ) - \frac{1}{n } k^2 ( q \cdot p ) } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_6 & = & \int d^{n}p\,d^{n}q\ , \frac { ( q \cdot k)^2 - \frac{1}{n } k^2 q^2 } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_7 & = & \int d^{n}p\,d^{n}q\ , \frac { ( p \cdot k)^3 - \frac{3}{n+2 } k^2 p^2 ( p \cdot k ) } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_8 & = & \int d^{n}p\,d^{n}q\ , \frac { ( p \cdot k)^2 ( q \cdot k ) - \frac{3}{n+2 } k^2 p^2 ( q \cdot k ) } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_9 & = & \int d^{n}p\,d^{n}q\ , \frac { ( p \cdot k ) ( q \cdot k)^2 - \frac{1}{n+2 } k^2 [ 2 ( p \cdot q ) ( q \cdot k ) + q^2 ( p \cdot k ) ] } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \nonumber \\ { \cal h}_{10 } & = & \int d^{n}p\,d^{n}q\ , \frac { ( q \cdot k)^3 - \frac{3}{n+2 } k^2 q^2 ( q \cdot k ) } { [ ( p+k)^2+m_{1}^{2}]^2 \ , ( q^2 + m_{2}^{2 } ) \ , ( r^2 + m_{3}^{2 } ) } \end{aligned}\ ] ] as it can be seen in eqs . 5 , these scalar functions appear naturally in the tensorial decomposition . as discussed in the following section , the functions @xmath31 are only logarithmically divergent in the ultraviolet . because of this , they have very simple integral representations which can be used for numerical computation . once a way of calculating the special functions @xmath31 is available , the @xmath34 functions can easily be recovered by partial fractioning eqns . the partial fractioning generates essentially trivial products of one - loop tadpoles . the conversion formulae are given in appendix a. because the ultraviolet behaviour of the functions @xmath31 is logarithmic , fairly simple and symmetric integral representations can be found @xcite : @xmath35 % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_2 & = & \pi^4 k^2 \left [ - \frac{2}{\epsilon^2 } + \frac{1}{\epsilon } ( \frac{1}{2 } - 2 \gamma_{m_1 } ) + \frac{13}{8 } - \frac{\pi^2}{12 } + \frac{\gamma_{m_1}}{2 } - \gamma_{m_1}^2 - h_2 \right ] % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_3 & = & \pi^4 k^2 \left [ \frac{1}{\epsilon^2 } - \frac{1}{\epsilon } ( \frac{1}{4 } - \gamma_{m_1 } ) - \frac{13}{16 } + \frac{\pi^2}{24 } - \frac{\gamma_{m_1}}{4 } + \frac{\gamma_{m_1}^2}{2 } + h_3 \right ] % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_4 & = & \pi^4 ( k^2)^2 \left [ \frac{3}{2 \epsilon^2 } + \frac{1}{\epsilon } \frac{3 \gamma_{m_1}}{2 } - \frac{175}{96 } + \frac{\pi^2}{16 } + \frac{3 \gamma_{m_1}^2}{4 } + \frac{3}{4 } h_4 \right ] % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_5 & = & \pi^4 ( k^2)^2 \left [ - \frac{3}{4 \epsilon^2 } - \frac{1}{\epsilon } \frac{3 \gamma_{m_1}}{4 } + \frac{175}{192 } - \frac{\pi^2}{32 } - \frac{3 \gamma_{m_1}^2}{8 } - \frac{3}{4 } h_5 \right ] % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_6 & = & \pi^4 ( k^2)^2 \left [ \frac{1}{2 \epsilon^2 } - \frac{1}{\epsilon } ( \frac{1}{24 } - \frac{\gamma_{m_1}}{2 } ) - \frac{19}{32 } + \frac{\pi^2}{48 } - \frac{\gamma_{m_1}}{24 } + \frac{\gamma_{m_1}^2}{4 } + \frac{3}{4 } h_6 \right ] % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_7 & = & \pi^4 ( k^2)^3 \left [ - \frac{1}{\epsilon^2 } - \frac{1}{\epsilon } ( \frac{5}{24 } + \gamma_{m_1 } ) + \frac{287}{192 } - \frac{\pi^2}{24 } - \frac{5 \gamma_{m_1}}{24 } - \frac{\gamma_{m_1}^2}{2 } - \frac{1}{2 } h_7 \right ] % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_8 & = & \pi^4 ( k^2)^3 \left [ \frac{1}{2\epsilon^2 } + \frac{1}{\epsilon } ( \frac{5}{48 } + \frac{\gamma_{m_1}}{2 } ) - \frac{287}{384 } + \frac{\pi^2}{48 } + \frac{5 \gamma_{m_1}}{48 } + \frac{\gamma_{m_1}^2}{4 } + \frac{1}{2 } h_8 \right ] % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_9 & = & \pi^4 ( k^2)^3 \left [ - \frac{1}{3 \epsilon^2 } - \frac{1}{\epsilon } ( \frac{1}{24 } + \frac{\gamma_{m_1}}{3 } ) + \frac{95}{192 } - \frac{\pi^2}{72 } - \frac{\gamma_{m_1}}{24 } - \frac{\gamma_{m_1}^2}{6 } - \frac{1}{2 } h_9 \right ] % + { \cal o}(\epsilon ) \nonumber \\ { \cal h}_{10 } & = & \pi^4 ( k^2)^3 \left [ \frac{1}{4 \epsilon^2 } + \frac{1}{\epsilon } ( \frac{1}{96 } + \frac{\gamma_{m_1}}{4 } ) - \frac{283}{768 } + \frac{\pi^2}{96 } + \frac{\gamma_{m_1}}{96 } + \frac{\gamma_{m_1}^2}{8 } + \frac{1}{2 } h_{10 } \right ] % + { \cal o}(\epsilon ) \ ; \ ; .\end{aligned}\ ] ] here , @xmath36 is the space - time dimension , and @xmath37 ( @xmath38 is the euler constant , and @xmath39 is the t hooft mass ) . the special functions @xmath40 which appear in the formulae above are the finite part in the @xmath41 expansion of @xmath31 . they are defined by the following integral representations . except for special values of their arguments , they can not be further integrated into well - studied functions , such as the familiar polylogarithms , and as such , our strategy is to evaluate them numerically directly from their integral representations : @xmath42 \nonumber \\ h_3(m_1,m_2,m_3;k^2 ) & = & \int_0 ^ 1 dx \ , [ \tilde{g } ( x ) + \tilde{f_1 } ( x ) ] \ , ( 1-x ) \nonumber \\ h_4(m_1,m_2,m_3;k^2 ) & = & \int_0 ^ 1 dx \ , [ \tilde{g } ( x ) + \tilde{f_1 } ( x ) + \tilde{f_2 } ( x ) ] \nonumber \\ h_5(m_1,m_2,m_3;k^2 ) & = & \int_0 ^ 1 dx \ , [ \tilde{g } ( x ) + \tilde{f_1 } ( x ) + \tilde{f_2 } ( x ) ] \ , ( 1-x ) \nonumber \\ h_6(m_1,m_2,m_3;k^2 ) & = & \int_0 ^ 1 dx \ , [ \tilde{g } ( x ) + \tilde{f_1 } ( x ) + \tilde{f_2 } ( x ) ] \ , ( 1-x)^2 \nonumber \\ h_7(m_1,m_2,m_3;k^2 ) & = & \int_0 ^ 1 dx \ , [ \tilde{g } ( x ) + \tilde{f_1 } ( x ) + \tilde{f_2 } ( x ) + \tilde{f_3 } ( x ) ] \nonumber \\ h_8(m_1,m_2,m_3;k^2 ) & = & \int_0 ^ 1 dx \ , [ \tilde{g } ( x ) + \tilde{f_1 } ( x ) + \tilde{f_2 } ( x ) + \tilde{f_3 } ( x ) ] \ , ( 1-x ) \nonumber \\ h_9(m_1,m_2,m_3;k^2 ) & = & \int_0 ^ 1 dx \ , [ \tilde{g } ( x ) + \tilde{f_1 } ( x ) + \tilde{f_2 } ( x ) + \tilde{f_3 } ( x ) ] \ , ( 1-x)^2 \nonumber \\ h_{10}(m_1,m_2,m_3;k^2 ) & = & \int_0 ^ 1 dx \ , [ \tilde{g } ( x ) + \tilde{f_1 } ( x ) + \tilde{f_2 } ( x ) + \tilde{f_3 } ( x ) ] \ , ( 1-x)^3 \ ; \ ; . \label{eq : intrepr}\end{aligned}\ ] ] the four building blocks @xmath43 , @xmath44 , @xmath45 , and @xmath46 of these one - dimensional integral representations are given in appendix b. the ten special functions @xmath31 we introduced in the previous section , or , equivalently , their corresponding @xmath48 , are sufficient for treating all two - loop feynman graphs which may be encountered in renormalizable theories . to see this , it is useful to define the following auxiliary two - loop functions : @xmath49^{\alpha_{3 } } } ~~ .\ ] ] @xmath50 are trivially related to @xmath48 by a simple redefinition of the loop momentum ( see appendix a ) . the only reason for introducing them is for simplifying the discussion of this section ; some recursion relations can be written more compactly in terms of these scalar functions . to prove our assertion that the set of ten @xmath51 functions is sufficient for the case of renormalizable theories , we notice that not all functions @xmath50 of various indices are independent . there are recursion relations which relate functions of different indices . we will define the `` degree '' of a @xmath50 function as @xmath52 . the degree can be increased by differentiating with respect to the mass variables : @xmath53 and similarly for @xmath54 and @xmath55 . functions of the same degree are related by recursion relations obtained by differentiating with respect to the external momentum variable @xmath21 : @xmath56 { \cal p}^{a \ , b}_{\alpha_1 \ , \alpha_2 \ , \alpha_3 } + \frac{a k^2}{2 \alpha_1 } { \cal p}^{a-1 \ , b}_{\alpha_1 \ , \alpha_2 \ , \alpha_3 } \nonumber \\ { \cal p}^{a \ , b+1}_{\alpha_1 \ , \alpha_2 + 1 \ , \alpha_3 } & = & \frac{1}{2 \ ; \alpha_2 } \left [ 2 k^2 \frac{\partial}{\partial k^2 } - ( a+b ) \right ] { \cal p}^{a \ , b}_{\alpha_1 \ , \alpha_2 \ , \alpha_3 } + \frac{b k^2}{2 \alpha_2 } { \cal p}^{a \ , b-1}_{\alpha_1 \ , \alpha_2 \ , \alpha_3 } \ ; \ ; , \end{aligned}\ ] ] and @xmath57 { \cal p}^{a \ , b}_{\alpha_1 \ , \alpha_2 \ , \alpha_3 } = - 2 \alpha_3 \left [ k^2 { \cal p}^{a \ , b } _ { \alpha_1 \ , \alpha_2 \ , \alpha_3 + 1 } + { \cal p}^{a+1 \ , b } _ { \alpha_1 \ , \alpha_2 \ , \alpha_3 + 1 } + { \cal p}^{a \ , b+1}_{\alpha_1 \ , \alpha_2 \ , \alpha_3 + 1 } \right ] \ ; \ ; .\ ] ] when calculating a specific process , it is thus possible to just focus on the lowest order @xmath58 functions involved . all the other can be derived from these by differentiation , by using the relations above . at the same time , in renormalizable theories , there is a lower bound on the possible degree of @xmath58 functions involved in two - loop graphs , imposed by the dimension of the operators in the lagrangian . the minimal degree is one , and is attained for instance by a two - loop diagram such as the one shown in figure [ fig : deg1diagram ] . additional external legs can at most increase the degree of the @xmath58 functions involved . = 5 cm therefore , the most obvious choice of a basic set is the @xmath58 functions with @xmath59 , @xmath60 : @xmath61 however , we prefer to use the functions with @xmath62 , @xmath63 , @xmath64 instead : @xmath65 it turns out that this equivalent set of functions has simpler integral representations than the functions of eqs . [ eq : smallset ] . when needed , the eqs . [ eq : smallset ] functions can be derived from eqs . [ eq : largeset ] by partial @xmath4 : @xmath66 { \cal p}^{a \ , b}_{\alpha_1 \ , \alpha_2 \ , \alpha_3 } \right\ } \ ; \ ; , \end{aligned}\ ] ] because the set of functions of eqs . [ eq : largeset ] and the ten @xmath31 functions differ essentially by simple , one - loop tadpole contributions given in appendix a , this proves our assertion that the ten @xmath31 functions are sufficient for treating renormalizable theories at two - loop . we would like to stress that additional recursion relations and symmetries exist among this set of functions . an example of such a symmetry relation is given in eq . 16 of ref . @xcite , and additional ones may be present . such relations can be used , in principle , for restraining the number of master functions involved . however , the set of ten @xmath31 that we proposed has the advantage of having clean , symmetrical integral representations , where permutations of mass arguments are not involved , while symmetry relations obtained by loop momenta redefinitions often interchange the mass arguments . this simplicity is an advantage in practical calculations done by computer algebra , where a simpler algorithm may be preferable to a more complicated one which produces somewhat more compact results . thus , on general grounds , all tensor integrals @xmath67 with @xmath68 can be expressed by the standard set of ten functions via mass or momentum differentiation . the differentiations required can be carried out either by hand , or automatically , by using a computer algebra program such as mathematica or maple . however , in practical calculations of feynman graphs , before performing the tensor decomposition , it is advantageous to exploit all possible partial fractioning of the loop momenta , instead of resorting to recursion relations _ after _ the tensor decomposition . this can result in more compact results . this is in particular the case with the examples given in section 4 . there , the calculation can be simplified by noticing that in the feynman gauge , loop momenta @xmath4 and @xmath5 in the numerators come from fermion propagators , and there are no more than four of them . for those terms with four powers of @xmath4 and @xmath5 , one can algebraically rearrange them , so that two of these powers will contract as @xmath69 , @xmath70 or @xmath71 , and then perform partial fractioning of these terms , to reduce them into @xmath72 functions with @xmath73 . once the tensor reduction is done analytically , the original feynman diagram is expressed , according to figure [ fig : intosunset ] , as an integral over the set of remaining feynman parameters @xmath74 . the integrand itself consists of a sum of @xmath40 special functions and possibly trivial functions such as logarithms and rational functions of the kinematic variables @xmath17 , @xmath18 , @xmath19 , and @xmath75 . the @xmath40 functions themselves are given by the one - dimensional integral representations given in eqns . [ eq : intrepr ] . within our method , all these integration steps are performed numerically in general . therefore , it is natural to separate the numerical integration into two distinct steps : the routines to calculate the @xmath40 functions from their integral representations , and the final integration over the remaining feynman parameters @xmath74 . when performing the numerical integration , special care is needed to make sure that the integration is performed on the physical sheet . let us start with the integration of the @xmath40 functions . = 15 cm in figure [ fig : gfunction ] we plot the functions @xmath76 and @xmath77 as a function of the integration parameter @xmath78 . this is a typical behaviour above the threshold . at @xmath79 there is an integrable singularity of the logarithmic type , and similarly at @xmath80 . by mass or momentum differentiation the integrable nature of these singularities in @xmath78 is not changed ; this makes possible the treatment of more complicated topologies with additional external legs by using the same numerical approach . along the integration path there are two branching points , between which the function acquires an imaginary part . the integration path must be chosen such as to reproduce correctly this imaginary part . by continuing the integration parameter @xmath78 into the complex plane , we obtain a picture of the @xmath40 function s integrand as shown in figure [ fig : g3d ] . it is clear that if the position of the two branching points is known , a correct integration path can be calculated automatically by the computer program . finding the singularities involves some subtleties . = 14 cm adopting the notations introduced in appendix b , and inspecting the expressions of the functions @xmath76 , @xmath81 , @xmath82 and @xmath83 given there , the branching points we are looking for must be among the solutions of the equation : @xmath84 there are four solutions : @xmath85 \nonumber \\ x_{3,4 } & = & \frac{1}{2 \mu_{2}^{2 } } \ , [ \ , -a + b + \mu_{2}^{2 } \pm \sqrt{(a - b-\mu_{2}^{2})^{2 } - 4 b \mu_{2}^{2 } } \ , ] \ ; \ ; , \nonumber \\ \mu_{1,2}^{2 } & = & 1 - \kappa^{2 } \mp 2 \sqrt{- \kappa^{2 } } ~~ . \label{eq:4solutions}\end{aligned}\ ] ] to establish which two of the four solutions are the branching points we are looking for , it is useful to note that the causality of the green s functions can be expressed in at least two equivalent ways , which ought to lead to the same prescription for the integration path . the causality condition is expressed by the @xmath86 prescription in the feynman propagator , which means to shift all masses in the propagators @xmath87 . an equivalent way to impose causality is to calculate the euclidian green s functions and to go afterwards to physical momenta , approaching the cut on the positive real axis from above . this amounts to shifting the external momentum @xmath88 . these two prescriptions ought to be equivalent , and therefore have to fix the location of the physical singularities with respect to the real axis in the same way . for @xmath89 and @xmath90 , at @xmath91 both prescriptions lead to the same change , and therefore these are the singularities of the @xmath92 and @xmath93 functions . for @xmath94 and @xmath95 , the two prescriptions lead to opposite changes in the imaginary direction . since causality fixes the location of the singularities of the green s function uniquely , @xmath94 and @xmath95 can not correspond to real singularities of @xmath92 and @xmath93 . therefore @xmath92 and @xmath93 are analytical at these two points . the @xmath92 and @xmath93 functions themselves have only two branching points at @xmath89 and @xmath90 , because the singularities at @xmath94 and @xmath95 are compensating among the four terms of @xmath92 in eq . 26 . it is interesting to notice that the spurious solutions @xmath94 and @xmath95 correspond to the spurious , unphysical thresholds discussed in ref . @xcite in the context of the relation of the scalar sunset diagram with a generalization of hypergeometric functions . once the positions of these two singularities are known , the computer program can automatically compute an integration path which avoids them . because we are interested in a high accuracy and efficiency routine , we used an adaptive deterministic integration algorithm . such integration routines are very accurate provided that the integrand is smooth enough . the integrand itself is , of course , an analytic function along the complex integration path , and to preserve its smoothness it is advantageous to define a smooth integration path as well . we use an integration path defined in terms of higher - order spline functions such that both the path and its jacobian are smooth functions . a typical path is shown in figure [ fig : path ] . = 10 cm along these lines , a fully automatic computer program can be written , which first identifies the singularities , then calculates a suitable complex integration path , and then performs the numerical integration to obtain the @xmath40 functions starting from their integral representations given in eqs . [ eq : intrepr ] . by using adaptive integration routines , one typical evaluation of an @xmath40 function or of a derivative at eight digits takes of the order of 30 ms on a 600 mhz pentium processor . = 14 cm in practical calculations of feynman graphs , a number of mass or momentum derivatives of the functions @xmath40 become necessary . they can easily be obtained either by hand , or automatically , by computer algebra programs such as mathematica or maple . the numerical integration over the remaining feynman parameters @xmath74 is carried out along similar lines @xcite . the dimensionality of this final integration depends on the topology of the diagram and on the number of legs . with increasing complexity of the diagram , the methods we describe in this paper will still be applicable in principle , but in practice will result in larger integrands and a final numerical integration of higher dimension . thus its applicability will be limited in practice by the available computing power and by the ability of handling potentially large expressions which result from the tensor reduction in an error free way . here we would like to illustrate by means of a concrete example how the algorithm described above works in practice . examples involving two - point functions were given in ref . @xcite . here , we treat all two - loop three - point diagrams which contribute to the important physical process @xmath97 at @xmath98 . the diagrams involved are shown in figure [ fig : diagrams ] . in the complete calculation of this process , there are also self - energy type diagrams which contribute to the @xmath99 wave function renormalization . these two - point function graphs are simpler than the three - point graphs both analytically and numerically . within our method , they were calculated in ref . the process @xmath97 at @xmath98 at @xmath98 was calculated by means of a mass expansion in ref . @xcite . = 14.5 cm the two - loop methods we describe in this paper are intended primarily for the calculation of massive diagrams . however , sometimes massless particles are involved in a calculation and may lead to infrared singularities . we would like to stress that for purely massless calculations , such as qcd radiative corrections , more efficient methods are available already in the literature , which make use of the absence of masses @xcite . it is often possible in the massless case to carry out the calculation completely analytically . the discussion in this section applies to cases where the main difficulty is related to the presence of several masses , while the infrared structure is relatively simple . typical examples are the purely electroweak and the mixed electroweak - qcd radiative corrections , such as the @xmath97 at @xmath98 process discussed here . the diagrams @xmath100 , @xmath101 , @xmath102 , and @xmath103 are infrared divergent . the most common treatment of ir divergencies uses dimensional regularization to separate both the uv and the ir singularities . in our approach however , this is not directly possible because in the expressions of eqs . 7 the @xmath41 expansion already separated the uv singularities , while the possible ir singularities are still contained in the integral representations of the finite parts @xmath40 , with the intention of calculating these finite parts by numerical integration . therefore it is useful to separate first the infrared part of the two - loop diagrams in an analytically manageable form ( one - loop diagrams in this case ) . the `` real two - loop '' calculation which is left after this separation is then free of infrared singularities . the analytical separation of infrared divergencies is performed by noticing that the ir behaviour of the two - loop diagrams comes essentially only from the loop integration over the gluon propagator . then , the ir behaviour is left unchanged if the loop momentum on the propagator common to the two loop integrations ( @xmath104 ) is being `` frozen '' to the loop momentum of the ir - finite loop . this analytical separation of the ir behaviour is given in figure [ fig : ir ] for all ir divergent two - loop diagrams involved in this process . = 10 cm once the ir separation is performed , the ir finite part ( @xmath105 etc . ) can be calculated by numerical integration . for the numerical integration to be stable , one must make sure that the two components , _ e.g. _ @xmath100 and @xmath106 , have the same feynman parameterization such that the singularities cancel already before the feynman parameters integration . then , the ir divergent part(_e.g . _ @xmath106 ) can easily be treated analytically , in general by using dimensional regularization to regularize the ir singularities . for the case of the process considered in this example , a gluon mass regulator can also be used . this is possible because in this particular order in @xmath107 the ir structure is the same as in the abelian case , and a gluon mass regulator can be used without upsetting the slavnov - taylor identities . it is still an open question if such an analytical separation of the infrared singularities can _ always _ be performed such that the numerical integration can be carried out , in the case of other processes with potentially more complicated ir structure . this question clearly requires further investigation . we subtract the uv divergence of the diagrams of figure [ fig : diagrams ] by performing minimal subtractions of overall - divergences and sub - divergences . further , we extract the ir divergences analytically , according to the discussion of the previous section . the tensor decomposition results in a set of convergent integrals over the remaining feynman parameters , which can be carried out numerically . = 12.cm because in the top mass range of physical interest the diagrams are not close to a threshold , the top mass dependency in the relevant range is relatively smooth , as it can be seen in figure [ fig : diagramplots ] . we give numerical values for these diagrams in table 1 for three values of the top mass intermediary values can be well approximated by interpolation . in performing the numerical integration , we took @xmath108 gev , @xmath109 gev , and the effective electroweak mixing angle @xmath110 . for simplifying both the analytical and the numerical work , we neglected the @xmath99 mass , and in this limit all diagrams become proportional to @xmath111 . it is perfectly possible to keep the exact @xmath112 dependency throughout the whole calculation at the price of dealing with longer expressions and more computing time ; however the @xmath99 mass effect on the two - loop @xmath97 decay rate is expected to be very small . the numerical results are given in table 1 . as for the numerical accuracy and efficiency which can be attained , the results of table 1 are for a total of 26 individual feynman diagrams ( @xmath113 and @xmath114 exchange counted separately ) , each of them evaluated for 3 different values of the top mass . to carry out the numerical integration for this total of 78 two - loop individual feynman graph evaluations with an accuracy of @xmath115 , a total of 100 hours computing time on a 600 mhz pentium machine was used . higher numerical accuracy can be obtained by simply using more computing power . the analytical tensor decomposition part , performed by a form computer algebra program ( we also used a schoonship version with similar results ) , takes about one hour . .numerical values for the two - loop diagrams shown in figure [ fig : diagrams ] . @xmath116@xmath117 are the sums of the corresponding @xmath113 and @xmath114 exchange graphs . an overall color and coupling constant factor of @xmath118 is understood . the uv and ir divergences are removed as discussed in the text . the numerical integration accuracy is @xmath115 . the evaluation of a total of 78 feynman graph evaluations with this precision requires 100 hours computing time on a 600 mhz pentium machine . [ cols="<,^,^,^",options="header " , ] we described an algorithm for the tensor reduction of massive two - loop diagrams . it applies in principle to any massive two - loop graph , and it can be automatized in the form of a computer algebra program . the tensor decomposition algorithm results in a set of ten special functions @xmath40 which are defined in terms of one - dimensional integral representations . we described the numerical methods which are used for carrying out the remaining integrations . by applying the analytical reduction and numerical integration to an important three - point example , @xmath97 , we have shown that it can be used in realistic calculations , where several internal mass and external momenta scales are involved . this approach works for any such combination of kinematic variables , apart maybe from possible infrared complications . in the context of the @xmath97 example given in this paper , we discussed the analytical separation of the infrared divergencies . within our two - loop methods , if a process involves infrared singularities , these have to be dealt with in a special way because the numerical nature of our methods . it is an open question if this ir treatment can be generalized to other two - loop situations with potentially more complicated ir structure . * aknowledgements * the work of a.g . was supported by the us department of energy . the work of y .- p . y. was supported partly by the us department of energy . in section 2.2 we introduced two types of scalar functions which are involved in the tensor decomposition : @xmath31 and @xmath119 . then , we gave integral representations only for @xmath31 . the necessary @xmath120 scalar functions can be derived from the corresponding @xmath31 by partial fractioning : @xmath121 \nonumber \\ \tilde{{\cal p}}^{1 \ , 1}_{2 \ , 1 \ , 1 } & = & { \cal h}_5 + \frac{k^2}{2 n } \left\ { ( m_1 ^ 2 + m_2 ^ 2 - m_3 ^ 2 + k^2 ) { \cal h}_1 + 2 { \cal h}_2 - \tilde { \cal p}^{0 \ , 0}_{1 \ , 1 \ , 1 } \right . \nonumber \\ & & \left . \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; + t_2(m_1 ^ 2 ) \left [ t_1(m_2 ^ 2 ) - t_1(m_3 ^ 2 ) \right ] \right\ } \nonumber \\ \tilde{{\cal p}}^{0 \ , 2}_{2 \ , 1 \ , 1 } & = & { \cal h}_6 + \frac{k^2}{n } \left [ - m_2 ^ 2 { \cal h}_1 + t_2(m_1 ^ 2 ) t_1(m_3 ^ 2 ) \right ] \nonumber \\ \tilde{{\cal p}}^{3 \ , 0}_{2 \ , 1 \ , 1 } & = & { \cal h}_7 + \frac{3 k^2}{n+2 } \left [ - ( m_1 ^ 2 + k^2 ) { \cal h}_2 + \tilde{{\cal p}}^{1 \ , 0}_{1 \ , 1 \ , 1 } - 2 \tilde{{\cal p}}^{2 \ , 0}_{2 \ , 1 \ , 1 } \right ] \nonumber \\ \tilde{{\cal p}}^{2 \ , 1}_{2 \ , 1 \ , 1 } & = & { \cal h}_8 + \frac{3 k^2}{n+2 } \left [ - ( m_1 ^ 2 + k^2 ) { \cal h}_3 + { \cal p}^{0 \ , 1}_{1 \ , 1 \ , 1 } - 2 \tilde{{\cal p}}^{1 \ , 1}_{2 \ , 1 \ , 1 } \right ] \nonumber \\ \tilde{{\cal p}}^{1 \ , 2}_{2 \ , 1 \ , 1 } & = & { \cal h}_9 + \frac{k^2}{n+2 } \left [ ( m_1 ^ 2 + m_2 ^ 2 - m_3 ^ 2 + k^2 ) { \cal h}_3 - m_2 ^ 2 { \cal h}_2 + 2 \tilde{{\cal p}}^{1 \ , 1}_{2 \ , 1 \ , 1 } - \tilde { \cal p}^{0 \ , 1}_{1 \ , 1 \ , 1 } \right . \nonumber \\ & & \left . \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \;\ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \;\ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \;\ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \;\ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \;\ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; - 2 k^2 t_2(m_1 ^ 2 ) t_1(m_3 ^ 2 ) \right ] \nonumber \\ \tilde{{\cal p}}^{0 \ , 3}_{2 \ , 1 \ , 1 } & = & { \cal h}_{10 } + \frac{3 k^2}{n+2 } \left [ - m_2 ^ 2 { \cal h}_3 + k^2 t_2(m_1 ^ 2 ) t_1(m_3 ^ 2 ) \right ] \ ; \ ; \ ; \ ; , \end{aligned}\ ] ] where @xmath122 and @xmath123 are the euclidian one loop tadpole integrals : @xmath124 in section 2.4 , in order to simplify the discussion , we introduced the functions @xmath58 which are related to @xmath119 by a simple loop momentum shift : @xmath125 the functions with @xmath59 which appear in the relations 18 can be calculated by partial @xmath4 ( eq . 15 ) . for simplifying the notation , we omitted in the above formulae the mass and momentum arguments of the functions @xmath126 , @xmath127 , and @xmath128 , and understand that these arguments appear in this same order in all relations . the integral representations of the ten special functions functions @xmath40 , given in eqs . 8 , are built from the following functions : @xmath129 \nonumber \\ \tilde{f_2}(m_1,m_2,m_3;k^2;x ) & = & \frac{1}{3 } \left [ - \frac{2}{\kappa^2 } - \frac{1-\mu^2}{2 \kappa^2 } - \left ( \frac{1-\mu^2}{\kappa^2 } \right)^2 \right . \nonumber \\ & & \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \left . + y_1 ^ 3 \log{\frac{y_1}{y_1 - 1 } } + y_2 ^ 3 \log{\frac{y_2}{y_2 - 1 } } \right ] \nonumber \\ \tilde{f_3}(m_1,m_2,m_3;k^2;x ) & = & \frac{1}{4 } \left [ - \frac{4}{\kappa^2 } - \left ( \frac{1}{3 } + \frac{3}{\kappa^2 } \right ) \left ( \frac{1-\mu^2}{\kappa^2 } \right ) - \frac{1}{2 } \left ( \frac{1-\mu^2}{\kappa^2 } \right)^2 - \left ( \frac{1-\mu^2}{\kappa^2 } \right)^3 \right . \nonumber \\ & & \left . \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; \ ; + y_1 ^ 4 \log{\frac{y_1}{y_1 - 1 } } + y_2 ^ 4 \log{\frac{y_2}{y_2 - 1 } } \right ] \ ; \ ; , \end{aligned}\ ] ] in the above expressions , one special case must be treated separately , namely @xmath132 . the @xmath40 functions have a smooth limit for @xmath133 . for the purpose of numerical evaluation , it is useful to use a taylor expansion of the functions @xmath134 , @xmath135 , @xmath136 , and @xmath137 around @xmath138 for extremely small values of @xmath75 , where a direct evaluation by means of the exact expressions given above would be affected by large cancellations . it can be checked that this limit is regular and our approach reduces to the functions introduced by van der bij and veltman in ref . @xcite . z. bern , l. dixon , d.a . kosower , _ jhep * 0001 * ( 2000 ) 027 ; _ tausk , _ phys . b469 * ( 1999 ) 225 ; _ c. anastasiou , t. gehrmann , c. oleari , e. remiddi , j.b . tausk , _ hep - ph/0003261 ; _ v.a . smirnov , o.l . _ nucl . phys . * b566 * ( 2000 ) 469 ; _ j.a.m . vermaseren , s.a . larin , t. van ritbergen , _ phys . lett . * b405 * ( 1997 ) 327 ; _ _ phys . lett . * b404 * ( 1997 ) 153 ; _ _ phys b400 * ( 1997 ) 379 . _ r. harlander , t. seidensticker , m. steinhauser , _ phys . lett . * b426 * ( 1998 ) 125 ; _ j. fleischer , o.v . tarasov , f. jegerlehner , p. raczka , _ phys . lett . * b293 * ( 1992 ) 437 ; _ j. fleischer , f. jegerlehner , m. tentyukov , o. veretin , _ phys * b459 * ( 1999 ) 625 . _